Datasets:

TaylorAI

/

pubmed_commercial

like 12

PMC005xxxxxx/PMC5000870.txt

==== Front PlateletsPlateletsIPLTiplt20Platelets0953-71041369-1635Taylor & Francis 114880610.3109/09537104.2016.1148806ReviewSpecial Review: Platelet Half-LifeInherited thrombocytopenia: novel insights into megakaryocyte maturation, proplatelet formation and platelet lifespan Johnson Ben a Fletcher Sarah J. a Morgan Neil V. a * a Institute of Cardiovascular Sciences, College of Medical and Dental Sciences, University of Birmingham, UKCorrespondence: Dr Neil V. Morgan, Institute of Cardiovascular Sciences, College of Medical and Dental Sciences, University of Birmingham, B15 2TT, UK. E-mail: n.v.morgan@bham.ac.uk© Ben Johnson, Sarah J. Fletcher, & Neil V. Morgan This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. 17 8 2016 30 3 2016 27 6 519 525 20 10 2015 22 1 2016 10 12 2015 Published with license by Taylor & Francis.The Author(s)This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.Abstract The study of patients with inherited bleeding problems is a powerful approach in determining the function and regulation of important proteins in human platelets and their precursor, the megakaryocyte. The normal range of platelet counts in the bloodstream ranges from 150 000 to 400 000 platelets per microliter and is normally maintained within a narrow range for each individual. This requires a constant balance between thrombopoiesis, which is primarily controlled by the cytokine thrombopoietin (TPO), and platelet senescence and consumption. Thrombocytopenia can be defined as a platelet count of less than 150 000 per microliter and can be acquired or inherited. Heritable forms of thrombocytopenia are caused by mutations in genes involved in megakaryocyte differentiation, platelet production and platelet removal. In this review, we will discuss the main causative genes known for inherited thrombocytopenia and highlight their diverse functions and whether these give clues on the processes of platelet production, platelet function and platelet lifespan. Additionally, we will highlight the recent advances in novel genes identified for inherited thrombocytopenia and their suggested function. Keywords Inherited thrombocytopeniamegakaryocytesplateletsgene mutationsbleedingBritish Heart Foundation10.13039/501100000274FS/13/70/30521PG/13/36/30275Medical Research Council10.13039/501100000265Doctoral Training Grant ==== Body Introduction Inherited thombocytopenias (ITs) are a heterogeneous group of disorders characterized by a sustained reduction in platelet count manifesting as a bleeding diathesis. Since the discovery of disease inheritance patterns in disorders such as Bernard Soulier Syndrome (BSS), genetic studies of thrombocytopenia have been a vital tool in determining megakaryocyte and platelet physiology [1]. As a result of parallel whole exome and whole genome sequencing over the past 5–10 years, we are discovering increasing numbers of novel genes with a critical role in platelet physiology. However, are any of these genes giving us clues to platelet lifespan? Is IT predominately caused by a defect in the process of platelet production or is there more to the story than we currently know? To date, there are 31 genes suspected to cause 27 separate forms of inherited thrombocytopenia (Table I). The majority of patients with thrombocytopenia present secondary to syndromic disorders usually affecting the immune system or in cases of immune thrombocytopenic purpura (ITP). However, a number of cases present as a bleeding diathesis as a result of a reduced platelet count, sometimes alongside additional platelet dysfunction, indicating an effect solely on cells of the hematopoietic lineage. As a consequence of the varied phenotypic display and clinical presentation of inherited thrombocytopenias, there are several ways in which they are characterized. One such way, which has been favored following molecular characterization of the effect of mutations, is to group genes based upon their effect on megakaryocyte differentiation, platelet production and platelet function. Therefore, to understand whether a link to platelet lifespan can be established it is best to first consider all genes implicated in a reduced platelet count.Table I. Genetic causes of inherited thrombocytopenia, the encoded protein and their associated diseases. Grouped into their suggested role within megakaryopoiesis, platelet production or clearance/other. Area of mutational effect Gene Protein Associated disease References Megakaryopoiesis ANKRD26 Ankyrin repeat domain 26 ANKRD26-related thrombocytopenia [25]   ETV6 Transcription factor ETV6 THC5 [7]   FLI1 Friend leukaemia integration 1 transcription factor Paris Trousseau type thrombocytopenia/Jacobsen (11q23 del) [8]   FYB FYN-binding protein Novel thrombocytopenia [26]   GATA1 Erythroid transcription factor GATA1 related disease (XLT and XLTT) [9]   GFI1B Zinc finger protein Gfi-1b Grey Platelet Syndrome + novel thrombocytopenia [10]   HOXA11 Homeobox protein Hox-A11 Amegakaryocytic thrombocytopenia with radio-ulnar synostosis [74]   MPL Thrombopoietin receptor Congentical amegakaryocytic thrombocytopenia [12]   NBEAL2 Neurobeachin-like protein 2 Grey Platelet Syndrome [20]   RBM8A RNA-binding protein 8A Thrombocytopenia with absent radii [16]   RUNX1 Runt-related transcription factor Familial platelet disorder and predisposition to AML [11]   THPO Thrombopoietin Mild novel thrombocytopenia (heterozygous) [13] Platelet production ACTN1 Alpha-actinin-1 Bleeding disorder, platelet-type 15 [36]   CYCS Cytochrome C CYCS-related thrombocytopenia [45]   GP1BA Platelet glycoprotein 1b alpha chain Bernard–Soulier Syndrome + Platelet type von-Willebrand disease [75]   GP1BB Platelet glycoprotein 1b beta chain [76]   GP9 Platelet glycoprotein IX [77]   ITGA2B Integrin alpha-Iib ITGA2B/ITGB3-related thrombocytopenia [78]   ITGB3 Integrin beta-3 [79]   MKL1 MKL/myocardin-like protein 1 Thrombocytopenia with immunodeficiency [31]   MYH9 Myosin 9 MYH9 related disease [80]   PRKACG cAMP-dependant protein kinase catalytic subunit gamma Bleeding disorder, platelet-type 19 [39]   TUBB1 Tubulin beta-1 chain TUBB1-related macrothrombocytopenia [29]   WAS Wiskott–Aldrich syndrome protein Wiskott-Aldrich syndrome, X-linked thrombocytopenia [81] Platelet clearance/other ABCG5 ABC transporter G family member 5 Thromobocytopenia associated with sitosterolaemia [57]   ABCG8 ABC transporter G family member 8 [57]   ADAMTS13 GNE Disintegrin/metalloproteinase with thrombospondin motifs 13 Glucosamine (UDP-N-Acetyl)-2-Epimerase TTP, Upshaw–Schulman syndrome GNE myopathy with congenital thrombocytopenia [51] [65]   SLFN14 Schalfen family member 14 Novel thrombocytopenia [62]   STIM1 vWF Stromal interaction molecule 1 Von Willebrand factor Stormorken Syndrome Von Willebrand disease type 2B [82] [83] Genes that effect megakaryocyte differentiation and maturation Like all blood cells, megakaryocytes are derived from hematopoietic stem cells (HSCs) via progressively differentiated progenitor cells [2]. This is a process that is underpinned by the synthesis and synergistic effect of the cytokine thrombopoietin (TPO) through its receptor c-Mpl [3]. However, the commitment to differentiate and commit along the myelo-erythroid lineage is also highly regulated by transcription factors. One such transcription factor is GATA1, which is critically regulated transcriptionally [4]. Additional transcription factors such as Runt-related transcription factor 1, encoded by the gene RUNX1, ETV6 and FLI-1 in tandem with the transcriptional repressor GFI1B, ensure maturation of the megakaryocytes by binding critical promoter regions in crucial megakaryocyte-expressed genes [5, 6]. With the exception of Friend of Gata-1 (FOG1), the co-factor to GATA1, variants in the aforementioned hematopoietic transcription factors have previously been shown to cause inherited thrombocytopenia [7–11]. In addition, variants have also been observed in the TPO receptor gene MPL, as well as in its ligand, encoded by the gene THPO in one large Micronesian family with affected individuals displaying idiopathic aplastic anemia and mild thrombocytopenia [12, 13]. As the crucial role of TPO and transcription factors in megakaryocyte production is widely known, it is unsurprising that variants within these genes are commonly found within patients with inherited thrombocytopenia. However, clinical presentation does vary (most likely as a result of secondary effects of transcription factors and chemokines), but, all mutations are consistent in their ability to reduce proliferation of CD34+ cells by altering transcription profiles of genes crucial in megakaryocyte development [14]. In addition to the aforementioned genes involved specifically in the commitment to the proliferation towards megakaryocytes, a number of other genes have putative roles in differentiation and maturation. These include HOXA11, RBM8A, ANKRD26 and NBEAL2. Similar to FLI-1, ETV6 and AML-1, HOXA11 is a DNA-binding protein, which is expressed, like many other homeodomain box proteins, in human cord blood and hematopoietic precursor cells [15]. The specific molecular mechanism behind how variants in HOXA11 cause amegakaryocytic thrombocytopenia with radio-ulnar synostosis syndrome has yet to be established. However, it has been suggested that a point mutation in the third helix of the homeobox domain can lead to a reduction in CD61 expression in staurosporine-induced K562 cells [15]. RBM8A, the gene encoding the exon junction splicing complex component Y14, is interesting both in the genetic complexity of the disease and in its proposed molecular mechanism. Thrombocytopenia with absent radii (TAR) syndrome is known currently as a result of the co-inheritance of a variably sized deletion of the region surrounding 1q21.1, including RBM8A, alongside one of two relatively high-frequency SNPs within the regulatory region of RBM8A [16]. The clinical outcome is severe but heterogeneous and seems to be determined by a number of factors, including variable gene expression and possible epigenetic modifiers. Little is known on the specific cellular effect of the compound hypomorphic effect of mutations within RBM8A, but a reduction in megakaryocyte progenitors is observed and is thought to arise as a result of aberrant JAK2 signaling downstream of TPO [17, 18]. The production of intracellular granules, during megakaryocyte maturation, from the budding of small vesicles containing cargo from trans-Golgi network is crucial for the further differentiation of platelets [19]. Platelet α-granules play a crucial role in platelet function and amplification of activation during hemostasis. A number of ITs, including GATA1-related thrombocytopenia (RT) and GFI1B-RT, show reduced numbers of platelet α-granules. However, a complete absence of platelet α-granules is observed in patients with Grey Platelet Syndrome (GPS), a condition caused by genetic variations within NBEAL2 [20]. NBEAL2 localizes to the endoplasmic reticulum in platelets, where it is likely to function in membrane trafficking; however, the role of NBEAL2 in GPS and the subsequent reduction in platelet count is still largely unknown. Nbeal2−/- mice do recapitulate the phenotype of GPS, including an absence of platelet α-granules as well as delayed megakaryocyte maturation, decreased survival and decreased ploidy in cultured megakaryocytes [21]. Interestingly, however, a lack of Nbeal2 in mice does not affect initial α-granule formation, packaging or transport to the budding platelets within megakaryocytes [22]. Both GPS mice and patients do develop premature myelofibrosis within the bone marrow, which may be as a result of spontaneous granule release or “leaking” from the megakaryocytes [22–24]. Another gene where variants have been shown to cause a defect in megakaryocyte maturation resulting in reduced platelet α-granules is ANKRD26. Like RBM8A, variants affect gene expression and occur in the 5’-UTR of the gene, specifically in a short stretch of nucleotides from c.-134G to c.-113A [25]. In affected individuals, megakaryocytes are small, with hypolobulated nuclei as a result of dysmegakaryopoiesis. One of the most recently discovered genes known to cause inherited thrombocytopenia is FYB. Although the effect of the mutation has yet to be determined, a role in megakaryopoiesis has been suggested due to a reduced percentage of mature megakaryocytes in the bone marrow [26]. It seems then that the progression of HSCs through the megakaryocytic lineage to achieve mature megakaryocytes is a prime pathway to be disrupted by genetic mutation. Indeed, most variants not only involve altered expression or downstream signaling of the chemokine TPO, but many also affect additional signaling pathways, leading to a wide range of secondary abnormalities. Genes that affect pro-platelet formation and release Following megakaryopoiesis, the next phase in platelet production is pro-platelet formation and release of platelets from pro-platelet tips [27]. As the expression profile alters to accommodate new functional processes of the megakaryocyte, so do the genes in which mutations mediate thrombocytopenia. As megakaryocytes reach their mature state, they undergo fundamental changes to be able to release platelets into circulation via bone marrow sinusoids. These changes are underpinned by cytoskeletal rearrangements and subsequent cellular signaling. Causative mutations of IT are actually most commonly found to functionally disrupt these processes. The fundamental action leading to the release of platelets is the production of proplatelet extensions from a demarcated membrane system. Once in the vascular niche, mature megakaryocytes produce protrusions by microtubule sliding [28]. β1-Tubulin is the main isoform within human megakaryocytes and mutations within the gene encoding TUBB1 are known to cause an autosomal dominant form of inherited thrombocytopenia [29]. Like β1- tubulin, F-actin is present throughout proplatelets and allows proplatelet branching. These actin polymerization processes are promoted by nucleation factors such as WASp, encoded by WAS, which is selectively expressed in hematopoietic cells [30]. Mutations in WAS generally follow a genotype–phenotype correlation, where variants which abolish WASp expression or result in the expression of a truncated protein are associated with Wiskott–Aldrich Syndrome (WAS). Those causing decreased levels of WASp result in X-linked thrombocytopenia (XLT). Missense variants are most common in XLT and those occurring in the first three coding exons and the correlating EVH1/WH1 domain tend to cause the mild X-linked thrombocytopenia, which is not associated with the immunodeficiency observed in WAS. Actin polymerization and the production of F-actin has recently been shown to be affected in a patient with a homozygous nonsense mutation within the gene encoding megakaryoblastic leukemia 1; MKL1 [31]. In addition to severe immunodeficiency, the patient suffers from mild-to-moderate thrombocytopenia (platelet counts 50–150 × 109/l). The cause of this is thought to be a loss of filamentous actin production due to a reduction in globular actin (G-actin) coupled with a reduction in megakaryocyte migration, which phenocopies mice deficient for murine Mkl1 [32, 33]. Myosins, in particular, IIA and IIB, are also involved in proplatelet formation by interacting with actin filaments generating contractile forces. One myosin, the heavy chain of nonmuscle myosin IIA (NMMHC-IIA) encoded by the gene MYH9, is thought to play a slightly different role as mutations actually increase proplatelet formation [34]. Mutations causing MYH9-related disease therefore cause thrombocytopenia through the most proximal part of platelet formation, where a disruption in the Rho-Rho kinase-myosin IIA pathway leads to a lack of reactivation of NMMHC-IIA. This is a process that is required for the budding of preplatelets and subsequent separation into platelets; therefore, mutations in MYH9 cause a loss of induced fragmentation of preplatelets promoting the formation of a reduced number of large platelets [35]. Cross linking of actin filaments is required for their binding to the actin cytoskeleton. This is mediated in structures known as actin-binding domains (ABDs), which contain α-actinin. One of the four isoforms, which is expressed in megakaryocytes, is ACTN1. To date, 13 pathogenic missense variants have been found within the encoding gene ACTN1; all cause a mild dominantly inherited thrombocytopenia where megakaryocyte tips are reduced in number, increased in size and exogenous expression of the mutations inhibits actin filament assembly [36]. Filamins, via an N-terminal ABD, anchor the filaments to the cell membrane. Variations in the most abundant filamin in platelets, FLNA, are known to cause periventricular nodular heteropia as well as an isolated thrombocytopenia [37]. Some patient megakaryocytes from cultured peripheral CD34+ blood cells show frayed structures and signs of cytoplasmic degradation favoring abnormal distribution into incorrectly packaged platelet-like fragments. However, a population of platelets negative for FLNA was observed, indicating it may not be necessary in proplatelet formation, therefore, playing a noncrucial role [37]. There may remain an important role for FLNA in platelet biogenesis as its proteolysis is protected by phosphorylation at S2152 by cAMP-dependent protein kinase A, which is similar to the protection observed in GP1bβ [38]. The kinase consists of two catalytic subunits, the y-isoform of which is encoded by PRKACG. To date, only one variant has been found within PRKACG, which results in rapid degradation of FLNA and a reduction in the percentage of proplatelet-bearing megakaryocytes [39]. Complementary to the physical formation of proplatelets is the relay of extracellular signals to the cytoskeleton via membrane-bound receptors. Two main receptors are affected by mutations in inherited thrombocytopenia. The first is the receptor for Von-Willebrand factor (VWF), GP1b-IX-V. Consisting of four subunits, GP1bα, GP1bβ, GPIX and GPV, activation of the receptor via interaction of VWF with GP1bα relays signals to aid in proplatelet formation. Variants in the encoding genes, GP1BA, GP1BB and GP9, are known to cause both monoallelic and biallelic forms of BSS and lead to a reduction in proplatelet-forming megakaryocytes in culture [40, 41]. The second receptor known to be affected in the spectrum of inherited thrombocytopenias is the integrin GPIIb-IIIa, the receptor for fibrinogen. Encoded by the genes ITGA2B and ITGB3, mutations are associated with Glanzmann thrombasthenia (GT). GT often causes bleeding with no alteration in platelet count; however, gain of function variants exists that lead to a thrombocytopenia, suggesting a role for the interaction of fibrinogen for platelet biogenesis [42]. Recently, it has been suggested that cytoskeleton rearrangement can be impaired, arresting actin turnover at the stage of polymerization, by permanent triggering of the aIIbβ3-mediated outside-in signaling [43]. Murine megakaryocytes transduced with an ITGB3 variant (c.del647-683) generated proplatelets with a reduced number of large tips and barbell-proplatelets, suggesting impaired cytoskeletal remodeling as the cause of thrombocytopenia in patients with GT. One gene whose involvement in platelet production currently, through the intrinsic apoptosis pathway, is controversial is Cytochrome C (CYCS). CYCS initiates apoptosis through the activation of caspase 9 and subsequently caspase 3, a dysregulation of which was originally suggested as a feature of platelet production [44, 45]. To date, two variants (Y49H [p.48 in original publication] and G42S [p.41 in original publication]) have been observed in the apoptosis-linked gene CYCS [45, 46]. Indeed, electron microscopy of patient bone marrow shows intramedullary naked megakaryocyte nuclei, indicative of dysregulated megakaryopoiesis and premature platelet release. Furthermore, mouse lung fibroblasts transduced with CycsY48H and CycsG41S show increased apoptotic activity in response to staurosporine. It was therefore first believed that apoptosis had a fundamental role in the timing of pro-platelet release, hypothesizing that inhibition of actin polymerization, a process required for the formation of proplatelets, activates the apoptotic pathway [47]. However, platelet production can proceed regardless of both the intrinsic and extrinsic apoptotic pathways, suggesting there is more to the role of apoptosis in megakaryopoiesis than currently meets the eye [48]. Genes with a molecular function outside of platelet production It is clear and obvious that the fundamental processes governing the production of platelets are easily affected by deleterious mutations leading to thrombocytopenia. This is summarized in Figure 1. However, disease affecting megakaryocytes and the formation of platelets do not cover all of the genes currently known to cause thrombocytopenia. These alternative genes do not play role in megakaryopoiesis/thrombopoiesis. Their molecular causes of disease are unique or not currently known and it is best to consider them individually and separate from the aforementioned genes.Figure 1. Megakaryopoiesis, platelet production and other causes of IT. Differentiation from HSCs to platelets proceeds by a number of intermediate cell types, which leads to the formation of megakaryocytes which fragment via proplatelet formation to produce mature platelets. This process is driven by a number of genes encoded a number of transcription factors and proteins. Defects in these genes have been shown to give rise to thrombocytopenia by mostly affecting the relevant stage of platelet production they are labeled under. Variants that do not play a role in platelet production by megakaryopoiesis are included in the third subgroup entitled platelet removal, death and other. HSC: Hematopoietic stem cell, CMP: Common myeloid progenitor, MEP: Megakaryocyte-erythroid progenitor. VWF is particularly well known due to its importance in hemostasis, illustrated with the functional deficiency of vWF known as von Willebrand disease (VWD), a well-characterized and common collection of bleeding disorders. One particular type of VWD is known as type 2B, which refers to gain-of-function mutations that increase the affinity of vWF for GP1b on the platelet surface. The net effect of this is the presence of giant platelets due to continuous interactions between vWF and GP1bα disrupting megakaryopoiesis [49]. More interestingly, though, large multimers of vWF associated with platelets are also observed in VWD type 2B patients [50]. In vWF knock-in mice, these specific complexes are taken up efficiently by macrophages in the liver and spleen, indicating an increased clearance of platelets from the circulation. One interesting gene with a mechanism of pathogenesis for thrombocytopenia is ADAMTS13. Variants in the metalloproteinase ADAMTS13 are the genetic cause of Upshaw–Schulman syndrome, alternatively known as congenital thrombotic thrombocytopenic purpura (TTP) [51]. TTP is a life-threatening systemic illness characterized by hemolytic anemia, neurological symptoms, renal dysfunction as well as thrombocytopenia. Normally in-circulation ADAMTS13 is the von Willebrand factor-cleaving protease. However, loss-of-function mutations within ADAMTS13 and subsequent TTP result in vWF not being cleaved at the Tyr 842-Met 843 peptide bond resulting in the accumulation of large vWF multimers similar to a gain-of-function mutations in vWF[52]. Unlike VWD type 2B, the reduction in platelet count observed in TTP may be due to platelet aggregation under sheer stress because of an inability to cleave vWF [53]. This has the propensity to explain vWF–platelet aggregation in the arterioles and capillaries, a phenotype characteristic of TTP. An increased clearance of platelets is also present in patients with activating mutations of STIM1 within Stormorken syndrome and more recently York syndrome [54, 55]. Platelets in these patients are circulating in a preactivated state due to a constitutively active store operated Ca(2+) release-activated Ca(2+) (CRAC) channel and increase in Ca2+ entry. The net effect of which causes a reduction in the number of circulating platelets, but not due to an early ageing of platelets but an increased clearance of pre-activated platelets within the spleen, which is recapitulated in Stim1Sax/+ mice [56]. One interesting case of thrombocytopenia is as a secondary symptom with Sitosterolemia. Sitosterolemia is a rare, autosomal recessive disease caused by mutations in two adjacent ATP-binding cassette transport genes; ABCG5 and ABCG8 [57]. Hematological abnormalities, including thrombocytopenia, are a symptom of this complex disorder which results in decreased excretion and therefore accumulation of dietary sterols. ABCG5 and ABCG8, which normally act as a heterodimeric efflux pump, are not present on blood cells or platelets. The effects of mutations are therefore considered to be because of the relative toxicity they possess in high levels to cellular membranes [58, 59]. Interestingly, in ABCG5/ABCG8 knock-out mice, which are fed on high-plant-sterol diet, platelets are noted to be hyper-activatable and show impaired function [60]. One of the most recently discovered genes known to cause inherited thrombocytopenia is SLFN14. SLFN14 has recently been identified as an endoribonuclease, functioning to destroy mRNA, which cannot be correctly translated, and causes degradation of ribosomal subunits [61]. Three consecutive heterozygous missense mutations were identified in three unrelated families predicted to encode substitutions K218E, K219N and V220D [62]. Alongside the reduced number of platelets, they also show platelet dysfunction and an increased immature platelet fraction, a noninvasive measure indicative of increased platelet clearance. However, the exact mechanism through which mutations in SLFN14 mediate thrombocytopenia and aberrant platelet function is unknown. There is similar lack of knowledge for the molecular effect of one of the most recently reported causative genes of IT, GNE, Mutations within the gene encoding Glucosamine (UDP-N-Acetyl)-2-Epimerase/N-Acetylmannosamine kinase (GNE) are most noted as the molecular cause of Sialuria (OMIM#269921) and Hereditary Inclusion Body Myopathy (HIBM; OMIM#600737) [63, 64]. Recently, however, two related individuals presented with GNE myopathy associated with congenital thrombocytopenia [65]. Both patients suffered from a severe reduction in platelet count; however, the hematological effect of mutations within GNE has yet to be suggested. Clearly, it is not solely platelet production that is affected by deleterious mutation. Lifespan can be inadvertently altered by pre-activation, the production of aggregated multimers and by toxicity. However, with platelet lifespan so intrinsically controlled by levels of pro-apoptotic proteins versus pro-survival proteins, why is it that no variants have thus far been discovered in this specific pathway? For us to understand this further, we need to consider how platelet lifespan is controlled and how it could theoretically be disturbed. In platelets, it was long suggested that the Bcl-2 family of pro-survival proteins was required to suppress the pro-apoptotic signals of BAK and BAX [66]. It was Mason et al. who first established the link of degradation of Bcl-XL (encoded in humans by BCL2L1) to platelet death [67]. This subsequent loss of the restrained pro-apoptotic signal was therefore suggested as a molecular clock for platelet lifespan, an idea which was pioneered by the group of B. Kile. [67]. Megakaryocyte-specific deletion of Bcl-xL in mice does indeed profoundly reduce the platelet count to 2% of that observed in wild-type mice [68]. Additionally, platelet survival is markedly increased in Bak and Bax knock out and megakaryocyte-specific depleted mice [68]. So in a process so highly controlled by expression and protein levels, why are there not any variants in Bcl-xL or Bak/Bax that are known to cause thrombocytopenia? There are two possible explanations. One is that loss-of-function variants in BCL2L1 are lethal. While Bcl-XL is involved in preventing mature megakaryocytes from apoptosis it is not essential alone for their growth and development, whereas presence of either Bcl-XL or Mcl-1, another member of the pro survival Bcl-2 family, is required for organism viability [69]. However, Bcl-XL’s functions are not limited solely to platelet lifespan and it follows a wide expression pattern both in adult and in embryonic tissue [70, 71]. Interestingly, Jak2, which is thought to positively regulate Bcl-XL expression in megakaryoblastic cells, is associated with decreased survival when selectively deleted in mouse hematopoietic cells [69, 72]. Therefore, at this moment in time, it is difficult to rule out the possibility that an organism-specific effect may be occurring within humans where genetic variations in BCL2L1 are simply not viable. The other point to consider is the possibility that we haven’t determined a molecular cause of disease within BCL2L1 or any other genes involved in platelet lifespan simply due to their rarity. The most recent large-scale targeted genomics study into inherited thrombocytopenia revealed that only 50% of patients have a defined genetic etiology of disease [73]. With the progression of next-generation sequencing and the introduction of large-throughput whole exome and whole genome sequencing, new variants are being discovered in these previously undiagnosed patients [7, 62]. Inherited thrombocytopenias though, as a spectrum of disorders, are still fundamentally underdiagnosed. Yet, as our knowledge increases, we learn more about the effect of deleterious variants gaining insights into new mechanisms within platelet production, function and lifespan. Acknowledgments The work in the author’s laboratories is supported by the British Heart Foundation (PG/13/36/30275; FS/13/70/30521) and BJ is supported by an MRC Doctoral Training Grant (DTG) in Biomedical and Life Sciences. We thank Dr Steve Thomas for critical and constructive comments on our manuscript and Professor Ben Kile for his guidance regarding platelet lifespan. Declaration of interest The authors report no declarations of interest. ==== Refs References Bernard J SOULIER J. Sur une nouvelle variete de dystrphie thrombocytaire-hemorragipare congenitale Sem Hop Paris 1948 24 3217 3223 18116504 Ogawa M. Differentiation and proliferation of hematopoietic stem cells Blood 1993 81 11 2844 2853 8499622 Kaushansky K Lok S Holly RD Broudy VC Lin N Bailey MC Promotion of megakaryocyte progenitor expansion and differentiation by the c-Mpl ligand thrombopoietin Nature 1994 369 6481 568 571 8202159 Arinobu Y Mizuno S Chong Y Shigematsu H Iino T Iwasaki H Reciprocal activation of GATA-1 and PU.1 marks initial specification of hematopoietic stem cells into myeloerythroid and myelolymphoid lineages Cell Stem Cell 2007 1 4 416 427 18371378 Tijssen MR Ghevaert C. Transcription factors in late megakaryopoiesis and related platelet disorders J Thrombosis Haemostasis: JTH 2013 11 4 593 604 Foudi A Kramer DJ Qin J Ye D Behlich AS Mordecai S Distinct, strict requirements for Gfi-1b in adult bone marrow red cell and platelet generation J Exp Med 2014 211 5 909 927 24711581 Zhang MY Churpek JE Keel SB Walsh T Lee MK Loeb KR Germline ETV6 mutations in familial thrombocytopenia and hematologic malignancy Nat Genet 2015 47 2 180 185 25581430 Tunnacliffe A Jones C Le Paslier D Todd R Cherif D Birdsall M Localization of Jacobsen syndrome breakpoints on a 40-Mb physical map of distal chromosome 11q Genome Res 1999 9 1 44 52 9927483 Nichols KE Crispino JD Poncz M White JG Orkin SH Maris JM Familial dyserythropoietic anaemia and thrombocytopenia due to an inherited mutation in GATA1 Nat Genet 2000 24 3 266 270 10700180 Stevenson WS Morel-Kopp MC Chen Q Liang HP Bromhead CJ Wright S GFI1B mutation causes a bleeding disorder with abnormal platelet function J Thrombosis Haemostasis: JTH 2013 11 11 2039 2047 Song WJ Sullivan MG Legare RD Hutchings S Tan X Kufrin D Haploinsufficiency of CBFA2 causes familial thrombocytopenia with propensity to develop acute myelogenous leukaemia Nat Genet 1999 23 2 166 175 10508512 Ihara K Ishii E Eguchi M Takada H Suminoe A Good RA Identification of mutations in the c-mpl gene in congenital amegakaryocytic thrombocytopenia Proc Natl Acad Sci U S A 1999 96 6 3132 3136 10077649 Dasouki MJ Rafi SK Olm-Shipman AJ Wilson NR Abhyankar S Ganter B Exome sequencing reveals a thrombopoietin ligand mutation in a Micronesian family with autosomal recessive aplastic anemia Blood 2013 122 20 3440 3449 24085763 Songdej N Rao AK. Hematopoietic transcription factor mutations and inherited platelet dysfunction F1000prime Rep 2015 7 66 26097739 Horvat-Switzer RD Thompson AA. HOXA11 mutation in amegakaryocytic thrombocytopenia with radio-ulnar synostosis syndrome inhibits megakaryocytic differentiation in vitro Blood Cells Mol Diseases 2006 37 1 55 63 Albers CA Paul DS Schulze H Freson K Stephens JC Smethurst PA Compound inheritance of a low-frequency regulatory SNP and a rare null mutation in exon-junction complex subunit RBM8A causes TAR syndrome. Nat Genet 2012 44 4 435–439, S1–S2 Homans AC Cohen JL Mazur EM. Defective megakaryocytopoiesis in the syndrome of thrombocytopenia with absent radii Br J Haematol 1988 70 2 205 210 3191031 Fiedler J Strauss G Wannack M Schwiebert S Seidel K Henning K Two patterns of thrombopoietin signaling suggest no coupling between platelet production and thrombopoietin reactivity in thrombocytopenia-absent radii syndrome Haematologica 2012 97 1 73 81 21933853 Cramer EM Harrison P Savidge GF Wilbourn B Debili N Vainchenker W Uncoordinated expression of alpha-granule proteins in human megakaryocytes Prog Clin Biol Res 1990 356 131 142 1699235 Gunay-Aygun M Falik-Zaccai TC Vilboux T Zivony-Elboum Y Gumruk F Cetin M NBEAL2 is mutated in gray platelet syndrome and is required for biogenesis of platelet alpha-granules Nat Genet 2011 43 8 732 734 21765412 Kahr WH Lo RW Li L Pluthero FG Christensen H Ni R Abnormal megakaryocyte development and platelet function in Nbeal2(-/-) mice Blood 2013 122 19 3349 3358 23861251 Guerrero JA Bennett C van der Weyden L McKinney H Chin M Nurden P Gray platelet syndrome: proinflammatory megakaryocytes and alpha-granule loss cause myelofibrosis and confer metastasis resistance in mice Blood 2014 124 24 3624 3635 25258341 Larocca LM Heller PG Podda G Pujol-Moix N Glembotsky AC Pecci A Megakaryocytic emperipolesis and platelet function abnormalities in five patients with gray platelet syndrome Platelets 2015 26 8 751 757 25806575 Nurden AT Nurden P. The gray platelet syndrome: clinical spectrum of the disease Blood Rev 2007 21 1 21 36 16442192 Noris P Perrotta S Seri M Pecci A Gnan C Loffredo G Mutations in ANKRD26 are responsible for a frequent form of inherited thrombocytopenia: analysis of 78 patients from 21 families Blood 2011 117 24 6673 6680 21467542 Levin C Koren A Pretorius E Rosenberg N Shenkman B Hauschner H Deleterious mutation in the FYB gene is associated with congenital autosomal recessive small-platelet thrombocytopenia J Thrombosis Haemostasis: JTH 2015 13 7 1285 1292 25876182 Machlus KR Italiano JE Jr. The incredible journey: From megakaryocyte development to platelet formation J Cell Biol 2013 201 6 785 796 23751492 Bender M Thon JN Ehrlicher AJ Wu S Mazutis L Deschmann E Microtubule sliding drives proplatelet elongation and is dependent on cytoplasmic dynein Blood 2015 125 5 860 868 25411426 Kunishima S Kobayashi R Itoh TJ Hamaguchi M Saito H. Mutation of the beta1-tubulin gene associated with congenital macrothrombocytopenia affecting microtubule assembly Blood 2009 113 2 458 461 18849486 Massaad MJ Ramesh N Geha RS. Wiskott-Aldrich syndrome: a comprehensive review Ann New York Acad Sci 2013 1285 26 43 23527602 Record J Malinova D Zenner HL Plagnol V Nowak K Syed F Immunodeficiency and severe susceptibility to bacterial infection associated with a loss-of-function homozygous mutation of MKL1 Blood 2015 126 13 1527 1235 26224645 Gilles L Bluteau D Boukour S Chang Y Zhang Y Robert T MAL/SRF complex is involved in platelet formation and megakaryocyte migration by regulating MYL9 (MLC2) and MMP9 Blood 2009 114 19 4221 4232 19724058 Cheng EC Luo Q Bruscia EM Renda MJ Troy JA Massaro SA Role for MKL1 in megakaryocytic maturation Blood 2009 113 12 2826 2834 19136660 Chen Z Naveiras O Balduini A Mammoto A Conti MA Adelstein RS The May-Hegglin anomaly gene MYH9 is a negative regulator of platelet biogenesis modulated by the Rho-ROCK pathway Blood 2007 110 1 171 179 17392504 Spinler KR Shin JW Lambert MP Discher DE. Myosin-II repression favors pre/proplatelets but shear activation generates platelets and fails in macrothrombocytopenia Blood 2015 125 3 525 533 25395423 Kunishima S Okuno Y Yoshida K Shiraishi Y Sanada M Muramatsu H ACTN1 mutations cause congenital macrothrombocytopenia Am J Human Genet 2013 92 3 431 438 23434115 Nurden P Debili N Coupry I Bryckaert M Youlyouz-Marfak I Sole G Thrombocytopenia resulting from mutations in filamin A can be expressed as an isolated syndrome Blood 2011 118 22 5928 5937 21960593 Jay D Garcia EJ Lara JE Medina MA de la Luz Ibarra M. Determination of a cAMP-dependent protein kinase phosphorylation site in the C-terminal region of human endothelial actin-binding protein Arch Biochem Biophys 2000 377 1 80 84 10775444 Manchev VT Hilpert M Berrou E Elaib Z Aouba A Boukour S A new form of macrothrombocytopenia induced by a germ-line mutation in the PRKACG gene Blood 2014 124 16 2554 2563 25061177 Balduini A Malara A Pecci A Badalucco S Bozzi V Pallotta I Proplatelet formation in heterozygous Bernard-Soulier syndrome type Bolzano J Thrombosis Haemostasis: JTH 2009 7 3 478 484 Savoia A Kunishima S De Rocco D Zieger B Rand ML Pujol-Moix N Spectrum of the mutations in Bernard-Soulier syndrome Hum Mutat 2014 35 9 1033 1045 24934643 Kunishima S Kashiwagi H Otsu M Takayama N Eto K Onodera M Heterozygous ITGA2B R995W mutation inducing constitutive activation of the alphaIIbbeta3 receptor affects proplatelet formation and causes congenital macrothrombocytopenia Blood 2011 117 20 5479 5484 21454453 Bury L Falcinelli E Chiasserini D Springer TA Italiano JE Jr. Gresele P. Cytoskeletal perturbation leads to platelet dysfunction and thrombocytopenia in Glanzmann variants Haematologica 2015 Jiang X Wang X. Cytochrome C-mediated apoptosis Ann Rev Biochem 2004 73 87 106 15189137 Morison IM Cramer Borde EM Cheesman EJ Cheong PL Holyoake AJ Fichelson S A mutation of human cytochrome c enhances the intrinsic apoptotic pathway but causes only thrombocytopenia Nat Genet 2008 40 4 387 389 18345000 De Rocco D Cerqua C Goffrini P Russo G Pastore A Meloni F Mutations of cytochrome c identified in patients with thrombocytopenia THC4 affect both apoptosis and cellular bioenergetics Biochim Biophys Acta 2014 1842 2 269 274 24326104 Avanzi MP Izak M Oluwadara OE Mitchell WB. Actin inhibition increases megakaryocyte proplatelet formation through an apoptosis-dependent mechanism PloS One 2015 10 4 e0125057 25875470 Josefsson EC Burnett DL Lebois M Debrincat MA White MJ Henley KJ Platelet production proceeds independently of the intrinsic and extrinsic apoptosis pathways Nat Commun 2014 5 3455 24632563 Nurden P Debili N Vainchenker W Bobe R Bredoux R Corvazier E Impaired megakaryocytopoiesis in type 2B von Willebrand disease with severe thrombocytopenia Blood 2006 108 8 2587 2595 16720832 Casari C Du V Wu YP Kauskot A de Groot PG Christophe OD Accelerated uptake of VWF/platelet complexes in macrophages contributes to VWD type 2B-associated thrombocytopenia Blood 2013 122 16 2893 2902 23945153 Levy GG Nichols WC Lian EC Foroud T McClintick JN McGee BM Mutations in a member of the ADAMTS gene family cause thrombotic thrombocytopenic purpura Nature 2001 413 6855 488 494 11586351 Moake JL McPherson PD. Abnormalities of von Willebrand factor multimers in thrombotic thrombocytopenic purpura and the hemolytic-uremic syndrome Am J Med 1989 87 3N 9N 15N Tsai HM. Pathophysiology of thrombotic thrombocytopenic purpura Int J Hematol 2010 91 1 1 19 20058209 Nesin V Wiley G Kousi M Ong EC Lehmann T Nicholl DJ Activating mutations in STIM1 and ORAI1 cause overlapping syndromes of tubular myopathy and congenital miosis Proc Natl Acad Sci USA 2014 111 11 4197 4202 24591628 Markello T Chen D Kwan JY Horkayne-Szakaly I Morrison A Simakova O York platelet syndrome is a CRAC channelopathy due to gain-of-function mutations in STIM1 Mol Genet Metabol 2015 114 3 474 482 Grosse J Braun A Varga-Szabo D Beyersdorf N Schneider B Zeitlmann L An EF hand mutation in Stim1 causes premature platelet activation and bleeding in mice J Clinic Invest 2007 117 11 3540 3550 Berge KE Tian H Graf GA Yu L Grishin NV Schultz J Accumulation of dietary cholesterol in sitosterolemia caused by mutations in adjacent ABC transporters Science (New York, NY) 2000 290 5497 1771 1775 Matsuo M. ATP-binding cassette proteins involved in glucose and lipid homeostasis Biosci Biotechnol Biochem 2010 74 5 899 907 20460728 Su Y Wang Z Yang H Cao L Liu F Bai X Clinical and molecular genetic analysis of a family with sitosterolemia and co-existing erythrocyte and platelet abnormalities Haematologica 2006 91 10 1392 1395 17018391 Kanaji T Kanaji S Montgomery RR Patel SB Newman PJ. Platelet hyperreactivity explains the bleeding abnormality and macrothrombocytopenia in a murine model of sitosterolemia Blood 2013 122 15 2732 2742 23926302 Pisareva VP Muslimov IA Tcherepanov A Pisarev AV. Characterization of novel ribosome-associated endoribonuclease SLFN14 from rabbit reticulocytes Biochemistry 2015 54 21 3286 3301 25996083 Fletcher SJ Johnson B Lowe GC Bem D Drake S Lordkipanidze M SLFN14 mutations underlie thrombocytopenia with excessive bleeding and platelet secretion defects J Clinic Invest 2015 125 9 3600 3605 Seppala R Lehto VP Gahl WA. Mutations in the human UDP-N-acetylglucosamine 2-epimerase gene define the disease sialuria and the allosteric site of the enzyme Am J Human Genet 1999 64 6 1563 1569 10330343 Eisenberg I Avidan N Potikha T Hochner H Chen M Olender T The UDP-N-acetylglucosamine 2-epimerase/N-acetylmannosamine kinase gene is mutated in recessive hereditary inclusion body myopathy Nat Genet 2001 29 1 83 87 11528398 Izumi R Niihori T Suzuki N Sasahara Y Rikiishi T Nishiyama A GNE myopathy associated with congenital thrombocytopenia: a report of two siblings Neuromusc Disorders: NMD 2014 24 12 1068 1072 25257349 Wagner KU Claudio E Rucker EB 3rd Riedlinger G Broussard C Schwartzberg PL Conditional deletion of the Bcl-x gene from erythroid cells results in hemolytic anemia and profound splenomegaly Dev (Cambridge, England) 2000 127 22 4949 4958 Mason KD Carpinelli MR Fletcher JI Collinge JE Hilton AA Ellis S Programmed anuclear cell death delimits platelet life span Cell 2007 128 6 1173 1186 17382885 Josefsson EC James C Henley KJ Debrincat MA Rogers KL Dowling MR Megakaryocytes possess a functional intrinsic apoptosis pathway that must be restrained to survive and produce platelets J Exp Med 2011 208 10 2017 2031 21911424 Kodama T Hikita H Kawaguchi T Shigekawa M Shimizu S Hayashi Y Mcl-1 and Bcl-xL regulate Bak/Bax-dependent apoptosis of the megakaryocytic lineage at multistages Cell Death Different 2012 19 11 1856 1869 Michels J Kepp O Senovilla L Lissa D Castedo M Kroemer G Functions of BCL-X L at the interface between cell death and metabolism Int J Cell Biol 2013 2013 705294 23533418 Gonzalez-Garcia M Perez-Ballestero R Ding L Duan L Boise LH Thompson CB Bcl-XL is the major bcl-x mRNA form expressed during murine development and its product localizes to mitochondria Dev (Cambridge, England) 1994 120 10 3033 3042 Grisouard J Hao-Shen H Dirnhofer S Wagner KU Skoda RC. Selective deletion of Jak2 in adult mouse hematopoietic cells leads to lethal anemia and thrombocytopenia Haematologica 2014 99 4 e52 e54 24510341 Balduini CL Pecci A Noris P. Inherited thrombocytopenias: the evolving spectrum Hamostaseologie 2012 32 4 259 270 22972471 Thompson AA Nguyen LT. Amegakaryocytic thrombocytopenia and radio-ulnar synostosis are associated with HOXA11 mutation Nat Genet 2000 26 4 397 398 11101832 Ware J Russell SR Vicente V Scharf RE Tomer A McMillan R Nonsense mutation in the glycoprotein Ib alpha coding sequence associated with Bernard-Soulier syndrome Proc Natl Acad Sci USA 1990 87 5 2026 2030 2308962 Budarf ML Konkle BA Ludlow LB Michaud D Li M Yamashiro DJ Identification of a patient with Bernard-Soulier syndrome and a deletion in the DiGeorge/velo-cardio-facial chromosomal region in 22q11.2 Hum Mol Genet 1995 4 4 763 766 7633430 Wright SD Michaelides K Johnson DJ West NC Tuddenham EG. Double heterozygosity for mutations in the platelet glycoprotein IX gene in three siblings with Bernard-Soulier syndrome Blood 1993 81 9 2339 2347 8481514 Burk CD Newman PJ Lyman S Gill J Coller BS Poncz M. A deletion in the gene for glycoprotein IIb associated with Glanzmann’s thrombasthenia J Clinic Invest 1991 87 1 270 276 Loftus JC O’Toole TE Plow EF Glass A Frelinger AL 3rd Ginsberg MH. A beta 3 integrin mutation abolishes ligand binding and alters divalent cation-dependent conformation Sci (New York, NY) 1990 249 4971 915 918 Seri M Cusano R Gangarossa S Caridi G Bordo D Lo Nigro C Mutations in MYH9 result in the May-Hegglin anomaly, and Fechtner and Sebastian syndromes. The May-Heggllin/Fechtner Syndrome Consortium. Nat Genet 2000 26 1 103 105 10973259 Derry JM Ochs HD Francke U. Isolation of a novel gene mutated in Wiskott-Aldrich syndrome Cell 1994 78 4 635 644 8069912 Picard C McCarl CA Papolos A Khalil S Luthy K Hivroz C STIM1 mutation associated with a syndrome of immunodeficiency and autoimmunity New England J Med 2009 360 19 1971 1980 19420366 Cooney KA Nichols WC Bruck ME Bahou WF Shapiro AD Bowie EJ The molecular defect in type IIB von Willebrand disease Identification of four potential missense mutations within the putative GpIb binding domain. J Clinic Invest 1991 87 4 1227 1233

PMC005xxxxxx/PMC5000871.txt

==== Front PlateletsPlateletsIPLTiplt20Platelets0953-71041369-1635Taylor & Francis 114812910.3109/09537104.2016.1148129ArticlePlenary PaperHuman platelet activation by Escherichia coli: roles for FcγRIIA and integrin αIIbβ3 Watson Callum N. a Kerrigan Steven W. b Cox Dermot b Henderson Ian R. c Watson Steve P. a Arman Mònica a d * a Institute of Cardiovascular Sciences, College of Medical and Dental Sciences, University of Birmingham, Birmingham, UKb Cardiovascular Infection Group, Molecular and Cellular Therapeutics, Royal College of Surgeons in Ireland, Dublin, Irelandc Institute of Microbiology and Infection, College of Medical and Dental Sciences, University of Birmingham, Birmingham, UKd Centre for Cardiovascular and Metabolic Research, Hull-York Medical School, University of Hull, Hull, UKCorrespondence: Mònica Arman, Centre for Cardiovascular and Metabolic Research, Hull-York Medical School, Hardy Building, University of Hull, Cottingham Road, Hull, HU6 7RX, U.K. E-mail: Monica.Arman@hyms.ac.uk17 8 2016 30 3 2016 27 6 535 540 24 11 2015 22 1 2016 14 1 2016 © 2015 The Author(s). Published by Taylor & Francis.2015The Author(s)This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.Abstract Gram-negative Escherichia coli cause diseases such as sepsis and hemolytic uremic syndrome in which thrombotic disorders can be found. Direct platelet–bacterium interactions might contribute to some of these conditions; however, mechanisms of human platelet activation by E. coli leading to thrombus formation are poorly understood. While the IgG receptor FcγRIIA has a key role in platelet response to various Gram-positive species, its role in activation to Gram-negative bacteria is poorly defined. This study aimed to investigate the molecular mechanisms of human platelet activation by E. coli, including the potential role of FcγRIIA. Using light-transmission aggregometry, measurements of ATP release and tyrosine-phosphorylation, we investigated the ability of two E. coli clinical isolates to activate platelets in plasma, in the presence or absence of specific receptors and signaling inhibitors. Aggregation assays with washed platelets supplemented with IgGs were performed to evaluate the requirement of this plasma component in activation. We found a critical role for the immune receptor FcγRIIA, αIIbβ3, and Src and Syk tyrosine kinases in platelet activation in response to E. coli. IgG and αIIbβ3 engagement was required for FcγRIIA activation. Moreover, feedback mediators adenosine 5’-diphosphate (ADP) and thromboxane A2 (TxA2) were essential for platelet aggregation. These findings suggest that human platelet responses to E. coli isolates are similar to those induced by Gram-positive organisms. Our observations support the existence of a central FcγRIIA-mediated pathway by which human platelets respond to both Gram-negative and Gram-positive bacteria. Keywords Blood plateletsEscherichia coliFc gamma receptor IIAimmunitythrombosisBritish Heart Foundation10.13039/501100000274PG/13/42/30309This work was supported by the British Heart Foundation (PG/13/42/30309). ==== Body Introduction Platelets have been long known to be activated by bacteria [1]. This is likely to contribute to a balanced immune response [2], but it is also associated with pathological conditions such as infective endocarditis, atherothrombosis, and sepsis [3–5]. In the latter, disseminated microvascular thrombosis has a role in pathophysiology of sepsis and might be mediated through direct platelet–bacterium interactions. Recently, great emphasis has been placed on understanding the molecular mechanisms by which platelets are activated by bacterial cells. Elucidation of such mechanisms would provide opportunities to regulate them during infection. These mechanisms are diverse and include activation by whole bacteria or their released products [1, 6]. Despite multiple bacterial-strain specific molecular interactions, human platelet FcγRIIA is required for activation by a number of different Gram-positive species [7–15] and might contribute to the thrombotic complications found in infective diseases [15]. FcγRIIA is a low-affinity receptor for the constant region of IgGs that recognizes IgG-coated bacteria or their products through avidity. Upon ligand engagement, FcγRIIA signals through Src and Syk tyrosine kinases via a dual YxxL sequence known as an immunoreceptor tyrosine-based activation motif (ITAM) that is present in its cytoplasmic tail [16]. Gram-negative E. coli are commensal bacteria of the human and other mammalian gastrointestinal tracts. They rarely cause disease, except in cases of damaged gastrointestinal barriers or immunocompromised hosts. However, pathogenic strains of E. coli can cause three general clinical syndromes: enteric/diarrheal disease, urinary tract infections, and sepsis/meningitis [17]. In the latter, E. coli strains are the most common Gram-negative bacteria isolated from patients with bacteremia, sepsis, and neonatal meningitis [17–20], causing a major clinical burden and thousands of deaths per year. However, scarce information is available on the molecular interactions between E. coli and platelets [21]. The aim of this study was two-fold: to investigate human platelet activation by whole E. coli clinical isolates, and to investigate if FcγRIIA mediates platelet activation. Materials and methods Reagents All reagents were from described sources [14]. Fibrinogen was from Calbiochem (Merck Millipore, Nottingham, UK) and was depleted of IgGs by incubation with protein A (rec-Protein A-Sepharose 4B Conjugate, Life Technologies [Paisley, UK]). Bacterial culture and preparation E. coli strains, CFT073 (isolated from a patient with acute pyelonephritis and bacteremia [22]) and RS218 (isolated from a case of neonatal meningitis [23]), were cultured aerobically at 37°C overnight in an LB broth. Bacteria were washed and adjusted in PBS to an optical density (OD) of 1.6 at a wavelength of 600 nm. Bacteria were used at a 10-fold dilution in aggregation assays unless otherwise indicated. Assays of platelet function Platelet preparation from healthy volunteers was performed as previously described [14]. The study design was approved by the relevant ethics committee (ERN_11-0175). Platelet aggregation was assessed by light transmission in a PAP-8 aggregometer for up to 30 min. Time-matched controls were run alongside. Stimulation by cross-linking of FcγRIIA was performed by preincubation of platelets for 3 min with mAb IV.3 (4 μg/mL) followed by anti-mouse IgG F(ab’)2 (30 μg/mL). When indicated, concentrations of both mAb IV.3 and anti-mouse IgG F(ab’)2 were doubled or reduced to half. ATP release was assessed at the end of the recording using a luciferin–luciferase based assay [9]. Eptifibatide (9 μM), dasatinib (4 μM), and PRT-060318 (10 μM) were used at supramaximal concentrations. Cell lysates and protein phosphorylation studies were performed as previously published [14]. Statistical analysis Statistical analysis was performed using GraphPad (Prism). Data are presented as mean ± standard deviation (SD), and comparisons between mean values were performed using Student’s t-test or ANOVA when multiple samples were compared. p < 0.05 (two-tailed) was considered to be significant. Results and discussion E. coli bacteria stimulate αIIbβ3-mediated platelet aggregation via FcγRIIA and Src and Syk tyrosine kinases Previous studies have shown a characteristic pattern of platelet activation by Gram-positive bacteria, i.e. they induce “all-or-nothing” aggregation of platelets following a lag time that decreases with increasing concentrations of bacteria [14]. We hypothesized that Gram-negative organisms could trigger platelet aggregation in a similar manner. To test this, we investigated two blood-borne isolates of E. coli, CFT073, which was isolated from a patient with urinary tract infection and bacteremia, and RS218, which was isolated from a child with meningitis. We found that both strains induced “all-or-nothing” platelet aggregation in plasma after a lag phase (Figure 1 A.i and B.i). In contrast, as exemplified in Figure 1C for cross-linking of mAb IV.3 to cluster FcγRIIA, most platelet agonists cause rapid activation, which can give rise to partial aggregation when low concentrations of agonist are used. This suggests that bacteria have a unique positive feedback mechanism that gives rise to an “all-or-nothing” response.Figure 1. E. coli clinical isolates stimulating platelet aggregation in plasma. (A) Effect of E. coli CFT073 concentration on the lag time for the onset of platelet aggregation. (A.i) Bacterial suspensions at OD600nm 1.6 were used at different dilutions in platelet aggregation assays as indicated. Platelet-rich plasma (PRP) was 80% of the final volume and was kept constant in all reactions. (A.ii) Lag time for the onset of aggregation was measured in platelets from seven different donors using a 10-fold dilution of the bacterial suspension. Aggregation was observed in six out of seven donors, and lag times are indicated. Experiments were performed on different days (mean ± SD, n = 6). (B) Effect of E. coli RS218 concentration on the lag time for the onset of platelet aggregation. (B.i) Assays were performed as indicated in A.i. for E. coli RS218. (B.ii) Lag time for the onset of aggregation was measured as explained in A.ii. Aggregation was observed in six out of seven donors, and lag times are indicated. Experiments were performed on different days (mean ± SD, n = 6). (C) Effect of crosslinked mAb IV.3 concentration on platelet aggregation. In order to crosslink the FcγRIIA receptor, platelets were pre-incubated for 3 min with 2, 4 or 6 μg/mL mAb IV.3 followed by addition of 15, 30, or 45 μg/mL anti-mouse IgG F(ab’)2, respectively. Platelet-rich plasma (PRP) was 80% of the final volume and was kept constant in all reactions. The rest of the study was performed with the intermediate bacterial concentration; e.g. bacterial suspensions at OD600nm 1.6 were used at a 10-fold dilution in aggregation assays. Under these experimental conditions and performing the reactions in the presence of plasma, both strains induced platelet aggregation in six out of seven donors tested. For E. coli CFT073, the lag time for the onset of aggregation varied from 130 to 330 sec (mean ± SD: 200 sec ± 69, n = 6) (Figure 1A.ii). E. coli RS218 induced aggregation with lag times ranging from 120 to 270 sec (mean ± SD: 180 sec ± 58, n = 6) (Figure 1B.ii). As shown in Figure 2A and B, platelet aggregation to E. coli CFT073 and RS218 was blocked in the presence of the αIIbβ3 antagonist eptifibatide, which confirmed that the change in light transmission was due to αIIbβ3-mediated platelet–platelet binding rather than passive agglutination.Figure 2. E. coli-induced platelet aggregation in plasma depends on FcγRIIA and Src and Syk tyrosine kinases. (A) Effect of αIIbβ3 antagonist, eptifibatide, and anti-FcγRIIA mAb IV.3 on E. coli CFT073-induced platelet aggregation. PRP was incubated for 2 min with eptifibatide (9 μM) or 10 min with mAb IV.3 (10 μg/mL) or vehicle prior to addition of bacteria, and platelet aggregation was monitored by light transmission aggregometry. The results on the right-hand side graph are shown as mean ± SD of five independent experiments; *p < 0.05. One representative experiment is shown on the left. (B) Effect of αIIbβ3 antagonist, eptifibatide, and anti-FcγRIIA mAb IV.3 on E. coli RS218-induced platelet aggregation. The same experimental conditions as in Figure 2A were used for E. coli RS218. The results on the right-hand side graph are shown as mean ± SD of four independent experiments; *p < 0.05. One representative experiment is shown on the left. (C) Effect of the Src-tyrosine kinase inhibitor, dasatinib, and the Syk-tyrosine kinase inhibitor, PRT-060318, in E. coli CFT073-induced platelet aggregation in plasma. PRP was incubated for 2 min with dasatinib (4 μM) or PRT-060318 (10 μM) or vehicle prior to addition of bacteria, and platelet aggregation was monitored. The results on the right-hand side graph are shown as mean ± SD of four independent experiments; *p < 0.05. One representative experiment is shown on the left. (D) Effect of the Src-tyrosine kinase inhibitor, dasatinib, and the Syk-tyrosine kinase inhibitor, PRT-060318, in E. coli RS218-induced platelet aggregation in plasma. The same experimental conditions as in Figure 2C were used for E. coli RS218. The results on the right-hand side graph are shown as mean ± SD of four independent experiments; *p < 0.05. One representative experiment is shown on the left. To analyze whether FcγRIIA and its signaling pathway components, Src and Syk tyrosine kinases, have a role in E. coli induced platelet aggregation, specific inhibitors were employed. Aggregation induced by E. coli CFT073 or RS218 strains was abolished when platelets were pre-incubated with either mAb IV.3 (FcγRIIA inhibitor, Figure 2A and 2B), dasatinib (Src inhibitor, Figure 2C and 2D), or PRT-060318 (Syk inhibitor, Figure 2C and 2D) in the presence of plasma. Platelet activation by E. coli requires the interplay between αIIbβ3 and FcγRIIA A key role for FcγRIIA in platelet activation by E. coli is further supported by the observation that dense granule secretion was inhibited by mAb IV.3 (Figure 3A.i and A.ii) and that FcγRIIA became phosphorylated in response to both E. coli CFT073 and RS218 strains (Figure 3B). Furthermore, while washed platelets were not able to support bacteria-mediated aggregation, aggregation to E. coli RS218 was restored in the presence of human IgGs (e.g. pooled human IgGs purified from healthy donors) alone or with fibrinogen (Figure 3C.ii). Simultaneous addition of human IgGs and fibrinogen was necessary for E. coli CFT073 to induce aggregation in washed platelets (Figure 3C.i). These observations suggest that the initiating event in activation is engagement of FcγRIIA by plasma IgG bound to bacteria.Figure 3. Platelet activation by E. coli requires the interplay between αIIbβ3 and FcγRIIA. (A) αIIbβ3 and FcγRIIA mediate E. coli induced platelet secretion. Platelet-rich plasma was incubated for 2 min with eptifibatide (9 μM) or 10 min with mAb IV.3 (10 μg/mL) or vehicle prior to addition of bacteria, and platelet aggregation was monitored by light transmission aggregometry. Reactions were also performed by crosslinking FcγRIIA receptor. Platelets were pre-incubated for 3 min with 4 μg/mL monoclonal antibody (mAb) IV.3 followed by addition of 30 μg/mL anti-mouse IgG F(ab’)2. Supernatants were collected at time of full aggregation, or a parallel time point in the case of inhibition. Supernatants were analyzed for ATP release by luciferin–luciferase assay. ATP levels released by TRAP (100 μM) stimulated platelets were used to normalize data. The results are shown as mean ± SD of five independent experiments for E. coli CFT073 and RS218 and three independent experiments for crosslinked mAb IV.3; *p < 0.05. (B) E. coli induces tyrosine phosphorylation of FcγRIIA that is dependent on αIIbβ3. Aggregation reactions were performed in plasma in the presence or absence of eptifibatide (9 μM) and cell lysates collected at time of full aggregation, or equivalent times in the case of eptifibatide-treated samples where aggregation was inhibited. Immunoprecipitations for FcγRIIA were performed and tyrosine-phosphorylation detected by Western blot. Representative results of three independent experiments are shown. (C) hIgGs reconstitute aggregation to E. coli in washed platelets. Aggregation reactions to E. coli CFT073 and E. coli RS218 were performed in washed platelets in the presence or absence of IgG-depleted fibrinogen (1 mg/mL) and pooled human IgGs from healthy donors (hIgG) (0.1 mg/mL). Three independent experiments were performed per strain using different platelet donors. One representative experiment is shown for E. coli RS218. For E. coli CFT073, platelet aggregation was restored in two out of three platelet donors when supplementing the reactions with fibrinogen plus hIgG, and one representative experiment is shown. Research on Gram-positive bacteria has shown that platelet activation is often the result of multiple bacterium–platelet molecular interactions. These include the combination of bacterial strain-specific molecular interactions and shared IgG-FcγRIIA mediated signaling events [6, 14]. Among the former, strain-specific streptococci and staphylococci proteins are found that bind directly or indirectly (e.g. via fibrinogen) to platelet surface receptors such as αIIbβ3 or GPIb [6, 24]. Our results suggest that the mechanisms of platelet activation by E. coli might also have a strain-dependent component. However, future work is necessary to characterize the exact molecular interactions between platelets and these two E. coli strains, including the identification of potential bacterial components binding (directly or indirectly) to platelet surface receptors. Previous studies demonstrated an unexpected role for αIIbβ3 in controlling platelet dense granule secretion and FcγRIIA phosphorylation in response to a wide range of Gram-positive bacteria [13, 14]. We analyzed whether the same pattern of regulation could take place for E. coli. Indeed, E. coli CFT073 and RS218-induced dense granule secretion was inhibited by eptifibatide (Figure 3A.i and 3A.ii) demonstrating that secretion is dependent on αIIbβ3 engagement. In contrast, secretion induced by cross-linked mAb IV.3 was not decreased by eptifibatide (Figure 3A.iii). Moreover, bacteria-induced tyrosine phosphorylation of FcγRIIA was also dependent on αIIbβ3 engagement as observed by the inhibition of phosphorylation in the presence of eptifibatide (Figure 3B). The inability to detect secretion and FcγRIIA tyrosine phosphorylation in the absence of αIIbβ3 engagement might reflect the weak nature of the pathway initiated after FcγRIIA engagement by IgG-coated bacteria in the absence of feedback signals. The mechanism by which initial αIIbβ3 engagement takes place is, however, unclear. Strain-specific events mediating αIIbβ3 activation are thought to exist for Gram-positive bacteria. For most cases, it is likely that αIIbβ3 inside-out activation is achieved by FcγRIIA signaling, as well as by signaling from other strain-specific bacterial ligand–platelet receptor pairs. However, some bacteria such as Streptococcus gordonii DL1 and Staphylococcus aureus Newman can bind directly or indirectly (e.g. via fibrinogen) to αIIbβ3, which could facilitate αIIbβ3 activation [1, 10, 25–27]. Further investigations are necessary to characterize the mechanisms that lead to αIIbβ3 activation by E. coli CFT073 and RS218, including the identification of potential E. coli ligands binding to αIIbβ3. Platelet activation by E. coli is also dependent on secondary mediators ADP and TxA2 Platelet activation is reinforced by secondary mediators, which include release of stored ADP from dense granules and de novo synthesis of TxA2. Positive feedback mechanisms driven by secondary mediators are normally required for full and/or sustained platelet aggregation to low concentrations of agonists. This can be seen in platelets stimulated with an intermediate concentration of thrombin-related peptide (TRAP). Pre-treatment of platelets with the ADP-receptor P2Y12 antagonist, Cangrelor, and/or with cyclooxygenase inhibitor, indomethacin, did not affect the initial TRAP-induced aggregation but was followed by slow deaggregation that was not seen in controls (Figure 4C).Figure 4. Platelet activation by E. coli depends on secondary mediators ADP and TxA2. Platelet-rich plasma was incubated for 2 min with the cyclooxygendase inhibitor indomethacin (10 μM), ADP-receptor P2Y12 inhibitor Cangrelor (1 μM), or vehicle (DMSO) prior to addition of E. coli CFT073 (A), E. coli RS218 (B), or 50 μM TRAP (C), and platelet aggregation was monitored. The results are shown as mean ± SD of three independent experiments; *p < 0.05. In contrast, secondary mediators are key for platelet activation to Staphylococcus and Streptococcus strains [14]. Here, inhibitors were used to investigate a role for the two feedback agonists in E. coli-induced platelet aggregation. Activation in response to E. coli CFT073 and RS218 was completely abrogated in the presence of either Cangrelor or indomethacin, or by the combination of both (Fig 4A and 4B). Thus, platelet activation by E. coli bacteria is also dependent on ADP and TxA2. Altogether, these data suggest that the combination of FcγRIIA activation upon recognition of IgG-coated bacteria plus αIIbβ3 engagement (e.g. either through binding to bacteria or as a result of inside-out platelet signaling events as discussed before) results in a weak signal leading to release of ADP and TxA2. At this stage, feedback mechanisms are key in order to achieve full activation. Furthermore, ADP and TxA2 signal to neighboring (bacteria-free) platelets and induce αIIbβ3 inside-out activation and consequent platelet–platelet aggregation. Interestingly, FcγRIIA has also been shown to function as an adaptor protein amplifying αIIbβ3 signaling independent of extracellular IgG engagement [28, 29]. This suggests that FcγRIIA and αIIbβ3 could support both initial platelet–bacterium interaction and subsequent platelet–platelet aggregation by means of cooperative integrin/immunoreceptor tyrosine-based activation motif signaling. In summary, in this study, we provide evidence that E. coli induces activation of platelets through the same shared pathway described for various Gram-positive Staphylococcus and Streptococcus species [14]. This pathway involves IgG-dependent FcγRIIA activation of Src and Syk kinases, and is reinforced by αIIbβ3 engagement and secondary mediators. Despite the fact that FcγRIIA-mediated aggregation was previously observed for Gram-negative Helicobacter pylori [30] and Porphyromonas gingivalis [31], the signaling pathway and role of αIIbβ3 in activation has been only evaluated in few Gram-positive species [10, 13, 14]. The demonstration of a common mode of platelet activation to Gram-positive and Gram-negative species further identifies FcγRIIA as a candidate receptor for prevention of bacteria-mediated platelet activation in thrombosis and related disorders. We found that platelets from one out of seven donors did not respond to either E. coli CFT073 or E. coli RS218. Donor variation in bacteria-induced platelet aggregation is common and has been reported before [32, 33]. Future studies using a larger number of donors will be required in order to evaluate the potential effect of plasma IgG levels and FcγRIIA polymorphisms and/or surface expression levels on human platelet aggregation in response to E. coli. However, previous studies have not found a clear correlation between donor response and the above parameters in the case of Gram-positive bacteria [32, 33] and, for Streptococcus sanguinis, IgG levels can only partially account for donor variability [33]. As previously mentioned, individual bacterial strains can mediate platelet–bacterium interactions by multiple receptor–ligand pairs [1, 6], each one having a different contribution to the adhesion and/or platelet activation steps. Although we have found that FcγRIIA has a key role in human platelet activation by E. coli CFT073 and RS218, other platelet receptors might be simultaneously interacting with these strains, and this should be investigated in future studies. Furthermore, the bacterial components that are being targeted by IgGs-FcγRIIA remain to be identified. Previous literature shows that E. coli secreted Shiga-toxin, which is a virulent factor associated with hemolytic uremic syndrome [34], induces platelet activation [35]. And this might contribute to the formation of platelet thrombi in kidney glomerular capillaries, small arterioles, and arteries [34]. Additionally, human platelets bind lipopolysaccharide (LPS) from enterohemorrhagic E. coli via toll-like receptor (TLR) 4 and CD62, leading to cell activation [36]. TLR4-mediated platelet cytokine secretion has been described in response to E. coli LPS [37]. While E. coli CFT073 and RS218 do not produce Shiga-toxin, a role for E. coli CFT073 and RS218 cell wall LPS in platelet activation cannot be discarded, either in relation to FcγRIIA-mediated events (via IgG) or independently of the IgG receptor. As a first approach to test the role of TLR4 in platelet activation in response to our E. coli strains, we used an inhibitory anti-TLR4 antibody, HTA125, and found that it had no effect on platelet aggregation, and did not prolong the lag time response to either E. coli CFT073 or E. coli RS218 using two different donors (data not provided). This suggests that TLR4 is not essential for platelet activation by the bacterial strains tested. However, it is still possible that E. coli LPS-IgG immune complexes are formed that can engage platelet FcγRIIA directly. In any case, the exact role of E. coli CFT073 and RS218 LPS and platelet TLR4 in mediating platelet activation still needs to be deciphered. Interestingly, one previous study showed that FcγRIIA was required for platelet-mediated killing of IgG-opsonized E. coli K12 [38]. This suggests that platelet activation by bacteria might have different outcomes depending on the overall scenario: while unbalanced thrombi formation might have detrimental effects in cases such as HUS or sepsis, platelet activation by E. coli coated with IgG found in sera from healthy individuals (i.e. such ones used in this study) might contribute to host defense. Future investigations are necessary to further decipher the molecular interactions between E. coli and platelets, including potential synergism between IgG-FcγRIIA and LPS-TLR mediated pathways, and their role in homeostasis or disease. Furthermore, in view of the present study, care should be taken when using animal models for the study of platelet function during E. coli infections as FcγRIIA is not found in mice. Funding This work was supported by the British Heart Foundation (PG/13/42/30309). Declaration of interest The authors report no declarations of interest. ==== Refs References Cox D Kerrigan SW Watson SP. Platelets and the innate immune system: mechanisms of bacterial-induced platelet activation J Thromb Haemost 2011 9 1097 1107 21435167 Engelmann B Massberg S. Thrombosis as an intravascular effector of innate immunity Nat Rev Immunol 2013 13 34 45 23222502 Seymour GJ Ford PJ Cullinan MP Leishman S West MJ Yamazaki K. Infection or inflammation: the link between periodontal and cardiovascular diseases Future Cardiol 2009 5 5 9 19371196 Meier CR Derby LE Jick SS Vasilakis C Jick H. Antibiotics and risk of subsequent first-time acute myocardial infarction JAMA 1999 281 427 431 9952202 Que YA Moreillon P. Infective endocarditis Nat Rev Cardiol 2011 8 322 336 21487430 Kerrigan SW. The expanding field of platelet-bacterial interconnections Platelets 2015 26 293 301 25734214 Shannon O Hertzen E Norrby-Teglund A Morgelin M Sjobring U Bjorck L. Severe streptococcal infection is associated with M protein-induced platelet activation and thrombus formation Mol Microbiol 2007 65 1147 1157 17662041 Tilley D Arman M Smolenski A Cox D O’Donnell J Douglas C Watson S Glycoprotein Ibalpha Kerrigan S. FcgammaRIIa play key roles in platelet activation by the colonizing bacterium, Streptococcus oralis J Thromb Haemost 2013 11 941 950 23413961 Keane C Tilley D Cunningham A Smolenski A Kadioglu A Cox D Jenkinson HF Kerrigan SW. Invasive Streptococcus pneumoniae trigger platelet activation via Toll-like receptor 2 J Thromb Haemost 2010 8 2757 2765 20946179 Keane C Petersen H Reynolds K Newman DK Cox D Jenkinson HF Newman PJ Kerrigan SW. Mechanism of outside-in {alpha}IIb{beta}3-mediated activation of human platelets by the colonizing Bacterium, Streptococcus gordonii Arterioscler Thromb Vasc Biol 2010 30 2408 2415 21071690 Kerrigan SW Douglas I Wray A Heath J Byrne MF Fitzgerald D Cox D. A role for glycoprotein Ib in Streptococcus sanguis-induced platelet aggregation Blood 2002 100 509 516 12091342 Kerrigan SW Clarke N Loughman A Meade G Foster TJ Cox D. Molecular basis for Staphylococcus aureus-mediated platelet aggregate formation under arterial shear in vitro Arterioscler Thromb Vasc Biol 2008 28 335 340 18063809 Pampolina C McNicol A. Streptococcus sanguis-induced platelet activation involves two waves of tyrosine phosphorylation mediated by FcgammaRIIA and alphaIIbbeta3 Thromb Haemost 2005 93 932 939 15886812 Arman M Krauel K Tilley DO Weber C Cox D Greinacher A Kerrigan SW Watson SP. Amplification of bacteria-induced platelet activation is triggered by FcgammaRIIA, integrin alphaIIbbeta3, and platelet factor 4 Blood 2014 123 3166 3174 24642751 Arman M Krauel K. Human platelet IgG Fc receptor FcgammaRIIA in immunity and thrombosis J Thromb Haemost 2015 13 893 908 25900780 Brooks DG Qiu WQ Luster AD Ravetch JV. Structure and expression of human IgG FcRII(CD32) Functional heterogeneity is encoded by the alternatively spliced products of multiple genes. J Exp Med 1989 170 1369 1385 2529342 Kaper JB Nataro JP Mobley HL. Pathogenic Escherichia coli Nat Rev Microbiol 2004 2 123 140 15040260 Martin GS Mannino DM Eaton S Moss M. The epidemiology of sepsis in the United States from 1979 through 2000 N Engl J Med 2003 348 1546 1554 12700374 de Kraker ME Jarlier V Monen JC Heuer OE van de Sande N Grundmann H. The changing epidemiology of bacteraemias in Europe: trends from the European Antimicrobial Resistance Surveillance System Clin Microbiol Infect 2013 19 860 868 23039210 McPherson D Griffiths C Williams M Baker A Klodawski E Jacobson B Donaldson L. Sepsis-associated mortality in England: an analysis of multiple cause of death data from 2001 to 2010 BMJ Open 2013;3:e002586 10.1136/bmjopen-2013-002586 Clawson CC White JG. Platelet interaction with bacteria I. Reaction phases and effects of inhibitors. Am J Pathol 1971 65 367 380 4400052 Mobley HL Green DM Trifillis AL Johnson DE Chippendale GR Lockatell CV Jones BD Warren JW. Pyelonephritogenic Escherichia coli and killing of cultured human renal proximal tubular epithelial cells: role of hemolysin in some strains Infect Immun 1990 58 1281 1289 2182540 Silver RP Aaronson W Sutton A Schneerson R. Comparative analysis of plasmids and some metabolic characteristics of Escherichia coli K1 from diseased and healthy individuals Infect Immun 1980 29 200 206 6995336 Fitzgerald JR Foster TJ Cox D. The interaction of bacterial pathogens with platelets Nat Rev Microbiol 2006 4 445 457 16710325 Loughman A Fitzgerald JR Brennan MP Higgins J Downer R Cox D Foster TJ. Roles for fibrinogen, immunoglobulin and complement in platelet activation promoted by Staphylococcus aureus clumping factor A Mol Microbiol 2005 57 804 818 16045623 Miajlovic H Loughman A Brennan M Cox D Foster TJ. Both complement- and fibrinogen-dependent mechanisms contribute to platelet aggregation mediated by Staphylococcus aureus clumping factor B Infect Immun 2007 75 3335 3343 17438032 Petersen HJ Keane C Jenkinson HF Vickerman MM Jesionowski A Waterhouse JC Cox D Kerrigan SW. Human platelets recognize a novel surface protein, PadA, on Streptococcus gordonii through a unique interaction involving fibrinogen receptor GPIIbIIIa Infect Immun 2010 78 413 422 19884334 Boylan B Gao C Rathore V Gill JC Newman DK Newman PJ. Identification of FcgammaRIIa as the ITAM-bearing receptor mediating alphaIIbbeta3 outside-in integrin signaling in human platelets Blood 2008 112 2780 2786 18641368 Zhi H Rauova L Hayes V Gao C Boylan B Newman DK McKenzie SE Cooley BC Poncz M Newman PJ. Cooperative integrin/ITAM signaling in platelets enhances thrombus formation in vitro and in vivo Blood 2012 121 1858 1867 23264598 Byrne MF Kerrigan SW Corcoran PA Atherton JC Murray FE Fitzgerald DJ Cox DM. Helicobacter pylori binds von Willebrand factor and interacts with GPIb to induce platelet aggregation Gastroenterology 2003 124 1846 1854 12806618 Naito M Sakai E Shi Y Ideguchi H Shoji M Ohara N Yamamoto K Nakayama K. Porphyromonas gingivalis-induced platelet aggregation in plasma depends on Hgp44 adhesin but not Rgp proteinase Mol Microbiol 2006 59 152 167 16359325 Ford I Douglas CW Cox D Rees DG Heath J Preston FE. The role of immunoglobulin G and fibrinogen in platelet aggregation by Streptococcus sanguis Br J Haematol 1997 97 737 746 9217171 McNicol A Zhu R Pesun R Pampolina C Jackson EC Bowden GH Zelinski T. A role for immunoglobulin G in donor-specific Streptococcus sanguis-induced platelet aggregation Thromb Haemost 2006 95 288 293 16493491 Johnson S Waters A. Is complement a culprit in infection-induced forms of haemolytic uraemic syndrome? Immunobiology 2012 217 235 243 21852019 Karpman D Papadopoulou D Nilsson K Sjogren AC Mikaelsson C Lethagen S. Platelet activation by Shiga toxin and circulatory factors as a pathogenetic mechanism in the hemolytic uremic syndrome Blood 2001 97 3100 3108 11342436 Stahl AL Svensson M Morgelin M Svanborg C Tarr PI Mooney JC Watkins SL Johnson R Karpman D. Lipopolysaccharide from enterohemorrhagic Escherichia coli binds to platelets through TLR4 and CD62 and is detected on circulating platelets in patients with hemolytic uremic syndrome Blood 2006 108 167 176 16514062 Berthet J Damien P Hamzeh-Cognasse H Arthaud CA Eyraud MA Zeni F Pozzetto B McNicol A Garraud O Cognasse F. Human platelets can discriminate between various bacterial LPS isoforms via TLR4 signaling and differential cytokine secretion Clin Immunol 2012 145 189 200 23108090 Riaz AH Tasma BE Woodman ME Wooten RM Worth RG. Human platelets efficiently kill IgG-opsonized E coli. FEMS Immunol Med Microbiol 2012 65 78 83 22340259

PMC005xxxxxx/PMC5000873.txt

==== Front J Hunger Environ NutrJ Hunger Environ NutrWHENwhen20Journal of Hunger & Environmental Nutrition1932-02481932-0256Taylor & Francis 109514410.1080/19320248.2015.1095144ArticleArticlesComparing Prices for Food and Diet Research: The Metric Matters Jones N. R. V. a Monsivais P. a * a Centre for Diet and Activity Research, MRC Epidemiology Unit, University of Cambridge School of Clinical Medicine, Institute of Metabolic Science, Cambridge Biomedical Campus, Cambridge, UKCONTACT Pablo Monsivais pm491@medschl.cam.ac.ukCentre for Diet and Activity Research, MRC Epidemiology Unit, University of Cambridge School of Clinical Medicine, Box 285, Institute of Metabolic Science, Cambridge Biomedical Campus, CambridgeCB2 0QQ, UK.2 7 2016 25 4 2016 11 3 370 381 Published with license by Taylor & FrancisThe Author(s)This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.ABSTRACT An important issue in research into access to healthy food is how best to compare the price of foods. The appropriate metric for comparison has been debated at length, with proponents variously stating that food prices should be compared in terms of their energy content, their edible mass, or their typical portion size. In this article we assessed the impact of using different food price metrics on the observed difference in price between food groups and categories of healthiness, using United Kingdom consumer price index data for 148 foods and beverages in 2012. We found that the choice of metric had a marked effect on the findings and conclude that this must be decided in advance to suit the reason for comparing food prices. KEYWORDS Food pricesmeasurementresearch methodseconomicsfood securityaccessBritish Heart FoundationMR/K023187/1UK Clinical Research CollaborationMR/K023187/1National Institute for Health ResearchMR/K023187/1Economic and Social Research CouncilMR/K023187/1Medical Research CouncilMR/K023187/1Cancer Research UKMR/K023187/1The present study was undertaken by the Centre for Diet and Activity Research (CEDAR), a UKCRC Public Health Research Centre of Excellence. The authors gratefully acknowledge the funding from the British Heart Foundation, Cancer Research UK, Economic and Social Research Council, Medical Research Council, the National Institute for Health Research, and the Wellcome Trust, under the auspices of the UK Clinical Research Collaboration. The funding sources had no role in the design and conduct of the study or in the collection, management, analysis, and interpretation of the data. ==== Body Introduction An essential component of food access and food security is the notion that consumers have sufficient resources for a nutritious diet. 1 Thus, it is important to standardize methods for comparing the cost of different foods and beverages (hereafter collectively referred to as foods) to examine how prices may affect access to a nutritionally adequate diet. Having appropriate metrics for comparison is also important for developing public health policies that recommend substituting some foods for others, if they are to suitably account for the potential limiting factor of cost. Though food prices are currently monitored by governments, international agencies, and private organizations, such tracking is usually not suitable for comparing the costs of different foods in relation to their nutritional value. 2 This has implications for monitoring food security, given that the relationship between nutritional value and prices is a factor that will have a large impact on food security. An inadequate understanding of how the prices of different foods compare could also limit the effectiveness of food assistance and nutrition programmes, such as the Supplemental Nutrition Assistance Program in the United States or the UK’s Healthy Start, because these programs are predicated on the assistance being large enough to purchase foods that contribute to a healthy diet. As such, there is arguably a requirement to be able to fairly compare food prices, but there is an ongoing and vigorous debate as to the best way to make such comparisons, with different researchers favoring different metrics and finding alternatives to be misleading. 3 ,4 Though this is ostensibly an economic issue, the purpose of exploring it here is to improve the methodology available for research into the economic determinants of food insecurity, hunger, and malnutrition. In this article we consider earlier arguments for the different metrics of food price comparison and then apply these metrics to national food price data from the UK, exploring how the choice of metric can influence the findings of research on the question of whether healthier foods are more expensive. We then make suggestions as to how future research should express food prices to ensure that reported results are meaningful answers to the question being asked, providing reasons for and against using different metrics and examples of each. The debate so far The need for a common metric by which foods can be assessed is due to the different unit sizes of purchased goods, meaning that the price needs to be divided by some quality of the food in question so that it is comparable to others. For example, if we want to compare the price of an orange with the price of a bag of apples there are numerous ways to do this: we could compare the individual orange to the bag or in terms of mass, energy, typical portion size, or perhaps other quantities of the foods. Comparisons have typically been made by comparing the price divided by the energy content or mass of the food in question. 5 However, both of these approaches have received criticism 3 ,4 and neither has been adopted unanimously by researchers in this field. One of the earliest published efforts to consider the price of food in terms of its content was made by Atwater in 1894,6, 7 who recognized that the nutritive value of foods differed and that their cost ought to be framed in terms of the energy or nutrients they provide. The price per unit of energy metric has become a widely used approach for researchers interested in food prices and nutrition, with a range of different authors using it in different study designs and in different countries. 8 – 10 Moreover, the costing of foods in terms of price per unit energy in development economics suggests the utility of this metric for understanding how food prices affect diet in low-income populations. 4 ,11, 12 However, assessing price in terms of energy has been criticized on a number of grounds, namely, that people do not purchase foods in forms that are easily comparable in terms of calories, meaning that the price per calorie may not be relevant to consumer behavior. 13 Similarly, foods are not necessarily eaten in isoenergetic quantities, so comparisons between foods typically eaten in quantities that do not provide equal amounts of energy—for example, carrots and ice cream—is meaningless in practical terms. 14 Finally, there has been criticism of attempting to assess the relationship between how healthy a food is and its price when price is measured in terms of energy and the healthiness of a food is measured by its energy density, because this can lead to autocorrelation. 3 ,5 The critics of an energy-based metric instead often propose the use of mass as the most appropriate metric. 5 Proponents on both sides of the debate have engaged in studies using randomly generated data to investigate the apparent mathematical flaws in the other’s choice of metric, with opposing results. 7 ,15 A third alternative, the use of portion size as a price metric, has also been proposed on the grounds that comparisons of prices based on a fixed amount of calories or mass of food may have little behavioral relevance. Portion sizes are a better reflection of the quantities of foods that are typically consumed and these can vary substantially among different types of foods. 5 This method would allow for a more realistic comparison of different foods and for the cost figures to be more readily appraised. However, this metric requires up-to-date and accurate data concerning the quantity typically consumed, which may not always be available or appropriate for all populations, making it harder to adopt as a standard. In summary, this debate is sharply polarized and does not appear to be reaching a resolution. Accordingly, it is important to find common ground and identify instances where there is agreement about what most appropriate metric is the given the question being asked. Empirical analysis We analyzed government food price data from the UK to explore how the choice of food price metric can have on the results produced by research exploring which foods are more expensive. Methods Food price data We obtained national food price data used to calculate the UK Consumer Price Index (CPI) from the Office for National Statistics and matched them to a range of appropriate items in the National Diet and Nutrition Survey (NDNS), 16 using a method described previously. 17 Briefly, this was done by matching each CPI item to a range of NDNS foods deemed to be a good match. Following this, mean nutrient values were produced for each CPI item by calculating the mean of the nutrient values listed for the NDNS matches. To reflect the different ways in which a food may be prepared for consumption and the implications this would have for its nutrient content as listed in NDNS (which lists foods as consumed, after adjustments made in preparation), we weighted these mean values by the frequency with which each preparation method is recorded in NDNS, intending to accurately reflect the ways the foods are consumed by people in the UK. Prices were then adjusted for edible portion of each food using United States Department of Agriculture’s Handbook 102 18 to account for the fact that the price data were for foods as purchased and the NDNS data for foods as consumed. We assigned a portion size to each food item using portion sizes typically consumed in the UK, based on Wrieden and Barton. 19 Items are included in the CPI if they are frequently consumed by many households because the index aims to sample goods and services that are typical of expenditure in the UK. 20 As such, the foods included in it can reasonably be expected to represent those foods typically consumed. Classification of foods We determined which foods were more and less healthy using the Food Standards Agency’s WXYfm nutrient profiling model (hereafter referred to as the FSA Score), which provides a categorical definition of a food’s healthiness based upon energy, saturated fat, total sugar, sodium, fiber, protein, and fruit, vegetable, and nut content. 21 We assigned foods to food groups in line with the UK’s Eatwell Plate, a government-produced nutrition communication tool, using a table in The Livewell Report, which matched NDNS food categories to Eatwell food groups. 22 These steps resulted in a data set of 148 foods and beverages with information on their food groups, whether they are more or less healthy, and their mean 2012 price per unit of energy, per unit of mass, and per portion. Statistical analysis Tests for a significant difference in mean prices by group were conducted using a t test for comparing more and less healthy foods and analysis of variance for comparing the Eatwell food groups. All analyses were conducted using Stata (Ver SE 12.1). 23 Figures were produced using R (version 3.0.2 for Windows) and the ggplot2 package. 24 ,25 Results Table 1 reports the results of our analyses, containing the mean and standard error for each price metric by FSA score–defined healthiness and by food group. Prices were calculated on March 15, 2015, at an exchange rate of U.S.$1.47 per £1.The overall mean price per 100 g was £0.57 ($0.84), £0.50 ($0.74) per 100 kilocalories, and £0.47 ($0.69) per portion, with similar standard errors (0.05 for mass- and energy-based prices, 0.04 for portion-based prices). The results by FSA category show the impact of the choice of denominator on the observed relationship between how healthy a food is and its price: healthier foods were significantly more expensive in terms of energy (0.65 £/100 kcal ($0.96) versus 0.28 £/100 kcal ($0.41), P < .001) and per portion (0.55 £/portion ($0.81) versus 0.35 £/portion ($0.52), P < .035) yet significantly less expensive in terms of mass (0.46 £/100g ($0.68) versus 0.85 £/100g ($1.25), P < .001). When prices were examined by Eatwell group, the choice of price metric also exerted an effect on the results, with fruit and vegetables being the least expensive food group by mass (0.28 £/100g ($0.41)) and second least by portion (0.25 £/portion ($0.37)) yet the most expensive in terms of calories (0.83 £/100 kcal ($1.22)). Meat and other sources of a protein were the most expensive group by mass (0.93 £/100 g ($1.37)) and portion size (0.94 £/portion ($1.39)), yet in the middle of the price range given their energy content (0.52 £/100 kcal ($0.77)). In summary, these results show that there is a considerable difference in the cost of different foods depending on the price metric chosen.Table 1. Mean price, in UK Pounds Sterlinga, per 100 g, per 100 kcal, and per portion for foods and beverages by FSA score category and Eatwell food group, using prices from the Consumer Price Index for the third quarter of 2012, n = 148b.     Price per 100 g (£) Price per 100 kcal (£) Price per portion (£) n Mean Standard error Mean Standard error Mean Standard error Total 148 0.57 0.05 0.50 0.05 0.47 0.04 FSA score category  More healthy 87 0.46 0.05 0.65 0.08 0.55 0.07  Less healthy 61 0.85 0.09 0.28 0.03 0.35 0.04 P value of 2-tailed t test   <.001 <.001 .035 Eatwell food group  Bread, rice, potatoes, pasta 18 0.32 0.08 0.11 0.02 0.18 0.04  Fruit and vegetables 38 0.28 0.03 0.83 0.15 0.25 0.04  Milk and dairy foods 14 0.70 0.16 0.44 0.11 0.39 0.14  Meat, fish, eggs, beans, other sources of protein 36 0.93 0.08 0.52 0.06 0.94 0.11  Food and drinks high in fat and/or sugar 42 0.77 0.13 0.36 0.10 0.40 0.08 Pvalue of one-way analysis of variance   <.001 <.001 <.001 a$1.47 US per £ 1 UK on 15 March 2015. bFSA indicates Food Standards Agency. Discussion Our results show that the unit of comparison has an impact on which foods were the most expensive. The effect of this can be to completely change the relationship observed, which has considerable implications for any research on food costs or food security. This finding is in line with previous work,5, 14 which also found that fruits and vegetables went from being the most expensive to the least expensive food group depending on the price metric. As has been previously stated, 5 this finding is likely due to the low energy density of fruits and vegetables, which results in a low price per gram and a high price per calorie. Our findings have important implications for research comparing the price of different foods, given the impact that choosing an inappropriate metric would have on the results. Rather than promote any one food price metric to the exclusion of others, here we will suggest guidance as to when it is most appropriate to use the different metrics, given the research question being asked. Such guidance would enable researchers to select price metrics most appropriate for their research and also make the comparison of different studies easier, thus allowing information to be better pooled and used to inform food policy. When to use mass as a metric We now set out what we think are the circumstances in which each of the three price metrics assessed in this article are the most salient. When the research question makes a comparison between similar foods, the use of mass as the price comparison metric may be the most appropriate given that this allows the consumer to determine whether two products that will serve the same purpose within their diet differ in price. An example of this is the comparison between different packets of butter: both are nutritionally similar and will be served in similar portions but there may be a price difference between them. This metric seems most appropriate here because in some countries, including the UK 26 and Australia, 27 this unit price (price per unit weight) labeling is mandated and therefore is available to consumers when making decisions. However, the mass of the product bears little relationship to how it is consumed and the level of sustenance it provides, meaning that this metric is unlikely to be a meaningful way to compare very different types of food; for example, steak and lettuce. It should be noted that the price per unit of mass will not be available in all food outlets and in many countries and this should be taken into account by researchers deciding on a price comparison metric if they want it to have any behavioral relevance. When to use portions as a metric Dietary guidelines often include recommendations for consuming a specific number of portions of food per day or per week. Given the observed relationship between the price of more and less healthy foods, concerns may exist that the cost of recommended foods might prevent some groups from meeting these guidelines. 17 ,28 In such circumstances, the price-per-portion metric may be useful for estimating the likely impact on consumer costs, allowing guidelines to be adapted accordingly. For example, U.S. government analysis has estimated the cost of meeting the recommended 5 portions of fruit and vegetables per day. 29 Alternatively, when researchers wish to determine the cost of directly substituting one food for another, the use of portion sizes is appropriate given that the foods may be of different quantities of both energy and mass. This metric allows for a comparison to be made that is directly connected to how people typically consume the foods being compared. For example, when comparing the cost of serving of shepherd’s pie and a grilled cod fillet, the portion size is probably the most appropriate way to compare these foods, which differ in energy and nutrient contents but have a similar role in a meal. Studies have used this approach to examine the cost of substituting healthier foods for less-healthy foods in institutional settings 28 and the cost and nutritional effects of substituting fruit for fruit juice in the diets of children. 30 This approach is also likely to lead to results that are more readily interpretable, in contrast to prices being expressed in terms of energy or mass, which has advantages when it comes to dissemination to a lay audience. However, the use of a typical portion size relies on these data being available, which will not always be the case. When to use energy as a metric When addressing questions of public health and nutrition, pricing food on the basis of energy content appears to be the most appropriate approach because the comparison is between the sustenance the foods can contribute. This is based on the notion of physiological energy requirements for maintaining energy balance, with foods varying in the extent to which they contribute to achieving this energy target. In contrast, the mass and number of portions have no physiological basis. To give an example, this metric would be the best way to compare bread and apples, foods that are likely to be very different in terms how much energy they can provide when matched mass for mass or portion for portion. There are variations of this approach that can also be used, such as estimating the cost of obtaining a specified quantity of nutrients—for example, protein, 6 potassium, or fiber 31 ; however, energy is the key component of food required to sustain life in the short term, so for general questions concerning public health nutrition it will be suitable. However, as with portion sizes, the data required to calculate the price in terms of energy content will not always be available to researchers. Table 2 summarizes our considerations when selecting a food price metric. We hope that our suggestions as to the most appropriate metric for use in different situations may bring about an increased awareness of the notion that one metric cannot be used to answer all questions. However, these considerations are likely to be contested and we welcome debate over the advice we set out here.Table 2. Three commonly used food price metrics and indication of when their application is most appropriate. Food price metric Price per unit of mass Price per portion Price per unit of energy Best used for Comparing prices of nutritionally similar foods in the context of consumer choice, where mass is likely to be the only product information available to the purchaser. Comparing prices in the context of direct substitutions of one food for another. This is likely to be of use for analyzing food-based policies that promote such substitutions. Comparing prices in the context of public health and food security, where the quality of a diet consumed for survival is of concern. Potential concerns Limited relevance for assessing the cost of abating hunger or achieving sustenance. Requires accurate and appropriate portion size data for the population being studied and portions estimated for a different population may lead to inaccurate findings being reported if misapplied. Calorie information is not always available. Calorie-based comparisons might not be relevant for some consumer decision making. Beyond foods, it is important to consider the most appropriate metric with which to compare diets, because the overall diet is ultimately more important in terms of food security and long-term health than individual foods. We suggest that cost per unit of energy is the most appropriate metric for comparing the cost of different diets, given that the total diet should be within a specified level of energy intake based upon age, sex, basal metabolic rate, and physical activity levels, whereas the number of portions or the total mass of a diet can vary depending on the types of foods consumed, the typical portion size, and energy density of individual diets. This can be seen in the limited deviation in dietary energy observed in populations. Figure 1a shows the age- and sex-adjusted means of food energy consumed by adults in the UK by educational attainment based on data from the National Diet and Nutrition Survey (2008–2012). 32 Mean energy intake is tightly clustered across categories of educational attainment, with a maximum difference across groups of 168 kcal/day (10% variation). However, total mass of the diet (Figure 1b) varies substantially and is socially patterned, with higher income groups having a dietary mass that was 392 g/day (16%) higher than that of the lowest income group. The larger variation observed in dietary mass indicates that there is less physiological regulation of how much food mass can be consumed in contrast to energy, which appears to be more consistent across the population. Standardizing dietary costs for levels of energy intake also provides consistency with dietary guidance, in which recommended quantities of foods and nutrients are scaled in relation to energy intake.Figure 1. Bar chart indicating the age- and sex-adjusted (a) mean dietary energy intake and (b) mean dietary mass intake and 95% confidence interval by educational attainment using survey representative UK data for UK adults from the National Diet and Nutrition Survey (2008–2012) (n = 2083). 16 In the UK, United States, and other high-income countries, the burden of obesity and some chronic diseases is caused in part by the consumption of diets that contain excess energy but insufficient levels of other nutrients. 33 , 34 Because sufficient energy consumption alone is not enough to ensure good health in the long term, it would be important to identify ways of improving the nutrient to calorie ratio of the overall diet by consuming more nutrient-rich foods. Such foods might not necessarily be low cost on a per calorie basis but still be economical in terms of the cost per unit of nutrients. Studies of this kind in the United States have combined national food price data with nutrient composition data to identify low-cost, nutritious foods. 31 ,35 Though some have argued that the cost of foods expressed per unit of energy has little behavioral significance, 36 research indicates that lower income households tend to buy foods that are on average cheaper per calorie, perhaps meaning that there is a behavioral response to low prices for food energy. 37 , 38 Furthermore, longitudinal data indicate that, irrespective of their socioeconomic position, households shifted to purchasing cheaper calories during a period of rising food prices and falling real incomes. 39 This indicates that cutting food costs without lowering energy intake can lead to the substitution of foods that provide more energy for a given price, which has been suggested previously in simulation studies. 40 ,41 It is worth noting that our data are from the UK and that we have only considered the arguments in favor of different metrics in the context of a high-income country such as the UK, meaning that they may not necessarily apply to middle- or low-income settings. However, they could reasonably be expected to be appropriate for use in countries with similar economic and agricultural systems, particularly other countries within the European Union that are subject to the same laws governing agricultural subsidy and food production standards. The United Kingdom is a high-income country and despite the price of food rising by 7.7% since 2007, the mean percentage of household expenditure on food is 11.4% and overall UK prices are just 0.5% greater than the European Union average. 42 ,43 Users of our findings should assess whether or not the economic and food system context in which they are working is comparable to that of the UK. Conclusions In this article we have briefly reviewed the debate concerning the most appropriate price metric for the comparison of different foods and beverages, demonstrating the sizeable effect that metric choice can have on the results of price comparisons. In an attempt to outline guidance concerning the use of price comparison metrics, we have made suggestions as to when to apply the different approaches according to the question being addressed. Though there are numerous occasions where it is more appropriate to analyze food prices in terms of mass or portion size, when the research concerns public health and the ability to eat healthily, we argue that energy is often the most appropriate metric for assessing the cost of both foods and diets. Funding The present study was undertaken by the Centre for Diet and Activity Research (CEDAR), a UKCRC Public Health Research Centre of Excellence. The authors gratefully acknowledge the funding from the British Heart Foundation, Cancer Research UK, Economic and Social Research Council, Medical Research Council, the National Institute for Health Research, and the Wellcome Trust, under the auspices of the UK Clinical Research Collaboration. The funding sources had no role in the design and conduct of the study or in the collection, management, analysis, and interpretation of the data. ==== Refs References Food and Agriculture Organization Rome Declaration on World Food Security Available at: http://www.fao.org/docrep/003/w3613e/w3613e00.HTM Accessed April 21, 2015 Lee A Mhurchu CN Sacks G Monitoring the price and affordability of foods and diets globally. Obes Rev 2013 14 82 95 24074213 Frazão E. Reply to A. Drewnowski et al Am J Clin Nutr 2009 90 702 703 Drewnowski A. Reply to E. Frazão et al Am J Clin Nutr 2011 94 1655 Carlson A Frazão E. Are Healthy Foods Really More Expensive ? It Depends on How You Measure the Price Washington, DC: Economic Research Service 2012 Foods: Atwater W. nutritive value and cost USDA Farmers’ Bull 1894;23 Drewnowski A. The cost of U.S. foods as related to their nutritive value Am J Clin Nutr 2010 92 1181 1188 20720258 Monsivais P Mclain J Drewnowski A. The rising disparity in the price of healthful foods: 2004–2008 Food Policy 2010 35 514 520 25411518 Rao M Afshin A Singh G Mozaffarian D. Do healthier foods and diet patterns cost more than less healthy options ? A systematic review and meta-analysis BMJ Open 2013 3 Rydén PJ Hagfors L. Diet cost, diet quality and socio-economic position: how are they related and what contributes to differences in diet costs? Public Health Nutr 2011 14 1680 1692 21255480 Subramanian S Deaton A. The demand for food and calories J Polit Econ 1996 104 133 162 The World Bank Measuring poverty lines. Available at: http://info.worldbank.org/etools/docs/library/93518/Hung_0603/Hu_0603/Module3MeasuringPovertyLines.pdf Accessed March 15, 2013 Lipsky L. Reply to N. Darmon and M Maillot. Am J Clin Nutr 2010 9 1069 1070 Carlson A Frazão E. Food costs, diet quality and energy balance in the United States Physiol Behav 2014 1 11 Davis GC Carlson A. How spurious is the relationship between food price and energy density? A simple procedure and statistical test Paper presented at: Economics Association Annual Meeting; August 12–14 2012 Seattle, WA UK Data Service National Diet and Nutrition Survey, 2008–2011 Available at: http://discover.ukdataservice.ac.uk/catalogue/?sn=6533&type=Data catalogue. Accessed June 30, 2013 Jones NRV Conklin AI Suhrcke M Monsivais P. The growing price gap between more and less healthy foods: analysis of a novel longitudinal UK dataset PLoS One 2014 9 e109343 25296332 U.S. Department of Agriculture Agriculture Handbook 102: Food Yields Washington, DC: U.S. Department of Agriculture 1975 WL Wrieden KL Barton Calculation and Collation of Typical Food Portion Sizes for Adults Aged 19–64 and Older People Aged 65 and Over London, UK: Food Standards Agency 2006 Office for National Statistics Consumer Price Indices Technical Manual London, UK: Office for National Statistics 2012 Rayner M Scarborough P and the British Heart Foundation The UK Ofcom Nutrient Profiling Model Oxford, UK 2009 Available at: http://www.ndph.ox.ac.uk/bhfcpnp/about/publications-and-reports/group-reports/uk-ofcom-nutrient-profile-model.pdf Accessed February 11, 2016 Macdiarmid JI Kyle J Horgan G Livewell: A Balance of Healthy and Sustainable Food Choices London, UK: WWF-UK 2011 StataCorp Stata Statistical Software . Release 12 College Station, TX: StataCorp, LP 2011 R Core Team R: A Language and Environment for Statistical Computing Vienna, Austria: R Foundation for Statistical Computing 2008 Wickham H. ggplot2: Elegant Graphics for Data Analysis New York, NY: Springer Verlag 2009 Department for Business Innovation and Skills Progress on clearer pricing in supermarkets. Available at https://www.gov.uk/government/news/progress-on-clearer-pricing-in-supermarkets Accessed April 19, 2015 Australian Competition and Consumer Commission Grocery unit prices. Available at: https://www.accc.gov.au/consumers/groceries/grocery-unit-prices Accessed April 19, 2015 Mozaffarian RS Andry A Lee RM Wiecha JL Gortmaker SL. Price and healthfulness of snacks in 32 YMCA after-school programs in 4 U.S. metropolitan areas, 2006–2008 Prev Chronic Dis 2012 9 E38 22239753 Stewart H Hyman J JC Buzby Frazão E Carlson A. How Much Do Fruit and Vegetables Cost? Economic Information Bulletin No (EIB-71) Washington, DC: U.S. Department of Agriculture 2011 Monsivais P Rehm CD. Potential nutritional and economic effects of replacing juice with fruit in the diets of children in the United States Arch Pediatr Adolesc Med 2012 166 459 464 22566547 Drewnowski A Rehm CD. Vegetable cost metrics show that potatoes and beans provide most nutrients per penny PLoS One 2013 8 e63277 23691007 Department of Health National Diet and Nutrition Survey Years 1–4 2008/09–2011/12: User Guide London, UK: Public Health England 2011 Hill JO Wyatt HR Peters JC. Energy balance and obesity Circulation 2012 126 126 132 22753534 Zizza CA Duffy PA Gerrior SA. Food insecurity is not associated with lower energy intakes Obesity 2008 16 1908–1913 Drewnowski A. The nutrient rich foods index helps to identify healthy, affordable foods Am J Clin Nutr 2010 91 1095 1101 Lipsky LM Just DR Nansel TR Haynie DL. Fundamental misunderstanding of the relation between energy density (kcal/g) and energy cost ($/kcal) Am J Clin Nutr 2011 93 867 868 21325439 Appelhans BM Milliron B-J Woolf K Socioeconomic status, energy cost, and nutrient content of supermarket food purchases Am J Prev Med 2012 42 398 402 22424253 Barton KL Wrieden WL Sherriff A Armstrong J Anderson AS. Energy density of the Scottish diet estimated from food purchase data: relationship with socio-economic position and dietary targets Br J Nutr 2014 112 80 88 24804721 Griffith R Connell MO Smith K. Food Purchases and Nutrition Over the Recession London, UK: Institute for Fiscal Studies 2012 Darmon N Ferguson EL. A cost constraint alone has adverse effects on food selection and nutrient density: an analysis of human diets by linear programming. J Nutr 2002 132 3764 3771 Darmon N Ferguson EL Briend A. Impact of a cost constraint on nutritionally adequate food choices for French women: an analysis by linear programming J Nutr Educ Behav 2006 38 82 90 16595286 Office for National Statistics The UK has one of the fastest growing economies in the G7. Available at http://www.ons.gov.uk/ons/rel/elmr/gdp-and-the-labour-market/q1-2014–may-gdp-update/sty-gdp-g7-economies.html Accessed July 3, 2015 Department for Environment Food and Rural Affairs Food Statistics Pocketbook 2014 London, UK: Department for Environment, Food and Rural Affairs 2014

PMC005xxxxxx/PMC5001149.txt

==== Front Conserv PhysiolConserv PhysiolconphysconphysConservation Physiology2051-1434Oxford University Press 10.1093/conphys/cow024cow024ToolboxGet the most out of blow hormones: validation of sampling materials, field storage and extraction techniques for whale respiratory vapour samples Burgess Elizabeth A. *Hunt Kathleen E. Kraus Scott D. Rolland Rosalind M. John H. Prescott Marine Laboratory, New England Aquarium, 1 Central Wharf, Boston, MA 02110, USA*Corresponding author: John H. Prescott Marine Laboratory, New England Aquarium, 1 Central Wharf, Boston, MA 02110, USA. Tel: +1 (617) 226-2146. Email: eburgess@neaq.orgEditor: Steven Cooke 2016 26 8 2016 4 1 cow02429 9 2015 12 5 2016 14 5 2016 © The Author 2016. Published by Oxford University Press and the Society for Experimental Biology.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Respiratory vapor (blow) of cetaceans may contain vital physiologic data, yet fundamental methodological issues remain to be addressed for this nascent technique. We validated sample field storage; hormone extraction methods; and assay interference from sampling materials in order to ensure that reliable hormone data is obtained from whale blow. Studies are progressively showing that vital physiological data may be contained in the respiratory vapour (blow) of cetaceans. Nonetheless, fundamental methodological issues need to be addressed before hormone analysis of blow can become a reliable technique. In this study, we performed controlled experiments in a laboratory setting, using known doses of pure parent hormones, to validate several technical factors that may play a crucial role in hormone analyses. We evaluated the following factors: (i) practical field storage of samples on small boats during daylong trips; (ii) efficiency of hormone extraction methods; and (iii) assay interference of different sampler types (i.e. veil nylon, nitex nylon mesh and polystyrene dish). Sampling materials were dosed with mock blow samples of known mixed hormone concentrations (progesterone, 17β-estradiol, testosterone, cortisol, aldosterone and triiodothyronine), designed to mimic endocrine profiles characteristic of pregnant females, adult males, an adrenal glucocorticoid response or a zero-hormone control (distilled H2O). Results showed that storage of samples in a cooler on ice preserved hormone integrity for at least 6 h (P = 0.18). All sampling materials and extraction methods yielded the correct relative patterns for all six hormones. However, veil and nitex mesh produced detectable assay interference (mean 0.22 ± 0.04 and 0.18 ± 0.03 ng/ml, respectively), possibly caused by some nylon-based component affecting antibody binding. Polystyrene dishes were the most efficacious sampler for accuracy and precision (P < 0.001), but required an ethanol rinse for improved progesterone recovery (increased 81%; P < 0.001). Awareness of assay interference from exogenous materials is crucial to future studies. This study establishes critical groundwork to help ensure that hormones can be measured accurately in samples obtained from field collections of whale blow. Blowcetaceanhormone analysistechnique validationOffice of Naval Researchhttp://dx.doi.org/10.13039/100000006N000141310639 ==== Body Introduction As marine ecosystems are increasingly impacted by human activities, there is an urgent need to develop novel techniques for physiological assessment of living whales (Hunt et al., 2013). Endocrine information, in particular, can afford managers some insight into biological processes of conservation concern, namely reproduction and stress responses (Pukazhenthi and Wildt, 2004; Wikelski and Cooke, 2006). A major challenge in assessing large whales in this manner is that sample types conventionally used for hormone analysis (i.e. blood) are logistically impossible to collect from free-swimming individuals. However, whales breathe at the surface with huge tidal volumes (see Piscitelli et al., 2013), and the exhaled respiratory vapour or ‘blow’ may have a chemical composition that reflects components circulating in blood (see Aksenov et al., 2014), presenting a promising approach for non-invasive sampling of internal physiology. Field studies have shown that respiratory samples can be collected from whales by positioning a sampling device above (~0.5–1 m) the exhaling blowholes to catch a portion of the aerosol droplets (Hogg et al., 2009; Hunt et al., 2014a). Various collection materials have been used for this purpose, including nylon stocking, nylon veil, nylon mesh, cotton gauze, polypropylene containers and Petri dishes (Hogg et al., 2005, 2009; Acevedo-Whitehouse et al., 2010; Hunt et al., 2014a; Thompson et al., 2014). Moreover, several classes of steroid and thyroid hormones have been successfully detected in blow samples from North Atlantic right whales (Eubalaena glacialis; Hogg et al., 2009; Hunt et al., 2014a) and humpback whales (Megaptera novaeangliae; Hogg et al., 2009), as well as smaller cetaceans in captivity or under restraint in the wild, namely bottlenose dolphins (Tursiops truncatus; Hogg et al., 2005; Aksenov et al., 2014) and belugas (Delphinapterus leucas; Thompson et al., 2014). Recent analytical developments have shown that convenient and commercially available enzyme immunoassays are sensitive enough (picogram per millilitre level) to detect blow hormones (Hunt et al., 2014a; Thompson et al., 2014). Although sampling whale blow at sea can be challenging, it has proved feasible to collect samples from targeted individuals, as well as to obtain repeated samples over time (Hogg et al., 2009; Hunt et al., 2014a). These major sampling benefits, along with the prospect that hormone samples may provide insight into a range of biological questions (see Hunt et al., 2013), have encouraged researchers to begin exploring blow hormones in a number of cetacean species (e.g. Dunstan et al., 2012). However, important methodological issues still need to be addressed before quantification of hormone concentrations in blow samples can become a reliable technique (see Trout, 2008; Hunt et al., 2014a). The validity of hormonal analysis depends on accurate measurement of the hormone concentration in the designated sample matrix. Generally, most hormones are present in all individuals (of both sexes) at all times; what matters is the quantified concentration and/or relative levels in association with biological factors of interest. Biological variations in the concentrations of reproductive hormones may reflect sexual maturity, reproductive seasonality and/or the reproductive state of individuals (e.g. testosterone, progesterone and estradiol), whereas thyroid and adrenal hormones may reflect nutritional and metabolic processes (e.g. triiodothyronine) and/or activation of the stress system (e.g. cortisol and aldosterone; Norris and Carr, 2013). However, technical factors can play a crucial role in the successful evaluation of hormone concentrations. Fundamental concerns that may affect the resulting sample concentrations include the following: whether field storage conditions adequately preserve the sample hormone after collection (see Woods, 1975; Millspaugh and Washburn, 2004; Ziegler and Wittwer, 2005); the efficiency of the technique used to extract hormones from the sample (see Lynch et al., 2003); and whether there is assay interference from exogenous materials used in the collection and/or extraction process (e.g. cotton used in blow sample collection has been shown to interfere with some immunoassays; Hogg et al., 2009; Thompson et al., 2014). For blow hormone analysis, these technical considerations are especially relevant because of the relatively low concentrations of exhaled hormone metabolites and the need for a large surface area to collect respiratory droplets, as well as unknown variation in the whale's respiratory vapour flow. Currently, there is no consensus about the best procedures or precautions for blow sample storage in the field or sample preparation prior to hormone analysis. Evaluation of the factors that may distort or bias hormone concentrations is necessary to demonstrate the validity of any new method and must be achieved before a new technique merits general acceptance by the research community. In this study, we investigated the following technical issues pertaining to sample handling and processing (i.e. after collection and before immunoassay) that might influence the integrity of cetacean blow hormone results: (i) typical field storage of samples during daylong trips on small boats (i.e. insulated cooler with samples on ice); (ii) efficiency of hormone extraction methods; and (iii) assay interference from sampling and processing materials. Overall, our objective was to lay the foundation for application of hormone analysis of respiratory vapour to cetacean field research. Materials and methods Sampler types The following three materials were experimentally tested as potential sampler types for collecting whale respiratory vapour: (i) commercial nylon veil (ordinary bridal tulle fabric; ‘veil’ hereafter); (ii) laboratory-grade nitex nylon 110 mm mesh (Elko Filtering, Miami, FL, USA; ‘nitex mesh’ hereafter); and (iii) sterile polystyrene dishes (Corning® square 25 cm × 25 cm, CLS431111; Sigma-Aldrich, St Louis, MO, USA; ‘dish’ hereafter). To date, nylon (veil or stocking fabric) has been the most field-tested collection material for large whale blow (Hogg et al., 2009; Hunt et al., 2014a). However, more recent collection trials, using blow from captive cetaceans, demonstrated that nitex mesh had an improved performance in retaining sample volume (cf. nylon veil and cotton gauze), with reduced assay interference (cf. nylon stocking; Thompson et al., 2014). An alternative methodological approach is to avoid using fabric by collecting blow droplets onto a non-permeable surface, such as a polystyrene dish (see Acevedo-Whitehouse et al., 2010). Each of these materials requires modifications of sample processing and hormone extraction, which could potentially influence the resulting hormone concentrations. Our study design uses a holistic comparison of each sampler type (i.e. encompassing the sampling material, its associated outer storage container and an extraction protocol; see ‘Extraction techniques’ below) in a controlled experimental setting. For preparation of nylon sampler types, veil material was cut to 90 cm × 180 cm (folded 45 cm × 45 cm in eight-ply), and nitex mesh material was cut to 30 cm × 30 cm (single ply). Both veil and nitex mesh materials were thoroughly washed before use, using multiple wash cycles (Hogg et al., 2009; Hunt et al., 2014a) that involved soaking in warm soapy water for 10 min, then rinsing with tap water to remove soap, rinsing with distilled water (dH2O) twice, submerging and agitating in 70% ethanol (EtOH) for 15 min, and air drying. Nylon samplers were sealed in individual clean zip-type plastic bags, ready for use. Dishes had a clean, sterilized surface and lid that did not require any washing before experimental treatment. Experimental design and hormone treatments To determine the efficiency of hormone recovery from each sampler type, we prepared three different solutions with varying hormone ratios designed as mock ‘blow’ solutions containing a mixture of several different hormones, in addition to a fourth control solution consisting of dH2O only (i.e. no added hormone). Mixed hormone solutions were designed to mimic the general pattern of three hormone profiles of interest (adult male, pregnant female and adrenal glucocorticoid response), as well as to simulate a variety of hormone ratios across solutions. Hormone concentrations were prepared within a range of 0.1 (low) to 10 ng/ml (high) (Table 1), which is representative of serum concentrations in large whales (e.g. fin whale, Balaenoptera physalus, progesterone range 0.2–12 ng/ml; testosterone range 0.03–12 ng/ml; estradiol mean range 0.02–12 ng/ml; Kjeld et al., 2006). Preliminary data from cetacean blow in both mysticetes and odontocetes indicates that blow is likely to have similar hormone concentration ranges to plasma (Hunt et al., 2014a; Thompson et al., 2014). Table 1: Actual hormone concentrations (in nanograms per millilitre ± SD) in experimental treatment solutions (control, adult male, pregnant female and adrenal glucocorticoid response profiles) prepared as low (~0.1 ng/ml; light shade), medium (~1 ng/ml; medium shade) and high concentrations (~10 ng/ml; dark shade) of various hormones [testosterone (T), progesterone (P4), estradiol (E2), cortisol (F), aldosterone (ALD) and triiodothyronine (T3)]   Hormone concentration (ng/ml) Treatment solution T P4 E2 F ALD T3 Control 0.0 ± 0.0 0.0 ± 0.0 0.0 ± 0.0 0.0 ± 0.0 0.0 ± 0.0 0.0 ± 0.0 Adult male 11.5 ± 0.9 0.0 ± 0.0 0.1 ± 0.0 8.2 ± 0.7 0.8 ± 0.1 0.8 ± 0.1 Pregnant female 0.2 ± 0.0 8.5 ± 0.2 0.7 ± 0.1 0.6 ± 0.1 0.9 ± 0.2 0.6 ± 0.1 Adrenal glucocorticoid response 0.3 ± 0.0 0.0 ± 0.0 0.7 ± 0.0 8.4 ± 0.6 11.1 ± 1.5 0.7 ± 0.1 Assay limit of detection 0.03 0.05 0.03 0.05 0.01 0.07 For preparation of these treatment solutions, pure crystalline progesterone (catalogue no. P0130), 17β-estradiol (catalogue no. E8875), testosterone (catalogue no. T1500), cortisol (catalogue no. H4001), aldosterone (catalogue no. A9477) and triiodothyronine (catalogue no. T2877) were used (all from Sigma-Aldrich, St Louis, MO, USA). Stock solutions of each hormone were prepared in HPLC-grade EtOH and stored in Pyrex® 100 ml glass bottles. Final treatment solutions were then prepared in dH2O using the stock solutions to produce the desired combination of concentrations that were near the intended high (10.0 ng/ml), medium (1.0 ng/ml) or low (0.1 ng/ml) concentrations of each hormone (see Table 1). Triiodothyronine was the only hormone added at a uniform concentration to all mixed hormone solutions, because this hormone often has a narrow range of concentrations among individual cetaceans (e.g. St Aubin et al., 1996), especially when compared with steroids. Final concentrations of treatment solutions were expected to have minor deviation from target concentrations (see Table 1), as is typical when preparing hormone standards at the nanogram to picogram per millilitre level. In the experiment, the surface of each sampling material was dripped with 1.0 ml of a treatment solution (i.e. mixed-hormone ‘blow’ solution or control dH2O), simulating capture of respiratory vapour droplets from a whale. Previous field trials revealed that 1.0 ml is a typical ‘high-quality’ sample volume collected from a single exhalation of a large whale and that respiratory vapour samples from large whales typically are highly aqueous, with no visible lipid or mucoid fraction, even after centrifugation (Hunt et al., 2014a). For each treatment, we conducted eight replicates for each sampler type [n = 4 treatments (three mixed hormone solutions plus one control) × 8 replications = 32 samples for each sampler type]. This experiment was done in duplicate for the dish samplers in order to test two different methods of extracting (recovering) hormone from a flat dish surface (i.e. pipetting off droplets vs. rinsing the dish with 100% EtOH; (see ‘Extraction techniques’ below). Once treated with the mock ‘blow’ solution or control dH2O, the veil and nitex mesh samplers were sealed in an individual zip-type plastic bag, whereas dish samplers were covered with the lid and then sealed in a zip-type plastic bag. All samplers were then placed in a thick-walled (3.8 cm) cooler on icepacks for 6 h to simulate typical field storage conditions on a small research vessel during daylong sampling at sea. After 6 h, hormones were extracted from each sampler (total n = 128). Extraction techniques Hormones were extracted using different methods depending on the sampling material. Veil samplers (n = 32) were extracted by pouring 100 ml of 100% EtOH over each veil inside a 473 ml glass jar, the minimal size of container that allowed complete submergence of the eight-ply fabric in EtOH. The jar was sealed and hand shaken vigorously for 60 s, after which the liquid was decanted into 25 mm × 125 mm borosilicate glass tubes. These steps were repeated for a second rinse. An additional 20 ml of 100% EtOH was used to rinse the inside of the zip-type plastic bag that contained the sampler. The combined ~220 ml EtOH rinse (containing hormones) in glass tubes was evaporated to dryness under compressed air for 48 h and reconstituted in 1.0 ml of dH2O for analysis. Nitex mesh samplers (n = 32) were extracted by pouring 80 ml of 100% EtOH over each nitex mesh inside a 120 ml polypropylene jar. The solvent volume for nitex mesh was lower than that for veils owing to the smaller size of the nitex mesh; in both cases, enough EtOH was added to immerse the sampling material thoroughly. The jar was sealed and vigorously mixed on a plate-shaker for 1 h, after which the liquid was decanted into 25 mm × 125 mm borosilicate glass tubes. The nitex mesh was then put into a 50 ml Falcon® tube on top of two capped 2 ml microcentrifuge tubes (i.e. elevating the material off the bottom) and centrifuged at 4000 g for 15 min. It is noteworthy that the smaller fabric volume of the nitex mesh permitted controlled centrifugation (cf. hand-shaken veil samples, which were too large to centrifuge). Recovered fluid was added to the glass tubes. The zip-type plastic bag was also rinsed with 20 ml of 100% EtOH. The combined ~100 ml EtOH rinse in glass tubes was dried under compressed air for 24 h and reconstituted in 1.0 ml of dH2O. Dish samplers do not involve a fabric (unlike nitex mesh or veils); hence, the following two extraction methods were tested: (i) direct extraction by pipetting (n = 32; ‘pipetted dish’ hereafter); and (ii) an EtOH rinse (n = 32; ‘rinsed dish’ hereafter). The pipetting method was tested because it had the potential to minimize additional sample processing steps (e.g. dry-down and reconstitution) that might introduce noise to the hormone data. Pipetting was not attempted with the veil or nitex mesh because previous field trials showed that cetacean blow hormone adheres to fabric-type samplers; Hunt et al., 2014a). The pipetting method involved using a 1000 µl pipette to draw up all visible droplets from the surface of the dish and transfer them into a microcentrifuge tube for later assay. For the rinse method, 50 ml of 100% EtOH was poured over the dish, which was lidded and gently agitated on a plate-shaker for 30 min (i.e. to suspend any hormone that might have adhered to the dish surface). The EtOH rinse was then decanted into 25 mm × 125 mm borosilicate glass tubes, dried under compressed air for 24 h and reconstituted in 1.0 ml of dH2O. All samples (total n = 128) were stored frozen at −20°C until hormone analysis. Given that the containers used for storage and extraction of veil and nitex samplers were varied (mostly owing to the fabric size; see above), a supplemental experiment was conducted to investigate the influence of each container type [i.e. polypropylene (zip-type) bag, polypropylene jar and glass jar] as a source of assay interference. In brief, clean polypropylene bags, polypropylene jars and glass jars were individually rinsed with 100% EtOH (according to the extraction protocol above), and the resultant samples (n = 8 replicates for each container type) were analysed for progesterone, testosterone and cortisol (bracketing the range of polarity of steroids tested in this study; see ‘Hormone assays’ below). When tested in isolation, none of the bag or jar types used in this study produced interference effects (i.e. non-detectable hormone concentrations for all assays tested; data not shown). Hormone assays Enzyme immunoassays were used to quantify concentrations of testosterone (#K032-H1), progesterone (#K025-H1), estradiol (#K036-H1), cortisol (#K003-H1) and aldosterone (#052-H1; 2 h protocol; all from Arbor Assays, Ann Arbor, MI, USA). Radioimmunoassay was used to quantify triiodothyronine concentrations (#06B-254215; MP Biomedicals, Solon, OH, USA). These commercially available assay kits were selected based on previous successful use with respiratory vapour collected from free-swimming whales (see Hunt et al., 2014a). Assay procedures were performed according to the manufacturer's protocols. For further assay details and antibody cross-reactivities, see the enzyme immunoassay protocols from Arbor Assays (http://www.arborassays.com) and the radioimmunoassay protocol from MP Biomedicals (http://www.mpbio.com). Samples containing 10 ng/ml concentrations of particular hormones were assayed at a 1:20 dilution in dH2O for that hormone's assay. Pure mixed hormone ‘blow’ solutions and pure dH2O were distributed equally within and across assays to minimize potential effects of inter-assay variation. To monitor precision and reproducibility in our assays, low- (70–80% binding) and high-quality (20–30% binding) control samples were run on each plate (total n = 5 assays performed for each hormone). All assays were performed by the same person, with all samples, controls and standards assayed in duplicate, and results averaged accordingly. Intra-assay coefficient of variation averaged 4.0% across assays, calculated from the variation of measurements between duplicates. Inter-assay coefficient of variation was <7.0% (low control) and <5.0% (high control) for all six assays. Any sample with a coefficient of variation >10% was re-assayed. Final data were expressed as nanograms per millilitre of extracted sample. Statistical analyses Our analytical method focused on a comparison between apparent (measured) concentrations of samples and the known concentration of the treatment solution that had been dripped onto the sampler. The best sampling material and extraction technique would be that which demonstrated the accuracy and precision in measured hormone concentrations compared with the ‘actual’ concentration in pure solution after sample processing. We used simple descriptive statistics (mean ± SD) to summarize the data set. All data were analysed using SPSS statistical software (version 22.0 for Macintosh; IBM Corp., Armonk, NY, USA). Hormone data were log10-transformed to adjust for non-normal distributions based on skewness and kurtosis. Non-detectable values were substituted with half the limit of detection for that particular assay (see Table 1), in order to allow transformation of all data. Control treatment results were used to evaluate whether any sampling material and extraction technique interfered with the assay, i.e. causing spurious apparent hormone concentrations. Two-way analysis of variance was used to assess differences in hormone concentrations between treatment solutions and between sampler types. Where significant differences were detected, post hoc comparisons using Tukey's HSD test were performed to identify the source of variance. Assay results of the three mixed hormone solutions were considered as the ‘actual’ (known) concentrations applied in treatments. Student's paired t-test was used to detect differences in hormone concentrations between the pure mixed hormone solution and the resulting sample after experimental treatment. Accuracy was evaluated as the difference between the measured hormone concentration in the resulting sample and the known concentration of the mixed hormone solution. Precision was measured by the standard deviation among samples for each sampler type. The percentage recovery of each hormone in samples was calculated from the actual concentration expected, i.e. percentage recovery = measured concentration/actual concentration × 100. Hormone recoveries from each sampler type were compared using data from high-concentration samples (10 ng/ml), because these concentrations yielded the greatest assay reliability (i.e. near 50% bound on standard curve). Results for the two different extraction techniques tested on dish samplers (EtOH rinse vs. direct pipetting) were compared using a two-tailed Student's unpaired t-test. In order to evaluate the overall efficiency of each sampling material and respective extraction methods, we reduced the data for all six hormones using a multivariate principal components analysis. Before performing principal components analysis, the suitability of data for factor analysis was assessed. Inspection of the correlation matrix revealed the presence of coefficients of 0.3 and above. The Kaiser–Meyer–Olkin value was 0.5, and Bartlett's test of sphericity reached statistical significance (P < 0.001), supporting the factorability of the correlation matrix (Quinn and Keough, 2002). Principal components analysis revealed the presence of two components with eigenvalues exceeding 1, and an inspection of the scree plot showed a clear break after the second component. To aid in the interpretation of these two components, oblimin rotation was performed, with the two factors showing low inter-correlation (r = 0.21). The resulting eigenvector loadings associated with the new components were examined graphically to assess how each sampler type was able to distinguish between treatment solutions. For all analyses, P < 0.05 was considered as significant. Results All three treatment solutions had significantly different combinations of hormones (F2,72 = 32.9, P < 0.001; Table 1), demonstrating that the three hormone profiles, representing a (hypothetical) adult male, pregnant female and adrenal glucocorticoid response, were statistically distinguishable from each other. The immunoassays used had adequate sensitivity to measure quantities within the range of the lowest prepared hormone concentration used in this study (0.1 ng/ml). The only exception was the progesterone assay, which produced non-detectable results (i.e. between zero and the detection limit of the assay) for the low-level progesterone concentration (0.1 ng/ml) in the adult male and adrenal glucocorticoid response solutions when assayed without dilution. Nonetheless, all immunoassay results accurately differentiated between relative patterns of high, medium and low concentrations between treatment solutions for all hormones (all P < 0.001); excluding triiodothyronine hormone, which was prepared as a uniform concentration across treatment groups (F2,13 = 2.79, P = 0.21; Table 1). Distilled water, assayed as a pure solution, had non-detectable concentrations for all hormones under study, as expected (Table 1). All sampler types in the control treatment (no added hormone) produced some level of apparent (spurious) hormone in all immunoassays tested (Fig. 1). This measurable immunoreactivity was intrinsic to each sampler type, probably indicating interference with antibody binding in the assay. For all sampler types and hormones tested, assay interference levels averaged <0.3 ng/ml, with one extreme outlying result from a single veil sampler for estradiol (1.73 ng/ml). Levels were greatest for the progesterone assay (0.29 ± 0.26 ng/ml), followed by estradiol (0.15 ± 0.32 ng/ml) and cortisol (0.13 ± 0.10 ng/ml), and were very low for triiodothyronine (0.06 ± 0.07 ng/ml), testosterone (0.06 ± 0.04 ng/ml) and aldosterone (0.05 ± 0.06 ng/ml; F5,167 = 9.30, P < 0.001). However, these measures of assay interference were significantly different between sampler types (F3,167 = 24.13, P < 0.001; Fig. 1). The highest level of interference was measured from veil and nitex mesh samplers (mean across all hormones, 0.22 ± 0.04 and 0.18 ± 0.03 ng/ml, respectively), whereas rinsed dishes had lower levels (0.04 ± 0.01 ng/ml), and pipetted dishes demonstrated the least interference in hormone assays (0.01 ± 0.00 ng/ml). Assay interference from all sampler types and across all immunoassays demonstrated repeatability, as determined by low standard deviations between replicates (nitex mesh, SD < 0.11 ng/ml; veil, SD < 0.54 ng/ml; rinsed dish, SD < 0.06 ng/ml; and pipetted dish, SD < 0.01 ng/ml), i.e. considered to be background noise (Fig. 1). Figure 1: Quantification of assay interference from different sampling materials (veil, nitex mesh or dish) after extraction [using ethanol (EtOH) rinse or direct pipetting; n = 8 replicates for each sampler type], as determined for six immunoassays [testosterone (T), progesterone (P4), estradiol (E2), cortisol (F), aldosterone (ALD) and triiodothyronine (T3)]. For boxplots, the line inside the box indicates the median value, the height of the box encompasses the distance between the 25th and 75th quartiles, and the whiskers delineate extreme observations. Outliers are marked with a circle (>1.5 × interquartile range) and extreme outliers are marked with a star (>3 × interquartile range). Different letters denote a significant difference in resulting hormone measures between sampler types at P < 0.05. After storing samplers for 6 h in a cooler, all samples yielded hormone measurements that were statistically similar to actual concentrations of the pure mixed hormone solutions (t575 = −1.36, P = 0.18). For each sampler type, hormone concentrations differed significantly between samples treated with the mock pregnant female, adult male and adrenal glucocorticoid response solutions, as expected (all P < 0.05). However, the accuracy of observed hormone concentrations compared with expected levels was significantly different between sampler types (F3,575 = 24.09, P < 0.001). The resultant hormone measures most significantly affected by sampler type (i.e. P < 0.001) were progesterone (low concentration level, F3,60 = 19.32; high level, F3,28 = 92.94) and cortisol (low level, F3,28 = 101.04; high level, F3,60 = 9.63), as well as estradiol (medium level, F3,60 = 27.26) and triiodothyronine (medium level, F3,92 = 62.16; Fig. 2). Across all hormones and concentration levels, results were most accurate from rinsed dishes (range = −3.4 to 2.2 ng/ml), followed by nitex mesh (−4.2 to 3.1 ng/ml), then veil samples (−5.0 to 3.6 ng/ml), and the least accurate results from pipetted dishes (−7.9 to 3.8 ng/ml; F3,575 = 27.59, P < 0.001). The magnitude of change of sample hormone results attributable to the use of each sampler type was greater than the error expected owing to assay performance across runs (i.e. inter-assay coefficient of variation <7%). In addition, the direction of change (i.e. elevated or reduced recovery) in resulting measures for each sampler type also varied depending on the hormone type and concentration level (see Fig. 2). For dish samplers, hormone recovery on average was significantly improved when the EtOH rinse was applied (100 ± 3%) compared with directly pipetting the sample off the surface (88 ± 6%; t44 = −2.05, P = 0.04). In particular, progesterone recovery significantly increased, on average by 81%, when applying an EtOH rinse to the dish surface (t44 = −17.51, P < 0.001; all other hormones, P > 0.05; see Fig. 2). Recovery of sampled hormone from the nitex mesh samplers was on average 101 ± 3% and from veil samplers 97 ± 5%. Figure 2: Accuracy of hormone results (observed minus expected concentration, in nanograms per millilitre) using different sampling materials (veil, nitex mesh or dish) and respective extraction methods [ethanol (EtOH) rinse or direct pipetting]. Hormones measured were testosterone (T), progesterone (P4), estradiol (E2), cortisol (F), aldosterone (ALD) and triiodothyronine (T3). Values are means (±SD) across low (~0.1 ng/ml; open circles), medium (~1 ng/ml; shaded circles) and high (~10 ng/ml; filled circles) concentrations of mixed hormone treatment solutions. Zero values indicate that the observed concentration equals the actual concentration in pure treatment solutions; negative values indicate a deficit in observed hormone concentration; and positive values indicate a surplus. Red bars indicate the detectable level of immunoassay interference (see Fig. 1). Different letters denote a significant differences in resulting hormone measures between sampler types at P < 0.05 (N.S. denotes no statistical significance). Principal components analysis incorporating concentrations for all six hormones revealed two components, PC1 and PC2, explaining 43 and 31% of the variance, respectively. Both components showed a number of strong loadings, with five of six variables loading substantially on only one component. Cortisol and progesterone loaded (pattern coefficients >0.91; structure coefficients >0.92) strongly on PC1. Aldosterone, estradiol and testosterone (pattern coefficients >0.67; structure coefficients >0.71) loaded strongly on PC2. Each treatment group for each sampler type was plotted as a function of its loadings for PC1 and PC2 and compared with the expected levels in pure mixed hormone solutions. Multivariate principal components analysis demonstrated clear differences in the hormone profile of each treatment solution before sample processing (Fig. 3). After sample processing on different sampling materials (i.e. including effects of respective containers and extraction protocols), results for each sampler type maintained the distinct variation between treatment solutions. However, rinsed dishes produced more similar results to the levels in pure solution for all treatment groups compared with other sampler types (Fig. 3). Nonetheless, all the tested sampling materials and extraction methods demonstrated good precision in resulting hormone concentrations, as determined by low standard deviations between replicates (Fig. 3). Figure 3: Component matrix of the principal component analysis, showing the factor scores for each treatment solution (adult male, pregnant female and adrenal glucocorticoid profile) after processing on each sampler type [veil (EtOH), nitex mesh (EtOH), dish (pipette) or dish (EtOH)]. Expected levels in pure mixed hormone solutions before processing are marked with a cross. Values are means ± SD. Discussion Measurement of hormones in blow is a promising non-invasive approach to gather physiological data from free-swimming whales. However, reliable sample collection and hormone extraction techniques are pivotal to accurate analysis of blow hormones. Furthermore, these critical steps are also the most labour-intensive component of sample analysis and, possibly, the most error-prone part of the process (Lynch et al., 2003). Consequently, advances in blow hormone analysis are contingent on validation and improvements of technical methodologies; this is especially true given the low hormone concentrations in blow samples. Our study demonstrated that three sampling materials have the potential to influence the results of immunoassays for multiple hormone analyses at the levels conjectured for blow. Nonetheless, relative hormone patterns (i.e. low/medium/high levels for all six hormones) and mock physiological profiles (i.e. adult male, pregnant female and adrenal glucocorticoid response) were able to be identified correctly using any of the tested sampler types, suggesting that each sampler type may still be viable for collection of whale blow. Polystyrene dishes proved to be the most effective surface for sample accuracy and precision, yielding measures close to absolute values. However, samples collected on polystyrene dishes had to be extracted with an EtOH rinse, because direct pipetting reduced hormone recovery. Samples collected in the field can be vulnerable to artificial fluctuations in hormone concentration resulting from less than ideal sample storage and transportation methods (Woods, 1975; Whitten et al., 1998; Palme, 2005). Ideally, it is recommended to preserve samples immediately at sub-zero temperatures in order to minimize degradation of hormones (Woods, 1975; Whitten et al., 1998). Achieving freezing conditions can be logistically problematic in remote situations (Khan et al., 2002; Edwards et al., 2014); however, temporarily chilling samples in an insulated cooler is practical for small research vessels at sea. Our results showed that concentrations of six hormones (representing one thyroid hormone and all five classes of steroid hormone) did not change in samples stored in a cooler on ice for up to 6 h, suggesting that cold storage (~4°C) preserved sample integrity for a practical period of time. Previous studies on blow hormones have added a preservative (e.g. manganese chloride) to samples in order to prevent hormone degradation (Hogg et al., 2005, 2009). Our results, in agreement with Trout (2008), suggest that the addition of such an inhibitor to stabilize sample hormones may not be necessary. It should be noted that our experiment did not involve the complex biological matrix expected in exhaled breath condensate; in particular, naturally occurring bacteria (Acevedo-Whitehouse et al., 2010). The issue of bacterial metabolism of hormone samples is especially marked in excreted faeces, where gastrointestinal bacteria are abundant (Whitten et al., 1998). However, faecal hormone samples kept on ice immediately after defecation are considered to be relatively stable during the period of field collection (reviewed by Khan et al., 2002). In comparison, plasma steroids are more stable, with no change in concentrations for up to 72 h at room temperature and for at least a week at 4°C (Grant and Beastall, 1983; Bolelli et al., 1995). We theorize that hormones in blow are likely to have a stability closely resembling hormones in blood, and conclude that although immediate proper preservation (i.e. freezing samples) at the site of collection is preferable, the basic cold storage equipment tested here (i.e. thick-walled insulated cooler with ice packs) was adequate for blow sample preservation while conducting daylong fieldwork at sea. We emphasize that additional storage regimens, durations (i.e. longer-term storage >6 h) and preservation temperatures could be tested in follow-up studies, but the stability noted here is encouraging for proceeding with an insulated cooler for short-term storage, which is known to be feasible for most boat-based researchers. Extraction of hormones from the sample matrix is usually the first preparation step in quantifying hormone concentrations in biological samples. Many different extraction techniques (and solvents) have been developed to recover and extract hormone from various sample types (e.g. Rolland et al., 2005; Kellar et al., 2006; Hunt et al., 2014b), so proper selection and testing of the extraction procedure is essential. For the present study, extraction techniques were specific to each sampler type and were intended to maximize hormone recovery, with consideration for logistical efficiency (i.e. processing time and required resources) and sampler design characteristics (i.e. fabric cf. solid surface). Overall, the extraction methods tested in this study yielded valid concentrations of steroid and thyroid hormones in replicate samples, demonstrating consistent and repeatable results for all hormones. However, most notably, hormone recovery was reduced when a pipette was used to collect the sample directly off the dish surface. Our results demonstrate that rinsing the sampling surface with EtOH significantly improved sample hormone recovery and absolute value estimates, especially for progesterone. Progesterone is a less polar compound than the other hormones examined in this study (Westphal, 1986), and molecules may be particularly prone to adhering onto the surface of sampling material (i.e. rather than remaining suspended in the aqueous blow droplets). Hunt et al. (2014a) found a similar pattern for hormones in North Atlantic right whale blow that had been collected with nylon veils; in that study, the cortisol content of split samples was always greater when the veil material was rinsed with EtOH than when sample droplets were recovered by pipetting and/or centrifugation of the fabric. As this phenomenon has been now recorded with two different sampling materials, it seems necessary that blow sample processing methods should routinely use a solvent rinse on the sampling device to recover hormones, i.e. an EtOH or other solvent rinse, rather than relying on physical extraction of droplets from the sampler. Although rinsing is more time consuming (because of the dry-down phase) and costly (the chemical EtOH is required), it appears necessary for maximal isolation and/or concentration of the hormones of interest. This may prove crucial, considering that the collected blow specimen will probably be low volume, low concentration and diffusely distributed across the sampling material. Sampler materials used to absorb blow during sample collection and processing must be evaluated because exogenous substances can interfere with hormone analytical methods, sometimes leading to erroneous hormone concentrations (Shirtcliff et al., 2001; Tate and Ward, 2004; Bowen and Remaley, 2014). Of the sampler and extraction processes tested here, the veil and nitex mesh protocols (i.e. the sampler itself and the extraction process used to recover hormone) were more susceptible to apparent (spurious) hormone results in some assays, particularly progesterone, estradiol and cortisol. The container bags and jars, used in conjunction with these sampler types, produced no detectable interference effects when tested in isolation (see ‘Materials and methods’). Therefore, it is plausible that the measured immunoassay interference may be mostly attributed to nylon-related substances in the veil and nitex mesh materials, which may have leached into the sample during extraction, affecting antibody binding. Fortunately, immunoassay interference for veil and nitex mesh was consistent and generally low (<0.3 ng/ml in most cases); and, most crucially, these inherent levels did not impede the overall accuracy of hormone results in the sample. Furthermore, expected sample ratios (i.e. low/medium/high concentrations) were still correctly distinguished for all hormones. Thus, the levels of assay interference noted here, for all sampler types tested, can probably be regarded as tolerable ‘background noise’ that did not distort conclusions. Nonetheless, these results should caution investigators that materials (especially nylon) could potentially interfere with immunoassay results, depending on the hormone of interest. This study achieved the first step of demonstrating whether or not such interference occurs at all, but future studies could seek to partition the source of assay interference when further optimizing collection alternatives. We strongly recommend that researchers using novel sampler types should specifically test for assay interference from all exogenous materials used in the collection and extraction process, as well as routinely including ‘blank’ samples in their studies to serve as a negative control (i.e. sampling materials with no blow sample added that have been taken through the entire laboratory extraction process). Ultimately, it is crucial that researchers be aware of the level of hormone assay interference produced by materials used in sample collection and processing, and must carefully evaluate the suitability of new collection materials to ensure that signals of biological interest are still detectable and not misinterpreted. The discipline of non-invasive field endocrinology is still evolving for marine mammals, and the present study has presented important technical validations for sampling materials, the most practical storage to stabilize samples during daylong fieldwork, and efficient extraction techniques for hormone analysis of blow samples from any whale species. Using the techniques outlined in this study, relative patterns and near absolute values were measured, at levels conjectured for blow samples, for a mix of six hormones of interest in conservation physiology. With this technical information, researchers can make informed decisions about appropriate sampling materials, effective field storage and reliable sample preparation for integrity of blow hormone results. The present study, in combination with work by Hunt et al. (2014a), which places emphasis on validating the immunoassay antibodies for a given hormone and species, provides critical groundwork for dependable sample collection and analysis techniques that may ensure reliable hormone data obtained from whale blow. The next essential step for developing blow hormone analysis must be to determine an exact adjustment to correct for the unknown total volume and water content of the blow specimen collected, thereby permitting hormone concentrations in whale blow to be quantified precisely. Acknowledgements The authors would like to thank the Office of Naval Research (ONR) for supporting physiological research of large whales and for funding this methodological study (ONR award no. N000141310639). We appreciate the constructive feedback provided by three anonymous reviewers, which led to improvements in our manuscript. Funding This work was supported by the Office of Naval Research (ONR award no. N000141310639). ==== Refs References Acevedo-Whitehouse K , Rocha-Gosselin A , Gendron D (2010 ) A novel non-invasive tool for disease surveillance of free-ranging whales and its relevance to conservation programs . Anim Conserv 13 : 217 –225 . Aksenov AA , Yeates L , Pasamontes A , Siebe C , Zrodnikov Y , Simmons J , McCartney MM , Deplanque J-P , Wells RS , Davis CE (2014 ) Metabolite content profiling of bottlenose dolphin exhaled breath . Anal Chem 86 : 10616 –10624 .25254551 Bolelli G , Muti P , Micheli A , Sciajno R , Franceschetti F , Krogh V , Pisani P , Berrino F (1995 ) Validity for epidemiological studies of long-term cryoconservation of steroid and protein hormones in serum and plasma . Cancer Epidemiol Biomarkers Prev 4 : 509 –513 .7549807 Bowen RAR , Remaley AT (2014 ) Interferences from blood collection tube components on clinical chemistry assays . Biochem Med 24 : 31 –44 . Dunstan J , Gledhill A , Hall A , Miller P , Ramp C (2012 ) Quantification of the hormones progesterone and cortisol in whale breath samples using novel, non-invasive sampling and analysis with highly-sensitive ACQUITY UPLC and Xevo TQ-S. Waters – the science of what's possible: application note. Waters Corporation, Milford, MA, USA. http://www.waters.com/webassets/cms/library/docs/720004277en.pdf. Edwards KL , McArthur HM , Liddicoat T , Walker SL (2014 ) A practical field extraction method for non-invasive monitoring of hormone activity in the black rhinoceros . Conserv Physiol 2 : cot037; doi:10.1093/conphys/cot037 . Grant JK , Beastall GH (1983 ) Clinical Biochemistry of Steroid Hormones: Methods and Applications . Elsevier , New York . Hogg CJ , Vickers ER , Rogers TL (2005 ) Determination of testosterone in saliva and blow of bottlenose dolphins (Tursiops truncatus) using liquid chromatography–mass spectrometry . J Chromatogr B 814 : 339 –346 . Hogg CJ , Rogers TL , Shorter A , Barton K , Miller PJO , Nowacek D (2009 ) Determination of steroid hormones in whale blow: it is possible . Mar Mamm Sci 25 : 605 –618 . Hunt KE , Moore MJ , Rolland RM , Kellar NM , Hall AJ , Kershaw J , Raverty SA , Davis CE , Yeates LC , Fauquier DA et al (2013 ) Overcoming the challenges of studying conservation physiology in large whales: a review of available methods . Conserv Physiol 1 : cot006; doi:10.1093/conphys/cot006 . Hunt KE , Rolland RM , Kraus SD (2014 a) Detection of steroid and thyroid hormones via immunoassay of North Atlantic right whales (Eubalaena glacialis) respiratory vapour . Mar Mamm Sci 30 : 796 –809 . Hunt KE , Stimmelmayr R , George C , Hanns C , Suydam R , Brower H Jr, Rolland RM (2014 b) Baleen hormones: a novel tool for retrospective assessment of stress and reproduction in bowhead whales (Balaena mysticetus) . Conserv Physiol 2 : cou030; doi:10.1093/conphys/cou030 . Kellar NM , Trego ML , Marks CI , Dizon AE (2006 ) Determining pregnancy from blubber in three species of delphinids . Mar Mamm Sci 22 : 1 –16 . Khan MZ , Altmann J , Isani SS , Yu J (2002 ) A matter of time: evaluating the storage of fecal samples for steroid analysis . Gen Comp Endocrinol 128 : 57 –64 .12270788 Kjeld M , Ólafsson Ö , Víkingsson GA , Sigurjónsson J (2006 ) Sex hormones and reproductive status of the North Atlantic fin whale (Balaenoptera physalus) . Aquat Mamm 32 : 75 –84 . Lynch JW , Khan MZ , Altmann J , Njahira MN , Rubenstein N (2003 ) Concentrations of four fecal steroids in wild baboons: short-term storage conditions and consequences for data interpretation . Gen Comp Endocrinol 132 : 264 –271 .12812774 Millspaugh JJ , Washburn BE (2004 ) Use of fecal glucocorticoid metabolite measures in conservation biology research: considerations for application and interpretation . Gen Comp Endocrinol 138 : 189 –199 .15364201 Norris DO , Carr JA (2013 ) Vertebrate Endocrinology , Ed 5 Academic Press , New York . Palme R (2005 ) Measuring fecal steroids: guidelines for practical application . Ann NY Acad Sci 1046 : 75 –80 .16055844 Piscitelli MA , Raverty SA , Lillie MA , Shadwick RE (2013 ) A review of cetacean lung morphology and mechanics . J Morphol 274 : 1425 –1440 .24027086 Pukazhenthi BS , Wildt DE (2004 ) Which reproductive technologies are most relevant to studying, managing and conserving wildlife . Reprod Fertil Dev 16 : 33 –46 .14972101 Quinn GP , Keough MJ (2002 ) Experimental Design and Data Analysis for Biologists . Cambridge University Press , New York . Rolland RM , Hunt KE , Kraus SD , Wasser SK (2005 ) Assessing reproductive status of right whales (Eubalaena glacialis) using fecal hormone metabolites . Gen Comp Endocrinol 142 : 308 –317 .15935157 Shirtcliff EA , Granger DA , Schwartz E , Curran MJ (2001 ) Use of salivary biomarkers in biobehavioral research: cotton-based sample collection methods can interfere with salivary immunoassay results . Psychoneuroendocrinology 26 : 165 –173 .11087962 St Aubin DJ , Ridgway SH , Wells RS , Rhinehart H (1996 ) Dolphin thyroid and adrenal hormones: circulating levels in wild and semidomesticated Tursiops truncatus, and influence of sex, age, and season . Mar Mamm Sci 12 : 1 –13 . Tate J , Ward G (2004 ) Interferences in immunoassay . Clin Biochem Rev 25 : 105 –120 .18458713 Thompson LA , Spoon TR , Goertz CEC , Hobbs RC , Romano TA (2014 ) Blow collection as a non-invasive method for measuring cortisol in the beluga (Delphinapterus leucas) . PLoS ONE 9 : e114062 .25464121 Trout T (2008 ) Comment on “Determination of testosterone in saliva and blow of bottlenose dolphins (Tursiops truncatus) using liquid chromatography–mass spectrometry” . J Chromatogr B 863 : 192 –193 . Westphal UF. 1986 Steroid-Protein Interactions II (Monographs in Endocrinology) . Springer-Verlag , Berlin . Whitten PL , Brockman DK , Stavisky RC (1998 ) Recent advances in noninvasive techniques to monitor hormone-behavior interactions . Yearb Phys Anthropol 41 : 1 –23 . Wikelski M , Cooke SJ (2006 ) Conservation physiology . Trends Ecol Evol 21 : 38 –46 .16701468 Woods GF (1975 ) Chemical and microbiological transformation of steroids . In Cameron EHD, Hillier SG, Griffiths K, eds, Steroid Immunoassay . Alpha Omega, Cardiff, pp 5 –10 . Ziegler TE , Wittwer DJ (2005 ) Fecal steroid research in the field and laboratory: improved methods for storage, transport, processing, and analysis . Am J Primatol 67 : 159 –174 .16163716

PMC005xxxxxx/PMC5001150.txt

==== Front Conserv PhysiolConserv PhysiolconphysconphysConservation Physiology2051-1434Oxford University Press 10.1093/conphys/cow026cow026ReviewA perspective on physiological studies supporting the provision of scientific advice for the management of Fraser River sockeye salmon (Oncorhynchus nerka) Patterson David A. 1*Cooke Steven J. 2Hinch Scott G. 3Robinson Kendra A. 1Young Nathan 4Farrell Anthony P. 5Miller Kristina M. 61 Fisheries and Oceans Canada, Science Branch, Cooperative Resource Management Institute, School of Resource and Environmental Management, Simon Fraser University, Burnaby, BC, Canada V5A 1S62 Fish Ecology and Conservation Physiology Laboratory, Department of Biology and Institute of Environmental Science, Carleton University, Ottawa, ON, Canada K1S 5B63 Pacific Salmon Ecology and Conservation Laboratory, Department of Forest and Conservation Sciences, University of British Columbia, Vancouver, BC, Canada V6T 1Z44 Department of Sociology and Anthropology, University of Ottawa, Ottawa, ON, Canada K1N 6N55 Department of Zoology and Faculty of Land and Food Systems, University of British Columbia, Vancouver, BC, Canada V6T 1Z46 Fisheries and Oceans Canada, Science Branch, Pacific Biological Station, 3190 Hammond Bay Road, Nanaimo, BC, Canada V9T 6N7Editor: Craig Franklin * Corresponding author: Fisheries and Oceans Canada, Science Branch, Cooperative Resource Management Institute, School of Resource and Environmental Management, Simon Fraser University, Burnaby, BC, Canada V5A 1S6. Tel: +1 604 666 5671. Email: david.patterson@dfo-mpo.gc.ca2016 26 8 2016 4 1 cow02627 2 2016 30 5 2016 7 6 2016 © The Author 2016. Published by Oxford University Press and the Society for Experimental Biology.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Science advice based on physiology is supporting harvest decisions for sockeye salmon by providing a mechanistic understanding for in river mortality. This success is a function of political will, clarity in management objectives, and science management integration. Uncertainty in results and institutional caution are major challenges for using science advice. The inability of physiologists to effect change in fisheries management has been the source of frustration for many decades. Close collaboration between fisheries managers and researchers has afforded our interdisciplinary team an unusual opportunity to evaluate the emerging impact that physiology can have in providing relevant and credible scientific advice to assist in management decisions. We categorize the quality of scientific advice given to management into five levels based on the type of scientific activity and resulting advice (notions, observations, descriptions, predictions and prescriptions). We argue that, ideally, both managers and researchers have concomitant but separate responsibilities for increasing the level of scientific advice provided. The responsibility of managers involves clear communication of management objectives to researchers, including exact descriptions of knowledge needs and researchable problems. The role of the researcher is to provide scientific advice based on the current state of scientific information and the level of integration with management. The examples of scientific advice discussed herein relate to physiological research on the impact of high discharge and water temperature, pathogens, sex and fisheries interactions on in-river migration success of adult Fraser River sockeye salmon (Oncorhynchus nerka) and the increased understanding and quality of scientific advice that emerges. We submit that success in increasing the quality of scientific advice is a function of political motivation linked to funding, legal clarity in management objectives, collaborative structures in government and academia, personal relationships, access to interdisciplinary experts and scientific peer acceptance. The major challenges with advancing scientific advice include uncertainty in results, lack of integration with management needs and institutional caution in adopting new research. We hope that conservation physiologists can learn from our experiences of providing scientific advice to management to increase the potential for this growing field of research to have a positive influence on resource management. Migration mortalityscientific advicesockeye salmonthermal physiologyPacific Salmon Commission Southern Endowment Fund (NSERC)-DiscoveryNSERC-StrategicNSERC Ocean Tracking Network CanadaACCASPPlease write ACCASP in full.Genome BCPacific Salmon Foundation(DFO) Environment Watch Program ==== Body Introduction The recent expansion of knowledge in the field of conservation physiology (see Wikelski and Cooke, 2006; Cooke et al., 2013; Lennox and Cooke, 2014) has led to a growing recognition by researchers and managers alike of the potential benefit that physiological studies can have in addressing some of the key challenges in aquatic conservation (Young et al., 2006; Horodysky et al., 2015). The major advantage of applying physiological principles to management problems is the more complete mechanistic understanding (i.e. cause and effect) that traditionally comes from physiology (Cooke and O'Connor, 2010; Horodysky et al., 2015). Therefore, a defensible physiological explanation can be desired by managers to help inform decisions that have previously been based on simple correlative associations of different stressors and fish survival derived mainly from ecological studies (e.g. Macdonald et al., 2010). The challenge for researchers is the paucity of information on how best to transfer physiological research at the individual level to scientifically defensible predictions of population-level consequences for aquatic organisms that are desired by managers (Coristine et al., 2014). Much has been written about how different scientific disciplines create knowledge (Knorr Cetina, 1999; Becher and Trowler, 2001). In the life sciences, this process usually begins with a simple idea that, if deemed sufficiently interesting and promising, is refined using basic observations and preliminary research, further developed by testable hypotheses and in-depth descriptions through to quantifiable predictions, and finally, worked into general principles that allow us to make confident assertions about nature. At each step, the idea may be abandoned or changed, and very few scientific ideas evolve into general principles widely accepted by peers (Latour, 1987). Less familiar to researchers is the parallel, stepwise progression that fisheries managers commonly use ‘to learn’ about their environment in order to make informed decisions and develop policies (Fig. 1). For managers, the corollary process involves noting anomalies, documenting repeated patterns, articulating the management problem and surveying for solutions, defining the management objectives and constraints in relationship to the problem, communicating the potential mitigation measures with interested parties and developing general policy statements to cover future related issues. Managers often use different heuristics to find solutions, such as experiential learning and, as such, they do not explicitly need to engage in the scientific process to assist them (Fig. 1; Pullin et al., 2004). In order to begin to understand the knowledge-transfer challenges faced by researchers, it is important to note that these different learning methods used by managers and researchers can be completely independent, parallel processes that are designed to serve the needs of their respective vocations (Young et al., 2013). Figure 1: The learning methods used by managers and researchers that demonstrate the different and independent parallel activities designed to serve the information needs of their respective vocations. For management, the information is used for decision-making purposes and can lead to broad policy statements. For scientists, the information is used to increase scientific understanding that can potentially lead to general principles about the natural world. This is not meant to imply that managers do not use scientific learning methods as well. Researchers and managers are embedded in very different institutional environments, with different mandates, pressures and reward systems. Social science research has shown that scientists are most influenced by their community of peers. Career rewards in the form of research funding, promotions, reputation and influence are based on the system of peer review; a system that encourages scientists to speak predominantly amongst themselves and to communicate in ways that are most useful to peers rather than potential outside users of their knowledge and findings. Managers, on the contrary, are judged on their successful handling of real-world problems that involve a wide range of other actors, including politicians, bureaucrats, conservation groups, indigenous groups holding special rights and other users of the resources. In this complex institutional environment, managers draw on multiple forms of social and ecological knowledge, including personal and collective experience, professional judgement, intuition and direct observation, as well as scientific models and data (Pullin et al., 2004; Fazey et al., 2006). We therefore cannot assume that new science will ‘trickle down’ to management decision-making simply because of its quality or (presumed) relevance. Instead, we argue that better communication and coordination between researchers and managers can be achieved by integrating the stepwise pathways for increased understanding that we illustrate in Fig. 1. Scientific advice can be given to aid in the development of management policy or, more commonly, it can be given to aid in specific management decisions. The latter is the focus of this review. We have limited our discussion of scientific advice to the provision of scientifically defensible, transparent and reproducible information that can help to inform specific management decisions. Based on our experiences of providing scientific advice to aid in decision-making for management, we propose an idealized model of generating scientific advice using five levels that parallel the stages of learning activities and knowledge acquisition for management and science illustrated in Fig. 1. Each scientific advice level represents an increase in the veracity of the scientific advice. This continuum starts with Level 1 advice that provides simple educated thoughts about how the world might work and proceeds through to Level 5 advice that generates prescriptions using quantifiable predictions of outcomes that can be directly used in management decisions. We have modified Fig. 1, which illustrates how both groups seek to acquire information, reducing uncertainty and gaining insight, to reflect a more co-dependent activity relationship that can generate quality scientific advice (Fig. 2). Figure 2 presents the different levels of scientific advice that can result once a management problem can be posed accurately as a scientific question. The ability of science to respond at a given scientific advice level will depend on the state of knowledge and connectivity to management. The level of scientific advice is defined from the scientific perspective of increasing scientific defensibility (transparency, repeatability, strength of evidence in quantity and consistency) and reducing uncertainty (predictive power), as well as increasing the potential utility within management (level of integration with management problems, objectives and constraints). The communication pathways and management activities that facilitate connectivity are illustrated to emphasize that advancement beyond Level 3 scientific advice requires a more co-dependent relationship (Fig. 2). Figure 2: The idealized integration of management and scientific activities for the provision of different levels of scientific advice to aid decision-making. The dashed lines represent the communication between managers and researchers that is necessary to promote the advancement in the level of scientific advice provided (see ‘Communication’ section). The diagram is presented from the researcher's perspective of trying to advise managers. The level of scientific advice is a function of both the state of scientific knowledge determined by the scientific learning activity (denoted in italics) and the specific integration activity with management (denoted in bold). For example, prescriptions (i.e. Level 5 scientific advice) result from a combination of high-quality science (i.e. predictions) and joint communication activities with managers and affected user groups. There are an increasing number of examples of conservation physiology being used in the development of scientific advice for resource management (e.g. Cooke et al., 2012; Madliger et al., 2015), but there are very few details on the process and difficulties associated with integrating physiological results with scientific advice and management needs. In response, we provide our perspective of the overall integration process by providing detailed examples of successful and unsuccessful attempts at applying conservation physiology to fisheries management problems from our work. More specifically, we have drawn examples from our biological research to communicate the emerging role that physiology is having in raising the level of scientific advice being given to aid the management of Fraser River sockeye salmon (Oncorhynchus nerka) fisheries. This iconic salmon species is arguably the most culturally important and well-studied Pacific salmon in Canada (reviewed by Hinch et al., 2012; Johnson et al., 2012). The remarkable anadromous migrations of semelparous adult sockeye salmon have long fascinated physiologists independent of an explicit management problem, with early research being focused on energetics (Idler and Clemens, 1959; Brett, 1965), homing ability (Fagerlund et al., 1963) and senescence (McBride et al., 1963; Fagerlund, 1967; Donaldson and Fagerlund, 1970). More recently, the uncertainty surrounding the causes of declines in Fraser River sockeye salmon abundance caught national political attention, resulting in a major $30+ million federal judicial review called by the Prime Minister of Canada (Cohen, 2012). This latest judicial review is part of the long history of political interest in the migration problems of Fraser River sockeye salmon, starting with the Hells Gate slide (e.g. Ricker, 1947; Larkin, 1992; Fraser, 1995; Williams, 2005). It is against this interesting scientific and long-standing political background that our group focused our physiological research on major factors that can impact the upstream migration success of adult sockeye salmon returning to spawn in the Fraser River, with an aim to provide scientific advice relevant to the conservation of this important population. In the first part of this paper, we conduct an historic review of the biological research on the impacts of adverse environmental conditions (high discharge and high water temperature) on in-river mortality of Fraser River sockeye salmon. The aim is to characterize the five levels of scientific advice and to showcase the recent emergence of physiology in connection with the quality of scientific advice. This recap identifies the conditions in the management and science realms that we associate with success in providing scientific advice at higher levels. In the second part of this paper, we discuss the current scientific advice given to management based on the physiological research into the role of pathogens, sex and fisheries interactions in understanding in-river mortality. The purpose of the second section is to showcase both the breadth of success in using conservation physiology and the current challenges for advancing the utility of this information for management. Our goal is to provide other conservation physiologists with an understanding of the steps that are likely to be involved in gaining management uptake of their research, based on our successes and challenges using the rich history of biological and physiological research conducted on Fraser River sockeye salmon. We recognize that local geopolitical, sociocultural and institutional norms will influence the extent to which the specifics described here apply, but we submit that the lessons learned and general approach described herein should be of broad relevance to those working in different jurisdictions on different species and issues. Part I: levels of scientific advice for in-river mortality This section focuses on the historical progression of scientific advice related to water temperature and discharge that has been provided to management in response to high in-river mortalities of Fraser River sockeye salmon. For each level of scientific advice, we describe the type of scientific activity that has occurred and the resulting scientific advice that has been provided to managers through time. We also discuss the scientific and management activities that facilitate a higher level of scientific advice. These examples help to characterize our idealized model of scientific advice (Fig. 2). The earliest written records of run failure associated with Fraser River sockeye salmon date back to catch records from the Hudson Bay Company in the 1800s (Cooper and Henry, 1962). Although managers took note of such anomalies, they did not make any connection to specific environmental conditions, such as high discharge or water temperatures. The first series of notable run failures associated with environmental conditions for Fraser River sockeye salmon occurred in the early 1900s, with a series of major rockslides resulting from railroad construction in the Fraser Canyon at Hells Gate. This culminated in effectively delaying and, in some cases, completely blocking upstream migration for large numbers of migrating sockeye salmon (Thompson, 1945). The collapse of the record 1913 cohort of sockeye salmon 4 years later (catch declined from 32 million in 1913 to 7 million in 1917) was further confirmation of a major problem for management to consider (Roos, 1991). Since then, it has been recognized that a variable portion of Fraser River sockeye salmon that are estimated to enter the lower Fraser River are not accounted for at the spawning grounds after adjusting for estimates of in-river harvest. The source of this discrepancy is a combination of assessment errors in estimates of catch, and lower and upper-river estimates of abundance, as well as natural in-river mortality (Patterson et al., 2007b). Estimates of these discrepancies over the past 20 years have a cumulative total net loss of 18 million fish. In comparison, 80 million fish were harvested and 70 million escaped to spawn during the same time period (Pacific Salmon Commission, unpublished data). These in-river losses represent substantial foregone opportunities for First Nation communities to access salmon for food, social and ceremonial purposes, limitations for angler participation, millions of dollars in lost revenue to the salmon fishing industry, reductions in spawners for other ecosystem values, and loss of future recruitment. Therefore, in-river mortality associated with adverse migration conditions provides a clear problem, for which researchers can provide scientific advice on likely factors and mechanisms contributing to mortality and, thus, speak to the efficacy of potential mitigation measures. We will describe some of the research and information given at different levels of scientific advice over the years in relationship to in-river mortality. Level 1: notions The first level of scientific advice, and lowest level of veracity, in response to a management problem involves the simple generation of ideas based on the researcher's current knowledge base. This brainstorming is done without any direct empirical testing, before providing their best ‘notions’ to management, and is therefore subject to the potential cognitive frailties associated with any expert opinions (Sutherland and Burgman, 2015). We are not privy to the preliminary ideas that researchers shared (if any) with managers regarding the impact of environmental conditions on in-river mortality of Fraser River sockeye salmon in response to the first records of run failures in the 1800s. We speculate that any notions, if shared with management, had minimal impact on decisions, given that they are not recorded in any historic management documents. Level 2: observations The next level of scientific advice requires researchers to formulate a testable hypothesis regarding their ideas and start fundamental (also known as basic or discovery) research. The scientific advice at Level 2 is based on the scientific ‘observations’ derived from this basic research. In our example, observations of delays and downstream mortalities in the Fraser Canyon were meticulously recorded by field experts during the major rock slides from 1911 to 1914 (Thompson, 1945). It was theorized that the extreme hydraulic challenges created by the rock slides exceeded the swimming ability of most of the upstream-migrating sockeye salmon (Talbot and Jackson, 1950). This led fisheries biologists who worked for the management agencies responsible for the fisheries to test ideas regarding swimming performance and to assess passage ability directly using large-scale tagging projects (Thompson, 1945; Talbot and Jackson, 1950). Interestingly, not all fisheries scientists were in agreement with respect to the central role that migration barriers at Hells Gate had in limiting the productivity of stocks upstream of the barrier. Ricker (1947) had serious reservations about the quality of the tagging studies to the extent that he questioned the primacy of hydraulic barriers over fishery exploitation rates as being the primary cause for depressed sockeye salmon stocks. However, based on Thompson's (1945) interpretation of this preliminary research, management expeditiously responded by constructing fish passage facilities at Hells Gate in the late 1940s. The magnitude of the response by management to Level 2 scientific advice shows that even preliminary or contested research can have a high degree of impact on management decisions, but instances such as this are rare and usually applied at a small scale or a single site so that management can assess its effectiveness before widespread application (Gross, 2010). Prior to the 1960s, there is almost no explicit mention of thermal impacts in relationship to adverse migration conditions for sockeye salmon (Foerster, 1968) even though physiologists had long since recognized temperature as a principal factor in controlling biological processes and ultimate survival in fish (Fry, 1947). In Foerster's (1968) comprehensive review of sockeye salmon research and management, he mentions only 2 years with reports of in-river mortality, 1942 and 1958. Only the former event was linked explicitly to high temperatures, despite the fact that the latter was one of the warmest years on record for the Fraser River (Patterson et al., 2007a). The 1958 in-river loss estimate was 7.9% of the total run, considered above average at the time and enough for management to report in the annual summaries (IPFSC, 1959), but insufficient to warrant a management response. At this point, any scientific advice associated with thermal physiology would probably have triggered the curiosity of management. However, without an obvious link to a specific management problem (see proposed connectivity in Fig. 2), such as repeated observations of high mortality events with high water temperature, and with an absence of direct physiological research on adult sockeye salmon thermal tolerance, it probably would have had a minimal impact. Level 3: descriptions The third level of scientific advice is founded on the research results from more scientifically rigorous studies. This requires the researcher to work on generating biologically based descriptive models related to understanding the problem as identified by management and to present these model outputs (i.e. ‘descriptions’) to managers as scientific advice. In the case of the impacts of environmental conditions on in-river migration mortality, the continued challenge for researchers was to seek robust, scientific explanations for these in-river losses. In the 1960s, it became clear that high water temperatures, in addition to high discharge, during in-river migration were linked to natural mortality. It was the researchers within the agencies responsible for managing the fisheries who started a lot of the work on water temperature impacts on sockeye salmon physiology, examining the connections to disease (Colgrove and Wood, 1966; Williams, 1973), swimming ability (Brett, 1965) and upper thermal limits (Servizi and Jensen, 1977). This latter and oft-cited work on thermal tolerance was in direct response to a management request to determine the thermal impact of a major water diversion project in the Fraser River watershed. Unfortunately, the findings from this work probably set back further research on thermal physiology for several decades because of misinterpretation of the scientific results on the ecological relevance of upper lethal thermal limits. Further work on the impacts of high discharge conditions continued to stay ahead of temperature impacts. The involvement of our group with the research on Fraser River sockeye salmon migration problems in the 1990s was initially focused on high discharge and not thermal physiology. The aim was to describe swimming performance (Hinch et al., 1996), variation in swimming behaviour (Hinch and Rand, 1998) and, ultimately, migration survival (Rand and Hinch, 1998; Hinch and Bratty, 2000) associated with the hydrological challenges of the Fraser River. This work provided physiological evidence to explain why certain sections of the Fraser River were more challenging for fish to ascend than others, moving beyond unreliable observations of carcasses (Patterson et al., 2007b) and weak correlations of discharge with in-river loss estimates (Macdonald, 2000) to gain a better mechanistic understanding of in-river mortality. These physiological descriptions of the impact of high discharge on sockeye salmon were provided to management and influenced some of the early harvest decisions regarding adverse migration conditions (Macdonald, 2000; Macdonald et al., 2000). Moreover, the modelling of energy expenditures in relationship to encounter velocities led to more physiological research on the role of water temperature in salmon migration metabolism, an essential component of bioenergetics modelling for poikilotherms (Rand et al., 2006). Today, the direct and indirect physiological impacts of high water temperature comprise a large portion of the Level 3 scientific advice given to managers regarding the factors that impact in-river survival of Fraser River sockeye salmon. This research on thermal physiology has developed from historic pattern recognition of increasing high mortality events associated with high migration temperatures (Gilhousen, 1990) to a focus on elucidating physiological mechanisms that will reduce the uncertainty associated with understanding thermal-based mortality in wild salmon. Our group has been studying the myriad of ways that water temperature is impacting survival by measuring a variety of physiological responses, including swimming ability and behaviour (MacNutt et al., 2006), cardiorespiratory performance (Farrell et al., 2008; Eliason et al., 2011, 2013), disease progression (Wagner et al., 2005; Crossin et al., 2008; Mathes et al., 2010), genomic and cellular responses (Jeffries et al., 2012a, b) and metabolism (Clark et al., 2010). The overall scientific advice from this descriptive work is consistent; high water temperature and high discharge have a negative impact on many aspects of salmon physiology and, ultimately, the survival of sockeye salmon (reviewed by Cooke et al., 2012; Hinch et al., 2012). Managers are now well informed regarding the physiological impacts of high temperature and high discharge and can consider this information to account for in-river losses. In addition, the stage is set to move these descriptive research results to Level 4 scientific advice by aligning management objectives with predictive models that relate environmental conditions to in-river mortality. Level 4: predictions The fourth level of scientific advice presents predictions of different outcomes to aid management in decision-making. This means moving from descriptive models focused on an improved biological understanding to predictive models, complete with an estimate of uncertainty (Harwood and Stokes, 2003; Ascough et al., 2008). Ideally, this is a combination of more strategic research and different analytical approaches on the part of the researcher, and more precise feedback from managers regarding their objectives and constraints. Management feedback is required to provide a more thorough description of the fishery, including the following: the legal framework, i.e. defining the management actions the agency have regulatory control over (e.g. the spatial location of fisheries); the goals, i.e. articulating the objectives of the fishery (e.g. the alternative goals of maximizing total harvest vs. fishing opportunity); and the operational constraints, i.e. communicating the practical limitations of executing a fishery (e.g. the lead time required to open or close a fishery). Management needs to determine how the results generated from a quantitative model that can predict particular outcomes could be used under the existing operational and regulatory constraints. For Fraser River sockeye salmon, a major reason why the physiological research on the impacts of adverse migration conditions is currently used to support harvest management decision lies in the clarity of how the information fits into the overall management process. There is a clear legal mandate laid out in the 1999 Pacific Salmon Treaty (bilateral agreement between Canada and USA), in Article VI, Annex IV, Chapter 4 (10), to prevent overfishing by both countries. The treaty states that spawning escapement goals are a clear management objective of the agreement, and there is a clear mechanism to incorporate scientific advice into the process, as stated in Article VI, Annex IV, Chapter 4 (13b): ‘incorporate … management adjustments [harvest changes] that deal with environmental conditions [discharge, temperature] during in-river migration that could significantly impact the Fraser River Panel's ability to achieve spawning escapement objectives’ (Pacific Salmon Treaty, 1999). This legal background has helped to inform how the research is conducted and the types of questions asked in the development of both descriptive and predictive models. The descriptive models used in generating Level 3 scientific advice are about understanding biological relationships between adverse migration conditions and in-river mortality. The model selection criteria typically rely on biological realism (i.e. physiological support), model fit and model sensitivity. The main target audience for descriptive models is other researchers (i.e. for primary publication), whereas the predictive models, such as those described herein, are built for applied management purposes. As such, the developers of predictive models that forecast events for managers have to consider some key additional features in model selection, including model predictive power (e.g. bias and precision), forecasting constraints for predictor variables (e.g. water temperature and discharge forecasting) and management constraints (e.g. lead time to adjust fish harvest). The last of these features reflects the fact that predictive models are built to predict the outcome of different management responses within a realistic set of conditions. In the case of Fraser River sockeye salmon, in-season harvest adjustment models that use water temperature as a predictor variable must rely on forecasted temperatures (Hague and Patterson, 2014). This is to allow time to adjust harvest that normally occurs seaward (i.e. downstream and earlier) of the potentially damaging high water temperatures that the fish would subsequently experience in the river. In simple terms, throughout the fishing season the water temperature and discharge are forecasted, an estimate of loss is predicted by the models that use the forecasted environmental conditions, and harvest can be adjusted according to the expected losses (Hague and Patterson, 2007). These quantitative models that use water temperature to forecast in-river mortality rely on physiology as the primary rationale (Macdonald et al., 2010). The models work by quantifying the historic relationship between in-river loss and different metrics of water temperature and discharge. The water temperature metrics that are used reflect both mean temperature exposure and the threshold responses to high temperatures. The former is justified by physiological research on energy expenditures and disease progression (e.g. Wagner et al., 2005; Rand et al., 2006; Crossin et al., 2008), and the latter is supported by research on aerobic scope and cardiac failure (e.g. Farrell et al., 2008; Mathes et al., 2010; Eliason et al., 2011). Today, fisheries management is presented with an estimate of loss, with sufficient lead time to adjust harvest and proactively change the probability associated with achieving escapement goals (Macdonald et al., 2010), with the knowledge that the scientific advice is supported by internationally recognized physiological research. Level 5: prescriptions This fifth level of scientific advice requires the development of scientific prescriptions as part of an integrated management approach. At Level 4, researchers have collated and provided a synthesis of their results, including an appropriate disclosure of uncertainty for any predictions they make. Level 5 provides a prescription on how to use these predictions as part of a structured management decision process. This requires more communication with management and affected parties to explain the methods, biological rationale and uncertainty, as well as the strengths and limitations of the science and analytical techniques. Our group has had success in integrating scientific advice on environmental impacts into fisheries management decisions at this level through additional model performance evaluations and continual communication and engagement with managers and other interested parties (see ‘Communication’ section below). To get to this point, scientists and managers worked together to create a management prescription to outline how scientific advice on the impacts of water temperature and discharge on sockeye salmon mortality (in the form of model predictions) can be used in harvest planning. For Fraser River sockeye salmon, this involved adjusting harvest plans pre-season, using long-range forecasts of summer water temperatures and discharges in in-river loss models (Patterson and Hague, 2007) and, in-season, using forecasts of water temperature and discharge in similar in-river loss models (Macdonald et al., 2010). The biological rationale required to convince managers and educate user groups to support the use of these models, the harvest outcomes of which can have major financial and social consequences, is based in large part on physiological research. The totality of research used to support the numerical models includes >30 technical reports and >60 primary publications, the majority of which include physiology to seek mechanistic understandings of temperature- and discharge-related mortality (reviewed by Hinch et al., 2012; Johnson et al., 2012). Researchers were encouraged by managers to share physiological research and numerical modelling results with representatives of recreational, commercial, First Nations and conservation groups to facilitate acceptance of science-based prescription. All groups interested in the process were made aware of the research, had an opportunity to comment on preliminary results, and provided constructive feedback on new research ideas. When final decisions regarding harvest were being made, the research behind those decisions was not a surprise to those people who would be impacted by the harvest changes. Science is but one source of advice that managers will use in making choices (Rice, 2011); arguably, it should be the most transparent and repeatable. Part II: successes and challenges of advancing physiology-based advice There is a long history of researchers providing descriptive physiological results to management to help explain the impact of factors other than water temperature and discharge on the in-river mortality for Fraser River sockeye salmon. Early work that pre-dates our group includes field studies and laboratory experiments that examined physiological aspects of migratory difficulty and energy allocation (Gilhousen, 1980), cumulative stress (Fagerlund et al., 1995), disease progression (Colgrove and Wood, 1966) and suspended sediments (Servizi and Martens, 1987). In this section, we provide examples of our research group using physiological research on pathogens, sex and capture stress to gain a better understanding of in-river mortality in order to elucidate the challenges and successes in converting this type of work into advice for management. Pathogens The role of pathogens has long been associated with in-river mortality of sockeye salmon (Williams, 1973; St Hilaire et al., 2002; Jones et al., 2003; Miller et al., 2014). Early histological examinations of sockeye salmon that died prematurely in the river found a suite of different pathogens that vary annually (Wood, 1965; Williams, 1973), making it difficult to link a specific pathogen to the cause of death. Our more recent work on matching histopathology with host physiological response has provided a better understanding of the mortality associated with some pathogens (Wagner et al., 2005; Crossin et al., 2008; Bradford et al., 2010). In addition, we have started to examine transcriptional responses of sockeye salmon to different types of potential pathogens. The results from these studies have shown that genomic signatures associated with an immune response have the potential to predict migratory failure (Miller et al., 2009, 2011). The overall scientific advice from this work has confirmed the potentially important role that pathogens and associated diseases can play in accounting for in-river mortality, but the real utility for management has stalled at this descriptive stage. We are not yet in a position to recommend using pathogen loads, blood chemistry or gene expression patterns to predict fate at the population level for returning wild sockeye salmon given the high uncertainty (i.e. low overall variance in survival explained) in the results (Cooke et al., 2006; Miller et al., 2011). For example, in the study by Hammill et al. (2012), individuals from one of the three populations examined had a different suite of gene expression profiles that could predict fate. In order to potentially progress beyond Level 3 scientific advice, more research is being conducted using novel genomic approaches (e.g. Evans et al., 2011; Miller et al., 2014) to elucidate stock-specific differences in disease susceptibility that might arise from genetic and environmental influences, distance to the spawning grounds and the probability of pathogen exposure in order to reduce the high uncertainty associated with episodic disease events. More thought is being given by managers with respect to how these results could be used in management, presuming the uncertainty can be reduced. At this time, management acknowledges that there is a physiological explanation for disease-related mortality, but the utility of this scientific advice for predictive purposes will be contingent on the success of planned future work. Sex differences We are closer to making useful predictions for management in our next example, sex-specific mortality patterns. A common observation in our years of research examining factors related to in-river mortality is that female sockeye salmon suffer higher rates of mortality than males in response to stress (e.g. Patterson et al., 2004; Gale et al., 2014). The evidence for a sex bias is based on both laboratory holding studies and field telemetry studies for which we were able to document sex. In many of the holding studies, the mortality of females was twice that of males (Crossin et al., 2008; Gale et al., 2011; Robinson et al., 2013). The differences detected in the field tagging studies were not as large, but the mortality spread did become magnified with elevated water temperatures (Martins et al., 2012). Many of these studies also had supporting physiological measures that provided some potential mechanistic link to the observed differences in survival. Based on a combination of empirical data on sex-specific differences in mortality rates and physiological support in describing this mortality, we were confident in informing management that adult female sockeye salmon die at a higher rate than males in stressful conditions. As such, we unilaterally moved (i.e. without management support) from Level 3 scientific advice to work on population-level predictions consistent with our individual descriptive research. Unfortunately, the development of any predictive models using sex-specific mortality has stalled at Level 4 because of problems with both the science and management. From the scientific perspective, the patterns of differential mortality reported in our telemetric and holding studies were not reflected in the standard Fisheries and Oceans Canada (DFO; department responsible for managing Pacific salmon in Canada) annual stock assessment spawning ground enumeration studies, assuming a 50:50 ratio for salmon starting the spawning migration; however, annual information on variation in sex ratios of adults returning to the lower river are limited (Foerster, 1968). The ratios of male to female spawners do not appear to vary as a function of high migration temperatures as predicted by our research; the ratios only seemed to vary as a function of extreme high discharge years (Macdonald, 2000; D. Patterson, DFO, unpublished data). Hence, scaling up from research using individuals to the population-level responses did not occur. Moreover, for management, there is no clear way to use the sex-based differences without changing the management objectives. The present goals for achieving spawner escapements for Fraser River sockeye salmon are neutral to sex (i.e. no mention of sex-specific goals in the Pacific Salmon Treaty). We are now left with simply providing suggestions to both our peers and managers on how to proceed further with developing predictive models based on sex-based differences in survival. Researchers need to look more closely at the sex-ratio information collected on the spawning grounds and in downstream fisheries to determine the statistical power to detect varying levels of differential survival (i.e. effect size) given the current fisheries assessment methods. This is likely to be a common challenge in physiological research, because stock assessment information may not be collected at a sufficient level of precision to match the ability to detect a predicted response. More work is needed on the part of physiologists to determine whether experimenter effects of holding or tagging fish are confounding the survival estimates for females. Managers could also re-examine changes to the spawner goals that include female-specific targets. This example has shown that not all research will be useful to management immediately, and there is a risk in moving from Level 3 to Level 4 scientific advice if it is not done with close collaboration of science and management (Fig. 2). Fisheries interactions Our next example uses capture-and-release mortality research to show a more direct connection between the researchers and managers in moving advice from Level 3 to our current attempts to provide Level 5 scientific advice. During the past 10 years, we have focused a large portion of our efforts on understanding the fate of Pacific salmon released after capture, using physiology to elucidate the mechanisms behind lethal and sub-lethal responses (Raby et al., 2012, 2015a, b). Major findings from this work include the following: the functional basis for the differences in mortality associated with different gear types (Donaldson et al., 2011, 2012, 2013); the role of injury in causing physiological stress and mortality (Nguyen et al., 2014); the mixed benefits of using recovery methods (Donaldson et al., 2013; Robinson et al., 2013; Raby et al., 2015c); the among-population variation in mortality responses (Donaldson et al., 2012; Robinson et al., 2015); the negative impact of high temperature associated with capture and handling (Robinson et al., 2013; Gale et al., 2014); and the changes in stress responsiveness with maturation stage (Gale et al., 2011; Raby et al., 2013). All of this work is directly connected to the management problem of trying to describe mortality for post-season accounting of fishing impacts or to predict mortality for harvest planning purposes. The generic Level 3 scientific advice provided to management is that capture-and-release mortality can be understood better through physiology, but it is highly context dependent (Raby et al., 2015a), making it challenging to predict. Predictions of capture-and-release mortality associated with different fishing gear and at different temperatures have been generated based on field telemetry and holding studies that couple survival and physiology. For example, this work has shown that long-term mortality rates for beach-seined and angled sockeye salmon range from 20 to 30% during average summer water temperatures of 18°C (Donaldson et al., 2011, 2013). However, mortality will rise rapidly as capture and handling temperatures increase above 19°C, and if they persist above 21°C, there is almost 100% mortality within 4 days (Gale et al., 2011; Robinson et al., 2013). This advice has been presented to managers for potential use in harvest planning. As with research on thermal impacts on in-river mortality, there is a clear avenue where scientific advice on post-release mortality can be used by management. Each sockeye salmon fishery on the Fraser River has a post-release mortality rate value based on gear type and location (Fisheries and Oceans Canada, 2013). Therefore, management could update their post-release mortality estimates using scientific advice for the different sockeye salmon fisheries, warranting the transition from Level 4 to Level 5 scientific advice. The time line for managers and researchers to develop a management prescription and start applying new scientific advice at Level 5 will vary. The new research behind the post-release mortality rates for sockeye salmon (reviewed by Raby et al., 2015a) has not yet become a part of the management prescription. In this example, the reasons for delay are related to institutional caution and research uncertainty. Institutional resistance to change is common; managers can be cautious when faced with new information and may invoke processes to obtain feedback beyond science before proceeding (Young et al., 2013). Furthermore, there is still large uncertainty in the estimates we have derived, and the new mortality rates are, in some cases, considerably higher than those currently used. The higher rates are due, in large part, to the fact that our values are calculated using a longer period for monitoring mortality. The current rates used by management are typically based on 24 h post-release monitoring periods, compared with our estimates that are based on at least 96 h of post-release monitoring. This longer monitoring duration was part of a recent request by fisheries managers to include delayed mortality associated with fishing interactions. The current plan for incorporating new information as official scientific advice will require further meetings with various stakeholders and First Nations. As part of this plan for advancing the research on catch-and-release mortality, there is a formal request from the fisheries management sector to the science sector of DFO to develop scientific advice on updating post-release mortality rates using relatively new research. In Canada, we have a formal mechanism under the Canadian Science Advisory Secretariat that allows fisheries management to request formal scientific advice from their own agency. Our research team has been commissioned by the Canadian Science Advisory Secretariat to write the research document that will be the basis for this advice. The document will involve an in-depth look at the mechanistic (i.e. physiological) basis for how different factors impact fishing-related mortality, as well as a review of the mortality rates themselves. In other words, the official scientific advice will be based in large part on the ability of physiologists to explain why certain factors, such as injury, air exposure, handling time, capture time and revival methods, are important in generating estimates of fishing-related mortality (reviewed by Raby et al., 2015a). This information will be shared with all groups with a vested interest in the salmon fishery. Changes to mortality rates will not be made until all groups have had an opportunity to comment and process the new information; underscoring the central role that communication plays in any plan that provides scientific advice to management. Communication Scientific advice is not being given in a vacuum; meaning that this information will disseminate beyond the initial management audience, and therefore, it is important to be cognisant of the interpretation and use of this information by different groups. A communication plan is likely to be essential for success at the fifth level, but ideally it should be initiated the first time that managers and researches begin to exchange information and be the responsibility of both managers and researchers. The first priority for communicating research results is to satisfy management requirements, but researchers cannot be naïve to how other groups, such as the media, environmental groups, scientific community and fishing groups, can accidently misinterpret or deliberately spin the results. Misrepresentations can potentially derail the ability of managers to use physiological research, as exemplified earlier with the focus on upper lethal thermal limits (Servizi and Jensen, 1977), because they typically rely on information and feedback from these same outside groups to help make decisions (e.g. Cohen, 2012). It is worth repeating that scientific advice is only one source of information that managers rely on to make decisions (Rice, 2011). It is the responsibility of researchers to communicate clearly both the strengths and the limitations of the research to managers and a broader audience. In practical terms, this means a full disclosure of the uncertainty in the results, as well as clearly stating how the research does and/or does not relate to management or any other issue a stakeholder may decide to link it to (Regan et al., 2002; Harwood and Stokes, 2003). Managers, in turn, need to review the work critically, anticipate future criticisms and prepare researchers for stakeholder responses to their work. Unfortunately, the primary media used by researchers for communicating science (peer-reviewed journal articles) are likely to be one of the least effective or desirable methods for communicating this information to either management or other interested groups (Nguyen et al., 2012). Increasing the impact and relevance of physiological research will require not only good science but also effective communication with management and other interested parties. Synopsis The reasons for the success of our work include political motivation, funding, accountability, legal clarity, institutional environment, personal relationships and peer acceptance. For example, fisheries management recognized the potential problem of temperature-based in-river loss after a series of high-profile reviews in the early 1990s (Larkin, 1992; Fraser, 1995) and again in the early 2000s (Canadian Standing Committee on Fisheries and Oceans, 2005; Williams, 2005). Each of these government-commissioned reviews led to an increase in funding for sockeye salmon research on problems related to in-river mortality. Our research group benefited from the direct connection between a management problem and funding of physiologically based solutions. This funding included opportunities for traditional national science funds (e.g. Natural Sciences and Engineering Research Council of Canada; NSERC), strategic national science funds (e.g. NSERC-Strategic and NSERC-Network), internal government science agency funds, and competitive applied research funds from the Canada–US Pacific Salmon Treaty. Researchers were then held accountable to report both to their academic peers, through publications in scientific journals, and to the funding agencies via management. This dual accountability produced quality science that was also applied in nature. The legal framework of the Pacific Salmon Treaty also provided clarity of objectives and clear mechanisms to incorporate mitigation measures based on environmental impacts. Likewise, the post-release mortality rates in the official fishery planning documents of DFO (Fisheries and Oceans Canada, 2013) provide a clear outlet for using research results from the fisheries interaction work. The institutional environments of both groups also helped. The Canadian Science Advisory Secretariat process provides a direct link between scientific research, scientific advice and fisheries management. Management agencies also invite researchers to speak at annual meetings and organize strategic workshops on critical management issues. Our research group is also proactive and continues to hold annual workshops with managers, fishing sector representatives, First Nations and conservation groups (e.g. Hinch and Gardner, 2009) to encourage feedback. The benefits of these interactions are twofold. First, the people most affected by decisions that use our scientific advice were familiar with the research prior to it being used by management. Second, it fostered improved personal relationships among all groups, which improved overall trust, a key element in knowledge transfer (Young et al., 2013). The role that individual relationships can have to ensure success cannot be underestimated; this includes access to experts with the diverse skills required to tackle complex science questions. We were fortunate to know and be able to work with managers, ecologists, numeric modellers and physiologists. This collaboration has produced physiology publications that have been recognized by researchers internationally as major scientific contributions in the field of applied genomics (e.g. Miller et al., 2011) and oxygen limitation theory (Eliason et al., 2011). The scientific peer acceptance along with the institutional and situational conditions has been helpful in gaining support from management. The major challenges to providing scientific advice beyond Level 3, the simple descriptions of nature, are rooted in the different demands of science and management. Scientists are rewarded based in large part on developing novel scientific explanations for how the world works (Knorr Cetina, 1999; Provencal, 2011). Therefore, the interest of the researcher peaks at Level 3, where they can maximize the number of novel publications of different factors that influence survival in fish. Many researchers feel that simply by providing a scientific explanation for a specific management problem, they have completed their obligation to show the applied nature of their work (Blackman and Benson, 2012). They are often unaware of the extra work required to make confident predictions regarding future outcomes tailored to objectives of a specific fishery; this work requires more research and collaboration to reduce and quantify uncertainty to a level that managers or other knowledge practitioners want (Mergel et al., 2008). There is rarely any career advantage or reward for taking this path (Young et al., 2013). Conversely, managers must make decisions in a timely manner, and they are not specifically rewarded for knowing why a particular outcome has occurred, only that it is working. Therefore, the interest of managers peaks at Level 5, when there is a clear plan to use the information to aid in management. There is often no obligation, no accountability or no reward for managers that actively seek scientific advice at Level 3 and help to develop this research into Level 5 scientific advice. For the process to be effective, the researcher needs to put the management problem in the matrix of what is already known and unknown before consuming more time and money. For example, a synthesis of the existing knowledge base can be made to determine the level of uncertainty in management outcomes that will be reduced under different research projects. It is important to recognize that all of the advice provided, even at Level 1, is building a connection and trust among managers and researchers. Therefore, if a researcher is asked by management to provide scientific advice, it is important to do so expeditiously, along with a healthy dose of uncertainty (see Regan et al., 2002), before asking for funds and deferring an answer until they have the perfect descriptive or predictive model. Managers often need to act before scientific consensus is achieved (Ludwig et al., 1993). Likewise, researchers walking away from the problem after they have published a few descriptive papers on a subject will not go far in getting their science to be effective in management, given the large gap between Level 3 and Level 5. Managers and researchers alike could also benefit from seeing examples of where the long-term benefit of scientific advice at different levels has led to better management. There is lack of retrospective evaluations of the efficacy of scientific advice to improve management performance in fisheries. The temperature- and discharge-based mortality models are an exception in that we have performed retrospective evaluations of model performance at the behest of management (Cummings et al., 2011). We have also learnt from the development of the harvest adjustment models that most of these efforts will cost hundreds of thousands of dollars, take years of sustained effort to complete, use a team of researchers willing to meet with managers and stakeholders, require continuous model development and refinement and require extensive communications in the form of primary publications, technical reports, briefing notes, management meetings and presentations for public engagement, before management will have an updated prescription process to use new information supported by the physiological research. This is not meant as criticism of either side, but simply the reality of getting your specific prescription to market. Science will always be only a part of the decision-making process, but how big a part—the level of advice it can achieve—will depend on people, managers and researchers alike, who can help to promote the use of conservation physiology. Acknowledgements We would like to thank the hordes of students, colleagues, managers, First Nations, commercial fishers, recreational fishers, conservation groups and volunteers who have contributed to the scientific activities and integrative approaches to make this research possible. Special thanks to Steve Macdonald and the Pacific Salmon Commission staff for helping to pioneer some of our early attempts to transfer scientific advice to managers. We thank the EU COST Action on the Conservation Physiology of Marine Fishes for providing the opportunity to share our experiences. Funding This work was supported by Pacific Salmon Commission Southern Endowment Fund, Natural Sciences and Engineering Research Council of Canada (NSERC)-Discovery, NSERC-Strategic, NSERC Ocean Tracking Network Canada, Fisheries and Oceans Canada (DFO), Aquatic Climate Change Adaptation Services Program (ACCASP), Genome BC, Pacific Salmon Foundation and DFO Environmental Watch Program. ==== Refs References Ascough JC , Maier HR , Ravalico JK , Strudley MW (2008 ) Future research challenges for incorporation of uncertainty in environmental and ecological decision-making . Ecol Model 219 : 383 –399 . Becher T , Trowler P (2001 ) Academic Tribes and Territories: Intellectual Enquiry and the Culture of Disciplines . McGraw-Hill , London, UK . Blackman D , Benson AM (2012 ) Overcoming knowledge stickiness in scientific knowledge transfer . Public Underst Sci 21 : 573 –589 .23823166 Bradford MJ , Lovy J , Patterson DA (2010 ) Infection of gill and kidney of Fraser River sockeye salmon, Oncorhynchus nerka (Walbaum), by Parvicapsula minibicornis and its effect on host physiology . J Fish Dis 33 : 769 –779 .20626548 Brett JR (1965 ) The relation of size to rate of oxygen consumption and sustained swimming speed of sockeye salmon (Oncorhynchus nerka) . J Fish Res Board Can 22 : 1491 –1501 . Canadian Standing Committee on Fisheries and Oceans (2005 ) 2nd Report of the Standing Committee on Fisheries and Oceans: Here We Go Again…Or the 2004 Fraser River Salmon Fishery. Parliamentary Report, Ottawa, March 2005. 91 pp. Clark TD , Sandblom E , Hinc SG , Patterson DA , Frappell PB , Farrell AP (2010 ) Simultaneous biologging of heart rate and acceleration, and their relationships with energy expenditure in free-swimming sockeye salmon (Oncorhynchus nerka) . J Comp Physiol B Biochem Syst Environ Physiol 180 : 673 –684 . Cohen BI (2012 ) Commission of Inquiry into the Decline of Sockeye Salmon in the Fraser River: The Uncertain Future of Fraser River Sockeye, Volume 1: The Sockeye Fishery. Public Works and Government Services Canada, Ottawa, ON, Canada, pp 1–722. Colgrove DJ , Wood JW (1966 ) Occurrence and control of Flavobacterium columnare as related to Fraser River sockeye salmon . Int Pac Salmon Fish Comm Prog Rep 15 : 1 –51 . Cooke SJ , O'Connor CM (2010 ) Making conservation physiology relevant to policy makers and conservation practitioners . Conser Lett 3 : 159 –166 . Cooke SJ , Hinch SG , Crossin GT , Patterson DA , English KK , Shrimpton JM , Kraak GV , Farrell AP (2006 ) Physiology of individual late-run Fraser River sockeye salmon (Oncorhynchus nerka) sampled in the ocean correlates with fate during spawning migration . Can J Fish Aquat Sci 63 : 1469 –80 . Cooke SJ , Hinch SG , Donaldson MR , Clark TD , Eliason EJ , Crossin GT , Raby GD , Jeffries KM , Lapointe M , Miller K et al (2012 ) Conservation physiology in practice: how physiological knowledge has improved our ability to sustainably manage Pacific salmon during up-river migration . Philos Trans R Soc B Biol Sci 367 : 1757 –1769 . Cooke SJ , Sack L , Franklin CE , Farrell AP , Beardall J , Wikelski M , Chown SL (2013 ) What is conservation physiology? Perspectives on an increasingly integrated and essential science . Conserv Physiol 1 : cot001 ; doi:10.1093/conphys/cot001 .27293585 Cooper AC , Henry KA (1962 ) The history of the early Stuart sockeye run. New Westminster, BC: International Pacific Salmon Fisheries Commission Progress Report 10. Coristine LE , Robillard CM , Kerr JT , O'Connor CM , Lapointe D , Cooke SJ (2014 ) A conceptual framework for the emerging discipline of conservation physiology . Conserv Physiol 2 : cou033 ; doi:10.1093/conphys/cou033 .27293654 Crossin GT , Hinch SG , Cooke SJ , Welch DW , Lotto AG , Patterson DA , Jones SRM , Leggatt RA , Mathes MT , Shrimpton JM et al (2008 ) Exposure to high temperature influences the behaviour, physiology, and survival of sockeye salmon during spawning migrations . Can J Zool 86 : 127 –140 . Cummings JW , Hague MJ , Patterson DA , Peterman RM (2011 ) The impact of different performance measures on model selection for Fraser River sockeye salmon . N Am J Fish Manag 31 : 323 –334 . Donaldson EM , Fagerlund UH (1970 ) Effect of sexual maturation and gonadectomy at sexual maturity on cortisol secretion rate in sockeye salmon (Oncorhynchus nerka) . J Fish Board Can 27 : 2287 –2296 . Donaldson MR , Hinch SG , Patterson DA , Hills J , Thomas JO , Cooke SJ , Raby GD , Thompson LA , Robichaud D , English KK et al (2011 ) The consequences of angling, beach seining, and confinement on the physiology, post-release behaviour and survival of adult sockeye salmon during upriver migration . Fish Res 108 : 133 –141 . Donaldson MR , Hinch SG , Raby GD , Patterson DA , Farrell AP , Cooke SJ (2012 ). Population-specific consequences of fisheries-related stressors on adult sockeye salmon . Physiol Biochem Zool 85 : 729 –739 .23099469 Donaldson MR , Raby GD , Nguyen VN , Hinch SG , Patterson DA , Farrell AP , Rudd M , Thompson LA , O'Connor CM , Colotelo AH et al (2013 ) Evaluation of a simple technique for recovering Pacific salmon from capture stress: integrating comparative physiology, biotelemetry, and social science to solve a conservation problem . Can J Fish Aquat Sci 70 : 90 –100 . Eliason EJ , Clark TD , Hague MJ , Hanson LM , Gallagher ZS , Jeffries KM , Gale MK , Patterson DA , Hinch SG , Farrell AP (2011 ) Differences in thermal tolerance among sockeye salmon populations . Science 332 : 109 –112 .21454790 Eliason EJ , Clark TD , Hinch SG , Farrell AP (2013 ) Cardiorespiratory performance and blood chemistry during swimming and recovery in three populations of elite swimmers: adult sockeye salmon . Comp Biochem Physiol 166 : 385 –397 . Evans TG , Hammill EDD , Kaukinen K , Schulze AD , Patterson DA , English KK , Curtis JMR , Miller KM (2011 ) Transcriptomics of environmental acclimatization and survival in wild adult Pacific sockeye salmon (Oncorhynchus nerka) during spawning migration . Mol Ecol 20 : 4472 –4489 .21951593 Fagerlund UHM (1967 ) Plasma cortisol concentration in relation to stress in adult sockeye salmon during the freshwater stage of their life cycle . Gen Comp Endocrinol 8 : 197 –207 .6040052 Fagerlund UH , McBride JR , Smith M , Tomlinson N (1963 ) Olfactory perception in migrating salmon: III. Stimulants for adult sockeye salmon (Oncorhynchus nerka) in home stream waters . J Fish Board Can 20 : 1457 –1463 . Fagerlund UH , McBride JR , Williams IV (1995 ) Stress and tolerance In Groot C , ed., Physiological Ecology of Pacific Salmon . UBC Press , Vancouver, BC, Canada , pp 461 –503 . Farrell AP , Hinch SG , Cooke SJ , Patterson DA , Crossin GT , Lapointe M , Mathes MT (2008 ) Pacific salmon in hot water: applying aerobic scope models and biotelemetry to predict the success of spawning migrations . Physiol Biochem Zool 81 : 697 –708 .18922081 Fazey I , Fazey JA , Salisbury JG , Lindenmayer DB , Dovers S (2006 ) The nature and role of experiential knowledge for environmental conservation . Environ Conserv 33 : 1 –10 . Fisheries and Oceans Canada (2013 ) Pacific region integrated fisheries management plan, salmon – southern BC. http://waves-vagues.dfo-mpo.gc.ca/Library/348529.pdf Foerster RE (1968 ) The Sockeye Salmon, Oncorhynchus nerka Fish Res Board Can Bull 162 : 1 –422 . Fraser J (1995 ) Fraser River sockeye, 1994: problems and discrepancies: report of the Fraser River Sockeye Public Review Board . Canadian Government Publishing , Ottawa, Canada . Fry FEJ (1947 ) Effects of the environment on animal activity . Publ Ontario Fish Res Lab 68 : 1 –62 . Gale MK , Hinch SG , Eliason EJ , Cooke SJ , Patterson DA (2011 ) Physiological impairment of adult sockeye salmon in fresh water after simulated capture-and-release across a range of temperatures . Fish Res 112 : 85 –95 . Gale MK , Hinch SG , Cooke SJ , Donaldson MR , Eliason EJ , Jeffries KM , Martins EG , Patterson DA (2014 ) Observable impairments predict mortality of captured and released sockeye salmon at various temperatures . Conserv Physiol 2 : cou029 ; doi:10.1093/conphys/cou029 .27293650 Gilhousen P (1980 ) Energy sources and expenditures in Fraser River sockeye salmon during their spawning migration . Int Pac Salmon Fish Com Bull 22 : 1 –51 . Gilhousen P (1990 ) Prespawning mortalities of sockeye salmon in the Fraser River system and possible causal factors . Pac Salmon Fish Com Bull 26 : 1 –61 . Gross M (2010 ) Ignorance and Surprise: Science, Society, and Ecological Design . MIT Press , Boston, MA, USA . Hague MJ , Patterson DA (2007 ) Quantifying the sensitivity of Fraser River sockeye salmon (Oncorhynchus nerka) management adjustment models to uncertainties in run timing, run shape and run profile . Can Tech Rep Fish Aquat Sci 2776 : 1 –55 . Hague MJ , Patterson DA (2014 ) Evaluation of statistical river temperature forecast models for fisheries management . N Am J Fish Manage 34 : 132 –146 . Hammill E , Curtis JM , Patterson DA , Farrell AP , Sierocinski T , Pavlidis P , Hinch SG , Miller K (2012 ) Comparison of techniques for correlating survival and gene expression data from wild salmon . Ecol Freshw Fish 21 : 189 –199 . Harwood J , Stokes K (2003 ) Coping with uncertainty in ecological advice: lessons from fisheries . Trends Ecol Evol 18 : 617 –620 . Hinch SG , Bratty JM (2000 ) Effects of swim speed and activity pattern on success of adult sockeye salmon migration through an area of difficult passage . Trans Am Fish Soc 129 : 604 –612 . Hinch SG , Gardner J (2009 ) Conference on Early Migration and Premature Mortality in Fraser River Late-Run Sockeye Salmon: Proceedings. Vancouver, BC, Canada. Pacific Fisheries Resource Conservation Council, Vancouver, BC, Canada, pp 1–120. Hinch SG , Rand PS (1998 ) Swim speeds and energy use of river migrating adult sockeye salmon: role of local environment and fish characteristics . Can J Fish Aquat Sci 55 : 1821 –1831 . Hinch SG , Diewert RE , Lissimore TJ , Prince AMJ , Healey MC , Henderson MA (1996 ) Use of electromyogram telemetry to assess difficult passage areas for river migrating adult sockeye salmon . Trans Am Fish Soc 125 : 253 –260 . Hinch SG , Cooke SJ , Farrell AP , Miller KM , Lapointe L , Patterson DA (2012 ) Dead fish swimming: a review of research on the early migration and high premature mortality in adult Fraser River sockeye salmon Oncorhynchus nerka . J Fish Biol 81 : 576 –599 .22803725 Horodysky AZ , Cooke SJ , Brill RW (2015 ) Physiology in the service of fisheries science: why thinking mechanistically matters . Rev Fish Biol Fish 25 : 425 –447 . Idler DR , Clemens WA (1959 ) The energy expenditures of Fraser river sockeye salmon during the spawning migration to Chilko and Stuart lakes . Int Pac Salmon Fish Comm Bull 6 : 1 –80 . IPFSC (1959 ) International Pacific Fisheries Salmon Commission 1959 annual report. New Westminster, BC, Canada, pp 1–40. Jeffries KM , Hinch SG , Martins EG , Clark TD , Lotto AG , Patterson DA , Cooke SJ , Farrell AP , Miller KM (2012 a) Sex and proximity to reproductive maturity influence the survival, final maturation, and blood physiology of Pacific salmon when exposed to high temperature during a simulated migration . Physiol Biochem Zool 85 : 62 –73 .22237290 Jeffries KM , Hinch SG , Sierocinski T , Clark TD , Eliason EJ , Donaldson MR , Li S , Pavlidis P , Miller KM (2012 b) Consequences of high temperatures and premature mortality on the transcriptome and blood physiology of wild adult sockeye salmon (Oncorhynchus nerka) . Ecol Evol 2 : 1747 –1764 .22957178 Johnson JE , Patterson DA , Martins EG , Cooke SJ , Hinch SG (2012 ) Quantitative methods for analysing cumulative effects on fish migration success: a review . J Fish Biol 81 : 600 –631 .22803726 Jones SR , Prosperi-Porta G , Dawe SC , Barnes DP (2003 ) Distribution, prevalence and severity of Parvicapsula minibicornis infections among anadromous salmonids in the Fraser River, British Columbia, Canada . Dis Aquat Org 54 : 49 –54 .12718470 Knorr Cetina K (1999 ) Epistemic Cultures: How the Science Makes Knowledge . Harvard University Press , Cambridge, MA, USA . Larkin PA (1992 ) Analysis of possible causes of the shortfall in sockeye spawners in the Fraser River: a technical appendix In Pearse PH , ed., Managing Salmon on the Fraser . Fisheries and Oceans Canada , Vancouver, BC, Canada , pp 1 –33 . Latour B (1987 ) Science in Action: How to Follow Scientists and Engineers through Society . Harvard University Press , Cambridge, MA, USA . Lennox R , Cooke SJ (2014 ) State of the interface between conservation and physiology: a bibliometric analysis . Conserv Physiol 2 : cou003 ; doi:10.1093/conphys/cou003 .27293624 Ludwig D , Hilborn R , Walters C (1993 ) Uncertainty, resource exploitation, and conservation: lessons from history . Ecol Appl 3 : 547 –549 . McBride JR , Fagerlund UH , Smith M , Tomlinson N (1963 ) Resumption of feeding by and survival of adult sockeye salmon (Oncorhynchus nerka) following advanced gonad development . J Fish Board Can 20 : 95 –100 . Macdonald JS (2000 ) Mortality during the migration of Fraser River sockeye salmon (Oncorhynchus nerka): a study of the effect of ocean and river environmental conditions in 1997 . Can Tech Rep Fish Aquat Sci 2315 : 1 –120 . Macdonald JS , Foreman MGG , Farrell T , Williams IV, Grout J , Cass A , Woodey JC , Enzenhofer H , Clarke WC , Houtman R et al (2000 ) The influence of extreme water temperatures on migrating Fraser River sockeye salmon (Oncorhynchus nerka) during the 1998 spawning season . Can Tech Rep Fish Aquat Sci 2326 : 1 –117 . Macdonald JS , Patterson DA , Hague MJ , Guthrie IC (2010 ) Modeling the influence of environmental factors on spawning migration mortality for sockeye salmon fisheries management in the Fraser River, British Columbia . Trans Am Fish Soc 139 : 768 –782 . MacNutt MJ , Hinch SG , Lee CG , Phibbs JR , Lotto AG , Healey MC , Farrell AP (2006 ) Temperature effects on swimming performance, energetics, and aerobic capacities of mature adult pink salmon (Oncorhynchus gorbuscha) compared with those of sockeye salmon (Oncorhynchus nerka) . Can J Zool 84 : 88 –97 . Madliger CL , Cooke SJ , Crespi EJ , Funk JL , Hultine KR , Hunt KE , Rohr JR , Sinclair BJ , Suski CD , Willis CKR et al (2015 ) Success stories and emerging themes in conservation physiology . Conserv Physiol 4 : cov057 ; doi:10.1093/conphys/cov057 .27382466 Martins EG , Hinch SG , Patterson DA , Hague MJ , Cooke SJ , Miller KM , Robichaud D , English KK , Farrell AP (2012 ) High river temperature reduces survival of sockeye salmon (Oncorhynchus nerka) approaching spawning grounds and exacerbates female mortality . Can J Fish Aquat Sci 69 : 330 –342 . Mathes MT , Hinch SG , Cooke SJ , Crossin GT , Patterson DA , Lotto AG , Farrell AP (2010 ) Effect of water temperature, timing, physiological condition, and lake thermal refugia on migrating adult Weaver Creek sockeye salmon (Oncorhynchus nerka) . Can J Fish Aquat Sci 67 : 70 –84 . Mergel I , Lazer D , Binz-Scharf M (2008 ) Lending a helping hand: voluntary engagement in knowledge sharing . Int J Learn Change 3 : 5 –22 . Miller KM , Schulze AD , Ginther N , Li S , Patterson DA , Farrell AP , Hinch SG (2009 ). Salmon spawning migration: metabolic shifts and environmental triggers . Comp Biochem Physiol Part D Genomics Proteomics 4 : 75 –89 .20403740 Miller KM , Li S , Kaukinen KH , Ginther N , Hammill E , Curtis JMR , Patterson DA , Sierocinski T , Donnison L , Pavlidis P et al (2011 ) Genomic signatures predict migration and spawning failure in wild Canadian salmon . Science 331 : 214 –217 .21233388 Miller KM , Teffer A , Tucker S , Li S , Schulze AD , Trudel M , Juanes F , Tabata A , Kaukinen KH , Ginther NG et al (2014 ) Infectious disease, shifting climates, and opportunistic predators: cumulative factors potentially impacting wild salmon declines . Evol Appl 7 : 812 –855 .25469162 Nguyen VM , Rudd MA , Hinch SG , Cooke SJ (2012 ) Differences in information use and preferences among recreational salmon anglers: implications for management initiatives to promote responsible fishing . Hum Dimensions Wildl 17 : 248 –256 . Nguyen VM , Martins EG , Raby GD , Donaldson MR , Lotto AG , Patterson DA , Robichaud D , English KK , Farrell AP , Willmore WG et al (2014 ) Disentangling the roles of air exposure, gillnet injury, and facilitated recovery on the post-capture and release mortality and behavior of adult migratory sockeye salmon (Oncorhynchus nerka) in freshwater . Physiol Biochem Zool 87 : 125 –135 .24457927 Pacific Salmon Treaty (1999 ) http://www.psc.org/pubs/Treaty/Treaty.pdf Patterson DA , Hague MJ (2007 ) Evaluation of long range summer forecasts of lower Fraser River discharge and temperature conditions . Can Tech Rep Fish Aquat Sci 2754 : 1 –34 . Patterson DA , Macdonald JS , Hinch SG , Healey MC , Farrell AP (2004 ) The effect of exercise and captivity on energy partitioning, reproductive maturation and fertilization success in adult sockeye salmon . J Fish Biol 64 : 1039 –1059 . Patterson DA , Macdonald JS , Skibo KM , Barnes DP , Guthrie I, Hills J (2007 a) Reconstructing the summer thermal history for the lower Fraser River, 1941 to 2006, and implications for adult sockeye salmon (Oncorhynchus nerka) spawning migration . Can Tech Rep Fish Aquat Sci 2724 : 1 –43 . Patterson DA , Skibo KM , Barnes D , Hills JA , Macdonald JS (2007 b) The influence of water temperature on time to surface for adult sockeye salmon carcasses and the limitations in estimating salmon carcasses in the Fraser River, British Columbia . N Am J Fish Manage 27 : 878 –884 . Provencal J (2011 ) Extending the reach of research as a public good: moving beyond the paradox of “zero-sum language games.” . Public Underst Sci 20 : 101 –116 . Pullin AS , Knight TM , Stone DA , Charman K (2004 ) Do conservation managers use scientific evidence to support their decision-making . Biol Conserv 119 : 245 –252 . Raby GD , Donaldson MR , Hinch SG , Patterson DA , Lotto AG , Robichaud D , English KK , Willmore WG , Farrell AP , Davis MW et al (2012 ) Validation of reflex indicators for measuring vitality and predicting the delayed mortality of wild coho salmon bycatch released from fishing gears . J Appl Ecol 49 : 90 –98 . Raby GD , Cooke SJ , Cook KV , McConnachie SH , Donaldson MR , Hinch SG , Whitney CK , Drenner SM , Patterson DA , Clark TD et al (2013 ) Resilience of pink salmon and chum salmon to simulated capture stress incurred upon arrival at spawning grounds . Trans Am Fish Soc 142 : 524 –539 . Raby GD , Clark TD , Farrell AP , Patterson DA , Bett NN , Wilson SM , Willmore WG , Suski CD , Hinch SG , Cooke SJ (2015 a) Facing the river gauntlet: understanding the effects of fisheries capture and water temperature on the physiology of coho salmon . PLoS ONE 10 : e0124023 .25901952 Raby GD , Hinch SG , Patterson DA , Hills JA , Thompson LA , Cooke SJ (2015 b) Mechanisms to explain purse seine bycatch mortality of coho salmon . Ecol Appl 25 : 1757 –1775 .26591444 Raby GD , Wilson SM , Patterson DA , Hinch SG , Clark TD , Farrell AP , Cooke SJ (2015 c) A physiological comparison of three techniques for reviving sockeye salmon exposed to a severe capture stressor during upriver migration . Conserv Physiol 3 : cov015 ; doi:10.1093/conphys/cov015 .27293700 Rand PS , Hinch SG (1998 ) Swim speeds and energy use of river migrating adult sockeye salmon: simulating metabolic power and assessing risk of energy depletion . Can J Fish Aquat Sci 55 : 1832 –1841 . Rand PS , Hinch SG , Morrison J , Foreman MG , MacNutt MJ , Macdonald JS , Healey MC , Farrell AP , Higgs DA (2006 ) Effects of river discharge, temperature, and future climates on energetics and mortality of adult migrating Fraser River sockeye salmon . Trans Am Fish Soc 135 : 655 –667 . Regan HM , Colyvan M , Burgman MA (2002 ) A taxonomy and treatment of uncertainty for ecology and conservation biology . Ecol Appl 12 : 618 –628 . Rice JC (2011 ) Advocacy science and fisheries decision-making . ICES J Mar Sci 68 : 2007 –2012 . Ricker WE (1947 ) Hell's Gate and the sockeye . J Wildl Manage 11 : 10 –20 . Robinson KA , Hinch SG , Gale MK , Clark TD , Wilson SM , Donaldson MR , Farrell AP , Cooke SJ , Patterson DA (2013 ) Effects of post-capture ventilation assistance and elevated water temperature on sockeye salmon in a simulated capture-and-release experiment . Conserv Physiol 1 : cot015 ; doi:10.1093/conphys/cot015 .27293599 Robinson KA , Hinch SG , Raby GD , Donaldson MR , Robichaud D , Patterson DA , Cooke SJ (2015 ) Influence of post-capture ventilation assistance on migration success of adult sockeye salmon following capture and release . Trans Am Fish Soc 144 : 693 –704 . Roos JF (1991 ) Restoring Fraser River Salmon: a History of the International Pacific Salmon Fisheries Commission, 1937–1985 . Pacific Salmon Commission , Vancouver, BC, Canada , pp 1 –438 . St Hilaire S , Boichuck M , Barnes D , Higgins M , Devlin R , Withler R , Khattra J , Jones S , Kieser D (2002 ) Epizootiology of Parvicapsula minibicornis in Fraser River sockeye salmon, Oncorhynchus nerka (Walbaum) . J Fish Dis 25 : 107 –120 . Servizi JA , Jensen JO (1977 ) Resistance of adult sockeye salmon to acute thermal shock. International Pacific Salmon Fisheries Commission Progress Report 34, pp 1–11. Servizi JA , Martens DW (1987 ) Some effects of suspended Fraser River sediments on sockeye salmon (Oncorhynchus nerka) . In Smith HD , Margolis L , Wood CC , eds, Sockeye Salmon (Oncorhynchus nerka) Population Biology and Future Management. Canadian Special Publication of Fisheries and Aquatic Sciences 96, Fisheries and Oceans Canada, Ottawa, ON, Canada, pp 254–264. Sutherland WJ , Burgman M (2015 ) Policy advice: use experts wisely . Nature 526 : 317 –318 .26469026 Talbot GB , Jackson RI (1950 ) A biological study of the effectiveness of the Hell's Gate fishways. II. Variations in flow patterns at Hell's Gate and their relationships to the migration of sockeye salmon. International Pacific Salmon Fisheries Commission Bulletin 3 , International Pacific Salmon Fisheries Commission, New Westminster, BC, Canada , pp 1 –97 . Thompson WF (1945 ) Effect of the obstruction at Hell's Gate on the sockeye salmon of the Fraser River. International Pacific Salmon Fisheries Commission Bulletin 1 , International Pacific Salmon Fisheries Commission , New Westminster, BC, Canada , pp 1 –75 . Wagner GN , Hinch SG , Kuchel LJ , Lotto AG , Jones SRM , Patterson DA , Macdonald JS , Van Der Kraak G , Shrimpton JM , English KK et al (2005 ) Metabolic rates and swimming performance of adult Fraser River sockeye salmon (Oncorhynchus nerka) after a controlled infection with Parvicapsula minibicornis . Can J Fish Aquat Sci 62 : 2124 –2133 . Wikelski M , Cooke SJ (2006 ) Conservation physiology . Trends Ecol Evol 21 : 38 –46 .16701468 Williams IV (1973 ) Investigation of the prespawning mortality of sockeye in Horsefly River and McKinley Creek in 1969 . Int Pac Salmon Fish Comm Prog Rep 27 : 1 –42 . Williams B (2005 ) 2004 Southern salmon fishery post-season review, part 1. Fraser River sockeye report. http://waves-vagues.dfo-mpo.gc.ca/Library/314601.pdf Wood JW (1965 ) A report on fish disease as a possible cause of pre-spawning mortalities of Fraser River sockeye salmon Int Pac Salmon Fish Comm Bull 24 : 478 Young JL , Bornik ZB , Marcotte ML , Charlie KN , Wagner GN , Hinch SG , Cooke SJ (2006 ) Integrating physiology and life history to improve fisheries management and conservation . Fish Fish 7 : 262 –283 . Young N , Gingras I , Nguyen VM , Cooke SJ , Hinch SG (2013 ) Mobilizing new science into management practice: the challenge of biotelemetry for fisheries management, a case study of Canada's Fraser River . J Int Wildl Law Policy 16 : 328 –348 .

PMC005xxxxxx/PMC5001151.txt

==== Front Conserv PhysiolConserv PhysiolconphysConservation Physiology2051-1434Oxford University Press 2792850710.1093/conphys/cow025cow025Research ArticleLife history linked to immune investment in developing amphibians Woodhams Douglas C 1Bell Sara C 2Bigler Laurent 3Caprioli Richard M 4Chaurand Pierre 5Lam Brianna A 6Reinert Laura K 7Stalder Urs 3Vazquez Victoria M 8Schliep Klaus 1Hertz Andreas 1Rollins-Smith Louise A 7910Cooke Steven Editor1 Department of Biology, University of Massachusetts Boston, 100 Morrissey Blvd., Boston, MA 02125, USA2 College of Marine and Environmental Sciences, James Cook University, Townsville, QLD 4811, Australia3 Department of Chemistry, University of Zürich, Winterthurerstrasse 190, CH-8057 Zürich, Switzerland4 Mass Spectrometry Research Center and Department of Biochemistry, Vanderbilt University, Nashville, TN 37232-8575, USA5 Department of Chemistry, Université de Montréal, Montreal, QC, Canada H3T 1J46 Department of Biology, James Madison University, MSC 7801, Harrisonburg, VA 22807, USA7 Department of Pathology, Microbiology and Immunology, Vanderbilt University School of Medicine, Nashville, TN 37232-2363, USA8 Plant Biology Department, University of Georgia, Athens, GA 30602, USA9 Department of Biological Science, Vanderbilt University, Nashville, TN 37235-1634, USA10 Department of Pediatrics, Vanderbilt University School of Medicine, Nashville, TN 37232-2363, USACorresponding author: Department of Biology, University of Massachusetts Boston, 100 Morrissey Blvd., Boston, MA 02125, USA. Tel: +1 617 287 6679. Email: dwoodhams@gmail.com2016 26 8 2016 26 8 2016 4 1 cow02518 12 2015 09 5 2016 14 5 2016 © The Author 2016. Published by Oxford University Press and the Society for Experimental Biology.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Amphibian population declines are often associated with disease outbreaks at metamorphosis. We examine the ontogeny of innate immunity in a range of amphibians and find that skin peptide defenses may partially compensate for adaptive immune suppression during this sensitive stage. Skin defense trade-offs with development reflect amphibian life-history strategy. Abstract The broad diversity of amphibian developmental strategies has been shaped, in part, by pathogen pressure, yet trade-offs between the rate of larval development and immune investment remain poorly understood. The expression of antimicrobial peptides (AMPs) in skin secretions is a crucial defense against emerging amphibian pathogens and can also indirectly affect host defense by influencing the composition of skin microbiota. We examined the constitutive or induced expression of AMPs in 17 species at multiple life-history stages. We found that AMP defenses in tadpoles of species with short larval periods (fast pace of life) were reduced in comparison with species that overwinter as tadpoles and grow to a large size. A complete set of defensive peptides emerged soon after metamorphosis. These findings support the hypothesis that species with a slow pace of life invest energy in AMP production to resist potential pathogens encountered during the long larval period, whereas species with a fast pace of life trade this investment in defense for more rapid growth and development. Antimicrobial peptidesdisease ecologyinnate immunitylife-history strategyMALDI-TOF mass spectrometrytadpoles ==== Body Introduction Recent ecological theory suggests that trade-offs exist between host defenses and pace of life (Martin et al., 2007; Previtali et al., 2012; Sandmeier and Tracy, 2014; Sears et al., 2015). In animals with complex life histories, such as amphibians, slow pace of life often refers to growth to a large size over a relatively long larval period, including overwintering as larvae in some species (Stoks et al., 2006; Johnson et al., 2012). Amphibians with a fast pace of life are more likely to use behavioural defenses against parasitic trematodes than slow pace-of-life species, and they appear to invest less in costly immune defenses that would provide infection tolerance (Sears et al., 2015). Here, we examine an innate immune defense of amphibians, antimicrobial skin peptides, in relation to amphibian life history and pace of life. The cutaneous granular glands of many, but not all, amphibian species produce diverse gene-encoded antimicrobial peptides (AMPs) that are specific to species and populations (Apponyi et al., 2004; Woodhams et al., 2006a, b; 2010; Tennessen et al., 2009; Holden et al., 2015). Although peptide structure is clearly linked to amphibian phylogeny (Conlon et al., 2009b), peptide expression during ontogeny has not been carefully studied and may reflect ecological trade-offs between pace of life and immune investment (Todd, 2007). We hypothesized that slow pace-of-life species would produce and secrete effective AMPs during the larval period, whereas fast pace-of-life species would not. Little is known about the ontogeny of amphibian skin peptide defenses. By northern blot analysis and in situ hybridization, mRNA for the two most abundant antimicrobial peptides (magainin and PGLa) was first detected at the beginning of metamorphic climax in Xenopus laevis tadpoles. When whole animals were homogenized, mature peptides could be isolated (Reilly et al., 1994). In a similar study of Lithobates catesbeianus tadpoles, expression of mRNA for ranalexin was not detected by northern blot analysis in premetamorphic tadpoles (forelimbs had not emerged), but was detected in metamorphosing tadpoles and adults (Clark et al., 1994). In both studies, production of AMPs was localized to developing cutaneous granular glands (Clark et al., 1994; Reilly et al., 1994). Katzenback et al. (2014) found increasing mRNA expression of brevinin-1SY through tadpole development in Lithobates sylvaticus. A study of Rana ornataventris adds to the evidence that late stage (metamorphosing) tadpoles of some species, but not early stage tadpoles, produce AMPs that are also expressed by adults. Temporin-1O was produced at the end of pre-metamorphosis and increased in late stage tadpoles and adults of this species (Iwamuro et al., 2006). Wabnitz et al. (1998) observed small amounts of host defense peptides in protein extracts from larval Litoria splendida as early as 14 days after egg deposition, but the complete set of adult skin peptides was not detected until metamorphosis. An ultrastructural study of Phyllomedusa bicolor indicated that both mucous and granular glands were present in tadpoles, but the gland duct did not appear to develop until metamorphosis (Lacombe et al., 2000). Using mass spectrometry (MS), known antimicrobial peptides were not consistently expressed in Lithobates sphenocephalus until ~12 weeks after metamorphosis (Holden et al., 2015). Although all of these studies provide some information about when in ontogeny AMPs can be expressed, they do not address the question of whether living tadpoles secrete defensive peptides onto the skin to function in microbial defense. In this study, we used a combination of approaches to broadly assess the influence of host ontogeny and of interspecific variation in host life history on the expression of defensive skin peptides. Using matrix-assisted laser desorption time-of-flight (MALDI-TOF) MS, we tested for expression of AMPs in anuran species at developmental stages ranging from larvae to mature adults. In order to examine the influence of life-history trade-offs among species, we analysed AMP data from 17 anuran species and found that pace of life appears to influence immune investment. Our studies show that long-lived tadpoles of several species in two of six amphibian families sampled secrete a subset of the same active defensive peptides that are secreted by adults, but the full set of adult-type peptides does not emerge until metamorphosis. These studies also suggest that these innate immune defenses are an important aspect in the evolution of amphibian life-history strategy, and a trade-off exists between pace of life and investment in skin peptide defenses. Materials and methods Experimental amphibians Seventeen species of anuran amphibians were examined in these studies. Sample sizes, life stages sampled and body size of all individuals are presented in Tables 1 and 2. Although sampling occurred at different field and laboratory sites, the sampling and collection methods were consistent across species. Table 1: Skin peptide collections across anuran life-history stages Species Life-history stage (Gosner, 1960 stage) n Mass [g (mean ± SD)] Method of skin peptide inductiona Quantity of peptides [μg/g body mass (mean ± SEM)] Alytes obstetricans Tadpole (25) 25b 0.30 ± 0.09 Bath 1297.56 ± 571.93 Metamoprh 5 1.87 ± 0.18 Injection 2168.3 ± 410.4 Adult 8 7.08 ± 0.88 Injection 574.87 ± 216.02 Lithobates catesbeianus Tadpole (25–35) 7 21.99 ± 6.05 Bath 16.0 ± 3.3 Tadpole (36–39) 9 38.50 ± 1.84 Bath 22.2 ± 10.8 Tadpole (40–41) 9 44.18 ± 2.03 Bath 47.0 ± 15.7 Tadpole (42) 2 18.60 ± 3.08 Bath 230.3 ± 101.0 Metamorph 12 15.07 ± 2.76 Injection 66.4 ± 11.1 Subadult 5 45.97 ± 19.51 Injection 256.5 ± 72.9 Adult 10 100.1 ± 7.22 Injection 186.9 ± 61.4 Adult 6 14.95 ± 2.09 Injection 459.8 ± 184.3 Lithobates pipiens Tadpole (25) 9 2.44 ± 0.46 Bath 174.6 ± 8.5 Metamorph 11 1.84 ± 0.62 Bath 247.7 ± 76.9 Metamorph 3 2.44 ± 0.76 Injection 248.5 ± 105.9 Subadult 6 8.53 ± 2.42 Bath 64.5 ± 11.1 Adult 13 30.38 ± 4.07 Injection 437.45 ± 68.1 Adult 9 27.83 ± 8.54 Bath 42.7 ± 8.2 Litoria serrata Tadpole (25) 107c 1.33 ± 0.71 Bath 105.0 ± 26.2 Metamorph 8 0.25 ± 0.12 Bath 449.8 ± 131.9 Adult 24 4.88 ± 1.18 Bath 31.2 ± 4.5 Adult 20 4.06 ± 0.22 Injection 579.8 ± 50.5 Rana sierrae Tadpole (25) 25c 1.32 ± 0.23 Bath 171.9 ± 16.8 Metamorph 10 2.73 ± 0.44 Injection 1753.8 ± 596.0 Adult 30 34.59 ± 14.38 Injection 636.1 ± 80.1 Litoria nannotis Tadpole (25) 5 2.45 ± 0.43 Bath 240.3 ± 101.6 Mixophyes shevilli Tadpole (25) 10 1.53 ± 0.41 Bath 77.3 ± 62.6 Hyla cinerea Adult 10 6.91 ± 0.57 Injection 196.6 ± 42.1d Secretions were induced by norepinephrine in the bath (tadpoles) or by subcutaneous injection (post-metamorphosis). Note that the peptide quantity is comparable across samples collected by the same induction method. aBath indicates immersion for 15 min in 100 µM norepinephrine bitartrate; injection indicates subcutaneous administration of 10 nmol/g body mass norepinephrine bitartrate. bTwenty-five tadpoles; five groups of five tadpoles each. cOne hundred and seven tadpoles; 16 groups of 3–11 tadpoles each. dNo peptides were detected by mass spectrometry. Table 2: Skin peptides detected in larval and post-metamorphic amphibians of 17 species Species Number of tadpoles examined (Gosner, 1960 stage) Number of post-metamorphs examined Proportion of tadpole:adult AMPs detected Alytes obstetricans 25 (25) 13 0.88 Bufo bufo 30 (25) 15 0 Bufotes viridis 20 (25) 0 – Hyla meridionalis 15 (37–38) 0 – Hyla versicolor 4 (25) 0 – Lithobates catesbeianus 27 (25–42) 33 0.73 Lithobates pipiens 9 (25–41) 42 0.14 Litoria ewingii 10 (Anstis, 2002; stage 41) 12 0 Litoria serrata 107 (25) 52 0 Pelobates fuscus 11 (25–40) 9 0 Pelodytes punctatus 12 (37–41) 3 0 Pseudacris regilla 4 (25) 0a 0 Rana arvalis 12 (25) 12 0 Rana iberica 11 (25–38) 4 0 Rana sierrae 25 (25) 40 0.25 Rana temporaria 60 (25) 13 0 Rhinella marina 0 3 0 Abbreviation: AMPs, antimicrobial peptides. Proportions are based on data presented in Table 3. aPseudacris regilla AMPs were detected by RNA analysis (Robertson & Cornman, 2014). Litoria ewingii data are from Schadich et al. (2010). Tadpoles of northern leopard frogs, Lithobates pipiens, were obtained from a commercial supplier (NASCO, Fort Atkinson, WI, USA) and were maintained in groups of 10 in 16 litre tanks at Vanderbilt University. These and other species were sampled as tadpoles before limb development and as metamorphs immediately after tail resorption was complete. In July 2004, skin peptides from subadults were sampled in the field in Van Buren County, MI, USA. In September 2004, skin peptides from six adult L. pipiens were sampled from the same field location. Nine additional adult L. pipiens were obtained from Connecticut Valley Biological Supply Co., Southhampton, MA, USA, housed in 16 litre plastic containers and cared for as described below. Skin peptides from these frogs are described by Woodhams et al. (2006b) and Rollins-Smith et al. (2006). Tadpoles of American bullfrogs, Lithobates catesbeianus, were sampled in the field in Boulder, CO, USA in June 2013. An egg mass was collected from Davidson County, TN, USA in June 2004, and tadpoles were reared in the laboratory at Vanderbilt University. Tadpoles and adults were also obtained from Charles D. Sullivan Co., Inc. Newly metamorphosed L. catesbeianus were obtained from a commercial supplier (Rana Ranch, Twin Falls, ID, USA). After 1 year, the subadults were sampled in the laboratory. Adult frogs were sampled in the laboratory in May 2007 at the University of Georgia. At Sixty Lake Basin in the Sierra Nevada Mountains of CA, USA, skin peptides from adult mountain yellow-legged frogs, Rana sierrae, were collected in September 2004 (Rollins-Smith et al., 2006). At the same time and location, R. sierrae tadpoles were sampled for skin peptides. Metamorphs were sampled in the laboratory at James Madison University in January 2008 after being raised from an egg clutch salvaged from a drying pool in Sixty Lake Basin in the summer of 2007. Green-eyed treefrogs, Litoria serrata (formerly L. genimaculata), were sampled for skin peptides at Birthday Creek, near Paluma, Queensland, Australia in September 2005 (adults) and December 2006 (tadpoles). Data were also included from Woodhams (2003), including skin peptide samples from metamorphs raised in the laboratory at James Cook University from tadpoles collected at Birthday Creek, Queensland, Australia in February 2003. Common midwife toads, Alytes obstetricans, were raised at the University of Zurich through metamorphosis from egg clutches collected in Germany. Tadpoles and metamorphs were sampled for skin peptides in August 2008. Additional adults were sampled from Canton Basel, Switzerland in June 2009. Antimicrobial peptides were described by Conlon et al. (2009a). The five species described above were sampled as tadpoles, as recent metamorphs and as juveniles or adults for comparisons across life history, examination of constitutive skin peptide expression and quantification of induced skin peptide secretions. An additional 12 species were examined to compare more broadly between larval and adult stages (Table 2). The following species were raised in Switzerland from field-collected egg clutches: Bufo bufo and Rana temporaria (Woodhams et al., 2014), Bufotes viridis, Hyla meridionalis, Pelobates fuscus, Pelodytes punctatus, Rana arvalis and Rana iberica. Rhinella marina were sampled in Panama, Pseudacris regilla were sampled in California, and Hyla versicolor were sampled from Minnesota. In addition to these field-sampled amphibians, data are included from Litoria ewingii sampled in New Zealand (Schadich et al., 2010). An additional three species were sampled from only one life stage and include Litoria nannotis and Mixophyes shevilli tadpoles sampled in the field in Queensland, Australia, and Hyla cinerea adults obtained from PETCO Animal Supplies, Inc. and sampled at the University of Colorado, Boulder, CO, USA. Animal care Tadpoles were reared in dechlorinated tap water (changed twice weekly) and were fed boiled romaine lettuce. After forelimb emergence at stage 42 (Gosner, 1960), metamorphs were moved to 16 litre polystyrene containers set at an incline, with a small volume of water at one end so that the frogs could choose wet or dry conditions. All containers were sterilized with bleach and dried before use. Each newly metamorphosed juvenile was fed two or three vitamin-dusted crickets three times weekly. Skin peptide sampling Table 1 lists the methods of skin peptide induction and sampling for MS analyses for each species and life-history stage. Granular gland secretions were induced from post-metamorphic amphibians by administration of norepinephrine (bitartrate salt; Sigma, St Louis, MO, USA) by immersion in an aqueous solution or by subcutaneous injection (Rollins-Smith et al., 2002; Woodhams et al., 2006b). Tadpoles were exposed to an aqueous solution of norepinephrine, and when sampled, all were at developmental stages between stages 25 and 41 (Gosner, 1960). Metamorphs were sampled immediately after the tail was completely resorbed. After administration of norepinephrine, skin secretions were collected, partly purified and quantified as previously described (Rollins-Smith et al., 2006). Dry weight was measured for all A. obstetricans peptides. The total quantity of peptides recovered per gram body weight was determined for each sample. In addition to total skin secretion comparisons across life stages, the mixture of peptides locally present on different skin surfaces of adult frogs was sampled directly for MS analysis by applying 80 µm carbon-imbedded conductive polyethylene film (hereafter ‘peptide-absorbent film’; Goodfellow Cambridge Ltd, Cambridge, UK) as previously described (Woodhams et al., 2007). After norepinephrine induction of peptides in adult L. pipiens (n = 10), film was applied locally to the dorsal surface and dorsolateral ridge, ventral thigh and pelvic patch, ventral foot webbing and ventral gular skin for comparison of the relative intensity of peptide signals. Film was also applied locally to the dorsal and ventral surfaces of R. sierrae (n = 24) to detect constitutive expression of AMPs, and relative intensities were compared by Student's paired t-tests. Constitutive AMP expression from adults of several species has been described previously (Pask et al. 2012; Woodhams et al. 2012). Tadpoles of all species were blotted with peptide-absorbent film across the body and tail before norepinephrine induction to determine whether peptide expression was constitutive or required induction. Thus, constitutively expressed (ambient) peptides were sampled from all tadpoles within seconds of capture, before stimulation. Tadpoles were then sampled as described above for induced peptide expression. Analysis of skin peptides by mass spectrometry Skin peptide mixtures were analysed by MALDI-TOF MS following direct sampling with peptide-absorbent film (no purification or concentration steps) or after norepinephrine-induced peptide secretion, as previously described (Woodhams et al., 2007). Norepinephrine-induced peptide mixtures eluted from C-18 Sep-Paks were spotted onto the sample plate at 1 mg/ml before adding an equal volume of matrix. An Applied Biosystems Voyager DE-STR spectrometer was operated in reflector, delayed extraction and positive ion mode. For external calibration, a series of peptide standards (Sigma-Aldrich) was applied. Mass spectra were acquired across the range of m/z (mass to charge ratio) 600–10 000 and analysed after baseline subtraction and de-noising (smoothing) with Data Explorer v4.4 (Applied Biosystems). All samples from A. obstetricans were analysed by MALDI-TOF using an Autoflex I time-of-flight mass spectrometer (Bruker Daltonics GmbH, Bremen, Germany) equipped with a 337 nm nitrogen laser. A 20 µl sample solution was diluted with 20 µl of 0.1% trifluoroacetic acid, vortexed, and 1 µl was spotted onto a ‘Prespotted AnchorChip’ target prepared with α-cyano-4-hydroxycinnamic acid as matrix (CHCA; Bruker). Instrument calibration was obtained using signals from the HCCA matrix at m/z 379.09 and a mixture of standard peptides composed of bradykinin 1–7 (m/z 757.40), angiotensin II (m/z 1046.54), angiotensin I (m/z 1296.69), renin substrate (m/z 1758.93), adrenocorticotrophic hormone clip 18–39 (m/z 2465.20) and somatostatin 28 (m/z 3147.47), all obtained from the peptide calibration standard mix II (Bruker). Comparisons of peptide profiles were made among life-history stages of each species. Methods and MS–MS sequence confirmation of Rana sierrae tadpole AMP temporin-1M are provided in the Supplementary data as confirmation that the tadpole peptide has an identical structure to that found in adults. This peptide is also known to be antimicrobial and can inhibit the emerging pathogenic fungus Batrachochytrium dendrobatidis (Bd) at concentrations down to 6.25 μM (Rollins-Smith et al., 2006). Pace-of-life and immune function trade-offs The ability to express defensive peptides at the tadpole stage was determined and this trait mapped onto a phylogenetic tree. Many of the peptides detected on tadpoles (e.g. Supplementary Fig. S1) were previously shown to inhibit bacteria and Bd (Rollins-Smith et al., 2006). We determined whether this trait was exclusive to the amphibian lineage, or appears among several lineages in association with pace of life. For the 17 anuran species examined here and in previous studies, we determined whether maximal tadpole size differed between species capable of expressing peptides as tadpoles, and between species that typically overwinter throughout their range by Student's unpaired t-test. We used Fisher's exact test to determine whether the proportion of species capable of expressing peptides as tadpoles differed between fast and slow pace of life. We corrected for phylogeny by using a phylogenetic independent contrast with the pic function in the ape package of R statistical software, in a similar manner to Johnson et al. (2012). We performed a linear regression and phylogenetically independent contrast analysis using the proportion of AMPs expressed as tadpoles as the dependent variable. The independent variable was a categorical variable indicating whether an amphibian species typically overwinters. To determine whether occurrence of an overwintering tadpole stage can predict the likelihood of AMP production at the tadpole stage, we used a binary logistic regression. Some genera, including Bufo, Bufotes and Rhinella (family Bufonidae), did not have detectable AMPs, and a limited number of AMPs from Hyla have been described (Table 2; Erspamer et al., 1986; Conlon, 2011). In a preliminary study, we did not detect skin peptides from adult Hyla cinerea by MS (Table 1), although secretions had antifungal activity (Woodhams DC, Rollins-Smith LA, Voyles J, and Carey C, unpublished data). Results Antimicrobial peptides detected by direct matrix-assisted laser desorption time-of-flight mass spectrometry The expression of AMPs on the skin of live amphibians was detected by direct MALDI-TOF MS using samples from peptide-absorbent film blotted onto skin (Woodhams et al., 2007). Antimicrobial peptide signals were detected on all skin surfaces of adult L. pipiens induced with norepinephrine. The strongest relative intensity signals originated from films applied to the dorsal surface (Fig. 1). No significant differences were detected in the relative intensities of AMPs from dorsal or ventral surfaces of R. sierrae adults (temporin-1M, brevinin-1M, ranatuerin-2Ma and ranatuerin-2Mb; Student's paired t-tests, P-values > 0.05). The skin of all R. sierrae adults tested was heavily infected with the fungus Bd determined by qPCR (Briggs CJ & Vredenburg VT, unpublished data). In tadpoles, constitutive expression of peptides without norepinephrine induction was detected only in L. catesbeianus. Specifically, we observed mass signals for ranatuerin-2, -7, -8, -9 and ranalexin in this species. Figure 1: Peptide profiles from four skin surfaces of adult Lithobates pipiens upon induction of granular gland secretions. Representative spectra of 10 replicates are shown. Induced peptides examined by matrix-assisted laser desorption time-of-flight mass spectrometry After peptide induction, mass signals indicative of previously described AMPs (Table 3) were detected in skin secretions from metamorphs and adults of all five species tested and from tadpoles of A. obstetricans, L. catesbeianus, R. sierrae and L. pipiens (Fig. 2). Nearly identical profiles were observed between adults and metamorphs of each species. A subset of the post-metamorphic peptides was detected in secretions from the tadpoles as shown in Fig. 2, and the specific peptides detected in each species are listed in Table 3 (adult peptide profiles are shown in comparison with those of tadpoles in Fig. 2). In L. catesbeianus tadpoles, ranatuerin-2, -4, -6, -7, -8 and -9, ranalexin and palustrin-2CBa were found. Bradykinin (non-antimicrobial) and temporin-1M or temporin-1P were found in the secretions of tadpoles of R. sierrae and L. pipiens. Of the eight skin peptides described for A. obstetricans (Conlon et al., 2009a), tadpoles expressed all except alyteserin-2c. The sequence structure of temporin-1M recovered from R. sierrae tadpoles was confirmed by MS–MS (Supplementary material, Fig. S1). Some peptides listed in Table 3 were not detected in the present study and might have been missing in the sampled individuals (Tennessen et al., 2009) or might not be detectable by MALDI-MS owing to poor ionization. In general, species with rapid larval development (B. bufo, B. viridis, H. meridionalis, H. versicolor, L. pipiens, L. ewingii, L. serrata, P. regilla, R. arvalis, R. iberica, R. temporaria and R. marina) showed few or no AMP signals typical of adults, suggestive of little investment in skin peptide defenses before metamorphosis (Table 2). Figure 2: Representative skin peptide profiles of tadpoles and adults of five anuran species. Adult and metamorph spectra matched closely, and only adult profiles are displayed for clarity. Table 3: Skin peptides detected from amphibian life-history stages by matrix-assisted laser desorption time-of-flight mass spectrometry (MALDI-TOF MS), with citations provided for peptide descriptions Species Skin peptide Sequence Mono-isotopic mass (m/z) Signal detected by MALDI-TOF MS Reference Tadpole Metamorph Adult Alytes obstetricans Alytesin pEGRLGTQWAVGHLM-NH2 1535.8 X X X Erspamer et al. (1972) Alyteserin-2a ILGKLLSTAAGLLSNL.NH2 1582.1 X X X Conlon et al. (2009a) Alyteserin-2c ILGAILPLVSGLLSSKL.NH2 1605.0 X X Conlon et al. (2009a) Alyteserin-2b ILGAILPLVSGLLSNKL.NH2 1632.1 X X X Conlon et al. (2009a) Alyteserin-1c GLKEIFKAGLGSLVKGIAAHVAS.NH2 2263.5 X X X Conlon et al. (2009a) Alyteserin-1a GLKDIFKAGLGSLVKGIAAHVAN.NH2 2277.3 X X X Conlon et al. (2009a) Alyteserin-1b GLKEIFKAGLGSLVKGIAAHVAN.NH2 2291.4 X X X Conlon et al. (2009a) Alyteserin-1d GLKDIFKAGLGSLVKNIAAHVAN.NH2 2334.5 X X X Conlon et al. (2009a) Lithobates catesbeianus Temporin-CBa (ranatuerin-5) FLPIASLLGKYL.NH2 1333.8 X X Goraya et al. (1998); Hasunuma et al. (2010); Mechkarska et al. (2011) Temporin-CBf FLPIASMLGKYL.NH2 1351.8 Mechkarska et al. (2011) Temporin-CBb (ranatuerin-6) FISAIASMLGKFL.NH2 1396.8 X X X Goraya et al. (1998); Rollins-Smith et al. (2002); Mechkarska et al. (2011) Ranatuerin-7 FLSAIASMLGKFL 1396.8 X X X Goraya et al. (1998) Temporin-CBd (ranatuerin-8) FISAIASFLGKFL.NH2 1412.8 X X X Goraya et al. (1998); Mechkarska et al. (2011) Chensirin-2CBa IIPLPLGYFAKKP 1455.9 X X Hasunuma et al. (2010) Ranatuerin-9 FLFPLITSFLSKVL 1624.0 X X X Goraya et al. (1998) Brevinin-1CBa (ranalexin) FLGGLIKIVPAMICAVTKKC 2104.2 X X X Clark et al. (1994); Vignal et al. (1998); Rollins-Smith et al. (2002) Ranatuerin-1CBa (ranatuerin-1) SMLSVLKNLGKVGLGFVACKINKQC 2649.5 Goraya et al. (1998); Rollins-Smith et al. (2002); Mechkarska et al. (2011) Brevinin-1CBb (ranatuerin-4) FLPFIARLAAKVFPSIICSVTKKC 2651.5 X X X Goraya et al. (1998); Mechkarska et al. (2011) Ranatuerin-1CBb SMFSVLKNLGKVGLGFVACKVNKQC 2669.4 X X Mechkarska et al. (2011) Ranatuerin-2CBa (ranatuerin-2) GLFLDTLKGAAKDVAGKLEGLKCKITGCKLP 3186.8 X X X Goraya et al. (1998); Mechkarska et al. (2011) Palustrin-2CBa GFLDIIKDTGKEFAVKILNNLKCKLAGGCPP 3301.8 X X X Mechkarska et al. (2011) Ranatuerin-2CBc (ranatuerin-3) GFLDIINKLGKTFAGHMLDKIKCTIGTCPPSP 3414.8 X X Goraya et al. (1998); Mechkarska et al. (2011) Ranatuerin-2CBd GFLDIIKNLGKTFAGHMLDKIRCTIGTCPPSP 3442.8 Mechkarska et al. (2011) Lithobates pipiens Bradykinin RPPGFSPFR 1059.6 X X X Sin et al. (2008) Ranatensin-C TPQWATGHFM 1174.5 Erspamer et al. (1986) Ranatensin-C ZTPQWATGHFM 1303.2 Erspamer et al. (1984) Ranatensin QVPQWAVGHFM 1298.6 Nakajima et al. (1970) Temporin-1P FLPIVGKLLSGLL 1368.9 X X X Goraya et al. (2000); Rollins-Smith et al. (2002) Peptide leucine arginine (pLR) LVRGCWTKSYPPKPCFVR 2136.1 Salmon et al. (2001) Brevinin-1Pa FLPIIAGVAAKVFPKIFCAISKKC 2563.5 X X Goraya et al. (2000) Brevinin-1Pd FLPIIASVAANVFSKIFCAISKKC 2569.4 X X Goraya et al. (2000) Brevinin-1Pb FLPIIAGIAAKVFPKIFCAISKKC 2577.5 X X Goraya et al. (2000) Brevinin-1Pc FLPIIASVAAKVFSKIFCAISKKC 2583.5 X X Goraya et al. (2000) Brevinin-1Pe FLPIIASVAAKVFPKIFCAISKKC 2593.5 X X Goraya et al. (2000) Ranatuerin-2P GLMDTVKNVAKNLAGHMLDKLKCKITGC 3000.6 X X Goraya et al. (2000); Rollins-Smith et al. (2002); Chen et al. (2003) Ranatuerin-2Pa GFLSTVVKLATNVAGTVIDTIKCKVTGGCRK 3178.8 Chen et al. (2003); Vanhoye et al. (2003) Esculentin-2P GFSSIFRGVAKFASKGLGKDLARLGVNLVACKISKQC 3868.1 Goraya et al. (2000); Rollins-Smith et al. (2002) Litoria serrata Caerulein QQDYTGWMDF 1290.5 Rozek et al. (1998) Maculatin-2.1 GFVDFLKKVAGTIANVVT 1878.1 X X Rozek et al. (1998) Maculatin-1.1.1 FGVLAKVAAHVVPAIAEHF 1975.1 X X Rozek et al. (1998) Maculatin-1.1 GLFGVLAKVAAHVVPAIAEHF 2145.2 X X Rozek et al. (1998) Maculatin-1.2 GLFGVLAKVASHVVPAIAEHFQA 2360.3 Rozek et al. (1998) Maculatin 3.1 GLLQTIKEKLESLESLAKGIVSGIQA 2723.6 X X Rozek et al. (1998) Rozek et al. (1998) Rana sierrae Bradykinin RPPGFSPFR 1060.6 X X X Rollins-Smith et al. (2006) Temporin-1M FLPIVGKLLSGLL.NH2 1368.9 X X X Rollins-Smith et al. (2006) Temporin-1M (free acid) FLPIVGKLLSGLL 1369.9 X X Rollins-Smith et al. (2006) Ranatuerin-2Mb GIMDSVKGVAKNLAAKLLEKLKCKITGC 2929.6 X X Rollins-Smith et al. (2006) Ranatuerin-2Ma GLLSSFKGVAKGVAKDLAGKLLEKLKCKITGC 3273.9 X X Rollins-Smith et al. (2006) Pace-of-life and immune function trade-offs Of the 17 species examined in this analysis, four grow to a large size, develop slowly and usually overwinter as tadpoles (Fig. 3A). One species that can overwinter, P. fuscus, did not produce detectable AMPs in the tadpole stage, but peptides were detected by the same method in metamorphs. The other three overwintering species did produce detectable AMPs as tadpoles. Tadpoles that usually overwinter throughout their range reach a larger maximal size than those of typically non-overwintering species (Fig. 3A; Student's unpaired t-test, t15 = 6.143, P < 0.001). Likewise, species with tadpoles capable of expressing skin AMPs reach a significantly larger size as tadpoles than species that did not express skin AMPs as tadpoles (Fig. 3A; Student's unpaired t-test, t15 = 3.021, P = 0.009). A significantly higher proportion of slow pace-of-life species (three of four) were found to express peptides as tadpoles compared with fast pace-of-life species (one of 10; Table 3; Fisher's exact test, P = 0.041). The proportion of the AMP repertoire expressed as tadpoles in comparison with adults is on average 0.011 in non-overwintering species and 0.464 in overwintering species (Mann–Whitney U-test, P = 0.032). This characteristic is not exclusive to a single amphibian family (Fig. 3B), and a phylogenetic independent contrast using the pic function in the ape package of R statistical software showed that the trait of peptide expression at the tadpole stage differs by pace of life even after phylogenetic correction (F1,15 = 10.69, P = 0.00517). Thus, we propose that an overwintering tadpole stage predicts the likelihood that a species will produce AMPs as a tadpole (binary logistic regression, P = 0.021, odds ratio = 36). Figure 3: Species sampled and tadpole characteristics. (A) Maximal tadpole length as reported in species descriptions (AmphibiaWeb, 2016). Species with large tadpoles that tend to overwinter and are categorized here as ‘slow pace-of-life’ species (dark blue). (B) Neighbour-joining tree of taxa tested in this study. The analysis involved 29 nucleotide sequences of the 16S rRNA gene. The final data set consists of 439 aligned nucleotide positions. Distances were computed using the p-distance method and are in the units of the number of base differences per site. Analyses were conducted in MEGA7 (Kumar et al., 2016). Species with tadpoles expressing AMPs (green circles) are not exclusively found within a single family. Discussion Evolution of amphibian developmental strategies An amazing diversity of reproductive and developmental strategies exists among amphibian species that have implications for immune function and disease resistance (Todd, 2007; Gomez-Mestre et al., 2012). For example, some species have the capacity to overwinter in an aquatic tadpole stage, typically growing slowly to a large body size and requiring more permanent bodies of water. This may lead to exposure to a broad range of pathogenic organisms and greater resource allocation into immune defenses (e.g. Johnson et al., 2012). Other species, such as L. serrata and L. pipiens, can develop quickly in ephemeral water bodies or slow-moving creeks and may therefore devote fewer resources to immune function during the tadpole stage. Our data provide further support for the hypothesis of a trade-off between rapid growth and investment in immune defenses. We compare both among species that differ in life-history strategy and between developmental stages within a subset of those species. Relative to adult frogs, tadpoles either completely lacked or only expressed a reduced set of defensive skin peptides (Fig. 2). This was particularly evident in rapidly developing species that appear to invest little in tadpole AMP defense (Table 3). We did not detect AMPs in tadpoles of 12 species with short larval periods or in one species, P. fuscus, capable of overwintering in the larval stage. In contrast, tadpoles with a long larval period (A. obstetricans, L. catesbeianus and R. sierrae) expressed a subset of adult AMPs with known capacities for inhibiting amphibian pathogens, including Bd (Rollins-Smith and Conlon, 2005). Among 17 species, those with long-lived larval stages were most likely to produce AMPs (Table 2), suggesting that slower pace-of-life species invest more resources in AMP skin defense during the larval stage than fast pace-of-life species. As we did not test for overall immune function, it is possible that fast pace-of-life species have comparatively weak immune defenses or rely on alternative defenses during larval development, such as microbiota, that may be less energetically expensive (Kueneman et al., 2014). Although our sampling here of a few populations from either field or laboratory settings represents a broad preliminary survey, we expect that further studies may refine these results by testing members of additional families or across a variety of environmental conditions and populations among species. Some authors suggest that diseases and parasites (pathogenic fungi in particular; Green, 1999) are overlooked when explaining the diversity of amphibian developmental strategies (Todd, 2007). This may apply to trade-off strategies within species that have resulted in the evolution of flexible metamorphic timing. The timing of metamorphosis varies widely depending on larval conditions, including aquatic habitat (ephemeral or permanent), predation, density, competition, nutrition, pollutants and other factors (Smith-Gill and Berven, 1979; Werner, 1986; Alford and Harris, 1988). Developmental strategies may optimize growth in the least risky environment (Pechenik, 2006; Scott et al., 2007). In general, our data support the broad hypothesis that there may be a trade-off between larval growth and development of immune function, specifically AMP defenses. Studies by Groner et al. (2014) provide some support for this hypothesis, showing adjustments in skin peptide defense investment depending on stressors experienced early in ontogeny. Localized skin peptide expression We found antimicrobial peptides constitutively present on both dorsal and ventral surfaces of R. sierrae, and upon induction, AMPs were detected on all skin surfaces of adult L. pipiens (Fig. 1). A previous study showed that the amount of peptides recovered from resting and active (chased) L. pipiens was sufficient to inhibit Bd (Pask et al., 2012). Given that the strongest signals detected by direct MALDI-TOF on L. pipiens were associated with the dorsal surface, other skin surfaces may be slightly more prone to infection. The legs and feet of some amphibians have been shown to be most prone to infection by Bd (North and Alford, 2008), and this may also be influenced by the ontogeny of keratinized skin, developing first in the feet and hindlimbs at metamorphosis, or because these surfaces come into contact with the contaminated substrate more often than other skin surfaces (Marantelli et al., 2004; Weldon and Du Preez, 2006). There are several processes by which AMPs can be excreted. Granular glands of amphibian skin may not be able to discharge their AMP products fully onto the skin surface by a holocrine process until after development of the neuromuscular secretory apparatus and gland ducts in the epidermis (Delfino, 1980; Faszewski and Kaltenbach, 1995; Delfino et al., 1998, 2006; Lacombe et al., 2000). However, mature gland products may also be secreted by a merocrine process (either constitutive or induced exocytosis) and flow between epidermal cell layers and through epidermal interstices onto the skin surface before complete gland duct development (Delfino et al., 1998; Terreni et al., 2003; Quagliata et al., 2006). Some studies used transmission electron microscopy to examine discharge of granular glands (Delfino et al., 2006). Other studies used northern blotting, in situ hybridization or RT-PCR to detect mRNAs encoding AMPs (Clark et al., 1994; Reilly et al., 1994; Iwamuro et al., 2006; Katzenback et al., 2014). As none of these studies demonstrated secretion of active peptides from living tadpoles, we chose to use direct MALDI-TOF MS to examine AMP expression on tadpoles (Chaurand et al., 1999; Woodhams et al., 2007). By this method, we were able to detect constitutive expression of multiple AMP signals from the skin of L. catesbeianus tadpoles. Other components of amphibian skin defense The mucous layer covering amphibian skin is an ideal niche for many opportunistic pathogens because it contains mucopolysaccharides that are a potential nutrient source. Hence, defense of the skin is crucial for protection from many amphibian pathogens and involves both adaptive and innate immune defenses, including mucosal antibodies, epithelial barriers, phagocytic cells, AMPs, fatty acids, protease inhibitors and other factors (Carey et al., 1999; Rollins-Smith et al., 2009; Ramsey et al. 2010). In addition, symbiotic microbes can contribute to skin defense and may also produce antimicrobial metabolites (Brucker et al., 2008; Harris et al., 2009). Some of these components of amphibian immune defense may be interacting synergistically or antagonistically with AMPs during metamorphosis (Myers et al., 2012; Woodhams et al., 2014). Changes in skin defenses during ontogeny, as shown here, may partly explain corresponding shifts in skin microbiota (Kueneman et al., 2014, 2016; Krynak et al., 2015). Disease susceptibility at metamorphosis Several studies suggest that chytridiomycosis and other diseases can be most severe as amphibians undergo metamorphosis (Green et al., 2002; Bosch and Martínez-Solano, 2006; Carey et al., 2006; Garcia et al., 2006; Kriger and Hero, 2006; Rachowicz et al., 2006; Langhammer et al., 2014) or before they mature into adults (Abu Bakar et al., 2016). Experimentally reducing skin peptides in juvenile X. laevis caused increased susceptibility to Bd (Ramsey et al., 2010) and resulted in lethal chytridiomycosis in new metamorphs of L. pipiens (Pask et al., 2012). The mouthparts of larval amphibians can be infected by Bd, but tadpoles are largely immune to chytridiomycosis (Berger et al., 1998; Rachowicz and Vredenburg, 2004). The fungal pathogen Bd is thought to use keratin in tadpole mouthparts and adult amphibian skin. Upon metamorphosis, the fungus can spread from the infected mouthparts to the keratinized skin and lead to rapid mortality (Marantelli et al., 2004; Rachowicz and Vredenburg, 2004). Pathogenicity factors, including proteases that may degrade AMPs, are increased by exposure of Bd to thyroid hormone, which peaks during metamorphosis (Thekkiniath et al., 2013, 2015). Antimicrobial skin peptides could theoretically play a vital role in protecting some amphibians during this sensitive stage of development. However, if the new metamorphs experienced stress, such as limited food conditions, prior to metamorphosis, they might have fewer stored AMPs in granular glands available for defense. This might explain why newly metamorphosed L. sphenocephala raised in outdoor mesocosms were slow to develop an adult pattern of AMPs (Holden et al., 2015). Population level variation, not measured in this study, may also lead to variation in disease susceptibility at metamorphosis (Tobler and Schmidt, 2010; Bradley et al., 2015). The composition of peptides differed among life-history stages and species (Tables 2 and 3). Such variation in skin peptide defense provides a potential mechanism for differential colonization by microbiota among species and life stages (McKenzie et al., 2012; Kueneman et al., 2014, 2016; Krynak et al., 2015) and susceptibility to infection from a variety of pathogens leading to disease or malformation. A holistic measure of the mucosome, or the combined host- and microbiota-derived compounds in the mucus, can test function against pathogens and predict disease susceptibility (Woodhams et al., 2014). Many of the diseases impacting amphibian populations are transmitted by pathogens or parasites in the aquatic environment that interact with the skin mucosome, including ranavirus, Aeromonas hydrophila, Bd, Saprolegnia ferax, Anurofeca richardsi, Amphibiocystidium spp. and Ribeiroia ondatrae. Although the Gram-negative bacterium A. hydrophila is not inhibited by amphibian skin peptides tested to date (Rollins-Smith et al., 2002; Schadich and Cole, 2009; Tennessen et al., 2009), both ranaviruses and Bd can be inhibited (Rollins-Smith et al., 2002; Chinchar et al., 2004). The ability of tadpole peptides to inhibit other bacterial pathogens, protozoa or fungal infections, such as S. ferax (Romansic et al., 2006), remains to be tested. The contribution of larval skin defenses to differences among species in infection by the malformation-inducing trematode R. ondatrae (Johnson and Hartson, 2009) is also unknown. Lithobates catesbeianus is significantly more resistant to R. ondatrae infection than species such as P. regilla and B. americanus, which lack AMP defenses as tadpoles (Table 2; Johnson et al., 2013; Calhoun et al., 2016). Skin peptides tested here do not appear to explain resistance of H. versicolor and other Hyla species to R. ondatrae (Johnson and Hartson, 2009; LaFonte and Johnson, 2013). Ecologically, metamorphosis is a particularly vulnerable time for amphibians. Locomotion ability at metamorphosis lags below that of both tadpoles and adults, leaving metamorphs at greater risk of predation (Wassersug and Sperry, 1977). The type of predation, competition and environmental conditions, in addition to infection status, influence size at metamorphosis (Alford, 1999; Parris and Cornelius, 2004; Vonesh and Warkentin, 2006; Groner et al., 2014), and size at metamorphosis may influence subsequent survival (Pechenik, 2006; Scott et al., 2007). The timing of metamorphosis and stress during early ontogeny may have a significant influence on disease risk (Rollins-Smith, 1998; Groner et al., 2014). Pace of life appears to trade off with AMP immune investment in amphibian larvae. The repertoire of defensive peptides expressed on the skin changes with amphibian development such that at most a subset of the adult peptides occurs in tadpoles. Immunologically, the adaptive immune system reorganizes during metamorphosis, and innate skin peptide defenses may compensate for adaptive immune suppression, allow for restructuring of the microbiota upon metamorphosis and alter colonization resistance of parasites and pathogens. Supplementary material Supplementary material is available at Conservation Physiology online. Supplementary Material Supplementary Data Click here for additional data file. Acknowledgements The authors thank Cherie Briggs, Tate Tunstall and Vance Vredenburg for supplying adult R. sierrae, Jennifer Pryweller for collection of skin peptides from L. pipiens tadpoles in a preliminary study, and Jamie Voyles and Cynthia Carey for a preliminary study on Hyla cinerea AMPs. Josh Bakke, Simone Baumgartner, Jos Kielgast, Mary Stice, Leyla Davis and Ursina Tobler assisted with sampling for direct MALDI mass spectrometry analyses. Thanks to Josh Van Buskirk for access to tadpoles from Switzerland. Thanks to Bethany Hoye, Bryan LaFonte, Valerie McKenzie, Franklin Roman-Rodriguez and Holly Archer for tadpole sampling. Thanks to Reid Harris for facilitating R. sierrae experiments. Animal care and experimental procedures were approved by institutional animal care and use committees at Vanderbilt University, the University of Colorado and the University of Georgia. Scientific collection permits were provided by Michigan Department of Natural Resources, Tennessee Wildlife Resources Agency, Tennessee Department of Environment and Conservation, United States Department of the Interior National Park Service, California Department of Fish and Wildlife, Minnesota Department of Natural Resources, Queensland Parks and Wildlife Service, Autoridad Nacional del Ambiente, Panama and the Cantonal Veterinary Office of Zurich and Kanton Basel-Landschaft. Field work conformed to the field hygiene protocols proposed by Murray et al. (2011). Funding This work was supported by the National Heart, Lung, and Blood Institute's Immunology of Blood and Vascular Systems Training Grant 5T32 HL069765-05 to Jacek Hawiger, the Claraz Foundation, the Swiss National Science Foundation (31-125099 to D.C.W.), the United States National Science Foundation grants IOS-0520847, IOS-0619536 and IOS-1121758 to L.A.R.-S. and an Integrated Research Challenges in Environmental Biology grant IBN-9977063 to James P. Collins. ==== Refs References Abu Bakar A , Bower DS , Stockwell MP , Clulow S , Clulow J , Mahoney MJ (2016 ) Susceptibility to disease varies with ontogeny and immunocompetence in a threatened amphibian . Oecologia , Early View: 1 –13 DOI:10.1007/s00442-016-3607-4 Alford RA (1999 ) Ecology: resource use, competition, and predation In McDiarmid RW , Altig R , eds, Tadpoles: the Biology of Anuran Larvae . The University of Chicago Press , Chicago , pp 240 –278 . Alford RA , Harris RN (1988 ) Effects of larval growth history on anuran metamorphosis . Am Nat 131 : 91 –106 . Apponyi MA , Pukala TL , Brinkworth CS , Maselli VM , Bowie JH , Tyler MJ , Booker GW , Wallace JC , Carver JA , Separovic F et al (2004 ) Host-defence peptides of Australian anurans: structure, mechanism of action and evolutionary significance . Peptides 25 : 1035 –1054 .15203252 AmphibiaWeb: information on amphibian biology and conservation. [Web application ] (2016 ) AmphibiaWeb, Berkeley, CA, USA. http://amphibiaweb.org/ Anstis M (2002 ) Tadpoles of South-eastern Australia: A Guide with Keys . Sydney : Reed New Holland . Berger L , Speare R , Daszak P , Green DE , Cunningham AA , Goggin CL , Slocombe R , Ragan MA , Hyatt AD , McDonald KR et al (1998 ) Chytridiomycosis causes amphibian mortality associated with population declines in the rain forests of Australia and Central America . Proc Natl Acad Sci USA 95 : 9031 –9036 .9671799 Bosch J , Martínez I (2006 ) Chytrid fungus infection related to unusual mortalities of Salamandra salamandra and Bufo bufo in the Penalara Natural Park, Spain . Oryx 40 : 84 –89 . Bradley PW , Gervasi SS , Hua J , Cothran RD , Relyea RA , Olson DH , Blaustein AR (2015 ) Differences in sensitivity to the fungal pathogen Batrachochytrium dendrobatidis among amphibian populations . Conserv Bio 29 : 1347 –1356 .26219571 Brucker RM , Harris RN , Schwantes CR , Gallaher TN , Flaherty DC , Lam BA , Minbiole KPC (2008 ) Amphibian chemical defense: antifungal metabolites of the microsymbiont Janthinobacterium lividum on the salamander Plethodon cinereus . J Chem Ecol 34 : 1422 –1429 .18949519 Calhoun DM , Woodhams D , Howard C , LaFonte BE , Gregory JR , Johnson PT (2016 ) Role of antimicrobial peptides in amphibian defense against trematode infection . EcoHealth in press. DOI:10.1007/s10393-016-1102-3 . Carey C , Bruzgul JE , Livo LJ , Walling ML , Kuehl KA , Dixon BF , Pessier AP , Alford RA , Rogers KB (2006 ) Experimental exposures of boreal toads (Bufo boreas) to a pathogenic chytrid fungus (Batrachochytrium dendrobatidis) . EcoHealth 3 : 5 –21 . Carey C , Cohen N , Rollins L (1999 ) Amphibian declines: an immunological perspective . Dev Comp Immunol 23 : 459 –472 .10512457 Chaurand P , Stoeckli M , Caprioli RM (1999 ) Direct profiling of proteins in biological tissue sections by MALDI mass spectrometry . Anal Chem 71 : 5263 –5270 .10596208 Chen T , Farragher S , Bjourson A , Orr D , Rao P , Shaw C (2003 ) Granular gland transcriptomes in stimulated amphibian skin secretions . Biochem J 371 : 125 –30 .12413397 Chinchar VG , Bryan L , Silphadaung U , Noga E , Wade D , Rollins-Smith L (2004 ) Inactivation of viruses infecting ectothermic animals by amphibian and piscine antimicrobial peptides . Virology 323 : 268 –275 .15193922 Clark DP , Durell S , Maloy WL , Zasloff M (1994 ) Ranalexin: a novel antimicrobial peptide from bullfrog (Rana catesbeiana) skin, structurally related to the bacterial antibiotic, polymyxin . J Biol Chem 269 : 10849 –10855 .8144672 Conlon JM (2011 ) The contribution of skin antimicrobial peptides to the system of innate immunity in anurans . Cell Tissue Res 343 : 201 –212 .20640445 Conlon JM , Demandt A , Nielsen PF , Leprince J , Vaudry H , Woodhams DC (2009 a) The alyteserins: two families of antimicrobial peptides from the skin secretions of the midwife toad Alytes obstetricans (Alytidae) . Peptides 30 : 1069 –1073 .19463738 Conlon JM , Kolodziejek J , Nowotny N (2009 b) Antimicrobial peptides from the skins of North American frogs . Biochim Biophys Acta 1788 : 1556 –1563 .18983817 Delfino G (1980 ) L'attivita rigeneratrice del tratto intercalare helle ghiandole granulose cutanee dell ‘Ululone Bombina variegata pachypus (Bonaparte) (Anfibio, Anuro, Discoglosside); studio sperimentale al microscopio elettronico . Archo Ital Anat Embriol 85 : 283 –310 . Delfino G , Brizzi R , Alvarez BB , Kracke R (1998 ) Serous cutaneous glands in Phyllomedusa hypochondrialis (Anura, Hylidae): secretory patterns during ontogenesis . Tissue Cell 30 : 30 –40 .9569677 Delfino G , Drewes RC , Magherini S , Malentacchi C , Nosi D , Terreni A (2006 ) Serous cutaneous glands of the Pacific tree-frog Hyla regilla (Anura, Hylidae): patterns of secretory release induced by nor-epinephrine . Tissue Cell 38 : 65 –77 .16423375 Erspamer V , Erspamer GF , Inselvini M , Negri L (1972 ) Occurrence of bombesin and alytesin in extracts of the skin of three European discoglossid frogs and pharmacological actions of bombesin on extravascular smooth muscle . Br J Pharmacol 45 : 333 –348 .5048651 Erspamer V , Falconieri Erspamer G , Mazzanti G , Endean R (1984 ) Active peptides in the skins of one hundred amphibian species from Australia and Papua New Guinea . Comp Biochem Physiol 77C : 99 –108 . Erspamer V , Falconieri Erspamer G , Cei JM (1986 ) Active peptides in the skins of two hundred and thirty American amphibian species . Comp Biochem Physiol 85C : 125 –137 . Faszewski EE , Kaltenbach JC (1995 ) Histology and lectin-binding patterns in the skin of the terrestrial horned frog Ceratophrys ornata . Cell Tissue Res 281 : 169 –77 .7621521 Garcia TS , Romansic JM , Blaustein AR (2006 ) Survival of three species of anuran metamorphs exposed to UV-B radiation and the pathogenic fungus Batrachochytrium dendrobatidis . Dis Aquat Organ 72 : 163 –169 .17140139 Gomez-Mestre I , Pyron RA , Wiens JJ (2012 ) Phylogenetic analyses reveal unexpected patterns in the evolution of reproductive modes in frogs . Evolution 66 : 3687 –3700 .23206128 Goraya J , Knoop FC , Conlon JM (1998 ) Ranatuerins: antimicrobial peptides isolated from the skin of the American bullfrog, Rana catesbeiana . Biochem Biophys Res Commun 250 : 589 –592 .9784389 Goraya J , Wang Y , Li Z , O'Flaherty M , Knoop FC , Platz JE , Conlon JM (2000 ) Peptides with antimicrobial activity from four different families isolated from the skins of the North American frogs Rana luteiventris, Rana berlandieri and Rana pipiens . Eur J Biochem 267 : 894 –900 .10651828 Gosner KL (1960 ) A simplified table for staging anuran embryos and larvae with notes on identification . Herpetologica 16 : 183 –190 . Green AJ (1999 ) Implications of pathogenic fungi for life-history evolution in amphibians . Funct Ecol 13 : 573 –575 . Green DE , Converse KA , Schrader AK (2002 ) Epizootiology of sixty-four amphibian morbidity and mortality events in the USA, 1996–2001 . Ann NY Acad Sci 969 : 323 –339 .12381613 Groner ML , Rollins-Smith LA , Reinert LK , Hempel J , Bier ME , Relyea RA (2014 ) Interactive effects of competition and predator cues on immune responses of leopard frogs at metamorphosis . J Exp Biol 217 : 351 –358 .24115058 Harris RN , Brucker RM , Walke JB , Becker MH , Schwantes CR , Flaherty DC , Lam BA , Woodhams DC , Briggs CJ , Vredenburg VT et al (2009 ) Skin microbes on frogs prevent morbidity and mortality caused by a lethal skin fungus . ISME J 3 : 818 –824 .19322245 Hasunuma I , Iwamuro S , Kobayashi T , Shirama K , Conlon JM , Kikuyama S (2010 ) Expression of genes encoding antimicrobial peptides in the Harderian gland of the bullfrog Lithobates catesbeianus . Comp Biochem Physiol C Toxicol Pharmacol 152 : 301 –305 .20510387 Holden WM , Reinert LK , Hanlon SM , Parris MJ , Rollins-Smith LA (2015 ) Development of antimicrobial peptide defenses of southern leopard frogs, Rana sphenocephala, against the pathogenic chytrid fungus, Batrachochytrium dendrobatidis . Dev Comp Immunol 48 : 65 –75 .25218643 Iwamuro S , Nakamura M , Ohnuma A , Conlon JM (2006 ) Molecular cloning and sequence analyses of preprotemporin mRNAs containing premature stop codons from extradermal tissues of Rana tagoi . Peptides 27 : 2124 –2128 .16675060 Johnson PTJ , Hartson RB (2009 ) All hosts are not equal: explaining differential patterns of malformations in an amphibian community . J Anim Ecol 78 : 191 –201 .18671806 Johnson PTJ , Rohr JR , Hoverman JT , Kellermanns E , Bowerman J , Lunde KB (2012 ) Living fast and dying of infection: host life history drives interspecific variation in infection and disease risk . Ecol Lett 15 : 235 –242 .22221837 Johnson PTJ , Preston DL , Hoverman JT , Richgels KL (2013 ) Biodiversity decreases disease through predictable changes in host community competence . Nature 494 : 230 –233 .23407539 Katzenback BA , Holden HA , Falardeau J , Childers C , Hadj-Moussa H , Avis TJ , Storey KB (2014 ) Regulation of the Rana sylvatica brevinin-1SY antimicrobial peptide during development and in dorsal and ventral skin in response to freezing, anoxia and dehydration . J Exp Biol 217 : 1392 –1401 .24436376 Kriger KM , Hero JM (2006 ) Survivorship in wild frogs infected with chytridiomycosis . EcoHealth 3 : 171 –177 . Krynak KL , Burke DJ , Benard MF (2015 ) Larval environment alters amphibian immune defenses differentially across life stages and populations . PLoS ONE 10 : e0130383 .26107644 Kueneman JG , Parfrey LW , Woodhams DC , Archer HM , Knight R , McKenzie VJ (2014 ) The amphibian skin-associated microbiome across species, space and life history stages . Mol Ecol 23 : 1238 –1250 .24171949 Kueneman JG , Woodhams DC , Van Treuren W , Archer HM , Knight R , McKenzie VJ (2016 ) Inhibitory bacteria reduce fungi on early life stages of endangered Colorado boreal toads (Anaxyrus boreas) . ISME J 10 : 934 –944 .26565725 Kumar S , Stecher G , Tamura K (2016 ) MEGA7: Molecular Evolutionary Genetics Analysis version 7.0 for bigger datasets . Mol Biol Evol 3 : 1870 –1874 .54 . Lacombe C , Cifuentes-Diaz C , Dunia I , Auber-Thomay M , Nicolas P , Amiche M (2000 ) Peptide secretion in the cutaneous glands of South American tree frog Phyllomedusa bicolor: an ultrastructural study . Eur J Cell Biol 79 : 631 –641 .11043404 LaFonte BE , Johnson PT (2013 ) Experimental infection dynamics: using immunosuppression and in vivo parasite tracking to understand host resistance in an amphibian-trematode system . J Exp Biol 216 :3700 –3708 .23821713 Langhammer PF , Burrowes PA , Lips KR , Bryant AB , Collins JP (2014 ) Susceptibility to the amphibian chytrid fungus varies with ontogeny in the direct-developing frog, Eleutherodactylus coqui . J Wildlife Dis 50 : 438 –446 . McKenzie VJ , Bowers RM , Fierer N , Knight R , Lauber CL (2012 ) Co-habiting amphibian species harbor unique skin bacterial communities in wild populations . ISME J 6 : 558 –596 . Marantelli G , Berger L , Speare R , Keegan L (2004 ) Distribution of the amphibian chytrid Batrachochytrium dendrobatidis and keratin during tadpole development. Pacific Conserv Biol 10 : 173 –179 . Martin LB , Weil ZM , Nelson RJ (2007 ) Immune defense and reproductive pace of life in Peromyscus mice . Ecology 88 : 2516 –2528 .18027755 Mechkarska M , Ojo OO , Meetani MA , Coquet L , Jouenne T , Abdel-Wahab YH , Flatt PR , King JD , Conlon JM (2011 ) Peptidomic analysis of skin secretions from the bullfrog Lithobates catesbeianus (Ranidae) identifies multiple peptides with potent insulin-releasing activity . Peptides 32 : 203 –208 .21087647 Murray K , Skerratt L , Marantelli G , Berger L , Hunter D , Mahony M , Hines H (2011 ) Hygiene protocols for the control of diseases in Australian frogs. A report for the Australian Government Department of Sustainability, Environment, Water, Population and Communities . http://www.environment.gov.au/biodiversity/invasive-species/publications/hygiene-protocols-control-diseases-australian-frogs Myers JM , Ramsey JP , Blackman AL , Nichols AE , Minbiole KP , Harris RN (2012 ) Synergistic inhibition of the lethal fungal pathogen Batrachochytrium dendrobatidis: the combined effect of symbiotic bacterial metabolites and antimicrobial peptides of the frog Rana muscosa . J Chem Ecol 38 : 958 –965 .22914957 Nakajima T , Tanimura T , Pissano JJ (1970 ) Isolation and structure of a new vasoactive polypeptide . Fed Proc 29 : 282 . North S , Alford RA (2008 ) Infection intensity and sampling locality affect Batrachochytrium dendrobatidis distribution among body regions on green-eyed tree frogs Litoria genimaculata . Dis Aquat Organ 81 : 177 –188 .18998583 Parris MJ , Cornelius TO (2004 ) Fungal pathogen causes competitive and developmental stress in larval amphibian communities . Ecology 85 : 3385 –3395 . Pask JD , Woodhams DC , Rollins-Smith LA (2012 ) The ebb and flow of antimicrobial skin peptides defends northern leopard frogs (Rana pipiens) against chytridiomycosis . Glob Change Biol 18 : 1231 –1238 . Pechenik JA (2006 ) Larval experience and latent effects—metamorphosis is not a new beginning . Integr Comp Biol 46 : 323 –333 .21672745 Previtali MA , Ostfeld RS , Keesing F , Jolles AE , Hanselmann R , Martin LB (2012 ) Relationship between pace of life and immune responses in wild rodents . Oikos 121 : 1483 –1492 . Quagliata S , Malentacchi C , Delfino C , Brunasso AMG , Delfino G (2006 ) Adaptive evolution of secretory cell lines in vertebrate skin . Caryologia 59 : 187 –206 . Rachowicz LJ , Vredenburg VT (2004 ) Transmission of Batrachochytrium dendrobatidis within and between amphibian life stages . Dis Aquat Organ 61 : 75 –83 .15584413 Rachowicz LJ , Knapp RA , Morgan JA , Stice MJ , Vredenburg VT , Parker JM , Briggs CJ (2006 ) Emerging infectious disease as a proximate cause of amphibian mass mortality . Ecology 87 : 1671 –1683 .16922318 Ramsey JP , Reinert LK , Harper LK , Woodhams DC , Rollins-Smith LA (2010 ) Immune defenses against Batrachochytrium dendrobatidis, a fungus linked to global amphibian declines, in the South African clawed frog, Xenopus laevis . Infect Immun 78 : 3981 –3992 .20584973 Reilly DS , Tomassini N , Zasloff M (1994 ) Expression of magainin antimicrobial peptide genes in the developing granular glands of Xenopus skin and induction by thyroid hormone . Dev Biol 162 : 123 –133 .8125181 Robertson LS , Cornman RS (2014 ) Transcriptome resources for the frogs Lithobates clamitans and Pseudacris regilla, emphasizing antimicrobial peptides and conserved loci for phylogenetics . Mol Ecol Resour 14 : 178 –183 .24028389 Rollins-Smith LA (1998 ) Metamorphosis and the amphibian immune system . Immunol Rev 166 : 221 –230 .9914915 Rollins-Smith LA , Conlon JM (2005 ) Antimicrobial peptide defenses against chytridiomycosis, an emerging infectious disease of amphibian populations . Dev Comp Immunol 29 : 589 –598 .15784290 Rollins-Smith LA , Carey C , Longcore J , Doersam JK , Boutte A , Bruzgal JE , Conlon JM (2002 ) Activity of antimicrobial skin peptides from ranid frogs against Batrachochytrium dendrobatidis, the chytrid fungus associated with global amphibian declines . Devel Comp Immunol 26 : 471 –479 .11906726 Rollins-Smith LA , Woodhams DC , Reinert LK , Vredenburg VT , Briggs CJ , Nielsen PF , Conlon JM (2006 ) Antimicrobial peptide defenses of the mountain yellow-legged frog (Rana muscosa) . Dev Comp Immunol 30 : 831 –842 .16330099 Rollins-Smith LA , Ramsey JP , Reinert LK , Woodhams DC , Livo LJ , Carey C (2009 ) Immune defenses of Xenopus laevis against Batrachochytrium dendrobatidis Front Biosci 1 : 68 –91 . Romansic JM , Diez KA , Higashi EM , Blaustein AR (2006 ) Effects of nitrate and the pathogenic water mold Saprolegnia on survival of amphibian larvae . Dis Aquat Organ 68 : 235 –243 .16610589 Rozek T , Waugh RJ , Steinborner ST , Bowie JH , Tyler MJ , Wallace JC (1998 ) The maculatin peptides from the skin glands of the tree frog Litoria genimaculata: a comparison of the structures and antibacterial activities of maculatin 1.1 and caerin 1.1 . J Pept Sci 4 : 111 –115 .9620615 Salmon AL , Cross LJM , Irvine AE , Lappin TRJ , Dathe M , Krause G , Canning P , Thim L , Beyermann M , Rothemund S et al (2001 ) Peptide leucine arginine, a potent immunomodulatory peptide isolated and structurally characterized from the skin of the northern leopard frog, Rana pipiens . J Biol Chem 276 : 10145 –10152 .11099505 Sandmeier FC , Tracy RC (2014 ) The metabolic pace-of-life model: incorporating ectothermic organisms into the theory of vertebrate ecoimmunology . Integr Comp Biol 54 : 387 –395 .24760792 Schadich E , Cole ALJ (2009 ) Inhibition of frog antimicrobial peptides by extracellular products of the bacterial pathogen Aeromonas hydrophila . Lett Appl Microbiol 49 : 384 –387 .19728406 Schadich E , Cole A , Squire M , Mason D (2010 ) Skin peptides of different life stages of Ewing's tree frog . J Exp Zool A Ecol Genet Physiol 313A : 532 –537 . Scott DE , Casey ED , Donovan MF , Lynch TK (2007 ) Amphibian lipid levels at metamorphosis correlate to post-metamorphic terrestrial survival . Oecologia 153 : 521 –532 .17530291 Sears BF , Snyder PW , Rohr JR (2015 ) Host life history and host–parasite syntopy predict behavioural resistance and tolerance of parasites . J Anim Ecol 84 : 625 –636 .25583069 Sin Y , Zhou M , Chen W , Wang L , Chen T , Walker B , Shaw C (2008 ) Skin bradykinin-related peptides (BRPs) and their biosynthetic precursors (kininogens): comparisons between various taxa of Chinese and North American ranid frogs . Peptides 29 : 393 –403 .18164514 Smith-Gill SJ , Berven KA (1979 ) Predicting amphibian metamorphosis . Am Nat 113 : 563 –585 . Stoks R , De Block M , Slos S , Van Doorslaer W , Rolff J (2006 ) Time constraints mediate predator-induced plasticity in immune function, condition, and life history . Ecology 87 : 809 –815 .16676523 Tennessen JA , Woodhams DC , Chaurand P , Reinert LK , Billheimer D , Shyr Y , Caprioli RM , Blouin MS , Rollins-Smith LA (2009 ) Variations in the expressed antimicrobial peptide repertoire of northern leopard frog (Rana pipiens) populations suggest intraspecies differences in resistance to pathogens . Dev Comp Immunol 33 : 1247 –1257 .19622371 Terreni A , Nosi D , Greven H , Delfino G (2003 ) Development of serous cutaneous glands in Scinax nasica (Anura, Hylidae): patterns of poison biosynthesis and maturation in comparison with larval glands in specimens of other families . Tissue Cell 35 : 274 –287 .12921710 Thekkiniath JC , Zabet-Moghaddam M , Francisco SKS , Francisco MJS (2013 ) A novel subtilisin-like serine protease of Batrachochytrium dendrobatidis is induced by thyroid hormone and degrades antimicrobial peptides . Fungal Biol 117 : 451 –461 .23809655 Thekkiniath J , Zabet-Moghaddam M , Kottapalli KR , Pasham MR , Francisco SS , Francisco MS (2015 ) Quantitative proteomics of an amphibian pathogen, Batrachochytrium dendrobatidis, following exposure to thyroid hormone . PLoS ONE 10 : e0123637 .26046527 Tobler U , Schmidt BR (2010 ) Within- and among-population variation in chytridiomycosis-induced mortality in the toad Alytes obstetricans . PLoS ONE 5 : e10927 .20532196 Todd BD (2007 ) Parasites lost? An overlooked hypothesis for the evolution of alternative reproductive strategies in amphibians . Am Nat 170 : 793 –799 .17926300 Vanhoye D , Bruston F , Nicolas P , Amiche M (2003 ) Antimicrobial peptides from hylid and ranin frogs originated from a 150-million-year-old ancestral precursor with a conserved signal peptide but a hypermutable antimicrobial domain . Eur J Biochem 270 : 2068 –2081 .12709067 Vignal E , Chavanieu A , Roche P , Chiche L , Grassy G , Calas B , Aumelas A (1998 ) Solution structure of the antimicrobial peptide ranalexin and a study of its interaction with perdeuterated dodecylphosphocholine micelles . Eur J Biochem 253 : 221 –228 .9578480 Vonesh JR , Warkentin KM (2006 ) Opposite shifts in size at metamorphosis in response to larval and metamorph predators . Ecology 87 : 556 –562 .16602285 Wabnitz PA , Walters H , Tyler MJ , Wallace JC , Bowie JH (1998 ) First record of host defence peptides in tadpoles. The magnificent tree frog Litoria splendida . J Pept Res 52 : 477 –481 .9924992 Wassersug RJ , Sperry DG (1977 ) The relationships of locomotion to differential predation on Pseudacris triseriata (Anura: Hylidae) . Ecology 58 : 830 –839 . Weldon C , Preez Du LH (2006 ) Quantitative measurement of Batrachochytrium dendrobatidis in amphibian skin . Dis Aquat Organ 72 : 153 –161 .17140138 Werner EE (1986 ) Amphibian metamorphosis: growth rate, predation risk, and the optimal size at transformation . Am Nat 128 : 319 –341 . Woodhams DC (2003 ) The ecology of chytridiomycosis, an emerging infectious disease of Australian rainforest frogs . PhD thesis . James Cook University , Townsville, QLD, Australia . Woodhams DC , Rollins-Smith LA , Carey C , Reinert L , Tyler MJ , Alford RA (2006 a) Population trends associated with skin peptide defenses against chytridiomycosis in Australian frogs . Oecologia 146 : 531 –540 .16205955 Woodhams DC , Voyles J , Lips KR , Carey C , Rollins-Smith LA (2006 b) Predicted disease susceptibility in a Panamanian amphibian assemblage based on skin peptide defenses . J Wildlife Dis 42 : 207 –218 . Woodhams DC , Vredenburg VT , Simon M-A , Billheimer D , Shakhtour B , Shyr Y , Briggs CJ , Rollins-Smith LA , Harris RN (2007 ) Symbiotic bacteria contribute to innate immune defenses of the threatened mountain yellow-legged frog, Rana muscosa . Biol Conserv 138 : 390 –398 . Woodhams DC , Kenyon N , Bell SC , Alford RA , Chen S , Billheimer D , Shyr Y , Rollins-Smith LA (2010 ) Adaptations of skin peptide defences and possible response to the amphibian chytrid fungus in populations of Australian green-eyed treefrogs, Litoria genimaculata . Divers Distrib 16 : 703 –712 . Woodhams DC , Bell SC , Kenyon N , Alford RA , Rollins-Smith LA (2012 ) Immune evasion or avoidance: fungal skin infection linked to reduced defence peptides in Australian green-eyed treefrogs, Litoria serrata . Fungal Biol 116 : 1203 –1211 .23245614 Woodhams DC , Brandt H , Baumgartner S , Kielgast J , Kuepfer E , Tobler U , Davis LR , Schmidt BR , Bel C , Hodel S , Knight R , McKenzie V (2014 ) Interacting symbionts and immunity in the amphibian skin mucosome predict disease risk and probiotic effectiveness . PLoS One 9 (4 ): e96375 .24789229

PMC005xxxxxx/PMC5001152.txt

==== Front 10131695735616Stem Cell ResStem Cell ResStem cell research1873-50611876-77532595794610.1016/j.scr.2015.04.004nihpa809416ArticleChemoattraction of bone marrow-derived stem cells towards human endometrial stromal cells is mediated by estradiol regulated CXCL12 and CXCR4 expression Wang Xiuli Mamillapalli Ramanaiah *Mutlu Levent Du Hongling Taylor Hugh S. Department of Obstetrics, Gynecology and Reproductive Sciences, Yale School of Medicine, New Haven, CT, USA* Corresponding author at: Department of Obstetrics, Gynecology and Reproductive Sciences, Yale School of Medicine, 333 Cedar Street, P.O. Box: 208063, New Haven, CT 06520, USA. Fax: + 1 203 785 4713 ramana.mamillapalli@yale.edu (R. Mamillapalli).11 8 2016 24 4 2015 7 2015 26 8 2016 15 1 14 22 This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).Bone marrow derived cells engraft to the uterine endometrium and contribute to endometriosis. The mechanism by which these cells are mobilized and directed to the endometrium has not been previously characterized. We demonstrate that human endometrial stromal cells (hESCs) produce the chemokine CXCL12 and that bone marrow cells (BMCs) express the CXCL12 receptor, CXCR4. Treatment with physiological levels of estradiol (E2) induced both CXCL12 and CXCR4 expression in hESCs and BMCs, respectively. BMCs migrated towards hESCs conditioned media; a CXCR4 antagonist blocked migration indicating that CXCL12 acting through its receptor, CXCR4, is necessary for chemoattraction of BM cells to human endometrial cells. E2 increased both CXCL12 expression in endometrial cells and CXCR4 expression in BM cells, further enhancing chemoattraction. E2 induced CXCL12/CXCR4 expression in endometrium and BM, respectively, drives migration of stem cells to the endometrium. The E2-CXCL12/CXCR4 signaling pathway may be useful in determining treatments for endometrial disorders, and may be antagonized to block stem cell migration to endometriosis. ==== Body Introduction CXCR4 belongs to the CXC family of chemokine receptors. Interaction of CXCR4 with its ligand, stromal derived factor (SDF-1α, CXCL12) plays a key role in the mobilization and homing of stem cells (Hopman and DiPersio, 2014). CXCR4, expressed on the surface of stem cells, serves as a target for modulating migration (Lai et al., 2014). CXCL12 is produced by the stromal cells and endothelial cells of many organs including bone marrow (BM), endometrium, skeletal muscle, liver and brain (Sharma et al., 2011). In human endometrium, CXCL12 is expressed by stromal cells. Estradiol (E2) stimulates CXCL12 production from endometrial stromal cells (ESCs) (Ruiz et al., 2010; Tsutsumi et al., 2011) suggesting a role in stem cell recruitment to the uterus. BM-derived cells including hematopoietic stem cells (HSCs), mesenchymal stromal cells (MSCs), and endothelial progenitor cells (EPCs), significantly contribute to peripheral tissue repair and angiogenesis (Beauséjour, 2007). Therefore, factors influencing BM-derived cell migration and function are likely to have a broad impact. Overexpression of CXCR4 in stem cells (by cytokine induction or gene transfection) enhances MSCs homing in vivo to bone marrow as well as migration in vitro towards CXCL12 (Shi et al., 2007; Liu et al., 2013a; Marquez-Curtis et al., 2013; Hu et al., 2013). Recently it has been demonstrated that estrogen receptor (ER) is expressed in EPCs in vivo and in vitro (Baruscotti et al., 2010). EPCs proliferation is induced during the menstrual phase and the proliferation can be affected by estrogen through ERα activation (Foresta et al., 2010). These studies suggested the potential regulation of stem cells by sex steroids. Previous studies from our laboratory showed that BM-derived stem cells can engraft in the murine endometrium (Du and Taylor, 2007). We have shown that ischemia–reperfusion injury, toxicant exposure, and medications can alter the migration of BM-derived stem cells to the uterus, however the molecular mechanism responsible for the recruitment and engraftment of these cells is unknown (Zhou et al., 2011; Sakr et al., 2014; Lapidot, 2001). Here we report the effects of female sex hormones estradiol and progesterone on CXCR4 and CXCL12 expression, and the role of this chemokine and its receptor in migration of BMCs towards hESCs. Material and methods Cell culture Mouse bone marrow cells (mBMCs) were prepared from 8–10 weeks old female C57 BL/6 mice (Charles River Laboratories, Wilmington, USA) by flushing bone marrow from the tibia and femur, and filtering the marrow through sterile 70-μm nylon mesh. The filtered mBMCs were grown at a density of 2.5 × 106 cells/ml in DMEM/F-12 medium supplemented with 15% fetal bovine serum, containing penicillin (100 μg/ml) and streptomycin (100 μg/ml) (GIBCO-BRL, Rock-ville, USA). After 48 h the cells were gently washed with PBS and fresh medium added; the medium was subsequently changed for every 3–4 days until two weeks when the cells were used for experiments described below. Mouse uterine cells (mUCs) were prepared from 6–8 weeks old female C57 BL/6 mice by enzymatic digestion of the uterus in 0.125% type IA collagenase (Sigma, USA) for 1 h at 37 °C, and then filtered through a 70-μm filter. Human endometrial stromal cells (hESCs) were obtained from human endometria in the proliferative phase as described by Ryan et al. (1994). Both mUCs and hESCs were cultured in DMEM/F12 medium supplemented with 10% FBS and penicillin/streptomycin (100 μg/ml) for one week. The cells were then washed with PBS, trypsinized, plated and cultured for an additional 48 h before carry out the experiments. Experiments used to obtain the mouse and human cells were conducted under approved Yale Institutional Animal Care and Use Committee and Human Investigations Committee protocols, respectively. ABC-immunocytochemistry (ICC) and fluorescent ICC Cells grown (80% confluent) on glass microscope slides were fixed with freshly prepared 4% formaldehyde for 10 min and rinsed three times for 5 min each with PBS. The cells were blocked with 4% BSA in PBS for 30 min and incubated with the primary antibody (diluted in 1% BSA in PBS) in a humidified chamber overnight at 4 °C. For ABC-ICC, the cells were incubated with the secondary antibody in 1% BSA for 30 min at room temperature. The ABC staining and 3, 3′diaminobenzidine (DAB) kits (Vector Laboratories, USA) were used to visualize the immunocytochemical reaction under light microscope (Olympus BX41). For fluorescence-ICC, the cells were incubated with the secondary antibody in the dark for 30 min at room temperature and 4′, 6-diamidino-2-phenylindole (DAPI) (Vector laboratories, UK) was added on to the cells. The slides were examined under inverted fluorescence microscope (Axiovert 200 M, Zeiss Company, Germany). Flow cytometry After two weeks of culture, mBMCs were analyzed for mesenchymal stromal cells (MSCs), and endothelial progenitor cells (EPCs) by flow cytometry. The cells were incubated with the fluorescent-labeled antibodies against CD90, CD105, CD34, Flk-1 (BioLegend, San Diego, USA) and CD31 (eBiosciences, USA), or with isotype-matched irrelevant antibody (1 μg for 106 cells) for 30 min on ice in dark. The cells were then washed with flow cytometry staining buffer 3 times for 5 min at 3,000 rpm and the cell pellet was resuspended in 1 ml ice cold staining buffer for cell sorting. Flow acquisition was performed on LSRII Fortessa, LSRII, or FACSCalibur analyzers (BD Biosciences), and data were analyzed using Diva software (BD Biosciences, USA). Detection of CXCL12 by ELISA CXCL12α was assayed from the supernatants of cell cultures using ELISA kit (R & D Systems, USA) according to the manufacturer's instructions. mBMC, hESCs and mUC were cultured in DMEM/F12 supplemented with 10% FBS and 1% penicillin and streptomycin in a 6-well plate (1 × 105 cells/well). The supernatants were collected from 48 h old cell cultures. For steroid treatment, the 48 h old mBMC and hESCs cells were serum starved overnight and treated for 24 h with E2 or progesterone (P4) (Sigma, USA) at concentrations of 1 × 10−8, 1 × 10−7, 1 × 10−6 M. The supernatants were then collected. Migration assay The migration assay for mBMC and hESC cells was carried out using 8-μm pore size polycarbonate membrane (Millipore, USA). The serum free conditioned medium (600 μl) collected from 48 h old cultures from both cell types was added into the lower chamber and 200 μl of cells (5 × 104 cells) was placed into the upper insert. The cells in the upper insert were serum starved overnight and treated with either E2 for 24 h at 1 × 10−7 M, or AMD3100 antagonist of CXCR4 for 30 min before the migration assay. After 16 h in a humidified CO2 incubator at 37 °C, the non-migrating cells were scraped with a cotton swab from the top of the membrane. The cells migrating across the membrane were fixed, stained, and counted. Results are reported as chemotactic index (CI), defined as the number of cells migrating in response to the conditioned supernatants divided by number of cells that responded to the serum-free DMEM/F12 medium. Ethanol was used as a vehicle control to exclude the nonspecific effects on CXCR4. Detection of CXCR4 by Western blot Protein extracts (25 to 30 μg) from different cells, as well as treated mBMCs, were subjected to SDS-PAGE and immunoblotting using standard methods. Anti-CXCR4 and anti-α-tubulin antibodies used were from Santa Cruz Biotechnology (Dallas, USA) while secondary antibody conjugated with horseradish peroxidase was obtained from Cell Signaling (Beverly, Massachusetts). The CXCR4 protein band densities were quantified using Quantity One software from BioRad and relative band density was calculated as a ratio of sample to α-tubulin. Quantitative real-time RTPCR RNA was isolated using TRIzol (Invitrogen, Carlsbad, California) and purified on RNeasy minicolumns (QIAGEN, Valencia California), with on-column deoxyribonuclease digestion, as per the manufacturers' instructions. First-strand cDNA was reverse transcribed using iScript cDNA Synthesis Kit while iQ SYBR Green Supermix (Bio-Rad, Hercules, USA) based assays were performed for mCXCR4, mCXCL12 and α-tubulin for qRT-PCR analysis. The CXCR4 primers were as follows: forward 5′-TTTCAGATGCTTGACGTTGG-3′; and reverse 5′-GCGCTCTGCATCAGTGAC-3′; the CXCL12 primers were, forward 5′-ACTCACACTGATCGGTTCCA-3′ and reverse 5′-AGGTGCAGGTAGCAGTGACC-3′ and the primers for α-tubulin were, forward, 5′-ATGGAGGGGAATACAGCCC-3′ and reverse, 5′-TTCTTTGCAGCTCCTTCGTT-3′. For each experimental sample, a control without reverse transcriptase was run to verify that the amplification product arose from cDNA and not from genomic DNA. The relative expression levels normalized to α-tubulin, were determined using the comparative CT method (also known as the 2ΔΔCT method) (Applied Biosystems, 1997; Barr and Manning, 1999). Statistics Results are presented as the mean ± S.D. Statistical significance was determined using one-way ANOVA with the Newman–Keuls multiple comparisons test. All statistical analyses were carried out using Graph Pad Prism 4.00 for Macintosh (Graph-Pad Software for Science Inc., San Diego, CA, USA). Results Characterization of mBMC The CXCR4 and CXCL12 genes are remarkably conserved across diverse species. The human and murine CXCL12 differs by one amino acid and is cross reactive (Lapidot, 2001), providing us with an opportunity to conduct the study of murine CXCR4 with human CXCL12 signaling. The mBMCs were cultured for two weeks, washed with PBS, and trypsinized. The cell pellet was resuspended in FACS staining buffer and incubated with fluorescent labeled antibodies against CD90, CD105, CD34, CD31 and Flk-1. The cells were then and analyzed by FACS. As shown in Fig. 1A, 25.6% of mBMCs expressed CD90 while CD105 (Fig. 1B), CD34 (Fig. 1C), CD31 (Fig. 1D) and Flk-1 (Fig. 1E) were expressed on 20.7%, 67.8%, 60.5% and 68.5% of mBMCs respectively. CD90+ and CD105+ were considered MSC-specific surface markers while CD34+, CD31+, and Flk-1+ represented the EPC. Expression of CXCR4 and CXCL12 in mBMC, mUC and hESCs Cell lysates were prepared from 48 h old cells and 25 μg of protein from each cell type was subjected to SDS-PAGE followed by immunoblotting. As shown in Fig. 2A mBMCs had the highest CXCR4 expression among the three cell types while lowest expression was seen in hESCs. The differential expression of CXCR4 protein levels among these cell types was correlated with mRNA levels confirmed by qRT-PCR. The density of specific bands was quantified using Quantity One software. The relative band density was calculated as a ratio of sample to α-tubulin (data not shown). The CXCL12α was measured from the conditioned medium collected from the 48 h old cells using ELISA kit. As shown in Fig. 2B, CXCL12 was predominantly expressed in mBMCs; however hESCs also expressed CXCL12 at high level while mUCs showed very low level, CXCL12 expression. We localized the expression of CXCR4 in mBMCs with fluorescent ICC. As shown in Fig. 3, the CXCR4 is expressed intracellularly in 37.5% of mBMCs. Fig. 3A shows CD45 expression on mBMCs while 3B shows CXCR4 expression; 3C shows DAPI staining for nuclei and 3D shows the merge of CD45 and CXCR4 expression. CXCR4 expression is predominantly seen in the CD45 negative cells. Migration of mBMC and hESCs towards chemotactic activity of conditioned medium A migration assay was carried out to determine the chemotactic activity of CXCL12, using the conditioned media. We detected the migratory capacity of mBMCs towards hESC supernatant and hESCs towards mBMC supernatant. As shown in Fig. 4, hESC supernatant significantly induced the mBMC migration. Pretreatment of mBMCs with the CXCR4 antagonist AMD3100 blocked the mBMC migration in a dose-dependent manner, and 100 μg/ml AMD3100 completely abolished the mBMC migration. Effects of E2 and P4 on the expression of CXCR4 and CXCL12 in mBMC and hESCs qPCR analysis demonstrated that E2 caused a significant increase in mRNA expression levels of CXCR4 in mBMCs in a dose dependent manner at 6 h but at 24 h only physiological levels of E2 (10−7 M) continued to drive CXCR4 expression in mBMCs (Figs. 5 A & B). As shown in Fig. 5C, progesterone (P4) alone at a physiological concentration of 10−7 M also induced CXCR4 in mBMCs. The combination of E2 and P4 resulted in a similar level of expression as treatment with either sex steroid alone. Ethanol was used as a vehicle to determine the non-specific effects on the CXCR4 expression. Ethanol treated cells does not showed any change in the CXCR4 expression comparing to control cells which are not treated either by ethanol or E2 or P4. Western blot analysis confirmed that CXCR4 protein induction was significantly increased (2.68-fold) after treatment with 10−7 M E2 for 24 h (Fig. 5D) while E2 at a concentration of 10−8 M showed no induction. Conversely, E2 at a concentration of 10−6 M did not increase CXCR4 protein levels compared to untreated cells. In summary, physiological concentrations of E2 and P4 results in increased expression of CXCR4 in mBMCs. Based on the results of steroid-induced CXCR4 expression, physiological levels (10−7 M) of E2 and P4 were selected for examination of the effects of sex steroids on CXCL12 production. As shown in Fig. 5E, neither E2 nor P4 had any effect on CXCL12 production in mBMCs. However, in hESCs, E2 caused a significant increase in CXCL12 production compared to control and surprisingly P4 effectively inhibited E2-induced CXCL12 production in hESCs (Fig. 5F). In the both cell types mBMC and hESCs the protein levels were correlated to the mRNA levels confirmed by qRT-PCR. mBMC migration towards E2 and P4 treated hESC supernatants To determine if the enhanced CXCL12 and CXCR4 production induced by E2 would increase migration of BMCs to hESCs, we treated hESCs and mBMCs with E2 and performed a migration assay using the conditioned media, with and without the CXCR4 antagonist. The hESC supernatants were collected from 48 h old cultures. The mBMC in the upper insert were pretreated with 1 × 10−7 M E2 for 24 h after overnight serum starvation. Migration of mBMC was observed after 16 h. As shown in Fig. 6, mBMC migrated towards the E2-induced hESC supernatant in greater numbers compared to the untreated hESC supernatant. The number of migrated mBMCs decreased 42–51% when mBMCs were pretreated with the CXC4 antagonist AMD3100. The E2 induced migration of BMCs to hESCs was mediated by CXCL12/CXCR4. Discussion Bone marrow derived stem cells migrate to the uterine endometrium of both mice and humans (Du and Taylor, 2007; Taylor, 2004). This migration likely has a key role in the repair of the uterus after damage. Indeed, our group has previously demonstrated that migration and engraftment of BM derived stem cells to the uterine endometrium is increased after ischemic injury and decreased by environmental toxins such as tobacco (Zhou et al., 2011; Wolff et al., 2011). Further, BMC delivery to mice after injury improved reproductive performance (Alawadhi et al., 2014) and mBMCs express several nuclear receptors (Wang et al., 2006; Zhou et al., 2001; Jeong and Mangelsdorf, 2009). Characterization of the chemokines that regulate stem cell engraftment may allow increased engraftment of endogenous stem cells injury. It has been previously shown in other tissues that increased CXCL12 production at a site of injury enhances stem cell recruitment and promotes functional recovery (Liu et al., 2013b; Penn et al., 2012; Sundararaman et al., 2011; Unzek et al., 2007). Here we demonstrate that bone marrow cells will migrate towards endometrial cell conditioned media; this chemoattraction of CXCR4 expressing bone marrow cells is similarly mediated by CXCL12 production by endometrial cells. CXCL12 has been previously identified as an estrogen-regulated gene in estrogen receptor (ER)-positive ovarian and breast cancer cells (Jaerve et al., 2012). Here we show that in the endometrium, E2 significantly increased CXCL12 expression, suggesting a mechanism by which stem cells are recruited to the uterus in reproductive age women; it is likely that this recruitment ceases after menopause when the uterus is not longer needed for reproduction. Similarly, an increase in CXCR4 in bone marrow in response to estrogen enhances the mobility of these cells when needed for reproduction and in response to uterine signaling. Interestingly BM cells also produce CXCL12 at a high level. It is likely that local CXCL12 serves to retain these cells in the BM, preventing depletion. Elevated CXCL12 from the endometrium likely competes with BM derived CXCL12 as a chemoattractant for the BM stem cells (Segers et al., 2007). Elevated E2, which reaches the levels used here in the late proliferative phase, may ensure an adequate mobilization of stem cells near the time of ovulation and embryo implantation. P4 also induces the production of CXCL12 and may lead to further mobilization of stem cells in support of pregnancy. The regulation of stem cells by sex steroids is likely a widespread phenomenon. Nakada et al. (Nakada et al., 2014) showed that E2 promotes the HSCs self-renewal and the replicative activity of the HSC pool is augmented in female versus male mice. Li et al. (2013) reported that E2 enhanced the recruitment of BM-derived EPC into infarcted myocardium and induced CXCR4 expression in mice. Similarly, Foresta et al. (2010) have observed an increase in the number of CXCR4+ EPC during the ovulatory phase, which was likely caused by E2 activation. Sex steroid induced alterations in stem cell renewal and mobilization may underlie many sex specific differences in health and disease. In the ectopic endometrium of endometriosis, high E2 biosynthesis and low E2 inactivation lead to an excess of local E2 (Giudice and Kao, 2004). These provide a favorable condition for inducing BM-derived stem cell migration to normal and ectopic endometrium. Consistent with this theory, we have previously shown that stem cells are attracted to the ectopic endometrium (Sakr et al., 2014). The ectopic lesions compete for a limited pool of circulating BM-derived cells, depriving the uterus of the normal number of recruited stem cells. Insufficient uterine repair and regeneration may contribute to the infertility associated with endometriosis. The identification of CXCL12 as the primary chemokine that recruits BM-derived cells to the uterus may allow therapeutic use in endometriosis and other uterine disease to restore fertility. The expression of CXCL12 in mouse endometrial cells is far less than endometrial cells in humans. This may be the cause for the decrease in the number of stem cells recruited to the uterus in mouse. Moreover, mice do not menstruate and thereby may not be a need to attract new cells with every cycle while humans menstruate and have a greater need to regenerate the endometrium from stem cells. We conclude that estradiol plays a key role in normal and ectopic endometrium by augmenting the migration of BM-derived stem cells to the endometrium. Estradiol regulates stem cell migration by inducing CXCL12 expression by endometrial stromal cells and CXCR4 expression by BM-derived cells. Sex steroid induced stem cell recruitment may explain many health related sex differences. Estradiol or CXCL12/CXCR4 may prove to be useful therapeutic agents in stem cell mediated diseases. Acknowledgments We thank Yuping Zhou for assistance with western blots, Hanyia Naqvi for animal husbandry, Dr. Fei Li for mouse stem cell isolation, Dr. Sihyun Cho for qPCR analysis and Demetra Hufnagel for manuscript editing. Abbreviations hESCshuman endometrial stromal cells BMCsbone marrow cells CXCR4chemokine receptor type 4 E2estradiol CXCL12 or SDF-1αstromal derived factor BMbone marrow HSCshematopoietic stem cells ESCsendometrial stromal cells EPCsendothelial progenitor cells MSCsmesenchymal stromal cells mBMCsmouse bone marrow cells ERestrogen receptor Figure 1 Immunophenotypic characterization of mBMCs by FACS. Primary mBMCs were incubated with fluorescent-labeled CD90, CD105, CD34, CD31, and Flk-1 antibodies or with isotype-matched irrelevant antibody for FACS analysis. Experiments were performed 3 times with different samples and each time in duplicate. Results (% of cells) presented are the average of triplicates. Figure 2 Western blot analysis of CXCR4 and quantification of CXCL12 by ELISA. (A) The CXCR4 protein (25 μg) from mBMC, mUC or hESCs was subjected to 10%-SDS-PAGE and immunoblotted against anti-CXCR4 antibody. (B) The expression levels of CXCL12 quantified by ELISA from the supernatants of 48 h old mBMCs, hESCs and mUCs. The graph presents the mean of three separate experiments; error bars represent the ± S.D. * denotes a statistically significant difference (p b 0.05) vs mBMCs and # denotes a statistically significant difference (p b 0.02) vs hESCs. Figure 3 The detection of CXCR4 and CD45 in mBMCs with fluorescent ICC. Representative photomicrographs of fluorescent ICC for CD45 (A, green) and CXCR4 (B, red) expression in mBMCs (100x), DAPI-stained nuclei are depicted in blue (C), and merged CD45 and CXCR4 are shown in (D). Figure 4 The chemotactic activity of conditioned supernatants to mBMC and hESCs. The migratory capacity of mBMC towards hESC supernatant. Untreated or AMD100 treated for 30 min mBMCs were seeded (5 × 104/well) on the top of inserts and placed in a 24-well plate containing serum-free medium alone (control) or 48 h hESC serum-free supernatant. Data are shown as CI: cells migrating in response to the conditioned supernatants divided by cells responding to the serum-free medium. Bars represent the mean ± S.D. of four independent experiments, each performed in triplicate. * denotes a statistically significant difference (p b 0.05) vs control; # (p b 0.05) vs the cells without AMD3100 treatment. Figure 5 E2 and progesterone (P4) induced CXCR4 and CXCL12 expression in mBMC and hESCs. mBMCs were serum starved overnight prior to being stimulated with E2 or P4 in fresh serum-free DMEM/F12 medium for 6 h and 24 h and CXCR4 mRNA was analyzed by qRT-PCR. (A & B) E2-induced CXCR4 mRNA expression at 6 h and 24 h. (C) E2 alone or plus with P4-induced CXCR4 mRNA expression at 24 h. (D) Western blot analysis of E2-induced CXCR4 protein at 24 h. (E & F) mBMC and hESCs were serum starved overnight and supernatants were collected for ELISA analysis after the cells were stimulated with E2 or P4 for 48 h in serum-free DMEM/F12. The bars in each graph represent the mean ± S.D. of three individual experiments, each performed in triplicate. Statistical significance (*p b 0.05) is noted on the graphs (*p b 0.05 vs control, *p b 0.05 vs E2). Figure 6 Effects of E2 on migratory capacity of mBMC. Untreated or E2-treated mBMCs were seeded (5 × 104/well) on the top of inserts and placed in a 24-well plate containing serum-free DMEM/F12 or 48 h hESCs serum-free supernatant. The cells were pretreated with AMD3100 for 30 min before migration. Data are shown as CI: cells migrating in response to the conditioned supernatants divided by cells responding to the serum-free medium. Each bar represents the mean ± S.D. for data from three individual experiments and each experiment was performed in triplicate. * denotes statistical significance (p b 0.05) compared to control, and # denotes statistical significance (p b 0.05) compared to U1 or E2. ==== Refs References Alawadhi F Du H Cakmak H Taylor HS Bone marrow-derived stem cell (BMDSC) transplantation improves fertility in a murine model of Asherman's syndrome PLoS One 2014 9 5 12 Applied Biosystems ABI Prism 7700, User Bulletin 1997 2 Foster City, CA Barr A Manning D Manning DR G Proteins Techniques of Analysis 1999 227 245 CRC Press, Inc. Boca Raton, FL Baruscotti I Barchiesi F Jackson EK Imthurn B Stiller R Kim JH Schaufelberger S Rosselli M Hughes CC Dubey RK Estradiol stimulates capillary formation by human endothelial progenitor cells: role of estrogen receptor-{alpha}/{beta}, heme oxygenase 1, and tyrosine kinase Hypertension 2010 56 3 397 404 20644008 Beauséjour C Bone marrow-derived cells: the influence of aging and cellular senescence Handb. Exp. Pharmacol 2007 180 67 88 17554505 Du H Taylor HS Contribution of bone marrow-derived stem cells to endometrium and endometriosis Stem Cells 2007 25 8 2082 2086 17464086 Foresta C De Toni L Di Mambro A Ferlin A Perilli L Bertuzzi I Galan A Zuccarello D Role of estrogen receptors in menstrual cycle-related neoangiogenesis and their influence on endothelial progenitor cell physiology Fertil. Steril 2010 93 1 220 228 18990386 Giudice LC Kao LC Endometriosis Lancet 2004 13–19 364 1789 1799 Hopman RK DiPersio JF Advances in stem cell mobilization Blood Rev 2014 28 1 31 40 24476957 Hu C Yong X Li C Lü M Liu D Chen L Hu J Teng M Zhang D Fan Y Liang G CXCL12/CXCR4 axis promotes mesenchymal stem cell mobilization to burn wounds and contributes to wound repair J. Surg. Res 2013 183 1 427 434 23462453 Jaerve A Schira J Müller HW Concise review: the potential of stromal cell-derived factor 1 and its receptors to promote stem cell functions in spinal cord repair Stem Cell Transl. Med 2012 1 10 732 739 Jeong Y Mangelsdorf DJ Nuclear receptor regulation of stemness and stem cell differentiation Exp. Mol. Med 2009 41 8 525 537 19696553 Lai CY Yamazaki S Okabe M Suzuki S Maeyama Y Iimura Y Onodera M Kakuta S Iwakura Y Nojima M Otsu M Nakauchi H Stage-specific roles for CXCR4 signaling in murine hematopoietic stem/progenitor cells in the process of bone marrow repopulation Stem Cells 2014 32 7 1929 1942 24510783 Lapidot T Mechanism of human stem cell migration and repopulation of NOD/SCID and B2mnull NOD/SCID mice. The role of SDF-1/CXCR4 interactions Ann. N. Y. Acad. Sci 2001 938 83 95 11458529 Li H Liu J Ye X Zhang X Wang Z Chen A Zhou M Zhao Q 17β-Estradiol enhances the recruitment of bone marrow-derived endothelial progenitorcells intoinfarcted myocardium by inducing CXCR4 expression Int. J. Cardiol 2013 162 2 100 106 21636145 Liu N Tian J Cheng J Zhang J Migration of CXCR4 gene-modified bone marrow-derived mesenchymal stem cells to the acute injured kidney J. Cell. Biochem 2013a 114 12 2677 2689 23794207 Liu X Zhou C Li Y Ji Y Xu G Wang X Yan J SDF-1promotesendochondralbonerepairduringfracture healing at the traumatic brain injury condition PLoS One 2013b 8 1 Marquez-Curtis LA Gul-Uludag H Xu P Chen J Janowska-Wieczorek A CXCR4 transfection of cord blood mesenchymal stromal cells with the use of cationic liposome enhances their migration toward stromal cell-derived factor-1 Cytotherapy 2013 15 7 840 849 23623275 Nakada D Oguro H Levi BP Ryan N Kitano A Saitoh Y Takeichi M Wendt GR Morrison SJ Estrogen increases haematopoietic stem-cell self-renewal in females and during pregnancy Nature 2014 505 555 558 24451543 Penn MS Pastore J Miller T Aras R SDF-1 in myocardial repair Gene Ther 2012 19 583 587 22673496 Ruiz A Salvo VA Ruiz LA Báez P García M Flores I Basal and steroid hormone-regulated expression of CXCR4 in human endometrium and endometriosis Reprod. Sci 2010 17 10 894 903 20720261 Ryan IP Schriock ED Taylor RN Isolation, characterization, and comparison of human endometrial and endometriosis cells in vitro J. Clin. Endocrinol. Metab 1994 78 3 642 649 8126136 Sakr S Naqvi H Komm B Taylor HS Endometriosis impairs bone marrow-derived stem cell recruitment to the uterus whereas bazedoxifene treatment leads to endometriosis regression and improved uterine stem cell engraftment Endocrinology 2014 155 4 1489 1497 24484171 Segers VF Tokunou T Higgins LJ MacGillivray C Gannon J Lee RT Local delivery of protease-resistant stromal cell derived factor-1 for stem cell recruitment after myocardial infarction Circulation 2007 116 1683 1692 17875967 Sharma M Afrin F Satija N Tripathi RP Gangenahalli GU Stromal-derived factor-1/CXCR4 signaling: indispensable role in homing and engraftment of hematopoietic stem cells in bone marrow Stem Cells Dev 2011 20 6 933 946 21186999 Shi M Li J Liao L Chen B Li B Chen L Jia H Zhao RC Regulation of CXCR4 expression in human mesenchymal stem cells by cytokine treatment: role in homing efficiency in NOD/SCID mice Haematologica 2007 92 7 897 904 17606439 Sundararaman S Miller TJ Pastore JM Kiedrowski M Aras R Penn MS Plasmid-based transient human stromal cell-derived factor-1 gene transfer improves cardiac function in chronic heart failure Gene Ther 2011 18 867 873 21472007 Taylor HS Endometrial cells derived from donor stem cells in bone marrow transplant recipients JAMA 2004 292 81 85 15238594 Tsutsumi A Okada H Nakamoto T Okamoto R Yasuda K Kanzaki H Estrogen induces stromal cell-derived factor 1 (SDF-1/CXCL12) production in human endometrial stromal cells: a possible role of endometrial epithelial cell growth Fertil. Steril 2011 95 1 444 447 20869048 Unzek S Zhang M Mal N Mills WR Laurita KR Penn MS SDF-1 recruits cardiac stem cell-like cells that depolarize in vivo Cell Transplant 2007 16 879 886 18293886 Wang Q Yu JH Zhai HH Zhao QT Chen JW Shu L Li DQ Liu DY Dong C Ding Y Temporal expression of estrogen receptor alpha in rat bone marrow mesenchymal stem cells Biochem. Biophys. Res. Commun 2006 347 1 117 123 16806066 Wolff EF Gao XB Yao KV Andrews ZB Du H Elsworth JD Taylor HS Endometrial stem cell transplantation restores dopamine production in a Parkinson's disease model J. Cell. Mol. Med 2011 15 4 747 755 20406327 Zhou S Zilberman Y Wassermann K Bain SD Sadovsky Y Gazit D Estrogen modulates estrogen receptor alpha and beta expression, osteogenic activity, and apoptosis in mesenchymal stem cells (MSCs) of osteoporotic mice J. Cell. Biochem. Suppl. Suppl 2001 36 144 155 Zhou Y Gan Y Taylor HS Cigarette smoke inhibits recruitment of bone-marrow derived stem cells to the uterus Reprod. Toxicol 2011 31 2 123 127 20955787

PMC005xxxxxx/PMC5001161.txt

==== Front Regen Med ResRegen Med ResrmrRegenerative Medicine Research2050-490XEDP Sciences 2752979710.1051/rmr/160001-srmr160001-s10.1051/rmr/160001-sEditorialRegenerative Medicine Research: striving to better serve the emerging field Kang Y. James 12*1 Regenerative Medicine Research Center, Sichuan University West China Hospital, Sichuan PR China2 Pharmacology and Toxicology, University of Louisville School of Medicine, Louisville USA⁎ Corresponding author: rmr.editor@vip.163.com2016 12 8 2016 4 rmr/2016/01E128 6 2016 28 6 2016 © Y.J. Kang, published by EDP Sciences, 20162016Y.J. Kang, published by EDP SciencesThis is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. ==== Body Regenerative Medicine Research, previously published by BioMed Central, is now transferred to its new publisher, EDP Sciences. As the Editor-in-Chief, I am excited to announce the continuation of this new open access, online journal. It remains to publish research relating to both the fundamental and practical aspects of regenerative medicine, with a particular emphasis on translational studies. Regenerative medicine is an emerging field with the potential to empower medical research and practice. The process of tissue regeneration is essential to many organisms for a healthy life, from the highly studied axolotl, to us as human beings. In many ways regenerative medicine has become the leading edge of biomedical research and clinical practice, due to its key implications for the treatment of increasingly prevalent diseases, such as heart disease. The field is in the infant stage of its development but it is growing at a fast pace. However, comprehensive information or research data on regenerative medicine cannot presently be efficiently gathered, but can only be scrutinized through a diversity of different journals in various fields. Regenerative medicine is, by current definition [1], the process of creating living, functional tissues to repair or replace tissue or organ function lost due to age, disease, damage, or congenital defects. Obviously, this requires a multidisciplinary approach, and one that looks beyond tissue engineering and stem cells. Regenerative Medicine Research just meets this need and continues to find its own niche in biomedical science. As an open access journal, articles accepted for publication in Regenerative Medicine Research are made rapidly available online following acceptance. This ensures efficient dissemination of information and experimental data to scientific communities and the medical practice field. Importantly, a highly respected Editorial Board composed of internationally well-recognized experts in the field facilitates a fair, timely and rigorous peer review process. The Editorial Board, Publisher, and myself assure our authors and readers that we are committed to making Regenerative Medicine Research a preeminent platform for exchanging research protocols and experimental data in this exciting, emerging field. We look forward to receiving your high quality contributions to the journal. Editor-in-Chief Y. James KangCite this article as: Kang YJ (2016) Regenerative Medicine Research: striving to better serve the emerging field. Regen Med Res, 4, E1 ==== Refs 1 Mason C , Dunnill P ( 2008 ) A brief definition of regenerative medicine . Regen Med 3 (1 ), 1 –5 . Doi:10.2217/17460751.3.1.1. 18154457

PMC005xxxxxx/PMC5001165.txt

==== Front Front NutrFront NutrFront. Nutr.Frontiers in Nutrition2296-861XFrontiers Media S.A. 2761726110.3389/fnut.2016.00035NutritionPerspectivePotential Impact of Dietary Choices on Phosphorus Recycling and Global Phosphorus Footprints: The Case of the Average Australian City Metson Geneviève S. 1*†Cordell Dana 1Ridoutt Brad 231Institute for Sustainable Futures, University of Technology Sydney, Ultimo, NSW, Australia2Commonwealth Scientific and Industrial Research Organisation, Clayton South, VIC, Australia3Department of Agricultural Economics, University of the Free State, Bloemfontein, South AfricaEdited by: Philippe Hinsinger, INRA, France Reviewed by: Anthony Weatherley, University of Melbourne, Australia; Christian Folberth, International Institute for Applied Systems Analysis, Austria *Correspondence: Geneviève S. Metson, genemetson@gmail.com†Present address: Geneviève S. Metson, National Research Council, National Academies of Science, School of the Environment, Washington State University, Vancouver, WA, USA Specialty section: This article was submitted to Nutrition and Environmental Sustainability, a section of the journal Frontiers in Nutrition 26 8 2016 2016 3 3518 4 2016 09 8 2016 Copyright © 2016 Metson, Cordell and Ridoutt.2016Metson, Cordell and RidouttThis is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.Changes in human diets, population increases, farming practices, and globalized food chains have led to dramatic increases in the demand for phosphorus fertilizers. Long-term food security and water quality are, however, threatened by such increased phosphorus consumption, because the world’s main source, phosphate rock, is an increasingly scarce resource. At the same time, losses of phosphorus from farms and cities have caused widespread water pollution. As one of the major factors contributing to increased phosphorus demand, dietary choices can play a key role in changing our resource consumption pathway. Importantly, the effects of dietary choices on phosphorus management are twofold: First, dietary choices affect a person or region’s “phosphorus footprint” – the magnitude of mined phosphate required to meet food demand. Second, dietary choices affect the magnitude of phosphorus content in human excreta and hence the recycling- and pollution-potential of phosphorus in sanitation systems. When considering options and impacts of interventions at the city scale (e.g., potential for recycling), dietary changes may be undervalued as a solution toward phosphorus sustainability. For example, in an average Australian city, a vegetable-based diet could marginally increase phosphorus in human excreta (an 8% increase). However, such a shift could simultaneously dramatically decrease the mined phosphate required to meet the city resident’s annual food demand by 72%. Taking a multi-scalar perspective is therefore key to fully exploring dietary choices as one of the tools for sustainable phosphorus management. phosphorusdietfootprintsustainable resource userecyclingAustralia ==== Body Introduction: The Global Phosphorus Challenge Food production, and as such human diet, is the main driver of global phosphorus demand. Phosphorus is almost exclusively mined for food production (1), but globally, only 20% of what is mined as phosphorus is ultimately consumed as food (2). Phosphorus is essential for all life and has no substitute in plant and animal growth. However, different foods types require different amounts of phosphorus, which are sourced naturally from soils or in the form of fertilizers, feed, and food additives. Part of these different requirements stem from different crop and animal nutrient needs. These different requirements also stem from the additional losses1 of phosphorus that occur at each step of the food chain. Phosphorus can be lost or wasted through runoff and erosion from fields, as part of crop and food waste, as well as through the excreta of animals and humans. The process of having to grow animal feed crops, rearing animals, and converting them to human food requires substantially more land, energy, and water than the direct consumption of crops by humans (3), and this is also the case for phosphorus (4). Globally, 28% of the increased phosphate fertilizer consumption over 46 years (1961–2007) can be attributed to changes in diet, making it not only an important driver of global phosphorus flows but also a potential leverage point to increase phosphorus sustainability (4). As global population and per capita meat consumption increase, so will the demand for food and hence phosphorus, which could be unsustainable under a “business as usual” scenario (5). Phosphorus management is emerging as a serious global challenge to ensure affordable and accessible food and clean water. Mined phosphate rock, the source of most phosphorus fertilizers, is a non-renewable global resource, which is becoming an increasingly geopolitically, economically, and physically scarce resource (6). This makes short- and long-term availability and affordability for farmers (and thus consumers) a growing concern (7). We have already experienced some of the effects of short-term phosphorus scarcity, including phosphate fertilizer price spikes, during the 2008 food crisis where the most affected populations were, and continue to be, the ones who need phosphorus the most, such as farmers in developing countries who lack access to sufficient phosphorus to meet crop needs, especially in areas where highly weathered soils with high phosphorus fixation capacity make it more difficult for plants to access the little phosphorus that is applied (8–10). At the same time, losses of phosphorus to waterways, whether from agricultural fields or urban sewage, can cause severe water quality degradation, resulting in eutrophication, harmful algal blooms, and hypoxia that impair drinking water, recreational areas, and fisheries (11–13). Over 400 coastal water bodies globally are hypoxic because of excess phosphorus and nitrogen additions (14), while 60% of medium and large lakes and reservoirs are eutrophic (15). Current global phosphorus management, or lack thereof, is thus not sustainable (16, 17). The second UN Sustainable Development Goal2 highlights that long-term global food security and adequate human nutrition are dependent on agricultural and food chain practices and the resources those practices depend on. Although the goal itself does not mention phosphorus explicitly, as an essential agricultural input phosphorus should be considered in the way we meet this goal. Explicitly considering food system interventions that address phosphorus security will be essential to meeting food security and healthy environment objectives (18). In this paper, we explore one potentially important intervention to reduce the demand for mined phosphorus, reduce losses to waterways, and influence the recycling potential of waste: human diet. Here, we define diet as the quantity and diversity of food consumed per capita in 1 year, more specifically, the amount of food consumed in 16 basic food categories.3 We explore the implications of diet on phosphorus sustainability using the Australian context as a case study. Specifically, we investigated the impact of city residents shifting toward plant-based diets (see Box 1) on mineral fertilizer demand as well as on the potential supply of phosphorus from recycled wastes in urban areas.4 We choose the city as the point of intervention because most of the world’s population now live in cities, and in Australia, 90% already live in urban areas. Through this Australian case study, we demonstrate that a multi-scalar viewpoint, from local resource recovery to global mined phosphorus requirements, may change how one views human diet as a lever of change in a city. Box 1 Calculating phosphorus consumption and the phosphorus footprint associated with the Australian diet. To calculate phosphorus consumption in the average Australian diet, we multiplied national average per capita Australian food intake data by the average phosphorus concentrations of food groups (Eq. 1). More specifically, food intake data from the 2011 National Nutrition Survey were converted into 16 basic food equivalent categories using the same methods as with 1995 survey data described in Ridoutt et al. (19). The consumption of these 16 food groups, reported in weight per capita (e.g., kilogram of raw fruit equivalent per capita) were then multiplied by the average phosphorus content of each food group as documented in the Food Standards Australia New Zealand database (FSANZ, 20). (1) Pconsumed=∑ food groupsMassfood group×  Pconcentration (2) Pfootprint=∑ food groupsMassfood group×  Pfertilizer used In order to calculate conversion factors for phosphorus content in kilogram of phosphorus per kilogram of ingested food for the vegetable oil, tea and coffee, and alcoholic beverage food groups, an additional step was required. The concentration of phosphorus for these food groups was reported by phosphorus content in one liquid portion size by FSANZ, and thus, we used the density of water to calculate a conversion factor based on food intake weight. For the dairy food group, we used the density of milk (1.03 g/ml) to convert the phosphorus ingested per portion to an average phosphorus content conversion factor. We assume that 98% of phosphorus ingested is excreted (21). For the plant-based diet, we converted meat, dairy, eggs, and seafood food groups to pulses (i.e. beans or legumes) based on protein equivalent in order to keep the protein intake of the diet identical to 2011 reported levels (20). This may be a conservative estimate, as the average global citizen consumes a third more protein than actually required (22), much greater for Australia, hence a shift to a plant-based diet in Australia would not require replacement of all protein currently consumed. However, the nutritional value of plant-based protein vs. meat-based protein is still debated (23), and some sources recommend increasing protein intake to 125% of the omnivore daily recommendated values, if eating a plant-based diet (24). The phosphorus footprint calculations were based on the conversion factors developed in Metson et al. (4) (Eq. 2). Instead of the FAO food supply per capita used by Metson et al. (4), we used the 2011 Australian specific food intake plus food waste estimates [using the same methods as described in Ridoutt et al. (19)] to be consistent with the phosphorus consumption numbers described above. The food waste estimates were reported as a percentage for each food group intake weight and then added to intake. The current diet Australian footprint calculated here is lower than the value calculated in Metson et al. (4) (6.51 kg per capita). The discrepancy is likely due to the use of different data, even though, theoretically, both the bottom up (food intake to food grown used here) and the top down (food grown to dietary intake as used by FAO) should be roughly comparable. Footprint Analyses, Phosphorus, and Human Diet Environmental and ecological footprints are used to determine the resources needed to support particular consumption patterns, whether it is for an individual, an economic sector, or a community (25, 26). The intent is to better assess the impact of particular practices on sustainability, both from a resource use and pollution perspective. Dietary choices and agricultural practices, for example, have been studied for their water (27), energy and land (3), climate change contribution (28, 29), nitrogen (30), and phosphorus (4) footprints, both locally and globally. Footprint analyses date back to the 1990s (25), but phosphorus has been underrepresented in footprint analyses until recently [as presented in Metson et al. (31), for example]. Over the past 10 years, there have been many important contributions to measuring direct phosphorus flows through the food system (and non-food system) at the global (2), regional (32, 33), national (34–36), city (37, 38), and household (39) scales. More extensive meta reviews of phosphorus substance flow analyses at multiple scales can be found in Ref. (40, 41). Our understanding of the dietary phosphorus footprint has increased as concerns around local and global sustainable phosphorus management have grown along with an increased synthesis of disparate data. The phosphorus footprint of a country varies, largely due to the amount of meat consumption, as indicated by Figure 1 (4). We define the phosphorus footprint, as in Metson et al. (4), as the total amount of mined phosphorus required to produce the food consumed by one person in a particular country over 1 year. The average per capita human dietary phosphorus footprint has increased 38% globally since 1970 (4). There are important differences in the phosphorus footprint of countries, where countries with high Human Development Index (HDI) values typically consume larger amounts of animal products and have significantly higher phosphorus footprints than lower HDI counties. Although still not as high as the phosphorus footprints of the USA or Argentina, emerging economies have shown rapid phosphorus footprint increases over time (e.g., China’s phosphorus footprint has increased by 400% since 1970). Figure 1 Dietary phosphorus footprint associated with different food groups and selected countries demonstrating the important contribution of meat to the phosphorus footprint value and the large variability of the phosphorus footprint between countries. Australia, although not depicted here, has a phosphorus footprint of 6.51 kg per capita. Reproduced with permission from McGill University (42), with data based on Metson et al. (4). Some countries with poor nutritional status associated with undernourishment need to increase their caloric intake and diversify their diet, consequently increasing their phosphorus footprint. However, decreasing (or limiting) meat intake in already high phosphorus footprint countries would be an effective strategy to decrease mined phosphorus demand. For example, in the USA, moving from the current meat-intensive diet to a plant-based diet could potentially decrease the country’s phosphorus fertilizer demand by 44% (43). Similarly, a recent study of the Austrian food system revealed that moving toward healthier and less meat-intensive diet would decrease phosphorus demand by 20–25% and decrease phosphorus losses to waterways by 5–6% (44). It is clear that dietary choices have a large impact on the whole food production chain and thus on phosphorus sustainability globally. However, the effect of diet on phosphorus consumption and pollution depends on the particular context of analysis, such as type of livestock systems, source of phosphorus input into livestock systems (fertilized pastures, grain-fed, fodder, additives). For example, in Australia, 63% of the country’s phosphorus demand is associated with livestock production because the majority of animals are reared in fertilized pasture systems rather than intensive feedlots (45). The Impact of Diet on Australians’ Phosphorus Footprint and Waste In Australia, the amount of phosphorus mined to support the current average person’s diet (the phosphorus footprint) is over three times higher than that needed to support a plant-based diet. However, the amount of phosphorus consumed in the average Australian diet is slightly less than in a plant-based diet (Figure 2).5 More specifically, Australians currently ingest approximately 0.67 kg per capita each year in food but have a phosphorus footprint of 4.9 kg per capita, over seven times the amount ingested in food, according to our calculations. This Australian footprint value is similar to that of EU countries in 2007 (4). Figure 2 Comparison of the phosphorus footprint (black), phosphorus in consumed food (gray), and phosphorus in human excreta (white) in the current Australian diet and a hypothetical plant-based diet. A shift to a plant-based diet with the same protein content as the current average diet would potentially increase phosphorus ingested (and thus excreted) by 8% (to 0.73 kg per capita) but would decrease the footprint of each Australian to 1.35 kg per capita – a 72% decrease. Mihelcic et al. (46) reported that globally, the phosphorus excreted through urine and feces vary between 0.18 and 0.73 kg per capita per year where diets higher in vegetable proteins result in higher phosphorus concentrations excreted (following the same trend found in our calculations). Importantly, the difference in phosphorus footprint between a plant-based and a conventional (current) diet is greater than the difference between the phosphorus excreted with these different diets. That is, the impact potential of changing diets is significant for reducing mined phosphorus (footprint) and relatively insignificant for changing the phosphorus content of excreta. Potential Implications of Dietary Changes as a Strategic Intervention for Phosphorus Security There are a wide variety of interventions to move toward phosphorus security – changing diets is just one of these. These solutions range from recycling phosphorus in organic wastes, increasing efficiencies along the food chain to reduce phosphorus demand in all sectors, including mining and fertilizer production, agriculture, food production, and consumption (47, 48). These sustainability interventions can extend the longevity of finite phosphate resources, reduce phosphorus pollution, increase efficiency of the whole supply chain, and reduce communities’ and countries’ dependence on imported phosphate from geopolitically risky regions, buffering against price spikes and supply disruptions (8). Some policies, agricultural practices, and waste recovery strategies to increase phosphorus efficiency and recycling have been implemented locally [see some examples in Cordell et al. (49)]. However, selecting the appropriate suite of measures for each unique social–institutional and biophysical context is crucial (50, 51), and diet is no exception. Taking a multi-scalar and systems approach to examining diet as an intervention point is essential, especially for a city. The sustainability implications of Australian city residents shifting their current diets to plant-based diets are wide-ranging, in terms of phosphorus security benefits and geographical scales, as indicated in Table 1. Table 1 The multi-scale impact of changing diets in an average Australian city on phosphorus demand, pollution, and recycling. Scale of impact Local Regional Global Impact of changing diets at the city scale on: Phosphorus demand – – Can drastically reduce demand for mined phosphate (72% reduction), thereby extending the longevity of the world’s finite phosphate rock resources, and reducing Australia’s dependence on imported phosphate rock Phosphorus pollution No local pollution reduction potential in cities because negligible reduction in phosphorus content of city resident’s excreta/wastewater (8% increase) May be significant because less phosphorus is used in agriculture and livestock (due to the reduced phosphorus footprint), which implies less phosphorus flowing from agricultural soils into waterways in total (assuming other practices remain the same) Same as regional impact Phosphorus recycling Negligible/minimal changes in recycling potential in cities because no reduction in phosphorus content of city resident’s excreta/wastewater (8% increase) – – This Australian phosphorus footprint investigation highlights the multi-scalar implications of dietary choice. That is, the need to consider the role of urban consumer behavior beyond city boundaries in order to better account for the full impacts of local decisions, in this case particularly urban ones, on global and local long-term phosphorus security. In reality, the impacts of dietary choices on phosphorus demand and pollution are complex and require consideration of additional factors beyond categories of consumed food (e.g., meat, vegetables, and grains); however, there are little data available on such impacts. Consumer choices about food can affect farming practices, which in turn affect fertilizer demand as well as phosphorus losses. For example, deciding to adhere to an organic diet would influence the source and amount of fertilizers added to fields and losses to waterways (52, 53), while also potentially resulting in a healthier diet (54). In addition, the effect of dietary choices may also extend beyond encouraging best management practices on fields and pastures [see Sharpley and Wang (55), for examples of such practices] to include post-harvest losses of food and thus phosphorus. For example, eating a more local diet may minimize losses from storage and transport food spoilage. Local diets may also potentially make it easier to encourage and evaluate sustainable phosphorus farming practices [as shown through analyses of nitrogen in the food system locally (56) and difficulties of doing so globally (57)]. In other words, the qualitative decisions about diet, in addition to the quantitative decisions about food group consumption, may be important for phosphorus security. As such, we require more detailed information on farming practices, food supply chains, and dietary habits in order to make more locally accurate estimates of the role of dietary changes on mined phosphorus demand as well as phosphorus in waste streams that could potentially be recycled or affect waterways. Broader environmental and health assessment, which incorporates more than just phosphorus, is also necessary to support robust recommendations about the benefits of changes in production systems and diets. In summary, both phosphorus recycling and phosphorus demand management should be considered in evaluating and combining solutions to the phosphorus challenge; however, this may not always be obvious when studying phosphorus from an urban perspective. Shifting toward a more plant-based diet may not significantly impact urban phosphorus recycling potential but can have a large impact on the demand for mined phosphorus. As shown here using Australia as a case study, a footprint perspective can allow cities to evaluate potential interventions from a different perspective, allowing stakeholders to make more informed decisions about prioritizing solutions that holistically increase phosphorus security. Author Contributions GM and DC designed research and perspective, GM conducted analysis, BR provided diet composition data, GM lead writing effort with contributions from DC and BR. Conflict of Interest Statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. 1Here, by losses, we are referring to phosphorus that is applied to agricultural land that does not end up incorporated in human and animal tissues within a year. Some of these losses are permanent on human time scales, while others are potentially recoverable (especially phosphorus that may be stored in agricultural soils). 2https://sustainabledevelopment.un.org/?page=view&nr=164&type=230&menu=2059 3Vegetables, fruits, grains, meat, seafood, nuts and seeds, eggs, beans/peas/pulses, dairy, vegetable oil, tea and coffee, alcoholic beverages, cocoa, honey, and sugar. 4A related project specifically looks at Mapping Sydney’s Phosphorus Supply & Demand, see http://www.p-futurescities.net/sydney-australia/#MappingSydney 5See Box 1 for methods and data used to calculate the amount of phosphorus in food and excreta and the phosphorus footprint of Australians. Funding Dr. GM was supported by an Endeavour Research Fellowship from the Australian Government, and Dr. DC was supported by a Chancellor’s Postdoctoral Research Fellowship at the University of Technology Sydney. The food intake data used in this report were developed with funding to CSIRO from the Rural Industries Research and Development Corporation (RIRDC). ==== Refs References 1 Smil V Phosphorus in the environment: natural flows and human interferences . Annu Rev Energy Environ (2000 ) 25 :53 –88 .10.1146/annurev.energy.25.1.53 2 Cordell D Drangert J-O White S The story of phosphorus: global food security and food for thought . Glob Environ Change (2009 ) 19 :292 –305 .10.1016/j.gloenvcha.2008.10.009 3 Pimentel D Pimentel M . Sustainability of meat-based and plant-based diets and the environment . Am J Clin Nutr (2003 ) 78 :660S –3S .12936963 4 Metson GS Bennett EM Elser JJ The role of diet in phosphorus demand . Environ Res Lett (2012 ) 7 :044043 10.1088/1748-9326/7/4/044043 5 Cordell D Schmid-Neset D Whiteb D Drangerta J Preferred future phosphorus scenarios: a framework for meeting long-term phosphorus needs for global food demand . International Conference on Nutrient Recovery from Wastewater Streams (Vol. 2009 ), Vancouver : IWA Publishing (2009 ). 23 p. 6 Cooper J Lombardi R Boardman D Carliell-Marquet C The future distribution and production of global phosphate rock reserves . Resourc Conservat Recycl (2011 ) 57 :78 –86 .10.1016/j.resconrec.2011.09.009 7 Cordell D White S Peak phosphorus: clarifying the key issues of a vigorous debate about long-term phosphorus security . Sustainability (2011 ) 3 (10 ):2027 –49 .10.3390/su3102027 8 Cordell D Turner A Chong J The hidden cost of phosphate fertilizers: mapping multi-stakeholder supply chain risks and impacts from mine to fork . Glob Change Peace Secur (2015 ) 27 :323 –43 .10.1080/14781158.2015.1083540 9 Dumas M Frossard E Scholz RW . Modeling biogeochemical processes of phosphorus for global food supply . Chemosphere (2011 ) 84 :798 –805 .10.1016/j.chemosphere.2011.02.039 21463882 10 Sanchez PA Soil fertility and hunger in Africa . Science (2002 ) 295 :2019 10.1126/science.1065256 11896257 11 Anderson DM Glibert PM Burkholder JM Harmful algal blooms and eutrophication: nutrient sources, composition, and consequences . Estuaries (2002 ) 25 :704 –26 .10.1007/BF02804901 12 Bennett E Carpenter S Caraco N Human impact on erodable phosphorus and eutrophication: a global perspective . Bioscience (2001 ) 51 :227 –34 .10.1641/0006-3568(2001)051[0227:HIOEPA]2.0.CO;2 13 Dodds WK Bouska WW Eitzmann JL Pilger TJ Pitts KL Riley AJ Eutrophication of US freshwaters: analysis of potential economic damages . Environ Sci Technol (2008 ) 43 :12 –9 .10.1021/es801217q 19209578 14 Diaz RJ Rosenberg R . Spreading dead zones and consequences for marine ecosystems . Science (2008 ) 321 :926 –9 .10.1126/science.1156401 18703733 15 Sayers MJ Grimm AG Shuchman RA Deines AM Bunnell DB Raymer ZB A new method to generate a high-resolution global distribution map of lake chlorophyll . Int J Remote Sens (2015 ) 36 :1942 –64 .10.1080/01431161.2015.1029099 16 Carpenter SR Bennett EM Reconsideration of the planetary boundary for phosphorus . Environ Res Lett (2011 ) 6 :014009 10.1088/1748-9326/6/1/014009 17 Elser J Bennett E Phosphorus cycle: a broken biogeochemical cycle . Nature (2011 ) 478 :29 –31 .10.1038/478029a 21979027 18 Cordell D White S Life’s bottleneck: implications of global phosphorus scarcity and pathways for a sustainable food system . Annu Rev Environ Resour (2014 ) 39 :161 –88 .10.1146/annurev-environ-010213-113300 19 Ridoutt B Baird D Bastiaans K Darnell R Hendrie G Riley M Nutritional Sufficiency of the National Food Production System: Situation and Outlook . Government of Australia Rural Industries Research and Development Corporation (2014 ). RIRDC Publication No 14/074, RIRDC Project No PRJ-008806. 20 FSANZ . NUTTAB 2010 Online Searchable Database . Food Standards Australia New Zealand (2010 ). 21 Drangert J Fighting the urine blindness to provide more sanitation options . Water SA (1998 ) 24 :1 –8 . 22 World Resources Institute . Shifting diets for a sustainable food future . In: Ranganathan J Vennard D Waite R Dumas P Lipinski B Searchinger T , editors. Installment 11 of “Creating a Sustainable Food Future” . (2016 ). Available from: http://www.wri.org/sites/default/files/Shifting_Diets_for_a_Sustainable_Food_Future_0.pdf 23 Friedman M Nutritional value of proteins from different food sources. A review . J Agric Food Chem (1996 ) 44 :6 –29 .10.1021/jf9400167 24 Knurick JR Fladell L Johnston CS Efficacy of an evidence-based food guide for vegetarians and vegans . FASEB J (2013 ) 27 :1066 –1062 . 25 Fang K Heijungs R de Snoo GR Theoretical exploration for the combination of the ecological, energy, carbon, and water footprints: overview of a footprint family . Ecol Indic (2014 ) 36 :508 –18 .10.1016/j.ecolind.2013.08.017 26 Rees WE Ecological footprints and appropriated carrying capacity: what urban economics leaves out . Environ Urban (1992 ) 4 :121 –30 .10.1177/095624789200400212 27 Hoekstra AY Chapagain AK Water footprints of nations: water use by people as a function of their consumption pattern . Water Resour Manag (2007 ) 21 :35 –48 .10.1007/s11269-006-9039-x 28 González AD Frostell B Carlsson-Kanyama A Protein efficiency per unit energy and per unit greenhouse gas emissions: potential contribution of diet choices to climate change mitigation . Food Policy (2011 ) 36 :562 –70 .10.1016/j.foodpol.2011.07.003 29 Risku-Norja H Kurppa S Helenius J Dietary choices and greenhouse gas emissions assessment of impact of vegetarian and organic options at national scale . Prog Ind Ecol Int J (2009 ) 6 :340 –54 .10.1504/PIE.2009.032323 30 Galloway JN Winiwarter W Leip A Leach AM Bleeker A Erisman JW Nitrogen footprints: past, present and future . Environ Res Lett (2014 ) 9 :115003 10.1088/1748-9326/9/11/115003 31 Metson GS Smith VH Cordell DJ Vaccari DA Elser JJ Bennett EM Phosphorus is a key component of the resource demands for meat, eggs, and dairy production in the United States . Proc Natl Acad Sci U S A (2014 ) 111 :E4906 –7 .10.1073/pnas.1417759111 25385650 32 van Dijk KC Lesschen JP Oenema O . Phosphorus flows and balances of the European Union Member States . Sci Total Environ (2016 ) 542 :1078 –93 .10.1016/j.scitotenv.2015.08.048 26421756 33 Yuan Z Shi J Wu H Zhang L Bi J . Understanding the anthropogenic phosphorus pathway with substance flow analysis at the city level . J Environ Manage (2011 ) 92 :2021 –8 .10.1016/j.jenvman.2011.03.025 21489683 34 Antikainen R Lemola R Nousiainen JI Sokka L Esala M Huhtanen P Stocks and flows of nitrogen and phosphorus in the Finnish food production and consumption system . Agric Ecosyst Environ (2005 ) 107 :287 –305 .10.1016/j.agee.2004.10.025 35 Linderholm K Mattsson JE Tillman AM . Phosphorus flows to and from Swedish agriculture and food chain . Ambio (2012 ) 41 :883 –93 .10.1007/s13280-012-0294-1 22627872 36 Gumbo B Short-Cutting the Phosphorus Cycle in Urban Ecosystems . Institute for Water Education, UNESCO-IHE (2005 ). 37 Tangsubkul N Moore S Waite TD Incorporating phosphorus management considerations into wastewater management practice . Environ Sci Policy (2005 ) 8 :1 –15 .10.1016/j.envsci.2004.08.009 38 Montangero A Cau LN Anh NV Tuan VD Nga PT Belevi H . Optimising water and phosphorus management in the urban environmental sanitation system of Hanoi, Vietnam . Sci Total Environ (2007 ) 384 :55 –66 .10.1016/j.scitotenv.2007.05.032 17604824 39 Fissore C Baker L Hobbie S King J McFadden J Nelson K Carbon, nitrogen, and phosphorus fluxes in household ecosystems in the Minneapolis-Saint Paul, Minnesota, urban region . Ecol Appl (2011 ) 21 :619 –39 .10.1890/10-0386.1 21639033 40 Chowdhury RB Moore GA Weatherley AJ Arora M A review of recent substance flow analyses of phosphorus to identify priority management areas at different geographical scales . Resourc Conservat Recycl (2014 ) 83 :213 –28 .10.1016/j.resconrec.2013.10.014 41 Cordell D Neset TS Prior T The phosphorus mass balance: identifying ‘hotspots’ in the food system as a roadmap to phosphorus security . Curr Opin Biotechnol (2012 ) 23 (6 ):839 –45 .10.1016/j.copbio.2012.03.010 22503084 42 McGill University . Headway . (2013 ). Available from: http://www.mcgill.ca/research/files/research/mcgill_headway_7-1_spr13_web_opt1.pdf 43 MacDonald GK Bennett EM Carpenter SR Embodied phosphorus and the global connections of United States agriculture . Environ Res Lett (2012 ) 7 :044024 10.1088/1748-9326/7/4/044024 44 Thaler S Zessner M Weigl M Rechberger H Schilling K Kroiss H Possible implications of dietary changes on nutrient fluxes, environment and land use in Austria . Agric Syst (2015 ) 136 :14 –29 .10.1016/j.agsy.2015.01.006 45 Cordell D Jackson M White S Phosphorus flows through the Australian food system: identifying intervention points as a roadmap to phosphorus security . Environ Sci Policy (2013 ) 29 :87 –102 .10.1016/j.envsci.2013.01.008 46 Mihelcic JR Fry LM Shaw R . Global potential of phosphorus recovery from human urine and feces . Chemosphere (2011 ) 84 :832 –9 .10.1016/j.chemosphere.2011.02.046 21429554 47 Childers D Corman J Edwards M Elser J Sustainability challenges of phosphorus and food: solutions from closing the human phosphorus cycle . Bioscience (2011 ) 61 :117 –24 .10.1525/bio.2011.61.2.6 48 Cordell D White S Sustainable phosphorus measures: strategies and technologies for achieving phosphorus security . Agronomy (2013 ) 3 :86 –116 .10.3390/agronomy3010086 49 Cordell D Rosemarin A Schroder J Smit A . Towards global phosphorus security: a systems framework for phosphorus recovery and reuse options . Chemosphere (2011 ) 84 :747 –58 .10.1016/j.chemosphere.2011.02.032 21414650 50 Cordell D Neset T-S Phosphorus vulnerability: a qualitative framework for assessing the vulnerability of national and regional food systems to the multi-dimensional stressors of phosphorus scarcity . Glob Environ Change (2014 ) 24 :108 –22 .10.1016/j.gloenvcha.2013.11.005 51 Metson GS Iwaniec DM Baker LA Bennett EM Childers DL Cordell D Urban phosphorus sustainability: Systemically incorporating social, ecological, and technological factors into phosphorus flow analysis . Environ Sci Policy (2015 ) 47 :1 –11 .10.1016/j.envsci.2014.10.005 52 Mäder P Fliessbach A Dubois D Gunst L Fried P Niggli U . Soil fertility and biodiversity in organic farming . Science (2002 ) 296 :1694 –7 .10.1126/science.1071148 12040197 53 Senthilkumar K Nesme T Mollier A Pellerin S Regional-scale phosphorus flows and budgets within France: the importance of agricultural production systems . Nutr Cycl Agroecosys (2012 ) 92 :145 –59 .10.1007/s10705-011-9478-5 54 Strassner C Cavoski I Di Cagno R Kahl J Kesse-Guyot E Lairon D How the organic food system supports sustainable diets and translates these into practice . Front Nutr (2015 ) 2 :19 .10.3389/fnut.2015.00019 26176912 55 Sharpley A Wang X . Managing agricultural phosphorus for water quality: lessons from the USA and China . J Environ Sci (2014 ) 26 :1770 –82 .10.1016/j.jes.2014.06.024 25193824 56 Billen G Garnier J Thieu V Silvestre M Barles S Chatzimpiros P Localising the nitrogen imprint of the Paris food supply: the potential of organic farming and changes in human diet . Biogeosci Discuss (2011 ) 8 :10979 –1002 .10.5194/bgd-8-10979-2011 57 Oita A Malik A Kanemoto K Geschke A Nishijima S Lenzen M Substantial nitrogen pollution embedded in international trade . Nat Geosci (2016 ) 9 :111 –5 .10.1038/ngeo2664

PMC005xxxxxx/PMC5001191.txt

==== Front Front SurgFront SurgFront. Surg.Frontiers in Surgery2296-875XFrontiers Media S.A. 10.3389/fsurg.2016.00048SurgeryOriginal ResearchCancer Stem Cells in Glioblastoma Multiforme Bradshaw Amy 1Wickremesekera Agadha 12Brasch Helen D. 1Chibnall Alice M. 1Davis Paul F. 1Tan Swee T. 13*†Itinteang Tinte 1†1Gillies McIndoe Research Institute, Wellington, New Zealand2Department of Neurosurgery, Wellington Regional Hospital, Wellington, New Zealand3Wellington Regional Plastic, Maxillofacial and Burns Unit, Hutt Hospital, Wellington, New ZealandEdited by: Eberval Figueiredo, University of São Paulo, Brazil Reviewed by: A. Samy Youssef, University of Colorado, USA; Antonio Aversa Do Souto, National Institute of Cancer, INCA, Brazil *Correspondence: Swee T. Tan, swee.tan@gmri.org.nz†Equal senior authors. Specialty section: This article was submitted to Neurosurgery, a section of the journal Frontiers in Surgery 26 8 2016 2016 3 4806 7 2016 11 8 2016 Copyright © 2016 Bradshaw, Wickremesekera, Brasch, Chibnall, Davis, Tan and Itinteang.2016Bradshaw, Wickremesekera, Brasch, Chibnall, Davis, Tan and ItinteangThis is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.Aim To identify and characterize cancer stem cells (CSC) in glioblastoma multiforme (GBM). Methods Four-micrometer thick formalin-fixed paraffin-embedded GBM samples from six patients underwent 3,3-diaminobenzidine (DAB) and immunofluorescent (IF) immunohistochemical (IHC) staining for the embryonic stem cell (ESC) markers NANOG, OCT4, SALL4, SOX2, and pSTAT3. IF IHC staining was performed to demonstrate co-expression of these markers with GFAP. The protein expression and the transcriptional activities of the genes encoding NANOG, OCT4, SOX2, SALL4, and STAT3 were investigated using Western blotting (WB) and NanoString gene expression analysis, respectively. Results DAB and IF IHC staining demonstrated the presence of a CSC population expressing NANOG, OCT4, SOX2, SALL4, and pSTAT3 with the almost ubiquitous presence of SOX2 and a relatively low abundance of OCT4, within GBM. The expression of NANOG, SOX2 and, pSTAT3 but, not OCT and SALL4, was confirmed by WB. NanoString gene analysis demonstrated transcriptional activation of NANOG, OCT4, SALL4, STAT3, and SOX2 in GBM. Conclusion This study demonstrated a population of CSCs within GBM characterized by the expression of the CSC markers NANOG, SALL4, SOX2, pSTAT3 and OCT4 at the protein and mRNA levels. The almost ubiquitous presence of SOX2 and a relatively low abundance of OCT4 would support the putative existence of a stem cell hierarchy within GBM. glioblastoma multiformeembryoniccancerstem cellsexpressionhierarchy ==== Body Introduction Glioblastoma multiforme (GBM), a grade 4 astrocytoma, is the most aggressive primary brain tumor with a 5-year survival of 2%, despite intensive research (1, 2) and the tumor usually recurs following surgical resection, radiotherapy, and chemotherapy (3, 4). This poor prognosis has been attributed to the initiation, propagation, and differentiation of cancer stem cells (CSCs) (3, 5, 6). The CSC concept proposes that a cancer originates from a small population of CSCs either by the acquisition of mutations in normal embryonic stem cells (ESCs) or progenitor cells that were imbued with the abilities for uncontrolled growth and propagation (7, 8). While the tumorigenic CSCs are proposed to be the driving force behind tumor growth, the bulk of the tumor consists of non-tumorigenic cancer cells that have differentiated from these CSCs, leading to a vast cellular heterogeneity (8, 9). Furthermore, when exposed to certain epigenetic or environmental factors, the downstream cancer cells can be reprogramed to acquire stem cell properties (10). Embryonic stem cells, which were originally isolated from cells of the inner cell mass of an early mammalian embryo (blastocyst) (11), possess the ability for perpetual propagation and differentiation into all cell lineages (12). ESCs and their downstream progenitors express a variety of characteristic proteins, including cell surface markers and transcription factors (13). CSCs in GBM bear the characteristics of ESCs by their expression of similar proteins, making it possible to both characterize and isolate CSCs in GBM (13, 14). A recent report demonstrates an ESC-like signature in high grade (grades 3 and 4) gliomas with upregulation of NANOG, KLF4, OCT4, and SOX2 proteins (15). Upregulation of these proteins has been correlated with poorer survival in both high- and low-grade gliomas (15, 16), although inclusion of gliomas of different grades and their analysis as a single entity means that this study inherently lacks the specificity needed for the characterization of a unique CSC population in GBM. A recent review points to CSCs in GBM possessing a hierarchy, with overlapping phenotypes expressing upstream (ESC) and downstream (progenitor cell) markers (17). While many CSC markers have been associated with CSCs, there is evidence indicating that some proteins play a greater role than others in maintaining ESC capabilities. NANOG, OCT4, and SOX2 are transcription factors that have been proposed to function synergistically to maintain ESC pluripotency and self-renewal (18). Expression of this protein trio has also been linked to aggressiveness of GBM, and all three are expressed in most gliomas (19). SALL4 is another transcription factor responsible for zygotic survival and ESC pluripotency (20, 21). Additionally, SALL4 physically interacts with NANOG and is associated with OCT4 and SOX2 (22, 23). pSTAT3, a signaling and transcription activating molecule, is also involved in ESC pluripotency (24) and is proposed to engender expression of other ESC-associated proteins such as SOX2 and SALL4 (25). This study aimed to identify and characterize the CSC population within GBM, using the ESC markers SOX2, OCT4, pSTAT3, SALL4, and NANOG at both the transcriptional and translational levels. Materials and Methods Tissue Samples Six GBM tissue samples from three male and three female patients aged 42–81 (mean, 64.2) years were sourced from the Gillies McIndoe Research Institute Tissue Bank, and used in a study approved by the Central Health and Disabilities Ethics Committee (ref. no. 15CEN28). Histochemical and Immunohistochemical Staining Four-micrometer thick formalin-fixed paraffin-embedded sections of GBM from six patients were used for hematoxylin and eosin staining to confirm the diagnosis of GBM by an anatomical pathologist (HDB). 3,3-Diaminobenzidine (DAB) and immunofluorescent (IF) immunohistochemical (IHC) staining of these sections was then performed using the Leica Bond Rx auto-stainer (Leica, Nussloch, Germany) as previously described (26). DAB IHC staining for GFAP (cat# PA0026, Leica), NANOG (1:100; cat# ab80892, Abcam, Cambridge, UK), SOX2 (1:500; cat# PA-094, Thermo Fisher Scientific, Rockford, IL, USA), SALL4 (1:30; cat# CM385M-16, Cell Marque, Rocklin, CA, USA), pSTAT3 (1:100; cat# 9145, Cell Signaling Technology, Danvers, MA, USA), and OCT4 (1:1000; cat# ab109183, Abcam) diluted with Bond™ primary antibody diluent (cat# AR9352, Leica) was undertaken for all GBM tissue samples. IF IHC staining was performed on two representative GBM tissue samples from the original cohort of patients included in DAB IHC staining, using identical primary antibodies and concentrations with an appropriate fluorescent secondary antibody. All IF IHC-stained slides were mounted using Vectashield HardSet anti-fade mounting medium with 4,6-diamidino-2-phenylindole (cat# H-1500, Vector Laboratories, Burlingame, CA, USA). Positive control human tissues for the primary antibodies were seminoma for NANOG, SALL4, and OCT4, skin for SOX2, and tonsil for pSTAT3. A secondary and tertiary only negative control staining by omitting the primary antibodies was performed on a GBM sample randomly selected from the original cohort of GBM samples used for DAB IHC staining (Figure S1 in Supplementary Material). Image Analysis All DAB IHC-stained slides were visualized with an Olympus BX53 light microscope (Tokyo, Japan) and the images were captured with the CellSens 2.0 software (Olympus). IF IHC stained-slides were viewed and the images were captured using an Olympus FV1200 biological confocal laser-scanning microscope and processed with cellSens Dimension 1.11 software using 2D deconvolution algorithm (Olympus). Western Blotting Five snap-frozen GBM samples, from the original cohort of six patients included in DAB IHC staining, were washed in 1× PBS and homogenized in RIPA buffer (cat# R0278, Sigma-Aldrich, St Lewis, MA, USA) supplemented with Halt™ Protease and Phosphatase Inhibitor Cocktail (cat# 1861281, Thermo Scientific, Waltham, MA, USA) and dithiothreitol (DTT) (cat# DTT-RO, Sigma-Aldrich). Protein was precipitated using a Calbiochem® ProteoExtract® Protein Precipitation Kit (cat# 539180, EMD Millipore Corp, Billerice, MA, USA) for 1 h at −20°C, washed and re-suspended in 1× Laemmli sample buffer (cat# 161-0737, Bio-Rad, Hercules, CA, USA) with 1% DTT. Equal amounts of protein were heated at 85°C and separated on Bolt™ 4–12% Bis–Tris Plus gels (cat# NW04120BOX, Invitrogen, Carlsbad, CA, USA) via electrophoresis. Separated protein was transferred to a nitrocellulose membrane (cat# IB23001, Life Technologies, Carlsbad, CA, USA) and blocked in 1× tris-buffered saline (pH 7.4) containing 0.1% Tween-20 (TBST) containing 2% skim milk powder for 90 min at 4°C. Primary antibody probing for each CSC marker was overnight in TBST at 4°C with the following primary antibodies at the given concentrations: rabbit monoclonal anti-OCT4 (1:1000; cat# ab109183, Abcam), anti-pSTAT3 (1:2000; cat# ab9145, Abcam), anti-SOX2 (1:5000; cat# PA1-094, Thermo Fisher, Scoresby, VIC, Australia), and anti-NANOG (1:2000; cat# ab47102, Abcam). Secondary antibody probing was in 1× TBST for 60 min at 4°C with goat anti-rabbit horseradish peroxidase (HRP; 1:10,000; cat# A16110, Thermo Fisher). Beta-actin antibody probing was performed with the iBind™ Flex device (cat# SLF2000, Life Technologies) using primary mouse monoclonal anti-β-actin (1:2000 cat# ab8226, Abcam) and secondary donkey anti-mouse Alexa fluor 488 (1:2000 cat# A21202, Thermo Fisher). Clarity Western ECL (cat# 1705061, Bio-Rad) was used as the substrate for visualizing HRP-detected protein bands, and the Chemi Doc MP Imaging System (Bio-Rad) and Image Lab 5.0 software (Bio-Rad) were used for both HRP and fluorescent band detection and analysis. Positive controls for the markers examined were NTERA2 for OCT4; NTERA2 and human placenta for SOX2; mouse lung and human liver for pSTAT3; 3T3 cell lysate for NANOG. Negative controls were HeLa for OCT4; human placenta for pSTAT3; SY5Y for NANOG. No negative tissues or lysates could be found for SOX2. NanoString Gene Expression Analysis Total RNA was extracted from ~20 mg of six snap-frozen GBM tissue from the same cohort of patients included in DAB IHC analysis, using the MagJET RNA kit (cat# k2731, Thermo Scientific) and the Kingfisher Duo RNA extraction machine (Thermo Scientific). All samples were quantitated and quality controlled with the NanoDrop 2000 Spectrophotometer (Thermo Scientific) and the Qubit 2.0 Fluormeter (Thermo Scientific). The samples with A260/A230 ≥1.5 and A260/A280 ~2 were used for further analysis. The integrity of the RNA was assessed by New Zealand Genomics Ltd. (Dunedin, NZ) using Agilent 2100 BioAnalyzer (Agilent Technologies, Santa Clara, CA, USA). The isolated RNA was then subjected to NanoString nCounter™ Gene Expression Assay (NanoString Technologies, Seattle, WA, USA) as completed by New Zealand Genomics Ltd. (Dunedin, New Zealand), according to the manufacturer’s protocol. Probes for the genes encoding NANOG (XM_011520850.1), SOX2 (NM_003106.3), SALL4 (NM_020436.3), OCT4 (NM_001159542.1), and STAT3 (NM_139276.2) and the housekeeping gene GAPDH (NM_002046.3) were designed and synthesized by NanoString Technologies. Raw data were analyzed using nSolver™ software (NanoString Technologies) using standard settings and was normalized against the housekeeping gene. Results Histochemical and 3,3-Diaminobenzidine Immunohistochemical Staining Hematoxylin and eosin stain (Figure 1A) confirmed the diagnosis of grade 4 astrocytoma on all 6 GBM samples. Positive staining for pSTAT3 (Figure 1B, brown), SALL4 (Figure 1C, brown), and SOX2 (Figure 1D, brown) was observed in tumor cells and within areas of endothelial proliferation. pSTAT3 showed consistent nuclear staining throughout the sample (Figure 1B, brown), and SALL4 staining was localized predominantly to the nuclei of tumor cells but was mostly cytoplasmic in the proliferative endothelium (Figure 1C, brown). SOX2 staining showed strong nuclear staining in tumor cells that lessened in intensity within areas of endothelial proliferation, with a consistent, moderate level of cytoplasmic staining throughout the entire sample (Figure 1D, brown). NANOG was localized to the nuclei of tumor cells but was not present in the endothelium (Figure 1E, pink). OCT4 staining was scarce in all samples but showed differential staining patterns, with some cells exhibiting nuclear (Figure 1F, brown, left) and others cytoplasmic (Figure 1F, brown, right) staining. Figure 1 Representative image of H&E stained slides showing the presence of GBM (A) and DAB IHC-stained slides of GBM tissue demonstrating expression of CSC markers (B–F). pSTAT3 [(B), brown] was expressed in the nuclei of tumor and endothelium of the microvessels throughout the sample. SALL4 [(C), brown] was predominantly expressed on the nuclei of the tumor cells, and the cytoplasm of the endothelial cells lining the microvessels. SOX2 [(D), brown] displayed strong nuclear staining of the tumor cells, and moderate cytoplasmic staining of the endothelium of the microvessels. Nuclear expression of NANOG [(E), pink/red] was observed on the tumor cells, and to the lesser extent on the endothelium of the microvessels. Nuclear and cytoplasmic staining of OCT4 [(F), brown] was observed in few tumor cells. Cell nuclei were counterstained with hematoxylin [(A–F), blue]. Original magnification: 400X. Expected staining patterns for pSTAT3 (Figure S1A in Supplementary Material, brown), SALL4 (Figure S1B in Supplementary Material, brown), SOX2 (Figure S1C in Supplementary Material, brown), NANOG (Figure S1D in Supplementary Material, brown), and OCT4 (Figure S1E in Supplementary Material, brown) were demonstrated in the respective positive controls. Staining with the omission of the primary antibodies in a GBM sample provided an appropriate negative control (Figure S1F in Supplementary Material). Immunofluorescent Immunohistochemical Staining To investigate co-expression of the ESC markers, IF IHC staining was performed on two representative GBM samples used for DAB IHC staining. To identify GBM tumor cells, GFAP (Figures 2A–D, green) was utilized as a marker for glial cells (27, 28). A substantial number of GFAP+ cells also expressed SOX2 (Figure 2A, red, arrows), NANOG (Figure 2B, red, arrows), and pSTAT3 (Figure 2C, red, arrows) with relatively low expression of OCT4 (Figure 2D, red, arrow). To examine expression of SALL4, we counterstained the same samples for SOX2 (Figure 2E, red) and SALL4 (Figure 2E, green) demonstrating the expression of both markers in the same nuclei (Figure 2E, arrows). The vasculature (Figures 2A–C, arrows) did not demonstrate expression of GFAP (Figures 2A–C, green), as expected, along with minimal expression of the aforementioned ESC markers. Images of the individual stains are presented in Figure S2 in Supplementary Material. A GBM sample used as a negative control by omitting the primary antibodies, demonstrated the specificity of the antibodies used (data not shown). Figure 2 Representative images of IF IHC-stained sections of GBM tissue for ESC markers. SOX2 [(A), red, arrows], NANOG [(B), red, arrows], and pSTAT3 [(C), red, arrows] all showed nuclear expression on GFAP+ tumor cells [(A–C), green]. OCT4 [(D), red, arrows] staining was scarce and solely cytoplasmic in GFAP+ tumor cells. SALL4 [(E), green] and SOX2 [(E), red] were co-expressed (arrows) in the nuclei of some tumor cells, with SALL4 also staining SOX2–negative cells. Cell nuclei were counterstained with 4′, 6′-diamidino-2-phenylindole [(A–E), blue]. Scale bars: 20 μm. Western Blotting The presence of OCT4, NANOG, pSTAT3, and SOX2 in GBM samples was also examined by WB. OCT4 was below the detection level in all four GBM samples compared with NTERA2 cell lysate used as a positive control, which showed a band of approximately 46 kDa (Figure 3A). pSTAT3 was expressed in three out of five GBM samples at ~90 kDa (Figure 3B). NANOG was present in three out of four GBM samples, although multiple bands were detected with the antibody at approximately 40 and 35 kDa (Figure 3C). SOX2 was detected in four out of five samples and multiple bands at approximately 45 and 38 kDa were observed (Figure 3D). WB data for SALL4 has not been included due to antibody difficulties involving non-specific binding (data not shown). Figure 3 Western blots of five GBM tissue samples. OCT4 was not detected (A), pSTAT3 at ~90 kDa was found in three out of five samples (B), NANOG was present in all five samples with multiple bands at approximately 40 and 31 kDa (C). SOX2 was detected in four out of five samples with bands at approximately 45 and 38 kDa (D). NanoString Gene Expression Analysis mRNA quantification was performed for NANOG, OCT4, SALL4, STAT3, and SOX2 to investigate the presence of transcription activation of these markers in GBM. The expression values were normalized to that of the housekeeping gene GUSB and showed that all five markers were expressed in GBM samples (Figure 4). Figure 4 Relative expression of CSC mRNA transcripts in six GBM samples showing the presence of all 5 markers at varying levels. NANOG, OCT4, and SALL4 showed relatively low mRNA expression, while STAT3 and SOX2 displayed high levels of mRNA expression. Expression is depicted relative to the housekeeper GUSB. Discussion The CSC concept of cancer proposes that a tumor is generated by a small number of cells that possess the ability for indefinite self-renewal and differentiation into multiple cell types (7, 8). There is growing evidence in support of this concept, with CSCs being identified and characterized in many cancer types (17). The presence of CSCs in brain tumors was first reported by Singh et al. (6) and has been linked to tumor aggression and decreased life expectancy (8, 29). Furthermore, the presence of distinct ESC markers within GBM tumors has also been associated with poor outcomes (30–32). Activation of these protein markers in cancer cells imbues them with ESC characteristics such as indefinite self-renewal and pluripotency, and results in CSCs (17). We have demonstrated the expression of the ESC markers NANOG, SALL4, OCT4, SOX2, and pSTAT3 in GBM using DAB IHC staining (Figures 1A–F), WB (Figures 3A–D) and NanoString analysis (Figure 4). IHC staining showed the expression of all five ESC markers within the GBM samples examined (Figure 1), a finding that was corroborated by NanoString mRNA analysis in all six GBM samples. In this report, we have also demonstrated the expression of pSTAT3, NANOG, and SOX2 using WB analysis. Due to improperly functioning antibodies, the presence of SALL4 could not be determined by WB, and this warrants further investigation. It is intriguing that OCT4 was below detection levels by WB despite being observed by both IHC staining and NanoString analysis. As NanoString analysis showed relatively low transcript numbers for OCT4 (Figure 4) and DAB IHC staining showed OCT4 was expressed by very few cells within the tumor (Figure 1F), we infer that OCT4 was too low in abundance to be detected by WB. Possible reasons for this include sampling bias and/or relatively low levels of protein within the GBM tissues examined. This may in part be explained by the inherent intra-tumor heterogenity within GBM tumors, with previous studies demonstrating spectral expression patterns for both SOX2 and pSTAT3 including GBM (1, 33, 34). This is further supported by the well recognized inter-tumor heterogenity observed in GBM, which potentially provides each tumor with a unique stem cell signature (10, 17, 35, 36). This may also account for the variation of the number of detected bands for NANOG and SOX2 on WB (Figures 3C,D). Different-sized bands may represent detection of different protein isoforms that have undergone post-translational modification (PTM). Different sized variants of NANOG have been identified previously with native NANOG (NANOG a) at 34.2 kDa, NANOG b at 34.4 kDa, and NANOG c at 31.9 kDa (37). Furthermore, NANOG can undergo phosphorylation at several residues (38) and is also bound by the ubiquitin–proteasome system (39), both of which have the potential to alter band size on a WB. As PTM can alter protein function (39), the observation that three out of the four GBM samples analyzed contained only the 35 kDa variant (possibly corresponding to NANOG c with some PTMs, as the larger 40 kDa band in the positive 3T3 lysate control is likely to be NANOG a/b) indicates that the types of modified protein present within a tumor may be significant and that the presence of NANOG c in GBM may even be a predictor of tumor aggression. SOX2 can also undergo phosphorylation, SUMOylation, and glycosylation (40, 41). Similar to NANOG, SOX2 PTMs can upregulate or downregulate SOX2 function, thus influencing stem cell function (39, 41). It is therefore conceivable that multiple modified and/or isoforms of particular ESC markers may be present in any given tumor. This study demonstrates relative abundance of ESC markers OCT4, SOX2, SALL4, NANOG, and pSTAT3. Our data indicates the possibility that different isoforms or modified versions of SOX2 and NANOG may exist within different GBM tumors, and that a particular isoform present may influence tumor growth and tumor aggression in specific ways, although this remains a topic for further study. Additionally, in contrast to previous studies (15, 42), IHC staining and NanoString analysis in our study revealed relatively low expression levels of OCT4 at the transcriptional activation and corresponding protein levels. We speculate that the relatively low number of cells expressing OCT4 represent the most primitive stem cell population within GBM and that they may potentially give rise to the remaining down-stream cells within the GBM tumor. Similarly, the almost ubiquitous abundance of SOX2 within GBM suggests that this marker is expressed on the more differentiated cells reflecting SOX2 as a putative progenitor cell marker within the GBM samples used in this study. Further investigation in this area may lead to the possibility of tailoring future treatment of GBM by targeting the most primitive OCT4 + CSC subpopulation. Although a limitation of this study is the relatively small sample size, the novel findings we present lay a platform for future studies to better understand the precise role of CSCs in GBM. Ethics Approval Central Health and Disabilities Ethics Committee (ref. no. 15CEN28). Author Contributions TI and ST formulated the study hypothesis. TI, AW, and ST designed the study. TI, HB, AB, AW, PD, and ST interpreted the DAB IHC data. TI, AW, and ST interpreted the IF IHC data. AB performed WB analysis. AB, TI, AW, PD, and ST interpreted the WB data. AC processed the tissues for NanoString analysis and interpreted the data. AB drafted the manuscript. All authors commented on and approved the manuscript. Conflict of Interest Statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. TI, PD, and ST are inventors of the PCT patent application (No. PCT/NZ2015/050108) Cancer Diagnosis and Therapy. We thank Ms Liz Jones and Dr Jonathan Dunne of the Gillies McIndoe Research Institute for their assistance in IHC staining and advice regarding Western Blotting, respectively. Supplementary Material The Supplementary Material for this article can be found online at http://journal.frontiersin.org/article/10.3389/fsurg.2016.00048 Click here for additional data file. Click here for additional data file. ==== Refs References 1 Schmitz M Temme A Senner V Ebner R Schwind S Stevanovic S Identification of SOX2 as a novel glioma-associated antigen and potential target for T cell-based immunotherapy . Br J Cancer (2007 ) 96 (8 ):1293 –301 .10.1038/sj.bjc.6603802 17375044 2 Surawicz TS Davis F Freels S Laws ER Menck HR . Brain tumor survival: results from the National Cancer Data Base . J Neurooncol (1998 ) 40 (2 ):151 –60 .10.1023/a:1006091608586 9892097 3 Frosina G . Frontiers in targeting glioma stem cells . Eur J Cancer (2011 ) 47 (4 ):496 –507 .10.1016/j.ejca.2010.11.017 21185169 4 Stupp R Hegi ME Mason WP van den Bent MJ Taphoorn MJ Janzer RC Effects of radiotherapy with concomitant and adjuvant temozolomide versus radiotherapy alone on survival in glioblastoma in a randomised phase III study: 5-year analysis of the EORTC-NCIC trial . Lancet Oncol (2009 ) 10 (5 ):459 –66 .10.1016/S1470-2045(09)70025-7 19269895 5 Frosina G . The bright and the dark sides of DNA repair in stem cells . J Biomed Biotechnol (2010 ) 2010 :845396 .10.1155/2010/845396 20396397 6 Singh SK Clarke ID Terasaki M Bonn VE Hawkins C Squire J Identification of a cancer stem cell in human brain tumors . Cancer Res (2003 ) 63 (18 ):5821 –8 .14522905 7 Rajaraman R The origin of tumor cell heterogeneity . IUP J Biotechnol (2010 ) 4 (4 ):7 –43 . 8 Shipitsin M Polyak K . The cancer stem cell hypothesis: in search of definitions, markers, and relevance . Lab Invest (2008 ) 88 (5 ):459 –63 .10.1038/labinvest.2008.14 18379567 9 Adams JM Strasser A . Is tumor growth sustained by rare cancer stem cells or dominant clones? Cancer Res (2008 ) 68 (11 ):4018 –21 .10.1158/0008-5472.CAN-07-6334 18519656 10 Safa AR Saadatzadeh MR Cohen-Gadol AA Pollok KE Bijangi-Vishehsaraei K Glioblastoma stem cells (GSCs) epigenetic plasticity and interconversion between differentiated non-GSCs and GSCs . Genes Dis (2015 ) 2 (2 ):152 –63 .10.1016/j.gendis.2015.02.001 26137500 11 Evans MJ Kaufman MH Establishment in culture of pluripotential cells from mouse embryos . Nature (1981 ) 292 (5819 ):154 –6 .10.1038/292154a0 7242681 12 Thomson JA Itskovitz-Eldor J Shapiro SS Waknitz MA Swiergiel JJ Marshall VS Embryonic stem cell lines derived from human blastocysts . Science (1998 ) 282 (5391 ):1145 –7 .10.1126/science.282.5391.1145 9804556 13 Zhao W Ji X Zhang F Li L Ma L . Embryonic stem cell markers . Molecules (2012 ) 17 (6 ):6196 –236 .10.3390/molecules17066196 22634835 14 Schoenhals M Kassambara A De Vos J Hose D Moreaux J Klein B . Embryonic stem cell markers expression in cancers . Biochem Biophys Res Commun (2009 ) 383 (2 ):157 –62 .10.1016/j.bbrc.2009.02.156 19268426 15 Elsir T Edqvist PH Carlson J Ribom D Bergqvist M Ekman S A study of embryonic stem cell-related proteins in human astrocytomas: Identification of Nanog as a predictor of survival . Int J Cancer (2014 ) 134 (5 ):1123 –31 .10.1002/ijc.28441 24037901 16 Holmberg J He X Peredo I Orrego A Hesselager G Ericsson C Activation of neural and pluripotent stem cell signatures correlates with increased malignancy in human glioma . PLoS One (2011 ) 6 (3 ):e18454 .10.1371/journal.pone.0018454 21483788 17 Bradshaw A Wickremesekera A Tan ST Peng L Davis PF Itinteang T . Cancer stem cell hierarchy in glioblastoma multiforme . Front Surg (2016 ) 3 :21 .10.3389/fsurg.2016.00021 27148537 18 Boyer LA Lee TI Cole MF Johnstone SE Levine SS Zucker JP Core transcriptional regulatory circuitry in human embryonic stem cells . Cell (2005 ) 122 (6 ):947 –56 .10.1016/j.cell.2005.08.020 16153702 19 Guo Y Liu S Wang P Zhao S Wang F Bing L Expression profile of embryonic stem cell-associated genes Oct4, Sox2 and Nanog in human gliomas . Histopathology (2011 ) 59 (4 ):763 –75 .10.1111/j.1365-2559.2011.03993.x 22014056 20 Zhang J Tam W-L Tong GQ Wu Q Chan H-Y Soh B-S Sall4 modulates embryonic stem cell pluripotency and early embryonic development by the transcriptional regulation of Pou5f1 . Nat Cell Biol (2006 ) 8 (10 ):1114 –23 .10.1038/ncb1481 16980957 21 Lim CY Tam W-L Zhang J Ang HS Jia H Lipovich L Sall4 regulates distinct transcription circuitries in different blastocyst-derived stem cell lineages . Cell Stem Cell (2008 ) 3 (5 ):543 –54 .10.1016/j.stem.2008.08.004 18804426 22 Yang J Chai L Fowles TC Alipio Z Xu D Fink LM Genome-wide analysis reveals Sall4 to be a major regulator of pluripotency in murine-embryonic stem cells . Proc Nat Acad Sci U S A (2008 ) 105 (50 ):19756 –61 .10.1073/pnas.0809321105 19060217 23 Wu Q Chen X Zhang J Loh Y-H Low T-Y Zhang W Sall4 interacts with Nanog and co-occupies Nanog genomic sites in embryonic stem cells . J Biol Chem (2006 ) 281 (34 ):24090 –4 .10.1074/jbc.C600122200 16840789 24 Inghirami G Chiarle R Simmons WJ Piva R Schlessinger K Levy DE New and old functions of STAT3: A Pivitol target for individualized treatment of cancer . Cell Cycle (2005 ) 4 (9 ):1131 –3 .10.4161/cc.4.9.1985 16082218 25 Niwa H Burdon T Chambers I Smith A . Self-renewal of pluripotent embryonic stem cells is mediated via activation of STAT3 . Genes Dev (1998 ) 12 (13 ):2048 –60 .10.1101/gad.12.13.2048 9649508 26 Tan EM Chudakova DA Davis PF Brasch HD Itinteang T Tan ST . Characterisation of subpopulations of myeloid cells in infantile haemangioma . J Clin Pathol (2015 ) 68 (7 ):571 –4 .10.1136/jclinpath-2014-202846 25834091 27 Jung C Foerch C Schänzer A Heck A Plate K Seifert V Serum GFAP is a diagnostic marker for glioblastoma multiforme . Brain (2007 ) 130 (12 ):3336 –41 .10.1093/brain/awm263 17998256 28 Kriegstein A Alvarez-Buylla A . The glial nature of embryonic and adult neural stem cells . Annu Rev Neurosci (2009 ) 32 :149 –84 .10.1146/annurev.neuro.051508.135600 19555289 29 Pallini R Ricci-Vitiani L Banna GL Signore M Lombardi D Todaro M Cancer stem cell analysis and clinical outcome in patients with glioblastoma multiforme . Clin Cancer Res (2008 ) 14 (24 ):8205 –12 .10.1158/1078-0432.CCR-08-0644 19088037 30 Zhang L Yan Y Jiang Y Cui Y Zou Y Qian J The expression of SALL4 in patients with gliomas: high level of SALL4 expression is correlated with poor outcome . J Neurooncol (2015 ) 121 (2 ):261 –8 .10.1007/s11060-014-1646-4 25359397 31 Zeppernick F Ahmadi R Campos B Dictus C Helmke BM Becker N Stem cell marker CD133 affects clinical outcome in glioma patients . Clin Cancer Res (2008 ) 14 (1 ):123 –9 .10.1158/1078-0432.CCR-07-0932 18172261 32 Vescovi AL Galli R Reynolds BA . Brain tumour stem cells . Nat Rev Cancer (2006 ) 6 (6 ):425 –36 .10.1038/nrc1889 16723989 33 Annovazzi L Mellai M Caldera V Valente G Schiffer D . SOX2 expression and amplification in gliomas and glioma cell lines . Cancer Genomics Proteomics (2011 ) 8 (3 ):139 –47 .21518820 34 Mizoguchi M Betensky RA Batchelor TT Bernay DC Louis DN Nutt CL . Activation of STAT3, MAPK, and AKT in malignant astrocytic gliomas: correlation with EGFR status, tumor grade, and survival . J Neuropathol Exp Neurol (2006 ) 65 (12 ):1181 –8 .10.1097/01.jnen.0000248549.14962.b2 17146292 35 Stieber D Golebiewska A Evers L Lenkiewicz E Brons NH Nicot N Glioblastomas are composed of genetically divergent clones with distinct tumourigenic potential and variable stem cell-associated phenotypes . Acta Neuropathol (2014 ) 127 (2 ):203 –19 .10.1007/s00401-013-1196-4 24154962 36 Morokoff A Ng W Gogos A Kaye A . Molecular subtypes, stem cells and heterogeneity: implications for personalised therapy in glioma . J Clin Neurosci (2015 ) 22 (8 ):1219 –26 .10.1016/j.jocn.2015.02.008 25957782 37 Das S Jena S Levasseur DN . Alternative splicing produces Nanog protein variants with different capacities for self-renewal and pluripotency in embryonic stem cells . J Biol Chem (2011 ) 286 (49 ):42690 –703 .10.1074/jbc.M111.290189 21969378 38 Xie X Piao L Cavey GS Old M Teknos TN Mapp AK Phosphorylation of Nanog is essential to regulate Bmi1 and promote tumorigenesis . Oncogene (2014 ) 33 (16 ):2040 –52 .10.1038/onc.2013.173 23708658 39 Cai N Li M Qu J Liu GH Izpisua Belmonte JC . Post-translational modulation of pluripotency . J Mol Cell Biol (2012 ) 4 (4 ):262 –5 .10.1093/jmcb/mjs031 22679102 40 Van Hoof D Munoz J Braam SR Pinkse MW Linding R Heck AJ Phosphorylation dynamics during early differentiation of human embryonic stem cells . Cell Stem Cell (2009 ) 5 (2 ):214 –26 .10.1016/j.stem.2009.05.021 19664995 41 Jang H Kim TW Yoon S Choi SY Kang TW Kim SY O-GlcNAc regulates pluripotency and reprogramming by directly acting on core components of the pluripotency network . Cell Stem Cell (2012 ) 11 (1 ):62 –74 .10.1016/j.stem.2012.03.001 22608532 42 Schiffer D Annovazzi L Cassoni P Valentini MC Mazzucco M Mellai M Glioblastoma stem cells: Conversion or reprogramming from tumor non-stem cells? J Stem Cell Res Ther (2015 ) 5 :315 10.4172/2157-7633.1000315

PMC005xxxxxx/PMC5001192.txt

==== Front Fisc StudFisc Stud10.1111/(ISSN)1475-5890FISCFiscal Studies0143-56711475-5890John Wiley and Sons Inc. Hoboken 10.1111/j.1475-5890.2016.12084FISC12084Original ArticleOriginal ArticlesThe Challenge of Measuring UK Wealth Inequality in the 2000s† The challenge of measuring UK wealth inequality in the 2000sCrossley Thomas F. O'Dea Cormac Alvaredo Facundo alvaredo@pse.ens.fr 1 Atkinson Anthony B. tony.atkinson@nuffield.ox.ac.uk 2 Morelli Salvatore salvatore.morelli@unina.it 3 1 Paris School of Economics; INET at the Oxford Martin School; Conicet2 Nuffield College; London School of Economics; INET at the Oxford Martin School3 CSEF – University of Naples ‘Federico II’; INET at the Oxford Martin School31 3 2016 3 2016 37 1 10.1111/fisc.2016.37.issue-1Household Wealth Data and Public Policy13 33 © 2016 The Authors. Fiscal Studies published by John Wiley & Sons Ltd. on behalf of Institute for Fiscal StudiesThis is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.Abstract The concentration of personal wealth is now receiving a great deal of attention – after having been neglected for many years. One reason is the growing recognition that, in seeking explanations for rising income inequality, we need to look not only at wages and earned income but also at income from capital, particularly at the top of the distribution. In this paper, we use evidence from existing data sources to attempt to answer three questions: (i) What is the share of total personal wealth that is owned by the top 1 per cent, or the top 0.1 per cent? (ii) Is wealth much more unequally distributed than income? (iii) Is the concentration of wealth at the top increasing over time? The main conclusion of the paper is that the evidence about the UK concentration of wealth post‐2000 is seriously incomplete and significant investment in a variety of sources is necessary if we are to provide satisfactory answers to the three questions. wealth inequalityestate multiplierinvestment incomeD3H2 source-schema-version-number2.0component-idfisc12084cover-dateMarch 2016details-of-publishers-convertorConverter:WILEY_ML3GV2_TO_NLMPMC version:4.9.4 mode:remove_FC converted:25.08.2016† Submitted August 2015. The copyright line for this article was changed on 25th May after original online publication. This paper is an outgrowth of the larger project ‘Top Wealth Shares in the UK since 1896’ by the same authors. They thank for helpful comments Daniel Heymann, Thomas Piketty, Gabriel Zucman, the co‐editor Thomas Crossley, an anonymous referee and participants at the Household Wealth Data and Public Policy Conference (Bank of England, March 2015), the 16th Trento Summer School (June 2015) and the workshop on measuring inequalities of income and wealth (Berlin, September 2015). The authors are particularly grateful to Alan Newman (Office for National Statistics, ONS), who shared the Wealth and Assets Survey (WAS)‐based results submitted to the OECD Wealth Database, and to Sian‐Elin Wyatt (ONS), who addressed their inquiries about the WAS response rates. This research received financial support from the Institute for New Economic Thinking (INET), the European Research Council (ERC) and the Economic and Social Research Council / Department for International Development (ESRC‐DFID) joint fund (grant ES/I033114/1). The usual disclaimer applies. ==== Body Policy points The evidence about the UK distribution of wealth post‐2000 is seriously incomplete. Significant investment in statistics is necessary if we are to be able to draw firm conclusions about the extent of wealth concentration. I. Wealth inequality under the spotlight The distribution of personal wealth is now receiving a great deal of attention – after having been neglected for many years. One reason is the growing recognition that, in seeking explanations for rising income inequality, we need to look not only at wages and earned income but also at income from capital. Income from interest, from dividends and from rents represents a minority of total personal income, but it is nonetheless a significant part, in particular, at the top of the distribution. Moreover, viewed from the side of National Accounts, the share of income from capital and rents has been increasing in recent decades. In many OECD countries, the ratio of total personal wealth to total personal income has been rising. One consequence is that the role of inherited wealth – declining for much of the 20th century – has, in a number of countries, begun to acquire greater significance. The recent attention to the distribution of wealth has led to evidence being sought on several key questions. The first is the extent of concentration of wealth at the top. What is the share of total personal wealth that is owned by the top 1 per cent, or by even smaller groups such as the top 0.1 per cent? The second question is whether wealth is much more unequally distributed than income. The OECD report, In It Together, stresses that ‘household wealth – in particular financial assets – is much more unequally distributed than income’.1 The third key question is whether wealth inequality is increasing over time. The OECD report says of the UK that ‘the financial crisis has exacerbated the concentration of wealth at the top’.2 How far is wealth inequality increasing? In this paper, we examine the evidence on these three questions for the United Kingdom, focusing on the period from 2000 onwards. The first prerequisite is to consider the range of sources of evidence about wealth‐holding. There has been disagreement in the UK literature about the level and trend in the distribution of wealth, and this disagreement stems in part from the use of different sources. Section I. summarises the main ‘windows’ through which we can observe the distribution of wealth in the UK, drawing attention to their strengths and weaknesses. The second prerequisite is to clarify definitions. There is no such thing as ‘the’ distribution of wealth. A figure for the share of the top 1 per cent could relate to the top 1 per cent of households, or of families, or of individuals. The share could relate to wealth excluding or including pension rights; the pension rights could include state pensions or be limited to private pensions. The 1 per cent could be limited to residents or could include those non‐domiciled. Section 5. of the paper sets out some of the key definitional issues. Having cleared the ground, we examine in Section 3. the light that existing evidence casts on the answers to the three key questions posed in the previous paragraph. This examination leads us to identify important ways in which there needs to be investment in improving the informational base about wealth‐holding in the UK. The main conclusions are summarised in Section 3.. II. Different windows on wealth There are four main potential sources of evidence about the distribution of personal wealth in the UK: administrative (tax) data on estates at death, which indirectly provide evidence about the wealth of the living, by applying (the inverse of) mortality multipliers differentiated by age, sex and wealth class; administrative (tax) data on investment income, which indirectly provide evidence about the wealth of the living, by applying yield multipliers; household surveys of personal wealth, such as the Wealth and Assets Survey (WAS) conducted by the Office for National Statistics (ONS); lists of large wealth‐holders, such as the Sunday Times ‘Rich List’, which has been compiled by Philip Beresford in the UK, and the Forbes List of Billionaires.3 There are in addition synthetic estimates that draw on two or more sources, such as those of Credit Suisse Research Institute (2014), which combines household survey data from the WAS with the number of Forbes billionaires, and Vermeulen (2015), who combines extreme observations on the number of billionaires as well as their wealth from the Forbes List with the WAS data. In all cases, the evidence about the distribution of wealth has to be considered in relation to the external control totals for population, based on demographic data, and for total personal wealth. Each of the sources is considered below, where we summarise the methods and their main strengths and weaknesses. 1. The multiplied estate data Estimates of the distribution of wealth based on administrative data from the taxation of the estates of those dying in a particular year are reached by applying the estate multiplier method. Estimates of the estate distribution are first obtained from a sample of the estates submitting an inheritance tax return.4 Subsequently, the method considers the grossed‐up population of decedents as a sample of the living population. The death rate, however, is clearly not random, as it varies substantially across age, gender, and social or wealth class, etc. One can nonetheless define death as ‘random’ within each specific age, gender, marital status, social or wealth class cell, and take each cell‐specific mortality rate as the ‘sampling rate’. Their inverse (‘estate multiplier’) can then be used to re‐weight the observations for decedents in order to obtain the distribution for the living population. Additional adjustments have to be made in order to control for individuals not covered by the estate tax statistics and total assets not represented in the estate data. These data on estates at death have long been used for economic research in the UK. Initially, they were employed to make estimates of total personal wealth. Baxter (1869) estimated the total wealth on the basis of the Probate Duty data, applying a multiplier of 30, which he took to be the cycle for each devolution of property. In terms of the distribution, such a single multiplier, referred to as a ‘unity multiplier’, means that this method yields estimates of the distribution of estates, not of wealth. It takes no account of the differential rates of death by wealth class. Following the proposal of Coghlan (in the discussion of Harris and Lake (1906)), Mallet (1908) applied to each estate a multiplier related to age at death. These ‘general mortality multipliers’ were subsequently refined in Mallet and Strutt (1915) to apply ‘social class multipliers’ allowing for the lower mortality of the upper and middle classes. Differentiation was also made later on the basis of gender. For many years, this has formed the basis for estate‐based estimates of the distribution of wealth in the UK. Applying multipliers yields estimates of the number of individuals owning wealth in particular ranges and the amounts of their wealth. The next step is to relate the numbers and amounts to external control totals. In the latter case, the totals come from elements in the national balance sheets. This method was developed in Atkinson and Harrison (1978) and by the Inland Revenue in its (revised) Series C introduced to cover the period from 1976. Series C was published on an annual basis until 2005. A new methodology has since been introduced by HM Revenue & Customs (which has replaced the Inland Revenue), with estimates being produced for three‐year averages 2001–03, 2005–07 and 2008–10.5 The details of the change are discussed further in Section 3..1. The estate‐based estimates have evident shortcomings. The first, and most obvious, is that the estate data as such do not cover the rights to occupational or state pensions.6 The second, equally obvious, is that the degree of concentration of wealth is likely to be understated on account of tax avoidance and evasion. Estate planning is certainly an effective way to reduce tax liabilities at death. In the UK, for instance, assets given away at least seven years before death are not subject to estate taxation. In statistical terms, this problem is mitigated by the fact that the recipients are also subject to the risk of dying, and their multiplied‐up wealth appears in the estimated wealth distribution. However, the donors are likely to be unrepresentative of their class (being less healthy) so that the mitigation is only partial.7 A second source of avoidance is provided by trusts (mainly discretionary trusts). Although the official Series C attempted to make allowance for excluded wealth in trusts, these adjustments were based on limited and increasingly dated information. Moreover, the validity of the estate multiplier method depends on the estate multipliers. The HMRC wealth model approximates the mortality risk of wealthy individuals with that of individuals who are owner‐occupiers taken from the ONS Longitudinal Study of social class and occupational mobility. This model was recently updated using the English Longitudinal Study of Ageing (ELSA) to better capture the relationship between housing wealth and mortality.8 Similarly, Kopczuk and Saez (2004) use the mortality of US college‐educated individuals as a proxy for that of wealthy individuals. However, Saez and Zucman (2014) report that this may not be a good approximation of mortality for wealthier people (above the 90th percentile of the wealth distribution), whose mortality rate is considerably lower. They argue that this mortality gap has been increasing over time in the US, biasing downward the evolution of the estate‐based wealth shares. The estate multiplier approach would clearly benefit from a fresh systematic investigation of mortality risk within the population according to social classes and income and wealth levels. 2. The multiplied investment income data The investment income method has been applied to the UK by Atkinson and Harrison (1974 and 1978, ch. 7), building on the work of Barna (1945) and Stark (1972). The underlying method has been described by Saez and Zucman (2014, p. 1) in their recent paper on the US as follows: Starting with the capital income reported by individuals to the Internal Revenue Service—which is broken down into many categories: dividends, interest, rents, profits, mortgage payments, etc.—for each asset class we compute a capitalization factor that maps the total flow of tax income to the total amount of wealth recorded in the Flow of Funds. We then combine individual incomes and aggregate capitalization factors by assuming that within a given asset class the capitalization factor is the same for everybody. For example, if the ratio of Flow of Funds fixed income claims to tax reported interest income is 50, then $50,000 in fixed income claims is attributed to an individual reporting $1,000 in interest. They use the income tax data and the national balance sheets (Flow of Funds). This may be contrasted with the ‘hybrid’ investment income method used by Atkinson and Harrison, where the yields are taken from external sources and weighted using asset composition data from the estate‐based wealth estimates, leading to a single capitalisation factor applied to total investment income reported in the Survey of Personal Incomes (SPI), based on income tax returns. The adoption of this hybrid approach reflected the fact that the income data in the UK were only tabulated according to broad categories.9 It is also the case that, where the estate‐based asset composition data include assets that do not generate income taxable under the income tax, these can be allowed for in calculating the overall multiplier. Such assets include owner‐occupied houses, non‐interest‐bearing bank accounts, non‐taxable fixed‐income claims, durable goods and collectibles. In contrast, Saez and Zucman make allowance for such assets making use of additional data sources: surveys, property tax records, etc; and they are also able to attach estimates of funded pension wealth. The theoretical basis for the investment income method and the potential bias in the estimation of wealth inequality are set out in Atkinson and Harrison (1978, ch. 7 and appendix VII), where two main sources of error are identified: the variation with the level of wealth of the rates of return to individual asset types, and the variation in the rate of return for a given asset and wealth level (idiosyncratic returns). The US estimates of Saez and Zucman (2014) represent an advance in that they employ data from foundations to demonstrate that returns are flat within asset classes (the overall yield rises with wealth on account of asset composition). In the case of the second source of error, the authors argue from an illustrative calculation that ‘idiosyncratic returns cannot create much bias’ (p. 16). The discussion in Atkinson and Harrison (1978) is more cautious, concluding that the upward bias in the measurement of wealth inequality ‘is large enough to be taken seriously but not sufficient to discredit the investment income method’ (p. 199). The investment income method has considerable advantages in that the underlying data relate to the living population and the method does not depend on assumptions about the differential mortality rates by wealth classes. Estimates employing the hybrid investment income method were made by Atkinson and Harrison (1978) for 1968–69 and 1972–73. Today, however, it does not seem possible to satisfactorily apply the method using the currently available data. The SPI micro data available to public users only provide four variables (aggregating many different types of capital incomes): (i) dividends; (ii) income from property; (iii) net interest from UK banks, building societies and other deposit takers; and (iv) other investment income. In order to apply the full investment income method, a more detailed version of the SPI micro data would be necessary. The information contained in the internal SPI looks more promising, but at the moment of writing we have not yet obtained effective access to the micro data.10 The application of the hybrid method, as in Atkinson and Harrison (1978), could be contemplated, but this requires a detailed breakdown of wealth by asset types and wealth ranges. The published information for years 2000 onwards11 only gives six categories of assets and two of liabilities. Again, to apply the investment income method, more detailed information is required. In our view, the investment income method should certainly be explored further, but for this it is necessary that the underlying data be available in a more detailed form. 3. Household surveys Household surveys are a quite different source of data, unaffected by problems of tax avoidance and tax evasion because unrelated to the operation of the tax system, and able to furnish information about pension entitlements. These surveys date back in the UK to the Oxford Savings Surveys in the 1950s.12 In the 1970s, the Royal Commission on the Distribution of Income and Wealth investigated the possible role of sample surveys of wealth‐holding, commissioning two small pilot surveys, but concluded that the results, ‘notably the particularly low response rate (around 50 per cent)’, did not justify the launching of a full‐scale survey.13 More recently, attitudes towards household surveys have changed. The British Household Panel Survey (BHPS) began collecting data on financial wealth in 1995. In 2000, the Office for National Statistics began to plan the longitudinal Wealth and Assets Survey (WAS), which was launched in 2006, funded by a consortium which also included (in 2012) the Department for Work and Pensions, HMRC, the Financial Conduct Authority and the Scottish Government.14 The first WAS spanned the period 2006–08, and subsequent waves have covered 2008–10 (Wave 2), 2010–12 (Wave 3) and 2012–14 (Wave 4, full results not yet published), covering only Great Britain. Does the renewed interest in household survey data on wealth reflect a resolution of the problem of low response rates? This does not appear to be the case. Wave 1 of WAS in 2006–08 achieved a response rate of 54.6 per cent – similar to that found in the 1970s. Since the WAS is a longitudinal survey, the calculation of the combined response rate over successive waves is not straightforward, as the ONS attempts to re‐contact previous wave non‐contacts and movers (household splitting) between waves. Figure 1 shows the absolute number of households eligible at each stage and the number cooperating. Waves 2 and 3 achieved higher response rates among those eligible, but this still left a final total of only 15,517 households, compared with an initial eligible sample of 55,835 in Wave 1. Wave 3 included a new ‘booster’ sample, with a response rate of 50.8 per cent. Figure 1 Eligible and cooperating households in each WAS wave Source: Data provided by Office for National Statistics. A low rate of response does not necessarily imply that the results on wealth shares are biased. On the other hand, there are a priori reasons to expect there to be differential non‐response by wealth classes. The feasibility studies in the 1970s found that ‘the indications were that non‐response would be higher among those groups with higher incomes and substantial investment income’.15 In order to mitigate this effect, the WAS made use of information available from the income tax records to flag addresses where at least one person was likely to have total financial wealth above a certain threshold, and these flagged addresses had a higher (two‐and‐a‐half or three times) chance of selection.16 However, the evidence gathered in Vermeulen (2015) suggests that the oversampling strategy has not been very effective. There were also problems of incomplete response. In the case of business assets, ‘a high percentage of those who said they held business assets failed to provide an estimate of the value of such assets’.17 This led to business assets being excluded from the estimates of total wealth. This omission is likely to be particularly important in the upper wealth ranges. The issues of non‐response and under‐reporting at the top mean, in our view, that the Wealth and Assets Survey – valuable as it is in covering the majority of the population – cannot, on its own, provide a fully satisfactory representation of the upper tail of the UK wealth distribution. 4. The Rich Lists Since 1989, the Sunday Times has published annually in April a ‘Rich List’ of the wealthiest people or families in Great Britain. The lists, compiled by Philip Beresford, appear as a supplement to the newspaper, and on occasion in extended book form.18 The methods used in constructing the lists are set out in ‘Rules of engagement’.19 The description emphasises that the estimates are ‘the minimum wealth … the actual size of their fortunes may be much larger’. The construction of the list draws on a wide range of public information, coming from a variety of sources. The estimates relate to identifiable wealth, such as land, property, art, or significant shares in publicly‐quoted companies, and in recent years have paid particular attention to liabilities (for example, where shares are used as collateral for loans). UK top wealth‐holders are also included in the global Forbes List of (Dollar) Billionaires, published annually by the business magazine since 1987. The list is compiled by reporters who ‘meet with the list candidates and their handlers and interview employees, rivals, attorneys and securities analysts. … We do attempt to vet these numbers with all billionaires. Some cooperate, others don't’.20 Nonetheless, it is not easy to validate the information. In summary, the Rich Lists provide valuable insight into the upper tail of the wealth distribution, but it is not easy to assess their representativeness. 5. Total personal wealth The shares of top wealth‐holders depend on the control total for personal wealth. In the early part of the period with which we are concerned here, HMRC provided a reconciliation of the wealth totals that is of central importance in understanding the estimates of wealth shares.21 This is particularly useful to control for the individuals as well as the total assets not represented within the estate data (‘excluded wealth’). The reconciliation begins with the total net wealth identified in the multiplied‐up estate data (‘identified wealth’), which was £3,432 billion in 2005, as shown in the online appendix. The first stage involves adjustment for under‐recording and differences in valuation in the estate data (for example, replacing the maturity value of a life assurance policy by its equity value). This increases the total in 2005 to £4,097 billion. To this is added the estimated value of the so‐called ‘excluded wealth’ – namely, that wealth not subject to estate taxation as well as the wealth of those not covered by the estate data. The so‐called ‘excluded wealth’ includes estimates of joint properties, small properties and trusts. The resulting total is £5,005 billion in 2005, and this is defined as ‘Series C marketable wealth’. The total is 46 per cent higher than ‘identified wealth’. The Series C total marketable wealth may be compared with the total sector (S.14 and S.15 combined) wealth in the National Accounts balance sheets. There are significant definitional differences. The first is that the National Accounts combine households with non‐profit institutions serving households (NPISH); the second is that the National Accounts balance sheets are defined on an end‐of‐year basis. The most important difference, however, is the inclusion in the National Accounts total of the value of funded pension rights (£1,213 billion in 2005). The aggregate value of all pension rights, funded and unfunded, occupational and state, is given as £2,999 billion in 2005. It is evident that the adjustments to the estate data, and whether or not pension wealth is included, make a significant difference to the control totals employed. III. Inequality of what among whom? The paper is concerned with the distribution of personal wealth, by which we mean the value of the total assets owned (directly or indirectly) by individuals, net of their debts. Assets include financial assets, such as bank accounts, stocks or bonds, and real assets, such as houses, business assets and consumer durables. As defined here, it does not include ‘human capital’ (the capitalised value of future earnings). The implementation of this concept does, however, raise a number of definitional issues, and these are resolved in different ways in different sources of evidence. 1. Geographical scope First, there is the geographical scope. The estimates discussed here relate either to the United Kingdom (tax‐based estimates) or to Great Britain (the WAS household survey), the latter excluding Northern Ireland. Northern Ireland accounts for 2.8 per cent of the UK total resident population. However, the Sunday Times Rich List has a different approach. It includes people who live and work in Britain, and people who are married to Britons, who have strong links with Britain, who have estates and other assets there, or who have backed British political parties, British institutions and British charities. It includes British citizens abroad. The population represented is therefore more extensive than that in the estate‐based estimates, or investment income data, or the WAS household surveys. 2. Unit of analysis The unit of analysis in the case of the estate‐based estimates and the investment‐income‐based estimates is the individual.22 Estates are naturally recorded on the death of an individual. Since 1990, the income tax has been levied on an individual basis, and hence the investment income data take this form. In contrast, the WAS survey data relate to the total wealth of the household, defined as a person or a group of people (family members and non‐relatives) living together in the same dwelling.23 In the case of the Rich Lists, the unit may be more extensive than the household. For example, in the 2014 Sunday Times list, the top entry was the Hinduja brothers; third was Lakshmi Mittal and family, which includes his son and daughter; the wealth of number 11 includes that of Galen Weston, his wife and his nephew, George Weston. There are often multiple generations, such as number 19 (Earl Cadogan and his son, Viscount Chelsea). What difference does the unit of analysis make to the estimated wealth shares? How can we compare the estate‐based estimates of individual wealth with the household wealth estimates in the WAS? If we treat all units as weighted equally (so no account is taken of household size), then the control total for households is smaller than that for individuals (by a factor 1/h, less than 1, where h is the average number of adults per household). In 2010 in the UK, the value of h is close to 2, and we take that value in the illustrative examples below. The impact of moving from an individual to a household basis depends on the joint distribution of wealth. Suppose first that in the top 1 per cent of individuals each person is married to someone with equal wealth. They then constitute the top 1 per cent of households (since h = 2) and have the same share of total wealth. On the other hand, to the extent that the top 1 per cent marry out of that group, the household‐based share of total wealth is reduced. Similarly, if the top 1 per cent of individuals are all single, then they account for 2 per cent of total households, and the share of the top 1 per cent is reduced, compared with that measured on an individual basis. The calculations in Atkinson and Harrison (1978, p. 248) suggest that, in the limiting cases of all single or of rich married to poor, the share of the top 1 per cent could be reduced by 4 to 5 percentage points when moving from the individual to the household distribution. In practice, the household‐based estimates are likely to be lower but by less than this amount. 3. Method of valuation A third set of definitional issues concerns the method of valuation, a topic that is often taken for granted. As the ONS says of national balance sheets,24 the wealth figures are taken to represent the ‘market value of the financial and non‐financial assets’, but the application of the market value approach raises a number of issues. Life assurance policies provide an illustration. This asset changes value on death: the maturity value recorded in the estate exceeds the value to the person alive. For this reason, HMRC in its Series C made adjustments. But the market value, in terms of what the policy would fetch if surrendered, falls short of the continuing value to the person. Atkinson and Harrison (1978, p. 5) distinguish between ‘realisation’ and ‘going concern’ valuations. Interpreted as what a person could realise by the sale of all assets, net of liabilities, the former coincides in principle – with exceptions such as life policies – with the value placed on an estate at death. The going‐concern valuation, however, could well be considerably higher. That there can be a significant difference may be seen from the example of household contents (durables, furniture, etc.), where the price obtained on sale is likely to fall considerably short of the value to a continuing household (or the replacement cost). A less common, but important, example of differences between realisation and going‐concern valuations is that of family businesses. Finally, there is the case of pension rights, where the realisation value may be zero, but they are of considerable value to a living person. The standard approach to handling these differences is by the exclusion or inclusion of classes of assets. The current HMRC estate‐based estimates exclude pension rights (private and state). The WAS estimates both include and exclude pension rights. The WAS estimates also exclude business assets. However, it seems preferable to adopt explicitly either a realisation or a going‐concern basis. IV. Wealth shares in the UK since 2000 We discuss in turn the different sources of evidence about the distribution of wealth in the UK and the conclusions that can be drawn about the three questions posed at the start of the paper. As noted above, no results are given using the investment income method, since we do not yet have access to the necessary data. 1. Estate‐data‐based estimates We begin with HMRC Series C, which covers the years from 2000 to 2005 (excluding 2004). The shares of the top 10 per cent, 5 per cent and 1 per cent are shown in Figures 2 and 3 and Table 1. The published data also include the share of the top 50 per cent, top 25 per cent and top 2 per cent; they do not break down the top 1 per cent. Figure 2 Estimates of the top 10 per cent wealth shares since 2000 Note: The WAS estimates relate to households and to Great Britain. Figure 3 Estimates of the top 1 per cent wealth shares since 2000 Note: The WAS estimates relate to households and to Great Britain. Table 1 Estimates of top wealth shares Per cent HMRC Series C Derived from HMRC new series ONS WAS for OECD ONS WAS for OECD incl. pension wealth Credit Suisse Vermeulen Top 10% Top 5% Top 1% Top 10% Top 5% Top 1% Top 10% Top 5% Top 1% Top 10% Top 5% Top 1% Top 10% Top 1% Top 5%, lower bound Top 5%, upper bound Top 1%, lower bound Top 1%, upper bound 2000 56.0 44.0 23.0 51.5 20.5 2001 54.0 41.0 22.0 51.6 20.5 2002 54.0 41.0 21.0 50.4 37.5 17.9 51.6 20.6 2003 53.0 40.0 19.0 51.7 20.7 2004 51.7 20.8 2005 54.0 40.0 21.0 51.9 20.8 2006 51.5 38.2 19.7 51.9 20.9 2007 42.1 29.4 12.2 38.9 26.3 10.0 52.0 21.0 2008 52.1 21.1 2009 53.7 40.1 20.4 43.4 30.8 14.0 40.0 26.9 10.6 52.4 21.4 31.0 35.0 14.0 18.0 2010 52.8 21.8 2011 46.6 34.2 17.5 40.3 28.0 12.7 53.1 22.2 2012 53.5 22.6 2013 53.6 22.8 2014 54.1 23.3 Source and Note: HMRC Series C – table 13.5 on the HMRC website at http://webarchive.nationalarchives.gov.uk/20120403124426/ http://www.hmrc.gov.uk/stats/personal_wealth/13‐5‐table‐2005.pdf. Derived from HMRC new series – estimated from table 13.1 on the HMRC website at https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/447352/table_13‐1.pdf. ONS WAS for OECD – estimates provided by the ONS to the OECD Wealth Database. Credit Suisse – Credit Suisse Research Institute, 2014. Vermeulen – Vermeulen, 2015. Vermeulen's top wealth shares derived uniquely from the WAS (before any combination with the Forbes Rich List) are 30 per cent for the top 5 per cent and 13 per cent for the top 1 per cent. John Wiley & Sons, Ltd.Series C ended in 2005. HM Revenue & Customs (2012) explains the main changes in methodology in the new estimates, replacing Series C. The most important are the move to producing estimates based on data averaged over three years, the most recent being 2008–10, in order to reduce sampling variation, and the adoption of new multipliers based on the variation of mortality with housing wealth.25 On the other hand, HMRC has dropped the adjustments made in Series C for the excluded wealth, for valuation and to a balance‐sheet basis, focusing exclusively on the identified wealth. The link with the National Accounts balance sheets has been broken, making very difficult the estimation of the total personal wealth for these years. HMRC's reasons for dropping these adjustments are described as follows: These adjustments were not based on robust data, and used operational adjustments or assumptions instead. We do not know how accurate these adjustments are or if they should be changing over time. The data on Adjusted and Marketable Wealth is sensitive to these assumptions and so it was decided that this data was not robust enough for us to continue to publish it.26 These concerns are understandable. For example, the Board of Inland Revenue (2000) describes how the estimate of excluded wealth in trusts was based on studies for two years (1976 and 1988) which were by then distant in time. Although small in total, the addition would be largely allocated to the upper wealth groups. A significant investment would no doubt have been required to bring the estimate up to date. It is, however, regrettable that such an investment, and investments in other elements, were not given priority and that, as a result, the estimates of the wealth distribution are now less complete. The ‘new HMRC estimates’27 show the numbers and total wealth of individuals by ranges of net unadjusted wealth. In particular, the estimated wealth is not corrected for potential under‐reporting and undervaluation as was done for Series C data. The same data are also presented in the form of decile shares, but these are of little interest since they relate only to those identified as wealth‐holders (in 2008–10, only 31 per cent of the total population aged 18 and over) and only to identified wealth. In order to render the estimates closer to those for earlier years, we have made two adjustments. First, we have expressed the numbers as a percentage of the total population aged 18 and over. Second, we have taken as the control total for wealth the sum of identified wealth plus excluded wealth as estimated by HMRC,28 where this includes an estimate of the wealth of the excluded population. This procedure is applied up to 2005, the last year for which the HMRC reconciliation exercise has been published, and for subsequent years is extrapolated in line with total personal wealth as estimated in the UK balance sheet.29 The results are the estimates ‘derived from HMRC new series’ shown in Table 1. Comparing the estimates in Figures 2 and 3 for the overlapping period 2001–03 (with data points for the new ones plotted at 2002), we can see that the new HMRC estimates are, as we would expect, lower than the earlier Series C estimates. The share of the top 1 per cent is 3.1 percentage points lower. It should be noted that there is a considerable margin of error around our estimated control total. Series C, indeed, cannot be directly compared to the new assembled series from HMRC due to the lack of adjustments for wealth valuations. What do the estate‐based estimates tell us about the three questions with which we began? First, they indicate that the distribution of wealth in the UK is highly concentrated. The top 1 per cent own between one‐fifth and one‐quarter of total personal wealth. For instance, adding 3.1 percentage points to the estimate for 2008–10 gives a figure roughly comparable to those for Series C of 23.5 per cent for the share of the top 1 per cent. Second, if we take the estimates in Table 1, then the share of the top 1 per cent in total net worth (of individuals) is around double the share of the top 1 per cent (again, of individuals) in total net income (income after deducting income tax), which in the first half of the 2000s was around 10 per cent.30 The share of the top 10 per cent in total wealth was at that time about 50 per cent higher than the share of the top 10 per cent in total net income. Of course, the top x per cent of wealth‐holders are not necessarily the same people as the top x per cent of income‐recipients. Third, there is some indication that the top shares in wealth were increasing between 2001–03 and 2008–10, but this may depend on the estimation of the wealth control total, which is now subject to higher uncertainty as explained above. We would therefore be cautious about drawing any firm conclusion in view of the need for a more robustly established control total for wealth. 2. Household‐survey‐based estimates The introduction of the Wealth and Assets Survey provides a new and independent source of evidence about the distribution of wealth in Great Britain. Figures 2 and 3 show the estimated shares of the top 10 and 1 per cent as supplied to the OECD by the ONS for the three periods covered (in each case shown at the mid‐point year – for example, 2007 for 2006–08). These shares relate to household wealth (each household weighted as 1), and are shown including and excluding pension rights (the latter data for 2010–12 are included in the OECD Wealth Distribution Database, labelled 2012). The first finding is that, in the case of the overlapping period 2008–10, the WAS estimates excluding pension wealth suggest a share of the top 10 or top 1 per cent that is considerably below the estate‐based estimates: 43 per cent for the top 10 per cent, compared with 54 per cent, and 14 per cent for the top 1 per cent, compared with 20 per cent. We have to take account of the fact that these estimates are household‐based and that the geographic coverage differs, but the difference is larger than could be explained in this way. Moreover, if pension wealth is included, then the gap is even wider. The share of the top 1 per cent is only 11 per cent, or virtually half that found in the estate‐based estimates. If the share of the top 1 per cent were as low as 11 per cent, then we would have to revisit the conclusion that wealth is much more unequally distributed than income: the share of the top 1 per cent in after‐tax income in 2009–10 averaged 10.7 per cent. These estimates for Great Britain are compared by the OECD with estimates for other countries based on sample surveys. It is interesting to begin with an earlier such comparison: that between Great Britain and the United States based on household surveys in the 1950s.31 This found that the distribution of wealth was significantly more unequal in Britain. Sixty years later, the OECD figures show that the reverse is the case: the share of the top 1 per cent in the US in 2010 is 36.6 per cent, or more than double the UK figure for 2009.32 This dramatic change warrants further investigation, as does the fact that the top 1 per cent wealth share in Great Britain is so much lower (even leaving aside pensions) than in Austria and the Netherlands (both 24 per cent) and Germany (25 per cent). The second finding is that the WAS‐based estimates supplied by the ONS to the OECD show a distinct upward trend. The share of the top 1 per cent in 2010–12 is 2.7 percentage points higher than that in 2006–08, when measured including pension wealth, and the increase is nearly double (5.3 percentage points) for the estimates excluding pension wealth. Such a striking conclusion also needs to be investigated further. 3. Combined with the Rich Lists The OECD (2015) refers to the problems with studying the upper tail of the wealth distribution using household surveys: ‘measuring wealth at the top of the wealth distribution through household surveys is intrinsically difficult, as wealthy households typically under‐report their wealth [and] household surveys suffer from varying degrees of non‐response [– the] bias is particularly large when looking at the top 1 per cent of the distribution’ (p. 251). This has led to attempts to use independent data from Rich Lists to ‘complete’ the survey data. Vermeulen (2015) has combined extreme wealth observations from the Forbes List of World Billionaires with the WAS data for 2008–10. He begins by noting ‘there is a substantial gap between the highest ranked survey household and the lowest ranked Forbes individual’ (p. 18). Fitting a Pareto upper tail, he finds that the share of the top 1 per cent rises by between 1 and 5 percentage points, depending on the threshold assumed for the Pareto distribution.33 The higher end of this range would go some way towards closing the gap between the household survey estimates and those for individual wealth‐holding based on the estate data for 2008–10. At the same time, we should note that those identified in the Forbes List may include people who are not UK residents. A Pareto extrapolation had been used earlier by Davies and Shorrocks in the estimates they have prepared for Credit Suisse.34 In effect, they use the total number of billionaires (but not their wealth) reported in the Forbes List to fit a Pareto distribution. It is the changing number of billionaires that drives the year‐to‐year changes shown in Figures 2 and 3 (the dashed series), since the distribution is otherwise based on the WAS 2006–08. As may be seen, these estimates suggest that the share of the top 1 per cent is close to the estate‐based estimates, and the share has increased by some 3 percentage points over the period from 2000 to 2014. The Rich Lists provide information on the shape of the upper tail of the wealth distribution that allows for a more detailed investigation of the distribution within the top 1 per cent. To date, official estimates of wealth concentration have not shown shares for groups smaller than the top 1 per cent (the same limitation applied to the findings in Atkinson and Harrison (1978)). The Sunday Times Rich List for 2010 headline has 1,000 people with £335.5 billion. These make up 0.004 per cent of total (GB) households and 5.3 per cent of total WAS non‐pension wealth. V. Conclusions In this paper, we have used evidence from existing data sources to attempt to answer the three questions set out at the beginning and to identify the need for further information. The UK wealth distribution is indeed highly concentrated. The estate‐based estimates (the former HMRC Series C, the unadjusted estimates and the new HMRC estimates, allowing for the understatement of concentration) suggest that the share of the top 1 per cent is between a fifth and a quarter of total personal wealth. The household survey data cannot be used on their own to investigate concentration at the top. When combined with information about the upper tail, the survey‐based estimates (excluding pension wealth) are below the estate‐based estimates of top shares, but we have to allow for the fact that the estimates relate to households rather than individuals. On the basis of the estate‐based estimates, wealth inequality at the top exceeds inequality in after‐tax income: the share of the top 1 per cent in total wealth is about double the share of the top 1 per cent in after‐tax income. Finally, the estimates provide some support for the view that wealth inequality increased in the UK over the first decade of the present century, but we believe that any definitive statement should await further investigation. Indeed, the evidence about the UK distribution of wealth post‐2000 is seriously incomplete and the main conclusion of the paper is that significant investment is necessary if we are to provide satisfactory answers to the three questions. Moreover, given the limitations of each of the different sources, it is important to make use of all available approaches. The estate‐based estimates remain, in our view, an essential element when studying top wealth‐holdings (and we do not believe that the HMRC official estimates should be discontinued as currently proposed), but there needs to be a renewed investigation of the mortality multipliers, the necessary adjustments and the reconciliation with the balance‐sheet information. The investment income method should be explored further, but for this it is necessary that the underlying data be available in a more detailed form. The issues of non‐response and under‐reporting at the top mean that the household surveys – valuable though they are in covering the majority of the population – need to be supplemented when considering the upper tail. Consideration needs to be given to the use of investment income data for this purpose, in addition to the Rich Lists. These recommendations require resources, but unless such work is undertaken we shall not be able to draw firm conclusions about the extent of wealth concentration, how it compares with that in other countries and whether it is increasing over time. Supporting information Disclaimer: Supplementary materials have been peer‐reviewed but not copyedited. Appendix Click here for additional data file. 1 OECD, 2015, p. 34. 2 OECD, 2015, p. 241. 3 There exists a fifth potential source of evidence about the distribution of personal wealth: administrative data on the wealth of the living derived from personal wealth taxes. However, in the UK there is no annual wealth tax, and the council tax cannot be used for this purpose. 4 Namely, the estates gaining a grant of representation (known as confirmation of executors in Scotland, and probate or letter of administration in the rest of the UK). The estates held by younger individuals are oversampled, while 100 per cent of the largest estates are included. With this procedure, estates of low value are generally excluded, as well as those held in trusts or in joint names passing to a surviving spouse or civil partner. The excluded estates accounted for approximately 70 per cent of total estates in 2008–10. 5 Available from the HMRC website at http://webarchive.nationalarchives.gov.uk/20120403124426/ http://www.hmrc.gov.uk/stats/personal_wealth/archive.htm (Series C) and https://www.gov.uk/government/collections/distribution‐of‐personal‐wealth‐statistics (new series). 6 Although, for many years, the Inland Revenue made estimates of the distribution of wealth including occupational pension rights (Series D) or both occupational and state pension rights (Series E) – see, for example, Inland Revenue Statistics 1996, tables 13.6 and 13.7. 7 See Atkinson and Harrison (1978, pp. 32–3). 8 This represents an improvement on the ground that homeownership can hardly identify wealth in a context of high and increasing homeownership rates. However, as acknowledged by HMRC, the data still present some problems to the extent that ELSA is designed to be representative of older households living in England only. 9 See Atkinson and Harrison (1978, p. 175). 10 This explains why, at this stage, we do not provide results based on the investment income method in Section 3.. 11 Table 13.1 on the HMRC website at http://webarchive.nationalarchives.gov.uk/20120403124426/ http://www.hmrc.gov.uk/stats/personal_wealth/archive.htm. 12 See Atkinson and Harrison (1978, appendix I). 13 Royal Commission on the Distribution of Income and Wealth, 1979, p. 117. 14 Office for National Statistics, 2012. 15 Office for National Statistics, 2009, p. 2. 16 Office for National Statistics, 2009, p. 119. 17 Office for National Statistics, 2009, p. 5. 18 Beresford, 1990, 1991 and 2006. 19 For example, page 91 of the Sunday Times Magazine, 18 May 2014. 20 Dolan, 2012. 21 Table 13.4 on the HMRC website at http://webarchive.nationalarchives.gov.uk/20120403124426/ http://www.hmrc.gov.uk/stats/personal_wealth/archive.htm. 22 This may clearly vary across countries depending on the nature of the tax unit. In the US, for instance, a tax unit is similar to a family unit as it contains singles or married couples with or without dependants. Therefore the estimates of top wealth shares for the US by Saez and Zucman (2014) based on the capitalisation method relate to tax units and not to individuals. 23 Notwithstanding this, the WAS aims to follow individuals rather than households. In the case that a household splits, with individuals living at different addresses, WAS interviews all of the original sample members in the next wave of the survey (Office for National Statistics, 2014). 24 http://www.ons.gov.uk/ons/rel/cap‐stock/the‐national‐balance‐sheet/2014‐estimates/index.html. 25 The next update, covering the period 2011–13, was scheduled for publication in September 2015. However, as we write, HMRC is proposing not to publish this or any further updates, the reason being that it ‘[does] not think that the HMRC Personal Wealth National Statistics, which are based on data from Inheritance Tax returns for estates requiring probate, can be reliably used to look at the distribution of wealth amongst all people in the UK’ (HM Revenue & Customs, 2015, p. 4). There are indeed limitations to the estate‐based estimates, as we have noted above and as has been extensively discussed in the literature over more than a century (a literature to which the Inland Revenue, the predecessor of HMRC, has been a major contributor). But, as we have emphasised, other sources of data on wealth‐holding have significant limitations regarding the coverage of top wealth‐holders. 26 HM Revenue & Customs, 2012, p. 17. 27 Table 13.1 on the HMRC website at https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/447352/table_13‐1.pdf. 28 Table 13.4 on the HMRC website at http://webarchive.nationalarchives.gov.uk/20120403124426/http:/www.hmrc.gov.uk/stats/personal_wealth/archive.htm. 29 Blue Book 2014, S1HN‐LE‐B90: total net worth of households and NPISH. 30 The shares of the top 10, top 5 and top 1 per cent of net income are provided in the online appendix. The numbers are different from those appearing in the World Top Incomes Database (WTID) simply due to the different definition of the population control total: adults aged 18 and over in this paper (for consistency with the wealth distribution estimates), and adults aged 15 and over in the WTID. 31 Lydall and Lansing, 1959. 32 Cowell (2013, p. 44) draws attention to the ‘surprising’ finding from the earlier BHPS survey in 2000 that the UK exhibits less wealth inequality than the US, Canada and Sweden. 33 See also Cowell (2013) on fitting a Pareto distribution to household survey data – in this case, data from the Luxembourg Wealth Study. 34 See, for example, Credit Suisse Research Institute (2014, p. 9), which describes the way in which their method has changed over time. ==== Refs References Atkinson , A. B. and Harrison , A. J. (1974 ), ‘Wealth distribution and investment income in Britain ’, Review of Income and Wealth , vol. 20 , pp. 125 –42 . Atkinson , A. B. and Harrison , A. J. (1978 ), Distribution of Personal Wealth in Britain, Cambridge : Cambridge University Press . Barna , T. (1945 ), Redistribution of Incomes 1937, Oxford : Oxford University Press . Baxter , R. D. (1869 ), The Taxation of the United Kingdom, London : Macmillan . Beresford , P. (1990 ), The Sunday Times Book of the Rich: Britain's 400 Richest People, London : Weidenfeld and Nicolson . Beresford , P. (1991 ), The Sunday Times Book of the Rich, London : Penguin . Beresford , P. (2006 ), The ‘Sunday Times’ Rich List 2006–2007: 5,000 of the Wealthiest People in the United Kingdom, London , A and C Black . Board of Inland Revenue (2000 ), Inland Revenue Statistics, London : HMSO . Cowell , F. A. (2013 ), ‘UK wealth inequality in international context’, in Hills J. , Bastagli F. , Cowell F. , Glennerster H. , Karagiannaki E. and McKnight A. (eds), Wealth in the UK: Distribution, Accumulation, and Policy, Oxford : Oxford University Press . Credit Suisse Research Institute (2014 ), Global Wealth Databook 2014, London . Dolan , K. A. (2012 ), ‘Methodology: how we crunch the numbers ’, Forbes Magazine , 7 March. Harris , W. J. and Lake , K. A. (1906 ), ‘Estimates of the realisable wealth of the United Kingdom based mostly on the Estate Duty Returns ’, Journal of the Royal Statistical Society , vol. 69 , pp. 709 –45 . HM Revenue & Customs (2012 ), ‘UK personal wealth statistics 2008 to 2010’. HM Revenue & Customs (2015 ), ‘Consultation about ceasing publication of HMRC's Personal Wealth National Statistics’, consultation document. Kopczuk , W. and Saez , E. (2004 ), ‘Top wealth shares in the United States, 1916–2000: evidence from estate tax returns ’, National Tax Journal , vol. 57 , pp. 445 –87 . Lydall , H. F. and Lansing , J. B. (1959 ), ‘A comparison of the distribution of personal income and wealth in the United States and Great Britain ’, American Economic Review , vol. 49 , pp. 43 –67 . Mallet , B. (1908 ), ‘A method of estimating capital wealth from the estate duty statistics ’, Journal of the Royal Statistical Society , vol. 71 , pp. 65 –101 . Mallet , B. and Strutt , C. (1915 ), ‘The multiplier and capital wealth ’, Journal of the Royal Statistical Society , vol. 78 , pp. 555 –99 . OECD (2015 ), In It Together: Why Less Inequality Benefits All, Paris : Organisation for Economic Cooperation and Development . Office for National Statistics (2009 ), Wealth in Great Britain: Main Results from the Wealth and Assets Survey 2006/08, London : ONS . Office for National Statistics (2012 ), Wealth and Assets Survey Review Report, London : ONS . Office for National Statistics (2014 ), Wealth in Great Britain Wave 3, 2010–2012, London : ONS . Royal Commission on the Distribution of Income and Wealth (1979 ), Report No. 7: Fourth Report on the Standing Reference, Cmnd 7595, London : HMSO . Saez , E. and Zucman , G. (2014 ), ‘Wealth inequality in the United States since 1913: evidence from capitalized income tax data’, National Bureau of Economic Research (NBER), Working Paper no. 20625. Stark , T. (1972 ), The Distribution of Personal Income in the United Kingdom 1949–1963, Cambridge : Cambridge University Press . Vermeulen , P. (2015 ), ‘How fat is the top tail of the wealth distribution?’, unpublished working paper.

PMC005xxxxxx/PMC5001193.txt

==== Front Food Energy SecurFood Energy Secur10.1002/(ISSN)2048-3694FES3Food and Energy Security2048-3694John Wiley and Sons Inc. Hoboken 10.1002/fes3.83FES383Original ResearchOriginal ResearchVertical farming increases lettuce yield per unit area compared to conventional horizontal hydroponics Touliatos Dionysios 1 Dodd Ian C. 1 McAinsh Martin 1 1 The Lancaster Environment CentreLancaster UniversityLancasterUK* Correspondence Dionysios Touliatos, The Lancaster Environment Centre, Lancaster University, Lancaster, UK. Tel: 00441524593985; Fax: 00441524593985; E‐mail: d.touliatos@lancaster.ac.uk 06 6 2016 8 2016 5 3 10.1002/fes3.2016.5.issue-3184 191 04 2 2016 21 4 2016 © 2016 The Authors. Food and Energy Security published by John Wiley & Sons Ltd. and the Association of Applied Biologists.This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.Abstract Vertical farming systems (VFS) have been proposed as an engineering solution to increase productivity per unit area of cultivated land by extending crop production into the vertical dimension. To test whether this approach presents a viable alternative to horizontal crop production systems, a VFS (where plants were grown in upright cylindrical columns) was compared against a conventional horizontal hydroponic system (HHS) using lettuce (Lactuca sativa L. cv. “Little Gem”) as a model crop. Both systems had similar root zone volume and planting density. Half‐strength Hoagland's solution was applied to plants grown in perlite in an indoor controlled environment room, with metal halide lamps providing artificial lighting. Light distribution (photosynthetic photon flux density, PPFD) and yield (shoot fresh weight) within each system were assessed. Although PPFD and shoot fresh weight decreased significantly in the VFS from top to base, the VFS produced more crop per unit of growing floor area when compared with the HHS. Our results clearly demonstrate that VFS presents an attractive alternative to horizontal hydroponic growth systems and suggest that further increases in yield could be achieved by incorporating artificial lighting in the VFS. Land use efficiencyplant factoryurban agriculturevertical column grower.Biotechnology and Biological Sciences Research CouncilBB/J012718/1 source-schema-version-number2.0component-idfes383cover-dateAugust 2016details-of-publishers-convertorConverter:WILEY_ML3GV2_TO_NLMPMC version:4.9.4 mode:remove_FC converted:24.08.2016 Food and Energy Security 2016 ; 5(3): 184 –191 ==== Body Introduction The global population is expected to reach 9 billion by 2050, a significant proportion of which will be urban dwellers, requiring a 70% increase in agricultural productivity (Corvalan et al. 2005; Tilman et al. 2011). Continued rural to urban migration is predicted to drive the expansion of urban landscapes and accelerate the loss of cultivated land surrounding towns and cities (Pandey and Seto 2015). Coupled with land degradation and loss of soil fertility, due to land‐use intensification and climate change, agricultural land is increasingly becoming a scarce resource (Foley et al. 2011; Lambin et al. 2013) in addition to being a threat for biodiversity (Chaplin‐Kramer et al. 2015). In light of the above, the need for innovation in land‐use efficiency for crop production is therefore increasingly important (Lambin and Meyfroidt 2011). Vertical farming has been proposed as an engineering solution to increase productivity per area by extending plant cultivation into the vertical dimension, thus enhancing land use efficiency for crop production (Eigenbrod and Gruda 2014). The large‐scale implementation of vertical farming involves stacking growth rooms, such as glasshouses and controlled environment rooms, on top of each other to construct food‐producing high‐rise buildings (Despommier 2011). The same concept can be applied at a smaller scale through vertical farming systems (VFS). These growth systems expand crop production into the vertical dimension to produce a higher yield using less floor area (Hochmuth and Hochmuth 2001; Resh 2012). Examples of VFS include the use of vertical columns (Linsley‐Noakes et al. 2006), vertically suspended grow bags (Neocleous et al. 2010), conveyor‐driven stacked growth systems (Mahdavi et al. 2012), A‐frame designs (Hayden 2006), and plant factory approaches (Kato et al. 2010). Although these studies have quantified crop production, there have been few direct comparisons with horizontal systems of similar cropping density and little information is available on whether vertical column systems present a viable alternative to horizontal crop production systems. In addition, previous yield comparisons of VFS with conventional horizontal systems have confounded other factors with crop orientation. For example, yield increases of 129–200% in VFS and increased profits of 3.6–5.5 US dollar·m−2 compared to conventional soil cultivation have been reported (Liu et al. 2004). However, their VFS utilized a soilless growing medium; rendering the comparison essentially invalid. Similarly, significantly higher yields have been reported for strawberry grown in a vertical column VFS compared to conventional grow bags and a multi‐tiered VFS (Ramírez‐Gómez et al. 2012); although no information was provided regarding the root zone volume of the growth systems. The aim of this study was to compare a vertical column VFS and a conventional horizontal hydroponic system (HHS) with similar fertigation regimes, root zone volumes, and planting densities to determine whether VFS represents a viable alternative to HHS. Lettuce was used as a model plant as it is widely grown in hydroponics as a rapidly growing leafy vegetable (Safaei et al. 2015) thereby avoiding some of the complexities of changes in crop biomass allocation during the reproductive process (Heller et al. 2014). The study was conducted indoors using only artificial lighting, as this is the dominant approach found in most urban vertical farming projects, especially in plant factory designs (Kang et al. 2014; He et al. 2015) and allows more precise control of environmental conditions (Poorter et al. 2012). Our results show that VFS increased lettuce yield per unit area compared to HHS and suggest that variation in light intensity between cropping systems of different spatial orientation could explain differences in crop yield. Materials and Methods Location The study was conducted in a 3.4 m × 4.15 m walk‐in Controlled Environment room (CE room) at the Lancaster Environment Centre (LEC, Lancaster University, UK). Illumination was provided by 12 400 W metal halide lamps (HQI‐T 400N; Osram, St Helens, UK) for a 16 hour photoperiod (06.00 h to 22.00 h). Highly reflective plastic film (LBS Horticulture Ltd, Lancashire, UK) was placed on the walls of the room in order to increase the diffusion of light. Room air temperature ranged between 16 and 18 °C and relative humidity ranged from 60% to 80%. Room temperature and humidity were recorded by an Ektron II C sensor (HortiMaX B.V., Pijnacker, the Netherlands), which was hanging from the ceiling in the middle of the CE room, at 1.83 m above the ground. The CE room accommodated 2 VFS and 2 HHS, with one of each arranged on each side of the room. Preliminary measurements of photosynthetic photon flux density (PPFD) (before and after installing the VFS) revealed no shading effect of the VFS on the PPFD in the HHS (data not shown). The vertical farming system Plants were grown vertically in upright cylindrical columns comprised of individual modular units stacked on top of each other to reduce the system footprint. Each modular unit consisted of two stackable elements: a growing container (10.5 cm high and 7.5 cm radius) and a spacing collar (20 cm high and 7.5 cm radius). Five growing containers and six spacing collars were sterilized in TriGene disinfectant (MediChem International Ltd., Sevenoaks, UK) prior to filling each container with 130 g ± 0.5 g medium grade perlite (LBS Horticulture Ltd, Lancashire, UK). Highly porous perlite was the substrate of choice in order to minimize the risk of root‐zone hypoxia and the resultant accumulation of ethylene within the airspace of the vertical column. The perlite was held in place by horticultural frost fleece (LBS Horticulture Ltd, Lancashire, UK) that was placed at the bottom of each growing container. The perlite in each container was levelled and seedlings were placed on the perlite at 90° to the horizontal. Each VFS contained 20 lettuce plants in total. Distance between the top of the VFSs and the light was 80 cm. The horizontal hydroponic system The HHS comprised of five cylindrical PVC pipes (45.5 cm high and 3.6 cm radius), which were sterilized in TriGene disinfectant (MediChem International Ltd), filled with 130 g ±0.5 g of perlite (LBS Horticulture Ltd, Lancashire, UK) and placed in parallel at 20 cm apart center to center. Horticultural frost fleece (LBS Horticulture Ltd, Lancashire, UK) was placed in the outlet of each pipe to hold the perlite in place. Each pipe held four plants placed in 4.4 cm square holes, in rows. In order to prevent the growth of algae black nylon fabric was used to cover the outlet channel of the system. Each HHS contained 20 lettuce plants in total. PVC pipes were mounted on commercial growth benches (90 cm from the ground and 130 cm from the lights); which was equivalent in height to VFS's layer 2 (Fig. 1). Figure 1 Schematic shows planting densities within the HHS and VFS. (A) Overhead view of HHS. (B) Side‐view of VFS. (C) Overhead view of VFS. The HHS occupied 0.4 m2 of growing floor area, whereas the VFS occupied 0.02 m2 per column of floor area. The grey rectangles show the exact position of the 400 W metal halide lamps above the growth systems. The grey circles within the VFS show the exact measurement positions of the Macam Q203 Quantum radiometer quantum sensor. Light measurements in the HHS were obtained directly above the plants, at 20 cm distance from the PVC pipes. Scale bar is 20 cm. VFS, vertical farming system; HHS, horizontal hydroponic system. Fertigation Each growth system was supplied with recirculated half‐strength Hoagland's solution (Hoagland and Arnon 1950) from a 18 L Titan PC4R Tank (Kingspan Environmental Ltd, Armagh, UK), through a 1.27 cm double‐walled PVC hose (LBS Horticulture Ltd). The composition of the nutrient solution was 0.5 mmol·L−1 NH4NO3, 1.75 mmol·L−1 Ca(NO3)2·4H2O, 2.01 mmol·L−1 KNO3, 1.01 mmol·L−1 KH2PO4, 0.5 mmol·L−1 MgSO4·7H2O, 1.57 μmol·L−1 MnSO4·5H2O, 11.3 μmol·L−1 H3BO3, 0.3 μmol·L−1 CuSO4·5H2O, 0.032 μmol·L−1 (NH4)6Mo7O24·4H2O, 1.04 μmol·L−1 ZnSO4·7H2O, and 0.25 mmol·L−1 NaFe EDTA. Nutrient solution Electrical Conductivity (EC) was 1 ± 0.2 dS·m−1. The nutrient solution in tanks was replaced weekly and 2 mol·L−1 H3PO4 was used to maintain a pH of 5.8 ± 0.2, which was checked daily. Hozelock 360° Micro Jet microsprinklers (Hozelock Limited, Aylesbury, UK) delivered the solution to the top layer of the VFS, allowing gravity‐driven drip‐irrigation of plants in growing modules below them in the column. In the HHS the nutrient solution was delivered to each pipe of the HHS by a microsprinkler. The effluent from the bottom layer of the VFS and from all PVC pipes of the HHS was subsequently returned to the tank and recirculated around the growing systems using a submersible aquarium water pump (All Pond Solutions Ltd, Middlesex, UK), capable of delivering a maximum of 3100 L·h−1. Hozelock Coupling 13 mm hose connectors (Hozelock Limited, Aylesbury, UK) were used to connect all hoses and pumps. Each pump was programmed to operate for 1 min every hour using a multi purpose electronic digital programmable timer (JoJo Waterproof Digital Outdoor Electrical Timer; AuctionZ Ltd, Bradford, UK). Consequently, the microsprinklers sprayed nutrient solution within the top layers of the VFS and within the horizontal layers of the HHS for 1 min every hour. Plant material Romaine lettuce (Lactuca sativa L. cv. ‘Little Gem’) seeds were sown in 84‐cell plug trays (tray dimensions: 53 cm × 31 cm × 5.5 cm) containing Levington M3 compost (Scotts UK, Ipswich, UK). Seedlings were watered daily with tap water and were grown at a PPFD of 200 μmol·m−2·s−1 over a 16 h photoperiod. Plants were transplanted 20 days after sowing at the four true leaf stage. Data collection and statistical analysis Photosynthetic photon flux density measurements were obtained using a Macam Q203 Quantum radiometer (Macam Photometrics LTD, Livingstone, UK) 1 and 5 weeks after the plants were transferred to the growth systems. The quantum sensor was placed in the middle of the spacing collar, in a 10 cm radius zone around the vertical column in the VFS, and was placed 20 cm above the PVC pipes in the HHS. This approach ensured consistency in light intensity measurements and avoided plant damage during sampling. Shoot fresh weight was measured immediately after harvest on week 5 using a 2 decimal point scientific balance. To compare average shoot fresh weight per growth system using Student's t‐test, the data were square root transformed, as they were not normally distributed (Table 1). Yield and number of plants per occupied growing floor area for growth systems was used to calculate the ratio of VFS to HHS (Table 1). Linear regression analysis was used to analyze the relationship between shoot fresh weight and vertical or horizontal layers within each growth system (Fig. 2). Significant differences in PPFD during week 1 within the VFS and the HHS were detected using one‐way ANOVA followed by Tukey post hoc analysis (Fig. 3). The relationship between PPFD during week 5 and shoot fresh weight was analyzed using linear regression analysis (Fig. 4), with P‐value <0.05 considered to indicate a statistically significant difference. All statistical tests were performed using “R” version 3.1.2 software (R Development Core Team 2014). Table 1 Comparison of the productivity of the vertical farming system (VFS) and horizontal hydroponic system (HHS) Parameter HHS VFS Result Shoot fresh weight (g) Mean ± SE (n = 40) 138 ± 6 95 ± 6 P < 0.001b Yield per occupied growing floor areaa (kg FW·m−2) 6.9 95 VFS/HHS = 13.8 Number of plants per occupied growing floor areaa (plant number m−2) 50 1000 VFS/HHS = 20 a HHS growing floor area: 0.4 m2, VFS growing floor area: 0.02 m2. b Student's t‐test on square root transformed data, t (78) = 5.656. John Wiley & Sons, LtdFigure 2 Linear regression analysis of shoot fresh weight versus layer in the VFS (solid line; closed symbols) and HHS (open symbols), respectively. When the linear regression was not significant the regression line was omitted. The regression equation, adjusted R 2 values and significance of the regression (P‐value) are reported at the top of the panel. VFS, vertical farming system; HHS, horizontal hydroponic system. Figure 3 PPFD within the vertical farming system (VFS; closed symbols) and the horizontal hydroponic system (HHS; open symbols) plotted against layers in the growth systems. Values indicated with different letters indicate statistically significant differences, whereas those marked with the same letters show statistically similar values. Error bars represent SE (n = 8). VFS, vertical farming system; HHS, horizontal hydroponic system; PPFD, photosynthetic photon flux density. Figure 4 Linear regression analysis of shoot fresh weight versus PPFD in the vertical farming system (VFS; solid line; closed symbols) and horizontal hydroponic system (HHS; open symbols), respectively. When the linear regression was not significant the regression line was omitted. The regression equation, adjusted R 2 values, and significance of the regression (P‐value) are reported at the top of the panel. VFS, vertical farming system; HHS, horizontal hydroponic system; PPFD, photosynthetic photon flux density. Results The VFS produced more crop per unit area compared to the HHS The VFS produced 13.8 times more crop, calculated as a ratio of yield (kg FW) to occupied growing floor area (m2). However, mean FW (g) for lettuce crops grown within the HHS was significantly higher than those grown within the VFS. Therefore, although the same number of plants was grown in each system the two HHS produced 1.7 kg more crop compared to the two VFS (5.5 and 3.8 kg of crop in total, respectively). Consequently, the higher productivity of the VFS, in terms of kg FW·m−2 growing floor area, can be attributed to the 20‐fold higher number of plants per growing floor area (Table 1). Yield decreased from top to base of VFS, whereas yield was uniform within the HHS Shoot fresh weight decreased from top to base of vertical columns of the VFS, whereas no gradient in productivity was observed between horizontal layers in the HHS. Crop productivity was uniform in the HHS with a range of 133 g within a normal distribution, whereas in the VFS, crop productivity had a range of 180 g within a nonnormal distribution with positive skewness (Sk = 1.035). Plants grown within the top layer of the VFS and within all layers of the HHS were of similar shoot fresh weight. However, in the middle and bottom layers (Layers 2–5) of the VFS, productivity decreased significantly. As a result, the bottom layer of the VFS produced 43% less shoot fresh weight than the top layer of the VFS (Fig. 2). Light intensity decreased from top to base of VFS Light intensity decreased significantly from top to base of vertical columns within the VFS, whereas no significant difference in PPFD was observed within the horizontal layers of the HHS. PPFD values varied between 491 and 134 μmol·m−2·s−1 from top to base of vertical column of the VFS and between 570 and 340 μmol·m−2·s−1 within the horizontal layers of the HHS. The top layer of the VFS received similar PPFD to all the horizontal layers of the HHS. However, within the vertical layers, as distance from the light source increased, there was a significant drop in PPFD values within the VFS. There was no significant difference in PPFD between layers 2 and 3 and between layers 3, 4, and 5 (Fig. 3). Light intensity influenced growth in the VFS but not in the HHS There was a significant positive relationship between shoot fresh weight and PPFD in the VFS, indicating that as light intensity increased so did crop productivity (Fig. 4). In contrast, there was no significant relationship between yield and PPFD within the HHS. Discussion Although it has been proposed that increases in yield per growing area can be achieved by extending plant cultivation into the vertical dimension using VFS (Eigenbrod and Gruda 2014), to date there is no conclusive evidence that this is indeed the case. However, our results show that crop productivity, defined as a ratio of yield to occupied growing floor area, is 13.8 times higher in VFS than the HHS. This is likely because by incorporating the vertical dimension into the growth environment, the VFS can grow 20‐fold more plants per unit area than the HHS (Table 1). However, these calculations are based on independent vertical columns and do not consider the effect of column spacing on yield per occupied growing floor area. For example, in high wire crop training systems high planting densities impose intense competition for light within the growth system (Pettersen et al. 2010). This is an important factor that needs to be considered in future studies, as spacing between vertical columns influenced crop productivity in VFS glasshouse trials (Liu et al. 2004). In contrast, the absolute yield of the HHS, in terms of shoot fresh weight, was higher than the VFS (Table 1). This can be explained by the significant decrease in PPFD from top to base of the vertical columns (Fig. 3) and significant causal relationship between shoot fresh weight and PPFD within the VFS (Fig. 4) that limited growth in the lower layers. Light intensity is one of the primary variables affecting lettuce yield and quality (Ferentinos et al. 2000; Son and Oh 2013; Ouzounis et al. 2015) and it is has been well documented that lettuce yield increases with increasing light intensity (Knight and Mitchell 1988; Frantz and Bugbee 2005). Therefore, since yield decreased from top to base of the vertical column, and yield was uniform within the HHS (Fig. 2), it was anticipated that the VFS would produce less crop in total than the HHS. Similarly, light intensity and shoot fresh weight were highly correlated and both decreased from top to base of vertical columns in a glasshouse (Liu et al. 2004). Light gradients from top to base of vertical column systems were also reported in glasshouse vertical strawberry cultivation (Ramírez‐Gómez et al. 2012). Therefore, our data suggest that top to base gradients in light intensity and shoot fresh weight limit plant growth in vertical columns in both indoor and glasshouse settings. Vertical light intensity gradients (e.g. Fig. 3) could be altered by natural illumination. In glasshouse trials with vertical columns, light intensity decreased from top to base of vertical columns, with lower PPFD values being recorded toward the northern side of columns compared to the southern side (Liu et al. 2004). Thus, natural illumination introduced an additional gradient in light distribution within the VFS. Future studies are therefore required to test whether natural illumination diminishes or exacerbates light intensity gradients. Light intensity in growth chambers is known to decrease as distance from the light source increases (Poorter et al. 2012) and this phenomenon partially explained the large variance in PPFD observed in vertical layers 2–5 within the VFS (Fig. 3). In addition, a “shading effect” within the VFS (Linsley‐Noakes et al. 2006) was due to higher positioned plants within the VFS obscuring lower positioned plants from the light source. Side‐on rather than top‐down illumination could potentially ameliorate the shading effect, consequently mitigating the gradient in crop productivity within the VFS. Side‐on illumination, also known as interlighting has improved light distribution within tall canopies and, in some cases, increased crop yield and light use efficiency (Olle and Viršile 2013). Interlighting with light‐emitting diodes (LEDs) ameliorated mutual shading within tomatoes at high planting density and increased tomato yield by 12–14% in comparison to the control (Lu et al. 2012). Overhead illumination combined with intracanopy lighting using HPS lamps increased cucumber yield in high‐wire crop training system by 11% compared to traditional overhead illumination (Pettersen et al. 2010). In contrast, there were no differences in productivity when comparing LED interlighting against overhead HPS in high‐wire tomato cultivation (Gómez et al. 2013). Similarly, interlighting by fluorescent tubes improved fruit quality but did not increase yield in high‐wire cucumber production in the glasshouse (Heuvelink et al. 2006). This variation in crop responses to interlighting may be due to the different environmental conditions and crop management applied. However, since vertical column systems share similar light distribution properties within the vertical plane to plants grown in high wire crop training systems (Hovi et al. 2004), side‐on illumination could potentially mitigate observed light gradients within the VFS. Interestingly, only 36% of the variation in shoot fresh weight within the VFS was explained by the gradient in PPFD (Fig. 4) with the remaining 64% of variance being attributed to putative temperature gradients and putative nutrient concentration gradients along the vertical column (Jones 2014). Nutrient concentration gradients within the gully of Nutrient Film Technique (NFT) systems have been claimed to influence crop uniformity within the NFT (Puerta et al. 2007). An important difference between the VFS and the HHS of this study was that each layer of the HHS received nutrient solution directly from the tank whereas, in the VFS nutrient solution was delivered to the top layer and was gravity‐driven drip‐fed to vertical layers beneath creating the potential for marked gradients in nutrient availability within the VFS. Identifying the physiological effects of this putative nutrient concentration gradient to growth within VFS is therefore an important area for future studies. To conclude, from a commercial point of view, the effects of gradients within the VFS on crop value will depend on how the crop is going to be processed and marketed. For example, if lettuce was grown to be sold as individual heads, then the nonuniform productivity of the VFS would be a potential weakness of the VFS over the HHS. However, if the crop was destined for precut salad bags then crop uniformity may be irrelevant while increased yield per unit area could be a significant advantage of the VFS. Therefore, crop utilization and marketability and an investigation of the cost‐to‐benefits ratio of these growing systems will be the ultimate criteria to decide whether VFS can provide an alternative to HHS. Conclusions Vertical column‐based VFS presented a viable alternative to conventional horizontal growth systems by optimizing growing space use efficiency, thereby producing more crop per unit area. Further increases in yield could be achieved by incorporating artificial lighting within the VFS to mitigate the observed PPFD gradient. Future studies could investigate the influence of putative additional gradients (such as root and canopy zone temperature and nutrient concentration gradients) within the VFS and interactions between vertical columns in terms of competition for PPFD in large‐scale vertical farming settings. Conflict of Interest None declared. ==== Refs References Chaplin‐Kramer , R. , R. P. Sharp , L. Mandle , S. Sim , J. Johnson , I. Butnar , et al. 2015 Spatial patterns of agricultural expansion determine impacts on biodiversity and carbon storage . Proc. Natl Acad. Sci. USA 112 :7402 –7407 .26082547 Corvalan , C. , S. Hales , and A. J. McMichael . 2005 Ecosystems and human well‐being: health synthesis . World Health Organization . Despommier , D. 2011 The vertical farm: controlled environment agriculture carried out in tall buildings would create greater food safety and security for large urban populations . J. Für Verbraucherschutz Leb. 6 :233 –236 . Eigenbrod , C. , and N. Gruda . 2014 Urban vegetable for food security in cities. A review . Agron. Sustain. Dev. 35 :483 –498 . Ferentinos , K. P. , L. D. Albright , and D. V. Ramani . 2000 SE – structures and environment: optimal light integral and carbon dioxide concentration combinations for lettuce in ventilated greenhouses . J. Agric. Eng. Res. 77 :309 –315 . Foley , J. A. , N. Ramankutty , K. A. Brauman , E. S. Cassidy , J. S. Gerber , M. Johnston , et al. 2011 Solutions for a cultivated planet . Nature 478 :337 –342 .21993620 Frantz , J. M. , and B. Bugbee . 2005 Acclimation of plant populations to shade: photosynthesis, respiration, and carbon use efficiency . J. Am. Soc. Hortic. Sci. 130 :918 –927 . Gómez , C. , R. C. Morrow , C. M. Bourget , G. D. Massa , and C. A. Mitchell . 2013 Comparison of intracanopy light‐emitting diode towers and overhead high‐pressure sodium lamps for supplemental lighting of greenhouse‐grown tomatoes . HortTechnology 23 :93 –98 . Hayden , A. L. 2006 Aeroponic and hydroponic systems for medicinal herb, rhizome, and root crops . HortScience 41 :536 –538 . He , J. , L. Qin , Y. Liu , and T. W. Choong . 2015 Photosynthetic capacities and productivity of indoor hydroponically grown Brassica alboglabra bailey under different light sources . Am. J. Plant Sci. 6 :554 . Heller , H. , A. Bar‐Tal , S. Assouline , K. Narkis , S. Suryano , A. de la Forge , et al. 2014 The effects of container geometry on water and heat regimes in soilless culture: lettuce as a case study . Irrig. Sci. 33 :53 –65 . Heuvelink , E. , M. J. Bakker , L. Hogendonk , J. Janse , R. Kaarsemaker , and R. Maaswinkel . 2006 Horticultural lighting in the Netherlands: new developments . In V International Symposium on Artificial Lighting in Horticulture. 711 :25 –34 . Hoagland , D. R. , and D. I. Arnon . 1950 The water‐culture method for growing plants without soil . Circ. Calif. Agric. Exp. Stn. 347 :2nd edn. Hochmuth , R. , and G. J. Hochmuth . 2001 A decade of change in Florida's greenhouse vegetable industry: 1991–2001 Pp. 280–283 in Proc. Fla. State Hort. Soc. Hovi , T. , J. Näkkilä , and R. Tahvonen . 2004 Interlighting improves production of year‐round cucumber . Sci. Hortic. 102 :283 –294 . Jr Jones , , J. B. 2014 Complete guide for growing plants hydroponically . CRC Press . Kang , J. H. , S. KrishnaKumar , S. L. S. Atulba , B. R. Jeong , and S. J. Hwang . 2014 Light intensity and photoperiod influence the growth and development of hydroponically grown leaf lettuce in a closed‐type plant factory system . Hortic. Environ. Biotechnol. 54 :501 –509 . Kato , K. , R. Yoshida , A. Kikuzaki , T. Hirai , H. Kuroda , K. Hiwasa‐Tanase , et al. 2010 Molecular breeding of tomato lines for mass production of miraculin in a plant factory . J. Agric. Food Chem. 58 :9505 –9510 .20695489 Knight , S. L. , and C. A. Mitchell . 1988 Effects of CO2 and photosynthetic photon flux on yield, gas exchange and growth rate of Lactuca sativa L. “Waldmann”s Green’ . J. Exp. Bot. 39 :317 –328 .11539044 Lambin , E. F. , and P. Meyfroidt . 2011 Global land use change, economic globalization, and the looming land scarcity . Proc. Natl Acad. Sci. USA 108 :3465 –3472 .21321211 Lambin , E. F. , H. K. Gibbs , L. Ferreira , R. Grau , P. Mayaux , P. Meyfroidt , et al. 2013 Estimating the world's potentially available cropland using a bottom‐up approach . Glob. Environ. Change 23 :892 –901 . Linsley‐Noakes , G. , L. Wilken , and S. de Villiers . 2006 High density, vertical hydroponics growing system for strawberries . Acta Hortic. 708 :365 . Liu , W. , D. K. Chen , and Z. X. Liu . 2004 High efficiency column culture system in China . Acta Hortic. 691 :495 –500 . Lu , N. , T. Maruo , M. Johkan , M. Hohjo , S. Tsukagoshi , Y. Ito , et al. 2012 Effects of Supplemental Lighting within the Canopy at Different Developing Stages on Tomato Yield and Quality of Single‐Truss Tomato Plants Grown at High Density . Environ. Control. Biol. 50 :1 –11 . Mahdavi , S. , M. Kafi , R. Naderi , and T. Sadat . 2012 Vertical mobile planting system consistent with the pattern of solar radiation and effects of system on light exposure and growth of Gerbera cut flowers (Gerbera jamesonii cv. Antibes), in greenhouse culture . J. Agric. Technol. 8 :1461 –1468 . Neocleous , D. , C. Kaittanis , N. Seraphides , and P. Polycarpou . 2010 Horizontal and vertical soilless growing systems under Cyprus conditions . J. Appl. Hortic. 12 :140 –144 . Olle , M. , and A. Viršile . 2013 The effects of light‐emitting diode lighting on greenhouse plant growth and quality . Agric. Food Sci. 22 :223 –234 . Ouzounis , T. , B. R. Parjikolaei , X. Fretté , E. Rosenqvist , and C. O. Ottosen . 2015 Predawn and high intensity application of supplemental blue light decreases the quantum yield of PSII and enhances the amount of phenolic acids, flavonoids, and pigments in Lactuca sativa . Front. Plant Sci. 6 . Pandey , B. , and K. C. Seto . 2015 Urbanization and agricultural land loss in India: comparing satellite estimates with census data . J. Environ. Manage. 148 :53 –66 , Land Cover/Land Use Change (LC/LUC) and Environmental Impacts in South Asia.24958549 Pettersen , R. I. , S. Torre , and H. R. Gislerød . 2010 Effects of intracanopy lighting on photosynthetic characteristics in cucumber . Sci. Hortic. 125 :77 –81 . Poorter , H. , F. Fiorani , M. Stitt , U. Schurr , A. Finck , Y. Gibon , et al. 2012 The art of growing plants for experimental purposes: a practical guide for the plant biologist . Funct. Plant Biol. 39 :821 –838 . Puerta , A. R. , S. Sato , Y. Shinohara , and T. Maruo . 2007 A modified nutrient film technique system offers a more uniform nutrient supply to plants . Horttechnology 17 :227 –233 . R Core Team 2014 R: A language and environment for statistical computing . R Foundation for Statistical Computing , Vienna, Austria Available at http://www.R-project.org/. Ramírez‐Gómez , H. , M. Sandoval‐Villa , A. Carrillo‐Salazar , and A. Muratalla‐Lúa . 2012 Comparison of hydroponic systems in the strawberry production . Acta Hortic. 947 :165 –172 . Resh , H. M. 2012 Hydroponic food production: a definitive guidebook for the advanced home gardener and the commercial hydroponic grower . CRC Press . Safaei , M. , J. Panahandeh , S. J. Tabatabaei , and A. R. M. Azar . 2015 Effects of different nutrients solutions on nutrients concentration and some qualitative traits of lettuce in hydroponics system . J. Sci. Technol. Greenh. Cult. 6 :Pe1 –Pe7 , En8. Son , K.‐H. , and M.‐M. Oh . 2013 Leaf shape, growth, and antioxidant phenolic compounds of two lettuce cultivars grown under various combinations of blue and red light‐emitting diodes . HortScience 48 :988 –995 . Tilman , D. , C. Balzer , J. Hill , and B. L. Befort . 2011 Global food demand and the sustainable intensification of agriculture . Proc. Natl Acad. Sci. USA 108 :20260 –20264 .22106295

PMC005xxxxxx/PMC5001194.txt

==== Front Geophys Res LettGeophys Res Lett10.1002/(ISSN)1944-8007GRLGeophysical Research Letters0094-82761944-8007John Wiley and Sons Inc. Hoboken 10.1002/2016GL068747GRL543952016GL068747First results from NASA's Magnetospheric Multiscale (MMS) MissionMagnetospheric PhysicsMagnetic ReconnectionMagnetospheric Configuration and DynamicsMagnetopause and Boundary LayersSolar Wind/Magnetosphere InteractionsSolar Physics, Astrophysics, and AstronomyMagnetic ReconnectionSpace Plasma PhysicsMagnetic ReconnectionTransport ProcessesResearch LetterResearch LettersSpace SciencesIon‐scale secondary flux ropes generated by magnetopause reconnection as resolved by MMS MMS Magnetopause Ion‐Scale Flux RopesEastwood Et Al.Eastwood J. P. 1 Phan T. D. 2 Cassak P. A. 3 Gershman D. J. 4 5 Haggerty C. 6 Malakit K. 7 Shay M. A. 6 Mistry R. 1 Øieroset M. 2 Russell C. T. 8 Slavin J. A. 9 Argall M. R. 10 Avanov L. A. 4 5 Burch J. L. 11 Chen L. J. 4 5 Dorelli J. C. 4 Ergun R. E. 12 Giles B. L. 4 Khotyaintsev Y. 13 Lavraud B. 14 15 Lindqvist P. A. 16 Moore T. E. 4 Nakamura R. 17 Paterson W. 4 Pollock C. 18 Strangeway R. J. 8 Torbert R. B. 10 11 Wang S. 4 5 1 Blackett LaboratoryImperial College LondonLondonUK2 Space Sciences LaboratoryUniversity of CaliforniaBerkeleyCaliforniaUSA3 Department of Physics and AstronomyWest Virginia UniversityMorgantownWest VirginiaUSA4 NASA Goddard Space Flight CenterGreenbeltMarylandUSA5 Department of AstronomyUniversity of MarylandCollege ParkMarylandUSA6 Department of Physics and AstronomyUniversity of DelawareNewarkDelawareUSA7 Department of PhysicsMahidol UniversityBangkokThailand8 Department of Earth, Planetary and Space SciencesUniversity of CaliforniaLos AngelesCaliforniaUSA9 Department of Climate and Space Sciences and EngineeringUniversity of MichiganAnn ArborMichiganUSA10 Institute for the Study of Earth, Oceans and SpaceUniversity of New HampshireDurhamNew HampshireUSA11 Southwest Research InstituteSan AntonioTexasUSA12 Laboratory for Atmospheric and Space PhysicsUniversity of Colorado BoulderBoulderColoradoUSA13 Swedish Institute of Space PhysicsUppsalaSweden14 Institut de Recherche en Astrophysique et PlanétologieUniversité de ToulouseToulouseFrance15 Centre National de la Recherche Scientifique, UMR 5277ToulouseFrance16 School of Electrical EngineeringRoyal Institute of TechnologyStockholmSweden17 Space Research InstituteAustrian Academy of SciencesGrazAustria18 Denali ScientificHealyAlaskaUSA* Correspondence to: J. P. Eastwood, jonathan.eastwood@imperial.ac.uk 18 5 2016 28 5 2016 43 10 10.1002/grl.v43.104716 4724 21 3 2016 28 4 2016 29 4 2016 ©2016. The Authors.This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.Abstract New Magnetospheric Multiscale (MMS) observations of small‐scale (~7 ion inertial length radius) flux transfer events (FTEs) at the dayside magnetopause are reported. The 10 km MMS tetrahedron size enables their structure and properties to be calculated using a variety of multispacecraft techniques, allowing them to be identified as flux ropes, whose flux content is small (~22 kWb). The current density, calculated using plasma and magnetic field measurements independently, is found to be filamentary. Intercomparison of the plasma moments with electric and magnetic field measurements reveals structured non‐frozen‐in ion behavior. The data are further compared with a particle‐in‐cell simulation. It is concluded that these small‐scale flux ropes, which are not seen to be growing, represent a distinct class of FTE which is generated on the magnetopause by secondary reconnection. Key Points Ion‐scale flux ropes are observed during magnetopause reconnection The largely force‐free flux ropes exhibit filamentary currents and nonideal ion behavior Small flux content and comparison with simulation indicate a secondary reconnection origin magnetic reconnectionmagnetopauseflux ropesecondary islandmagnetospheric multiscaleSTFC (UK)ST/G00725X/1NASANNX08AO83GNNX13AD72GNNX16AF75GNSFAGS‐1219382 source-schema-version-number2.0component-idgrl54395cover-date28 May 2016details-of-publishers-convertorConverter:WILEY_ML3GV2_TO_NLMPMC version:4.9.4 mode:remove_FC converted:25.08.2016 Eastwood , J. P. , et al. (2016 ), Ion‐scale secondary flux ropes generated by magnetopause reconnection as resolved by MMS , Geophys. Res. Lett. , 43 , 4716 –4724 , doi:10.1002/2016GL068747. ==== Body 1 Introduction The dayside magnetopause is an important location for studying magnetic reconnection in situ. Flux transfer events, typically identified as a bipolar signature in the component of the magnetic field normal to the magnetopause [Russell and Elphic, 1978], are a commonly observed feature of magnetopause reconnection [e.g., Rijnbeek et al., 1984; Wang et al., 2005, 2006; Kawano and Russell, 1997; Fear et al., 2008, 2009; Zhang et al., 2012]. Several competing models of flux transfer event (FTE) formation have been developed to date, based on transient/temporally varying reconnection [Scholer, 1988; Southwood et al., 1988], multiple, and possibly sequential, X‐line reconnection [Lee and Fu, 1985; Raeder, 2006], and spatially limited reconnection [Russell and Elphic, 1978]. Experimental analysis shows that FTEs are often more accurately described as flux ropes [e.g., Xiao et al., 2004; Eastwood et al., 2012]. The reported flux content of these structures is of the order of 1–10 MWb [e.g., Rijnbeek et al., 1984; Hasegawa et al., 2006], with sizes of the order of 1 to a few Earth radii (RE). More generally, flux ropes, and ion‐scale flux ropes in particular, are an important component of reconnection‐driven particle acceleration models [Drake et al., 2006a; Chen et al., 2008], arise naturally in 2‐D and fully 3‐D simulations of reconnection [Drake et al., 2006b; Daughton et al., 2011a, 2011b], and may even also modulate the reconnection rate itself [Karimabadi et al., 2007]. Efforts to understand the physics of magnetopause flux ropes and their internal structure have, to some extent, been somewhat limited by the time resolution of the available measurements, particularly plasma moments measured with a cadence of several seconds which may be comparable to the duration of some flux rope observations. Furthermore, insufficiently resolved observations may place an artificially high limit on the minimum flux rope size and thus impede attempts to accurately determine their size distribution [Fermo et al., 2011]. Here we present novel measurements from the four‐spacecraft Magnetospheric Multiscale (MMS) mission [Burch et al., 2015] of two flux ropes observed sequentially in a magnetopause reconnection jet. These flux ropes are small (radius ~ 7 magnetosheath ion inertial lengths (d i) ~ 0.17 RE) with duration of a few seconds in the data. The MMS tetrahedron was at 10 km (~0.14d i) scale, and the multipoint observations are used, in combination with computer simulation, to study their size, current density filamentation and structure, and nonideal plasma behavior. 2 Observations The MMS observations were made on 16 October 2015, 13:04:05 UT to 13:04:55 UT, shortly before MMS encountered the electron dissipation region at 13:07:02 UT [Burch et al., 2016]. Figures 1a–1n present MMS3 flux gate magnetometer (FGM) [Russell et al., 2014] and fast plasma experiment (FPI) [Pollock et al., 2016] observations using Geocentric Solar Ecliptic (GSE) coordinates. On this scale, data from the individual satellites appear identical. The data were captured at high time resolution: FPI moments are constructed from all‐sky electron and ion distributions at 30 ms and 150 ms cadence, respectively. Figure 1 (a–n) Time series of ion and electron spectrograms, magnetic field components, and strength, density, speed, ion and electron velocity, current density, and ion and electron temperature; (o, p) MMS tetrahedron configuration; and (q) cartoon of encounter. Data are shown in GSE. Figures 1o and 1p show that MMS was located duskward of the Sun‐Earth line, at [8.3, 8.5, −0.7] RE. The tetrahedron size was ~10 km, with MMS1, MMS2, and MMS3 approximately forming an equilateral triangle confined to the x GSE‐y GSE plane and MMS4 located below this plane at smaller values of z GSE. MMS was initially located in the magnetosphere (low density, northward magnetic field) and started to cross the magnetopause at 13:04:15 UT (when the plasma density began to increase, and Bz decrease), eventually reaching the magnetosheath flow proper at 13:04:40 UT. During this crossing, an enhanced flow was observed in the −vz direction when compared to the later magnetosheath interval, and MMS encountered two flux transfer events. These events are marked by vertical lines corresponding to the interval 13:04:26 UT to 13:04:32 UT (event #1) and 13:04:32 UT to 13:04:38 UT (event #2). They are identified more specifically as flux ropes (“FR#1” and “FR#2”) because of the correlated increase in |B| and By, centered on the bipolar (negative/positive) signature in Bx. Figure 1q provides a summary of the observations and their context. The red dashed line shows the trajectory of MMS relative to the magnetopause and the flux ropes. Initially located on the magnetospheric side, MMS skirted the magnetospheric edge of the first flux rope, since Bz was small but remained positive. During the encounter with the second flux rope, a bipolar signature in Bz was observed in conjunction with the bipolar signature in Bx. Furthermore, these reversals are coincident with the peak in |B| and By, indicating that when crossing the flux rope MMS cut through the center of the structure. The sense of the bipolar signatures in both events is consistent with southward flux rope motion, confirming that the two flux ropes are observed sequentially. The trajectory of the second event allows a force‐free flux rope model [e.g., Eastwood et al., 2012, and references therein] to be applied (time interval 13:04:32.75 UT to 13:04:35.5 UT) which gives an axis orientation of [−0.012, 0.989, −0.149] GSE and provides reasonable qualitative agreement to the features in the magnetic field time series (Figure 2l). Figure 2 (a–k) Time series of magnetic field components and strength, density, speed, ion and electron velocity, and current density for the first flux rope event in GSE; (l–v) same data products for the second flux rope event; (w–y) electron distribution function cuts in the v para/v perp plane at times corresponding to the vertical lines in Figures 2l–2v. Four‐spacecraft timing analysis [e.g., Schwartz, 1998] can also be applied to gain quantitative information about the orientation and motion of these structures relative to the spacecraft. In the case of a cylindrically symmetric flux rope, it can be shown that the times at which each spacecraft observes the peak field strength define a plane which is perpendicular to the tetrahedron direction of motion (in the flux rope frame) and contains the axis of the flux rope. The timing analysis result thus defines the orientation of the spacecraft trajectory through the flux rope in the plane perpendicular to the flux rope axis. Consequently, the perpendicular motion of the flux rope axis relative to the tetrahedron can be computed. FR#1 is found to be moving at v 1 ~ 223 km s−1 along n1 = [−0.375, −0.128, −0.918] GSE relative to MMS. FR#2 is found to be moving at v 2 ~ 264 km s−1 along n2 = [−0.784, −0.075, −0.617] GSE relative to MMS. The motion of both is consistent with the features of the time series in Figures 1a–1m, since MMS crossed from the leading edge of the flux rope on the magnetospheric side to the trailing edge of the flux rope on the magnetosheath side. Consequently, from the perspective of MMS the flux rope appeared to move in the −x GSE and −z GSE directions. The velocity components along x GSE and z GSE are v 2,x ~ −207 km/s and v 2,z ~ −162 km/s. Thus, FR#2 is moving slightly more slowly than FR#1 along the magnetopause in the −z GSE direction. These results, together with flux rope fitting (Figure 2l) showing the axis of the second flux rope pointing predominantly in the +y GSE direction, indicate that the GSE coordinate system is sufficient to understand the main features of the data shown in Figure 1. The duration of FR#1 (corresponding to the interval of enhanced |B| from 13:04:27 to 13:04:31) is ~4 s. The length of the chord through FR#1 is estimated to be ~ 890 km ~ 12d i since the magnetosheath density ~ 9.8 cm−3 with a corresponding ion inertial length d i = 73 km. Regarding FR#2, the fact that Bx and Bz reversed sign at the time that |B| peaked is evidence that MMS passed very close to its central axis. The duration of FR#2 (13:04:32.4–13:04:36.6) thus corresponds to a diameter of ~1100 km ~ 15d i. Both flux ropes are therefore of similar size. Finally, we note that the reconnection jet speed in the −z GSE direction is comparable to the timing‐derived speeds, and there is no evidence for converging flow in the z GSE direction across each island. Such observations (which are rare) would be associated with an “active” flux rope [Hasegawa et al., 2010; Øieroset et al., 2011, 2014]. 3 Current Density Structure and Filamentation The current density can be calculated using the curlometer technique [e.g., Robert et al., 1998; Dunlop et al., 2002], which assumes that the magnetic field varies linearly across the spacecraft tetrahedron. It is best suited to investigating phenomena where the scale size of the tetrahedron is much less than the structure under consideration. This criterion is well satisfied here, as the spacecraft separation is more than an order of magnitude less than the structure size. The high quality of the FPI data is such that the current density can also be calculated directly according to the formula J = n e e(vi − ve) where n e is the electron number density, e is the elementary coulomb charge, and vi and ve are the ion and electron velocities (we assume a quasi‐neutral proton plasma). Here the ion moments are interpolated onto the electron moments to compute J at the highest possible time resolution. Extremely good agreement between both techniques is found over the whole interval (Figures 1i–1k). Both methods reveal that there is considerable structure and variability in the current density time series. This is not immediately evident from the magnetic field time series itself, which appears rather smooth. The very good agreement between the two independent methods confirms that these features are physical. Current densities calculated using particle data from the other three satellites show similarly good agreement. Figure 2 shows the current density in each flux rope in more detail. The components of the current density show that the current is also predominantly in the +y GSE direction. As a more general observation, in FR#1, the peak current density is smaller compared to that in FR#2. This is most probably associated with the different trajectory that MMS took through each structure. The agreement between the two methods of calculating the current density is excellent. This validates the interpolation of the ion data, and the comparison reveals a fundamental feature of the plasma now resolved by MMS: the ion velocity time series is indeed considerably smoother and less variable than the electron velocity time series; this is not simply because it is sampled at lower time resolution. Note that the FPI current density is calculated at the vertex of the tetrahedron (here at MMS3), whereas the curlometer represents an “average” current density. The fact that MMS3 is north of the tetrahedron barycenter means that the FPI current time series tends to lead the curlometer time series, as the structures are moving from north to south. In fact, there is evidence that in places the current density gradients are sharper in the FPI data than in the curlometer. This may indicate that there is current density structure on scales below the tetrahedron scale. Figures 2g and 2r show the curlometer‐derived current density within each flux rope parallel and perpendicular to the magnetic field. It is clear that the current density, which is carried by the electrons, is predominantly field aligned. Correspondingly, the J × B force is found to be small, and this directly confirms the validity of the force‐free model as applied to FR#2. An interesting experimental observation is that the thermal pressure is not negligible (Figure 1n), and superthermal electrons are present in the flux ropes (Figure 1b). However, the thermal pressure is relatively uniform and isotropic throughout. This demonstrates that force‐free solutions are not restricted to low β plasmas. We take advantage of FR#2's largely force‐free nature to estimate its flux content using an analytic formula [Eastwood et al., 2012]. Here for a model peak field strength of 54.4 nT and a radius of 550 km, the flux content is Φ = 0.4158[2πB 0 R f 2 J 1(2.40482)] = 22.3 kWb. This is considerably less than the typically observed FTE flux content of ~ 1 MWb [e.g., Rijnbeek et al., 1984; Hasegawa et al., 2006]. We now examine the features of the current density time series more closely. Within FR#1, the current density in the leading edge is initially relatively weak and reduces to nearly zero at 13:04:29.1 UT. The current density then immediately increases and is enhanced over the interval where |B| is maximized until 13:04:29.9 UT where it again reduces to close to zero. A third enhancement in current density then occurs which persists until 13:04:30.5 UT. In FR#2, there is an initial channel of strong parallel current of duration ~0.15 s which peaks at 13:04:33.35 UT. This current filament is on the “magnetospheric” edge of the flux rope. The duration corresponds to a thickness of 40 km or 0.54d i. The current density then reduces to close to zero, before increasing to a peak just after 13:04:34 UT which corresponds to the maximum field strength in the core of the flux rope. The current density then falls to zero at just after 13:04:34.4 UT before increasing and then decreasing again. A further current filament is seen between 13:04:35.75 and 13:04:36.3 UT which sits near the “magnetosheath” edge of the flux rope. The electron populations that carry these currents are resolved by MMS. For example, Figures 2w–2y show three examples of electron distributions captured before and during FR#2, presented in the v para/v perp plane, and measured at 30 ms cadence with the midtime corresponding to the vertical lines in Figures 2l–2v. Whereas the distribution in Figure 2w is taken at a quiet time before FR#2 to provide context, in both the current filament and at the center of the flux rope (Figures 2x and 2y) the current is carried by a relatively cold population moving antiparallel to the magnetic field. There is also the signature of a hotter population moving parallel to the magnetic field, which indicates the presence of magnetospheric plasma and is consistent with the expected 3‐D topology of magnetopause flux ropes. 4 Nonideal Plasma Behavior In both these flux rope events, there is predominantly parallel current. We now examine evidence for nonideal behavior associated with differential perpendicular motion, that is, deviations from E + vi,e × B = 0. Note that a deviation in only one component is required for the vector E + v × B to be nonzero. Figure 3 compares data from the spin plane electric field double probes (shown in black) [Lindqvist et al., 2014] with −vi × B and −ve × B (shown in red). The data are presented in the spacecraft coordinate system, which is close to GSE. Although only MMS3 data are shown, it should be emphasized that the same features are visible on all four MMS satellites. Overall, there is good agreement between the time series (disregarding small systematic constant offsets), but there are intervals where there is significant deviation in the ion plasma, whereas the electron plasma remains frozen in. Figure 3 (a–i) MMS3 time series of magnetic field components and strength, density, ion and electron velocity, and spin plane electric field; (j–r) same data products for the second flux rope event. Data are shown in spacecraft coordinate system. For example, in FR#1, such deviations are seen at 13:04:29.1–13:04:29.2 UT and 13:04:30.0–13:04:30.2 UT (marked by vertical lines). These intervals are adjacent to the enhanced central current density channel in FR#1. Throughout the flux rope Ex is negative. During both deviations Ex is more negative than −(vi × B)x and so since the magnetic field points predominantly in the +By direction, this indicates that the ions are moving more slowly than the electrons in the axial direction (~ −z GSE) direction. In FR#2 a first deviation in ion behavior is seen at 13:04:33.6–13:04:33.9 UT. This deviation occurs after the strong parallel filamentary current at the magnetospheric edge of the flux rope (dashed line). The first interval of non‐frozen‐in ion behavior is therefore inside the flux rope. Again, Ex is negative, and Ex is more negative than −(vi × B)x, so this similarly indicates that the ions are moving more slowly. A second deviation is seen at 13:04:35.75–13:04:36.25 UT. There is a negative/positive bipolar signature in Ex. This deviation is at the same time, and of similar duration, as the current filament on the magnetosheath edge of the flux rope, and so this motion is associated with the small perpendicular current observed there (Figure 2r). In contrast to the ions, the agreement between E x,y and (−ve × B) x,y is extremely good throughout both events, showing that the electrons are largely frozen in throughout the encounter with both flux ropes. However, there is tentative evidence for non‐frozen‐in electron behavior, based on the deviation in E x,y at 13:04:27.4 UT (dash‐dotted line). This will be the subject of future investigation. 5 Comparison With Simulations To better understand the MMS data, and FR#2 in particular, Figure 4 shows the results of a particle‐in‐cell simulation of reconnection performed using the P3D code [Zeiler et al., 2002], described in the supporting information. Figure 4 Particle‐in‐cell simulation of reconnection with cuts through a flux rope at z = −6.35. See text for details. Figures 4a and 4b show the number density and the magnetic field strength. The magnetosheath (high density, low magnetic field strength) is above the current sheet where the magnetic field points to the left (Bz < 0) and the “magnetosphere” (low density, high magnetic field strength) is below where the magnetic field points to the right (i.e., Bz > 0). We focus on the flux rope that has formed on the current sheet between z = −4 and z = −10. This flux rope, comparable in size to those observed by MMS, is moving to the left as the simulation evolves (equivalent to moving southward along the −z GSE direction). Figures 4c and 4d show maps of the Jx and Jy current density (red: positive, blue: negative). Jy, the out‐of‐plane current density, is enhanced within the island but stronger on the magnetospheric edge. Jx is enhanced around the edge of the island. Figure 4e shows the magnetic field profile along a vertical cut at z = −6.35. In this plot the magnetosphere is on the left‐hand side. Figure 4f shows the profile of Jy total (black), ion current density Jiy (red), and electron current density Jey (blue). On the magnetospheric edge there is a spike (thickness ~0.4) in the current density carried by the electrons whose magnitude, converted to SI units, is 810 nAm−2, which is very similar to that observed (Figure 2s). Further analysis shows that this current spike is largely field aligned. This feature is identified as the counterpart of the current spike at the start of FR #2 at 13:04:33.35 UT. Referring more closely to this feature in Figure 2s, it can be seen that there is also a component of the current in the −x GSE direction. This is on the magnetospheric side of the flux rope near its leading edge where B xGSE is negative and corresponds to the blue region at the bottom left edge of the flux rope in Figure 4c. Figure 4e shows that By > 0 inside the simulated flux rope, and so Jx is expected to be negative to the left of the flux rope center and positive to the right which is observed in Figure 4c. We therefore conclude that the filamentary current observed by MMS at the leading edge of FR#2 corresponds to the thin parallel current layer that forms in the simulation at x = −0.6 in Figures 4d and 4f. Finally, Figure 4g shows (E + vi × B)x/Ex. This immediately reveals that there are regions of non‐frozen‐in ion flow confined within the flux rope and that this is not homogenous. This is confirmed by Figure 4h which is a cut along z = −6.35 showing the nature of the frozen in ion (red) and electron (blue) flow. The ions are not frozen in, except for a region near the center of the flux rope. This is qualitatively consistent with the MMS observations. 6 Discussion and Conclusions The MMS observations reveal in new detail the existence of ion‐scale flux rope FTEs entrained in a magnetopause reconnection jet. The current, carried by the electrons, is largely parallel, and so the overall structure is well described by a relatively simple force‐free model. However, multispacecraft analysis reveals that the flux rope current density is highly structured with filament thicknesses that can be comparable to the ion inertial length. There are also regions of nonideal ion flow adjacent to these current layers and interior to the flux ropes. The second event was more conducive to comparison with simulations because of the MMS trajectory through the flux rope center. The simulation is 2.5‐D but nevertheless reproduces many of the key observed features of size, current filamentation, and non‐frozen‐in ion behavior. However, the core field is not very strongly enhanced, because of pressure balance with compressed plasma in the flux rope core that cannot escape. Also, in contrast to the observations, the simulated flux ropes are only found to be force‐free in the core region. Future comparisons with fully 3‐D simulations should even more accurately reproduce MMS observations. FR#2 is found to have a flux content of 22.3 kWb which is considerably less than typically observed FTEs. A natural question to ask is whether these FTEs, which are small, are in the process of growing into larger structures. We do not observe a converging v zGSE flow that would indicate the presence of active X lines, and thus flux being added to the structure. On the other hand, if these structures were simply to expand to typical FTE dimensions, retaining typical flux content, they would require an initial core field strength in excess of 1000 nT which is not observed. Thus, we conclude that these events are different in nature from previously reported magnetopause flux rope events. It is a well‐known feature of simulations and also magnetotail observations that flux ropes may be generated as a consequence of secondary reconnection processes at a preexisting X line [e.g., Daughton et al., 2011a]. In the simulation presented here, although the flux ropes are transient and arise not long after the initialization of the code, their formation is related to the onset of secondary instabilities in the vicinity of an X line and so provide insight into the structure and properties of flux ropes produced in this way. Given the ion‐scale size, the small and stable flux content, the filamentary current structure, and the non‐frozen‐in plasma properties, together with the comparison with simulation, the analysis presented here suggests that these flux ropes are produced by secondary reconnection processes occurring in the vicinity of a single X line. This is not strictly the same as any of the three previously identified mechanisms discussed in section 1. Subsequent analysis of the MMS data set should reveal further information about the occurrence and statistical properties of such ion‐scale flux ropes. Supporting information Supporting Information S1 Click here for additional data file. Acknowledgments Simulations and analysis were performed at NCAR‐CISL and at NERSC. For MMS data visit https://lasp.colorado.edu/mms/sdc/public/. The simulation data used to produce the results of this paper are available from the authors. This work was supported by STFC (UK) (ST/G00725X/1), NASA (NNX08AO83G, NNX13AD72G, and NNX16AF75G), and NSF (AGS‐1219382). IRAP contribution to MMS was supported by CNES. ==== Refs References Burch , J. L. , T. E. Moore , R. B. Torbert , and B. L. Giles (2015 ), Magnetospheric multiscale overview and science objectives , Space Sci. Rev. , 1 –17 , doi:10.1007/s11214-015-0164-9. Burch , J. L. , et al. (2016 ), Electron‐scale measurements of magnetic reconnection in space , Science , doi:10.1126/science.aaf2939. Chen , L.‐J. , et al. (2008 ), Observation of energetic electrons within magnetic islands , Nat. Phys. , 4 , 19 –23 . Daughton , W. , V. Roytershteyn , H. Karimabadi , L. Yin , B. J. Albright , B. Bergen , and K. J. Bowers (2011a ), Role of electron physics in the development of turbulent magnetic reconnection in collisionless plasmas , Nat. Phys. , 7 (7 ), 539 –542 . Daughton , W. , V. Roytershteyn , H. Karimabadi , L. Yin , B. J. Albright , S. P. Gary , and K. J. Bowers (2011b ), Secondary island formation in collisional and collisionless kinetic simulations of magnetic reconnection , AIP Conf. Proc. , 1320 (1 ), 144 –159 , doi:10.1063/1.3544319. Drake , J. F. , M. Swisdak , H. Che , and M. A. Shay (2006a ), Electron acceleration from contracting magnetic islands during reconnection , Nature , 443 , 553 –556 .17024088 Drake , J. F. , M. Swisdak , K. M. Schoeffler , B. N. Rogers , and S. Kobayashi (2006b ), Formation of secondary islands during magnetic reconnection , Geophys. Res. Lett. , 33 , L13105, doi:10.1029/12006GL025957. Dunlop , M. W. , A. Balogh , K.‐H. Glassmeier , and P. Robert (2002 ), Four‐point cluster application of magnetic field analysis tools: The curlometer , J. Geophys. Res. , 107 (A11 ), 1384, doi:10.1029/2001JA0050088. Eastwood , J. P. , T. D. Phan , R. C. Fear , D. G. Sibeck , V. Angelopoulos , M. Oieroset , and M. A. Shay (2012 ), Survival of flux transfer event (FTE) flux ropes far along the tail magnetopause , J. Geophys. Res. , 117 , A08222, doi:10.1029/2012JA017722. Fear , R. C. , S. E. Milan , A. N. Fazakerley , E. A. Lucek , S. W. H. Cowley , and I. Dandouras (2008 ), The azimuthal extent of three flux transfer events , Ann. Geophys. , 26 , 2353 –2369 . Fear , R. C. , S. E. Milan , A. N. Fazakerley , K.‐H. Fornacon , C. M. Carr , and I. Dandouras (2009 ), Simultaneous observations of flux transfer events by THEMIS, Cluster, Double Star and SuperDARN: Acceleration of FTEs , J. Geophys. Res. , 114 , A10213, doi:10.1029/12009JA014310. Fermo , R. L. , J. F. Drake , M. Swisdak , and K.‐J. Huang (2011 ), Comparison of a statistical model for magnetic islands in large current layers with Hall MHD simulations and Cluster FTE observations , J. Geophys. Res. , 116 , A09226, doi:10.1029/2010JA016271. Hasegawa , H. , B. U. Ö. Sonnerup , C. J. Owen , B. Klecker , G. Paschmann , A. Balogh , and H. Rème (2006 ), The structure of flux transfer events recovered from Cluster data , Ann. Geophys. , 24 , 603 –618 . Hasegawa , H. , et al. (2010 ), Evidence for a flux transfer event generated by multiple X‐line reconnection at the magnetopause , Geophys. Res. Lett. , 37 , L16101, doi:10.1029/2010GL044219. Karimabadi , H. , W. Daughton , and J. Scudder (2007 ), Multi‐scale structure of the electron diffusion region , Geophys. Res. Lett. , 34 , L13104, doi:10.1029/2007GL030306. Kawano , H. , and C. T. Russell (1997 ), Survey of flux transfer events observed with the ISEE 1 spacecraft: Dependence on the interplanetary magnetic field , J. Geophys. Res. , 102 (A6 ), 11,307 –11,313 , doi:10.1029/97JA00481. Lee , L. C. , and Z. F. Fu (1985 ), A theory of magnetic flux transfer at the Earth's magnetopause , Geophys. Res. Lett. , 12 , 105 –108 , doi:10.1029/GL012i002p00105. Lindqvist , P.‐A. , et al. (2014 ), The spin‐plane double probe electric field instrument for MMS , Space Sci. Rev. , 1 –29 , doi:10.1007/s11214-014-0116-9. Øieroset , M. , et al. (2011 ), Direct evidence for a three‐dimensional magnetic flux rope flanked by two active magnetic reconnection X lines at Earth's magnetopause , Phys. Rev. Lett. , 107 , 165007. Øieroset , M. , D. Sundkvist , C. C. Chaston , T. D. Phan , F. S. Mozer , J. P. McFadden , V. Angelopoulos , L. Andersson , and J. P. Eastwood (2014 ), Observations of plasma waves in the colliding jet region of a magnetic flux rope flanked by two active X lines at the subsolar magnetopause , J. Geophys. Res. Space Physics , 119 , 6256 –6272 , doi:10.1002/2014ja020124. Pollock , C. , et al. (2016 ), Fast plasma investigation for magnetospheric multiscale , Space Sci. Rev. , doi:10.1007/s11214-016-0245-4. Raeder , J. (2006 ), Flux transfer events: 1. Generation mechanism for strong southward IMF , Ann. Geophys. , 24 , 381 –392 . Rijnbeek , R. P. , S. W. H. Cowley , D. J. Southwood , and C. T. Russell (1984 ), A survey of dayside flux transfer events observed by ISEE 1 and 2 magnetometers , J. Geophys. Res. , 89 , 786 –800 , doi:10.1029/JA089iA02p00786. Robert , P. , M. W. Dunlop , A. Roux , and G. Chanteur (1998 ), Accuracy of current density determination, in Analysis Methods for Multi‐spacecraft Data , edited by Paschmann G. and Daly P. W. , pp. 395 –418 , International Space Science Institute , Bern . Russell , C. T. , and R. C. Elphic (1978 ), Initial ISEE magnetometer results: Magnetopause observations , Space Sci. Rev. , 22 , 681 –715 . Russell , C. T. , et al. (2014 ), The magnetospheric multiscale magnetometers , Space Sci. Rev. , 1 –68 , doi:10.1007/s11214-014-0057-3. Scholer , M. (1988 ), Magnetic flux transfer at the magnetopause based on single X‐line bursty reconnection , Geophys. Res. Lett. , 15 , 291 –294 , doi:10.1029/GL015i004p00291. Schwartz , S. J. (1998 ), Shock and discontinuity normals, Mach numbers and related parameters, in Analysis Methods for Multi‐Spacecraft Data , edited by Paschmann G. and Daly P. W. , pp. 249 –270 , International Space Science Institute , Bern . Southwood , D. J. , C. J. Farrugia , and M. A. Saunders (1988 ), What are flux transfer events? Planet. Space Sci. , 36 , 503 –508 . Wang , Y. L. , et al. (2005 ), Initial results of high‐latitude magnetopause and low‐latitude flank flux transfer events from 3 years of Cluster observations , J. Geophys. Res. , 110 , A11221, doi:10.1029/2005JA011150. Wang , Y. L. , R. C. Elphic , B. Lavraud , M. G. G. T. Taylor , J. Birn , C. T. Russell , J. Raeder , H. Kawano , and X. X. Zhang (2006 ), Dependence of flux transfer events on solar wind conditions from 3 years of Cluster observations , J. Geophys. Res. , 111 , A04224, doi:10.1029/2005JA011342. Xiao , C. J. , Z. Y. Pu , Z. W. Ma , S. Y. Fu , Z. Y. Huang , and Q.‐G. Zong (2004 ), Inferring of flux rope orientation with the minimum variance analysis technique , J. Geophys. Res. , 109 , A11218, doi:10.1029/2004JA010594. Zeiler , A. , D. Biskamp , J. F. Drake , B. N. Rogers , and M. A. Shay (2002 ), Three‐dimensional particle simulations of collisionless magnetic reconnection , J. Geophys. Res. , 107 (A9 ), 1230, doi:10.1029/2001JA000287. Zhang , H. , M. G. Kivelson , V. Angelopoulos , K. K. Khurana , Z. Y. Pu , R. J. Walker , R. L. McPherron , T. S. Hsu , Q. G. Zong , and T. Phan (2012 ), Generation and properties of in vivo flux transfer events , J. Geophys. Res. , 117 , A05224, doi:10.1029/2011JA017166.

PMC005xxxxxx/PMC5001196.txt

==== Front Prev Vet MedPrev. Vet. MedPreventive Veterinary Medicine0167-58771873-1716Elsevier Scientific Publishing S0167-5877(16)30222-710.1016/j.prevetmed.2016.07.017ArticleThe attitudes of owners and veterinary professionals in the United Kingdom to the risk of adverse events associated with using non-steroidal anti-inflammatory drugs (NSAIDs) to treat dogs with osteoarthritis Belshaw Zoe z.belshaw.97@cantab.net⁎Asher Lucy 1Dean Rachel S. Centre for Evidence-based Veterinary Medicine, School of Veterinary Medicine and Science, University of Nottingham, Sutton Bonington Campus, Loughborough, Leicestershire, LE12 5RD UK⁎ Corresponding author. z.belshaw.97@cantab.net1 Present address: Centre for Behaviour and Evolution, Institute of Neuroscience, Newcastle University Henry Wellcome Building, Framlington Place, Newcastle, NE2 4HH, UK. 01 9 2016 01 9 2016 131 121 126 4 2 2016 26 7 2016 31 7 2016 © 2016 The Authors2016This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).Highlights • Many dogs with osteoarthritis receive NSAIDs for long time periods. • The risk of NSAID-related adverse events concerns veterinary surgeons and dog owners. • Many strategies are adopted to minimise the risk of these adverse events. • These strategies may compromise animal welfare through inadequate analgesia. • The evidence base to support these strategies is inadequate and should be addressed. Non-steroidal anti-inflammatory drugs (NSAIDs) are commonly prescribed by veterinary surgeons for the treatment of canine osteoarthritis, and affected dogs may receive these drugs for long periods of time. Whilst short term administration of NSAIDs to dogs is linked to adverse events such as gastrointestinal haemorrhage and renal injury, reports of adverse events associated with their long-term administration are limited in the veterinary literature. This study aimed to investigate the attitudes towards the long term use of NSAIDs for canine osteoarthritis held by three groups who manage osteoarthritic dogs in the United Kingdom: dog owners, veterinary surgeons and veterinary nurses. A qualitative methodology was adopted, using semi-structured interviews and focus groups. Thematic analysis of these data identified three themes: awareness of potential risks; recognition of adverse events; and influence of risk perception on the use of NSAIDs. Awareness of, and concern about, the risk of adverse events associated with NSAID administration to dogs with osteoarthritis was high in all groups, with veterinary surgeons being one of a variety of information sources used by owners to acquire this knowledge. Veterinary surgeons described difficulty in recognising, managing and avoiding adverse events associated with NSAIDs. When adverse events occurred, a wide range of management approaches were adopted ranging from a brief drug respite to permanent cessation of administration of any NSAIDs to that dog. Commonly employed approaches to minimise risk included dose reduction and screening blood tests. This study describes a high level of concern about the risks associated with long term NSAID administration to dogs with osteoarthritis and highlights a diverse range of strategies employed to minimise these risks. The evidence base for these strategies is poor, and this may present a risk to animal welfare if the affected dogs are not receiving adequate analgesia. In order to address this, more accurate and comprehensive data must be supplied to both veterinary professionals and owners on the true frequency of adverse events associated with long term administration of veterinary NSAIDs and how best to avoid them. Keywords Non-steroidal anti-inflammatory drugAdverse eventRiskQualitativeDogOsteoarthritis ==== Body 1 Introduction Non-steroidal anti-inflammatory drugs (NSAIDs) are commonly used to treat dogs with osteoarthritis (Sanderson et al., 2009). Systematic reviews have assessed the efficacy (Aragon et al., 2007, Sanderson et al., 2009) and safety of veterinary NSAIDs (Innes et al., 2010, Monteiro-Steagall et al., 2013). However, studies reporting long term safety are lacking (Innes et al., 2010), and reporting of adverse events may be incomplete, both in clinical trials and passive surveillance (Hunt et al., 2015). Concern about adverse events associated with the administration of NSAIDs to small animals remains a barrier to their prescription by veterinary surgeons (Bell et al., 2014). The attitudes of pet owners regarding NSAID safety have not previously been reported in peer reviewed literature, but survey data has found a link between safety concerns and poor compliance (Zoetis Inc., 2013). This may have serious implications for the welfare of animals not receiving prescribed analgesia. The attitudes of patients (Carnes et al., 2008, Milder et al., 2011a, Laba et al., 2013) and healthcare professionals (Mikhail et al., 2007, Cavazos et al., 2008, Braund and Abbott, 2011) to the use and safety of NSAIDs for human osteoarthritis are well documented. Several studies have demonstrated an important role for doctors in educating patients about potential adverse events associated with NSAIDs (Mikhail et al., 2007, Schmitt et al., 2011). The prescribing practice of healthcare professionals is directly influenced by their experiences, leading to avoidance of treatments that have caused adverse events in their own patients (Gabbay and May, 2004, Cavazos et al., 2008). Conversely, patients reliant on NSAIDs for analgesia may ignore the risks if they have not personally experienced an adverse event (Milder et al., 2011b), though perceived risks have been associated with the decision to discontinue NSAID treatment (Laba et al., 2013). Qualitative research uses a range of approaches including interviews and focus groups to explore motivations, attitudes and opinions (Bryman, 2012). A qualitative approach is particularly useful in areas where little previous research has been conducted and results can be used to inform subsequent quantitative studies (e.g. Wiseman-Orr et al., 2004). Analysis of data derived by qualitative research takes many forms. Thematic analysis (described by Braun and Clarke, 2006) is widely used in healthcare research due to its flexible methodology. The process is slow and iterative with transcribed lines of text coded for their meaning, with codes then collated into larger themes. The reports generated do not feature numbers but instead provide an analytic narrative which makes an argument related to the research question. A growing body of literature derived from qualitative research exists in veterinary medicine (e.g. Coyne et al., 2014, Horseman et al., 2014, Page-Jones and Abbey, 2015). The aim of this study was to use a qualitative approach to characterise the attitudes of dog owners, veterinary surgeons and veterinary nurses in England and Scotland towards the safety of NSAIDs used to treat osteoarthritis in dogs. The objective was to perform in-depth qualitative interviews with dog owners and focus groups with veterinary surgeons and veterinary nurses to capture a wide range of experiences of using NSAIDs to treat dogs with osteoarthritis. 2 Materials and methods This work forms a part of a larger qualitative study exploring the experiences of dog owners, veterinary surgeons and veterinary nurses in the United Kingdom managing dogs with osteoarthritis. Only results relevant to the aims of this study will be reported. Data were collected using semi-structured interviews for dog owners and focus groups for veterinary professionals (veterinary surgeons and nurses). Ethical approval for the study was granted by the ethics committee at the School of Veterinary Medicine and Science, University of Nottingham. Reporting follows the COREQ checklist (Tong et al., 2007). 2.1 Owner interviews Interviews were conducted between February and August 2014. The inclusion criteria for interviewees were: a) ownership of a dog currently treated or managed for osteoarthritis affecting at least one limb due to any underlying aetiology; AND b) ownership of a dog at least five years of age at the time of the interview; AND c) the owner(s) and dog must live in the United Kingdom. Recruitment was based on a purposive sampling frame constructed by the authors (available on request) containing both dog and owner variables to capture the widest possible range of experiences. Most interviewees were recruited by placing posters in the waiting rooms of a convenience sample of 10 veterinary practices in England and Scotland. All practices had previously agreed to collaborate with the Centre for Evidence-based Veterinary Medicine, University of Nottingham. Posters asked owners of older dogs with osteoarthritis to contact the lead author by email or telephone to share their experiences. Additional interviewees were recruited by snowball sampling or through the authors’ networks. Owners who expressed an interest in participation were sent information about the purpose of the study including details of the interviewer’s background as a veterinary surgeon and owner of a dog with osteoarthritis. If they agreed to participate, an interview date was then arranged by email or telephone. Incentives to participate were not provided. Interviews were conducted in the owners’ homes by one researcher (ZB) who had received training in qualitative research from the Health Experiences Research Group, University of Oxford. A semi-structured interview guide (available on request), piloted before use, was used to explore owner experiences. Pertinent to this manuscript, interviewees were asked about their experiences of treating their dog’s osteoarthritis, and any previous experience of osteoarthritis in humans or other species. Where owners had more than one dog with osteoarthritis, they chose either to focus on one dog or described their experiences with more than one dog. Interviews were conducted until data saturation was reached (see below). 2.2 Veterinary focus groups Focus groups were conducted between August and December 2014. The inclusion criterion was any practice from which owners had been recruited; veterinary nurses and veterinary surgeons that performed consultations with owners of dogs with osteoarthritis within that practice were then invited to participate. A purposive sampling frame ensured inclusion of a range of practice sizes, locations and types. Meetings were arranged through contact with one member of staff at the practice who then recruited others to participate; attendance was voluntary. Details of the purpose of the study were available in advance. Food was provided as an incentive to attend meetings and to ease discussion (Braun and Clarke, 2013). Focus groups were conducted on the practice premises for convenience and to provide a safe setting for open discussion. The focus group for veterinary nurses was conducted at a different time to that of the veterinary surgeons within the same practice to ensure both groups were comfortable to describe their experiences. All focus groups were conducted by one researcher (ZB). A semi-structured question list was used to prompt the participants but where possible discussion was allowed to proceed with minimal interruption. Questions focused on the diagnosis and treatment of osteoarthritis in dogs. 2.3 Thematic analysis Contextual field notes were made during the interviews and focus groups. Interviews and focus groups were audio recorded and professionally transcribed verbatim. Transcripts were not returned to the participants. Transcribed interviews and focus groups were read and reviewed several times by the lead author. Thematic analysis was performed following the six step plan described by Braun and Clarke (2006) using the organisational support of nVivo (nVivo v10, QSR). Constant comparison was used to ensure all opinions were included (Braun and Clarke, 2006, Coyne et al., 2014). Thematic analysis was performed in tandem with data collection. Data saturation for interviews and focus groups was defined as the point at which no additional themes emerged as a result of analysing new transcripts; at this point recruitment for further participants was halted. For the purpose of this secondary analysis, all extracts from both interviews and focus groups that described use of NSAIDs were scrutinised, and coded into anticipated and emergent themes, as described by Ziebland et al. (2004). Statistical analysis was not performed as the qualitative purposive sampling methodology aimed to capture a wide range of experiences rather than to represent a population (Ziebland et al., 2004, Silverman, 2013). 3 Results Forty-nine owners expressed interest in the study. Seventeen were either ineligible to participate, declined participation when provided with study details or did not have the time to be interviewed during the study period. Thirty-two interviews were conducted, capturing the views of 40 owners about 35 dogs with osteoarthritis treated at seventeen different veterinary practices. Interviews ranged from 52 to 170 min in length. All dogs discussed by owners had received at least one of five different NSAIDs (carprofen, cimicoxib, firocoxib, meloxicam, robenacoxib) for their osteoarthritis. Five focus groups, each of approximately 60 min length, were run in four veterinary practices totalling 31 participants. Four focus groups were conducted with veterinary surgeons. A single focus group was conducted with veterinary nurses in the only practice providing nurse clinics for dogs with osteoarthritis. Thematic analysis of the coded excerpts relating the use of NSAIDs identified three themes: awareness of potential risks; recognition of adverse events; and the influence of risk perception on the use of NSAIDs. Due to the word limits of this publication, exemplary quotations are provided to illustrate each theme. 3.1 Awareness of potential risks Almost all owners were aware of one or more “side effects” associated with the use of NSAIDs in dogs. Many drew analogies between NSAIDs used in humans and those prescribed for their dog. Rarely, owners had personal experience of adverse events due to NSAIDs in people such as stomach ulceration. A few owners had been told about adverse events associated with NSAID affecting the dogs of friends or colleagues, and a couple had experienced adverse events with previous dogs. Most frequently, owners expressed concern about NSAID administration causing “organ damage” and less often gastrointestinal injury including vomiting, diarrhoea and ulceration. Some owners identified specific organs such as the liver or kidneys which could be harmed. Adverse events were thought to be associated with both longevity of dosing and high drug doses. Some owners recalled warnings provided by the veterinary surgeon at the time of initial NSAID administration. Often these included a comparison with human NSAIDs, which may have assumed that owners would have knowledge of their associated adverse events, and advice to give the medication with food to avoid stomach ulcers. Very rarely, owners recalled being provided with printed information. A few owners asked the veterinary surgeon about the potential harms if no details were given. Drug data sheets were used as a source of information by a couple of owners; others remarked that the data sheet was not always available. One owner highlighted the importance of this:Now, the vet didn't tell me this, but I always read the leaflet anyway, and it said with this one [robenacoxib] that it's better to wait at least thirty minutes after food before you give the tablet. [Owner 5] The internet was a source of information about NSAIDs for many owners. Most described performing a search using Google using terms such as “dog arthritis” to look for information about treatment options. A few performed a more focused search to look for information about adverse events associated with particular prescription treatments. Their motivation in all cases was to seek information to supplement that provided by their veterinary surgeon. Several owners described finding websites which alarmed them, particularly regarding reports of deaths associated with carprofen. A few described their shock that a veterinary surgeon would prescribe a drug associated with death, though many were aware that they should not trust everything they read on the internet:The vets didn't tell me anything about it, and I'm afraid I looked it up on the internet, and found that Rimadyl does disagree with a lot of dogs. There's an awful lot of nasty stories out there about poisoning and all the rest but fortunately I'm well aware of internet world. [Owner 11] A few owners described the newer generation NSAIDs as safer based on fewer reports of adverse events on the internet; for example, one owner who was very concerned about carprofen thought cimicoxib had no potential for harm. Several owners said they would appreciate more guidance from their veterinary surgeon about how to interpret what they read online or which sites to trust. Many said they would always want to check what they read on the internet with their veterinary surgeon. Surprisingly, a couple of owners were unaware of any adverse events associated with the drugs, emphasising the important role of the veterinary surgeon in educating owners:No, I'm not aware of any side effects [associated with carprofen]. I think if there were side effects [my vets] would tell me so I tend not to Google dogs. [Owner 30] All veterinary professionals were aware of the potential for adverse events with NSAIDs; concerns about hepatic and renal compromise and gastrointestinal ulceration were commonly mentioned. Many veterinary surgeons discussed the safety advice they gave to owners; typically, this matched what the owners described. Most veterinary surgeons and veterinary nurses commented that time pressure in the osteoarthritis consultation was a barrier to a longer discussion. This was compounded by many diagnoses of osteoarthritis being made during a consultation for another problem. Some veterinary professionals thought most of their owners were aware of the risks associated with NSAID administration, whilst others were less sure:I don't think many of them know in advance. … I don't think in reality before you start them on these things, I don't think many clients have much of an inkling as to the potential side effects [Veterinary surgeon 1, Focus Group 2] I think a lot of owners are aware that, particularly the well-known Metacam, a lot of people are aware that it can be detrimental to the health. [Veterinary nurse 3, Focus Group 4] 3.2 Recognition of adverse events Some owners reported that their dog had experienced one or more episodes of gastroenteritis whilst receiving an NSAID. Typically this involved vomiting with or without blood, and/or diarrhoea containing fresh blood or melena. Often, gastroenteritis occurred soon after treatment inception but some dogs experienced signs after prolonged NSAID dosing. Some dogs experienced a single episode of vomiting or diarrhoea whilst others exhibited clinical signs for longer.[My vet] said ‘Let's try Previcox.' and then she was then having problems with it, she was bringing up frothy stuff, but it had little streaks of blood in it. [Owner 15] Owners reported that gastroprotectants such as ranitidine and sucralfate were commonly prescribed after such episodes, and sometimes their use was long-term if NSAIDs were continued. Once signs had abated, some dogs were re-introduced to the same NSAID and experienced no further adverse events, whilst others were switched to a different NSAID or an alternative analgesic such as tramadol. This appeared to be determined by the individual approach of the veterinary surgeon in charge of that case. Only a couple of dogs experienced haematemesis on multiple different NSAIDs; most dogs appeared to tolerate a different NSAID much better.She started on Previcox. And that gave her a bleed … Yeah, her poos came all suddenly full of blood. I stopped it immediately, took her back [to the vets], we gave her a break, and then changed her to Rimadyl, and she's been okay on that ever since. [Owner 16] Other adverse events associated by owners with NSAIDs were reported much less frequently. No owners reported renal or hepatic compromise as a result of the drugs, though one dog was investigated for hepatic side effects as a result of increasing liver enzymes on routine biochemistry.About three months later, because they're monitoring her regularly, [her liver values had] gone down again, then they went up again, then they'd gone down again, so we'd decided that she's just erratic. But it was then the vet said “Oh dear, we'd better have a biopsy.” and all the rest of it. Ooh! Panic! It was very expensive… I'm glad we did it because we ruled out she hadn't got cancer or anything. But I still don't know whether it was the Rimadyl that pushed them up a bit. [Owner 7] Veterinary professionals disagreed both within and between practices about the incidence of adverse events, with the veterinary nurses perceiving the incidence to be highest. However, most agreed that adverse events were typically associated with post-operative NSAID use rather than with dogs receiving NSAIDs for osteoarthritis.I think it’s one of these where there is no real number. The clients say to you, “What is the risk of vomiting and diarrhoea?” And normally I say about one in 40. I don’t think it’s any less frequent than that because I think it’s one of those things where probably each of us will diagnose it, to some extent or another, at least once a week. [Veterinary surgeon 1, Focus Group 1] I think sometimes we're attributing [gastrointestinal] signs to the non-steroidals and they're probably not purely caused by non-steroidals. The number of animals that we see with [gastrointestinal] upsets, which look like non-steroidal side effects, where they haven't taken any, is probably bigger than the ones we see that do. So, but obviously you have to err on the side of caution, you have to assume it is the medication that could cause the side effects…. [Veterinary surgeon 3, Focus Group 2] It was apparent that reporting of adverse events associated with NSAIDs to the Veterinary Medicine Directorate was extremely uncommon, with difficultly recognising true adverse events being one reason for this. Many veterinary surgeons associated adverse events with poor owner compliance, citing examples of incorrect dosing, the drugs being given without food or continued administration to an ill dog as common causes. No veterinary surgeons could recall a case of severe hepatic or renal toxicity related to NSAIDs. 3.3 Influence of risk perception on the use of NSAIDs There was a clear impact on owners and veterinary professionals of the awareness of potential adverse events related to NSAID administration. Many strategies were employed to minimise risks. Several owners stated that their dog should or could not be treated with NSAIDs. Typically this was based on the dog having experienced an adverse event whilst receiving an NSAID. Most veterinary surgeons described their frustration at being unable to persuade the owners of some dogs with osteoarthritis to ever use NSAID treatments. A few owners held extremely strong beliefs about the association between adverse events and specific NSAIDs.I'm absolutely anti Rimadyl. It's on my notes that never, ever give my dogs Rimadyl. And then when I find they’ve had Rimadyl I go mad. My parents' cocker spaniel died of platelet eruption on Rimadyl, at seven… It might be good pain relief but I will not have it. [Owner 3] Some veterinary surgeons elected to avoid NSAIDs in animals they perceived to be at high risk of adverse events, typically due to slight abnormalities on blood tests. The alternative treatment prescribed was usually tramadol, though several veterinary surgeons questioned its evidence base as an analgesic for canine chronic pain:I don’t know what the studies are about what to commonly do, but if anything I think as a practice we’re quite overly cautious I think about [using NSAIDs] − which is a good thing. Sometimes I think am I being overly cautious by starting on tramadol instead of on Loxicom or something; I’m not too sure. [Veterinary surgeon 5, Focus Group 1] Several veterinary surgeons and owners described certain NSAIDs as being “safer” than others; cimicoxib, firocoxib and robenacoxib were typically mentioned in this context. A few veterinary surgeons used one of these drugs as a first line for safety reasons, whilst others switched if a dog experienced something believed to be an adverse event. The subjective nature of these decisions was frequently clear when examples were discussed:… if [gastroenteritis] came on almost instantaneously, within days of starting carprofen, and the animal was really ill, then I would be really ‘Maybe we can‘t really use non-steroidals with this.' But if on the other hand it came on after a few weeks, and it wasn't particularly bad, then yeah, take it off the carprofen for a while, give it a wash-out period, get his tummy back to being normal and then try on Previcox or Cimalgex. [Veterinary surgeon 1, Focus Group 2] Many owners and veterinary surgeons described their desire to reduce the dose of NSAIDs. Some owners asked their veterinary surgeon to reduce the dose and many veterinary surgeons routinely lowered the daily dose over time. Veterinary surgeons often expressed concern about owners taking NSAID dose reduction decisions into their own hands. This was in sharp contrast to their attitude to other treatments such as tramadol and supplements where veterinary surgeons were typically happy for owners to modify the dose. A less common strategy was to reduce dosing frequency to alternate days. Most veterinary surgeons expressed uncertainty about whether, when and how to reduce the dose; within a practice there were frequently several methods. Many acknowledged that they did not know how best to proceed but most seemed confident that lower doses were effective.A lot of them do seem fine…if you've got a thirty-kilo dog, they seem quite happy on a twenty-kilo dose. And the owners see all the benefit from it… [Veterinary surgeon 2, Focus Group 5] Some owners reported that their dogs underwent regular screening blood tests whilst on NSAIDs; others could not recall their dog ever having been tested. Blood tests were described by owners as both providing comfort and as a source of major concern. One owner, interviewed whilst waiting for blood test results, described how her worry about adverse events had led to a blood test and her subsequent concern about what the results might mean:I was getting a bit twitchy… Because now he has Onsior forties every day now…. We did a blood test on Thursday, we won't get the results back til Monday, to see how his insides are coping with his meds. And if they're compromised I'm not quite sure where we go from there. [Owner 14] Almost all veterinary surgeons reported that they performed routine blood tests on dogs receiving NSAIDs. The frequency varied widely within and between practices with some veterinary surgeons insisting that dogs received a blood test before starting treatment and others happy to wait several months before testing. Most veterinary surgeons made decisions on the basis of basic blood biochemistry; liver function assessment by bile acid stimulation was not described. Several veterinary surgeons acknowledged that the results of blood tests rarely altered the dog’s treatment, and a few were unsure of how often tests should be performed:Personally I think it’s the right thing to do the bloods, and I do do the bloods, but whether or not we do it for the right reasons or whether or not there’s much evidence that it’s actually necessary − I think this is where there is a lack of knowledge around it. I totally recommend people do do them and I do them longer term; but often it’s not going to change what I do. More often it’s just a monitoring thing. [Veterinary surgeon 2, Focus Group 1] Veterinary surgeons recognised the need to balance quality with quantity of life when treating osteoarthritis, though several favoured treatments they perceived to be safer over ones known to be effective in older patients. Despite the risks, most owners continued to give their dog an NSAID, albeit often at a low dose. Many owners described the challenges of trading off pain relief and happiness with the risk of side effects. Opinions varied on striking the right balance:So as far as I'm concerned, if he's comfortable I'd rather him be comfortable, and die of liver failure at the age of ten than be in pain but live ‘til twelve. [Owner 20] It would be nice to have eventually painkillers that would not have any side effects. … That would be fantastic, rather than always think ‘Yeah, well, put up with being a bit stiff because I'm worried about painkillers' [Owner 16] 4 Discussion This qualitative analysis of data derived by interviews with dog owners and focus groups with veterinary professionals demonstrates high awareness and concern about potential adverse events associated with veterinary NSAIDs. This study demonstrates the potential this concern has to affect both prescription of NSAIDs by veterinary surgeons and compliance by owners. It highlights several major issues. Owners not satisfied with the information provided by their veterinary surgeon seek additional guidance, but trustworthy sources are difficult to find. Veterinary surgeons are unsure of the true incidence of adverse events associated with NSAIDs because they are hard to differentiate from concurrent disease. Veterinary surgeons are aware of little evidence to support their decisions to reduce the dose of NSAIDs or to perform regular routine blood tests. Some owners did not recollect a warning from their veterinary surgeon about the potential for adverse events associated with NSAIDs, and others did not think the information provided was adequate. Based on the focus groups data, it is likely that many veterinary surgeons do warn owners about at least some of the potential adverse events, though many commented on significant time pressures in the consulting room. Concerns about the length (Everitt et al., 2013, Robinson et al., 2014) and complexity (Robinson et al., 2015) of veterinary consultations have previously been raised and are likely to contribute to owners being unable to recall important information. Similar concerns have been raised by doctors prescribing NSAIDs (Mikhail et al., 2007). Many owners turned to the internet to supplement their knowledge. Most who did this recalled using broad search terms which brought up many websites. Little research has been conducted into the use of the internet by pet owners (Kogan et al., 2008) though veterinary surgeons in the United Kingdom believe it to be widespread and detrimental to pet health (British Veterinary Association, 2014). This study confirms the conclusions of Kogan et al. (2008) that owners would prefer to receive information from their veterinary surgeon and would like more advice on which websites to trust. Many owners of dogs with osteoarthritis expressed concern about the safety of NSAIDs and some preferred to risk efficacy by under-dosing their dog rather than risk adverse events. This is in contrast to the use of NSAIDs in adult human patients with osteoarthritis where qualitative research typically finds patients are unaware of or do not consider the risks (Milder et al., 2011a, Milder et al., 2011b, Schmitt et al., 2011). The decision making role of dog owners is similar to that of parents of young children. Research into attitudes surrounding the risks of childhood vaccination (Wroe et al., 2005) discovered that some parents viewed harms that occurred as a result of a decision not to immunise to be more acceptable than those associated with adverse drug events from vaccination. It is probable that a similar phenomenon occurs in dog owners. This is an area which warrants further research as it has clear implications for compliance and animal welfare as dogs in pain go untreated. Hunt et al. (2015) report on the frequency of adverse events following NSAID administration reported to the Veterinary Medicines Directorate. It found emesis was the most frequently reported adverse event associated with oral NSAIDs in dogs. Renal and hepatic insufficiently was reported at a low frequency. In our focus groups, some veterinary surgeons linked uncertainty about the incidence of adverse events to their cautious behaviour regarding dosing and monitoring. This highlights the importance of the publication and dissemination of accurate data reporting the incidence of adverse events associated with veterinary medicines. Hunt et al. (2015) recognised that the reported frequency of emesis does not match that of clinical studies. One cause of under-reporting identified in the current study was difficulty in determining the significance of gastroenteritis in dogs receiving NSAIDSs. Clearer guidance may be required on reporting of suspect adverse events to the Veterinary Medicines Directorate. Several veterinary surgeons commented on the poor evidence base for reduced dosing or monitoring of NSAIDs. Little has been published on either subject in veterinary medicine. A study by Wernham et al. (2011) examined the efficacy of meloxicam in dogs with osteoarthritis at a progressively reduced dose. Many dogs dropped out of the study as their owners perceived analgesia was inadequate but the authors concluded that dose reduction may be effective for some individuals. Use of the lowest effective dose is in dogs advocated by some authors for safety reasons (Lomas and Grauer, 2015) and is advised for doctors using NSAIDs in humans with osteoarthritis (Zhang et al., 2008). More research is urgently needed into the efficacy of low-dose NSAIDs in dogs with naturally occurring osteoarthritis. Many veterinarians performed regular blood tests on dogs receiving NSAIDs but specific guidance on monitoring of hepatic and renal parameters are rarely provided in companion animal NSAID datasheets (National Office for Animal Health (NOAH), 2014). Given the low frequency of renal and hepatotoxicity (Hunt et al., 2015) and the potential costs and morbidity associated with phlebotomy, a stronger evidence-base is required to guide decision making around biochemical monitoring. These results should not be interpreted as representative of the opinions of all dog owners or veterinary professionals in the United Kingdom as the aim was to report the widest possible breadth of experience. However, as this area has not previously been explored using qualitative research, we highlight important issues and provide a foundation for future studies. It is possible that the method of owner recruitment introduced respondent bias in selecting owners who were highly committed to their dogs, but the range of interviewees and attitudes obtained was very broad. A failsafe formula does not exist to identify data saturation (Ziebland and McPherson, 2006) and there is always a risk that additional participants could have articulated previously unrepresented attitudes. However, additional themes did not emerge in either the final focus group or the last four interviews so it is likely that data saturation had occurred within the population available for inclusion. The inclusion criteria excluded owners whose dogs were untreated. Exploring the attitudes of those owners was outside the scope of this work but would be valuable future research. As with all qualitative research, analysis by others may have led to alternative themes. Coding was not replicated, as advocated by Morse (1997). The findings of this study should be of value to anyone interested in improving analgesia in dogs with osteoarthritis. The significant barriers to compliance identified could be overcome by provision of more accurate and comprehensive data to both veterinary professionals and owners on the true frequency of adverse events associated with veterinary NSAIDs and how best to avoid them. Conflicts of interest None of the authors declares any conflicts of interest relevant to this paper. Acknowledgements Zoe Belshaw is undertaking a PhD funded by the Biotechnology and Biological Sciences Research Council (BBSRC; grant number BB/J014508/1) and the Centre for Evidence-based Veterinary Medicine. The Centre for Evidence-based Veterinary Medicine is supported by an unrestricted grant from the University of Nottingham and Elanco Animal Health. ==== Refs References Aragon C.L. Hofmeister E.H. Budsberg S.C. Systematic review of clinical trials of treatments for osteoarthritis in dogs J. Am. Vet. Med. Assoc. 230 2007 514 521 17302547 Bell A. Helm J. Reid J. Veterinarians' attitudes to chronic pain in dogs Vet. Rec. 175 2014 428 25028465 Braun V. Clarke V. Using thematic analysis in psychology Qual. Res. Psychol. 3 2006 77 101 Braun V. Clarke V. Interactive Data Collection 2: Focus Groups Successful Qualitative Research 2013 SAGE Publications Ltd London, UK 107 133 Braund R. Abbott J.H. Nonsteroidal antiinflammatory drugs (NSAIDs) and paracetamol for acute musculoskeletal injuries: physiotherapists' understanding of which is safer more effective, and when to initiate treatment Physiother. Theory Pract. 27 2011 482 491 21548816 British Veterinary Association, 2014 Available at: https://www.bva.co.uk/News-campaigns-and-policy/Newsroom/News-releases/98%E2%80%93of-small-animal-vets-say-owners-consult-Dr-Google-,-with-80-seeing-pets-brought-in-too-late/(Accessed 17th September 2015). Bryman A. The nature of qualitative research Social Research Methods 2012 Oxford University Press Oxford, UK 380 413 Carnes D. Anwer Y. Underwood M. Harding G. Parsons S. Influences on older people's decision making regarding choice of topical or oral NSAIDs for knee pain: qualitative study Br. Med. J. 336 2008 142 145 18056742 Cavazos J.M. Naik A.D. Woofter A. Abraham N.S. Barriers to physician adherence to nonsteroidal anti-inflammatory drug guidelines: a qualitative study Aliment. Pharmacol. Ther. 28 2008 789 798 19145734 Coyne L.A. Pinchbeck G.L. Williams N.J. Smith R.F. Dawson S. Pearson R.B. Latham S.M. Understanding antimicrobial use and prescribing behaviours by pig veterinary surgeons and farmers: a qualitative study Vet. Rec 2014 593 25200432 Everitt S. Pilnick A. Waring J. Cobb M. The structure of the small animal consultation J. Small Anim. Pract. 54 2013 453 458 23888879 Gabbay J. May A. l. Evidence based guidelines or collectively constructed mindlines? Ethnographic study of knowledge management in primary care Br. Med. J. 329 2004 1013 15514347 Horseman S.V. Rowe E.J. Huxley J.N. Bell N.J. Mason C.S. Whay H.R. The use of in-depth interviews to understand the process of treating lame dairy cows from the farmers’ perspective Anim. Welf. 23 2014 157 165 Hunt J.R. Dean R.S. Davis G.N. Murrell J.C. An analysis of the relative frequencies of reported adverse events associated with NSAID administration in dogs and cats in the United Kingdom Vet. J. 206 2015 183 190 26361747 Innes J.F. Clayton J. Lascelles B.D.X. Review of the safety and efficacy of long-term NSAID use in the treatment of canine osteoarthritis Vet. Rec. 166 2010 226 230 20173106 Kogan L.R. Goldwaser G. Stewart S.A. Schoenfeld-Tacher R. Sources and frequency of use of pet health information and level of confidence in information accuracy, as reported by owners visiting small animal veterinary practices J. Am. Vet. Med. Assoc. 232 2008 1536 1542 18479246 Laba T.L. Brien J.A. Fransen M. Jan S. Patient preferences for adherence to treatment for osteoarthritis: the MEdication Decisions in Osteoarthritis Study (MEDOS) BMC Musculoskelet. Disord. 2013 Lomas A.L. Grauer G.F. The renal effects of NSAIDs in dogs J. Am. Anim. Hosp. Assoc. 51 2015 197 203 25955147 Mikhail S.S. Zwar N.A. Vagholkar S. Dennis S.M. Day R.O. Non-steroidal anti-inflammatory drugs in general practice: a decision-making dilemma Med. J. Aust. 187 2007 160 163 17680741 Milder T.Y. Lipworth W.L. Williams K.M. Ritchie J.E. Day R.O. It looks after me: how older patients make decisions about analgesics for osteoarthritis Arthritis Care Res. 63 2011 1280 1286 Milder T.Y. Williams K.M. Ritchie J.E. Lipworth W.L. Day R.O. Use of NSAIDs for osteoarthritis amongst older-aged primary care patients: engagement with information and perceptions of risk Age Ageing 40 2011 254 259 21187281 Monteiro-Steagall B.P. Steagall P.V. Lascelles B.D. Systematic review of nonsteroidal anti-inflammatory drug-induced adverse effects in dogs J. Vet. Intern. Med. 27 2013 1011 1019 23782347 Morse J.M. Perfectly healthy, but dead: the myth of inter-rater reliability Qual. Health Res. 7 1997 445 447 National Office for Animal Health (NOAH), 2014. Available at: http://www.noahcompendium.co.uk/Compendium/Overview/-21789. html (Accessed 18th September 2015). Page-Jones S. Abbey G. Career identity in the veterinary profession Vet. Rec. 2015 Robinson N.J. Dean R.S. Cobb M. Brennan M.L. Consultation length in first opinion small animal practice Vet. Rec. 175 2014 486 25261270 Robinson N.J. Brennan M.L. Cobb M. Dean R.S. Capturing the complexity of first opinion small animal consultations using direct observation Vet. Rec. 2015 Sanderson R.O. Beata C. Flipo R.M. Genevois J.P. Macias C. Tacke S. Vezzoni A. Innes J.F. Systematic review of the management of canine osteoarthritis Vet. Rec. 164 2009 418 424 19346540 Schmitt M.R. Miller M.J. Harrison D.L. Farmer K.C. Allison J.J. Cobaugh D.J. Saag K.G. Communicating non-steroidal anti-inflammatory drug risks: verbal counseling, written medicine information, and patients' risk awareness Patient Educ. Couns. 83 2011 391 397 21129892 Silverman D. Credible Qualitative Research Interpreting Qualitative Data 2013 SAGE Publications Ltd London, UK 351 395 Tong A. Sainsbury P. Craig J. Consolidated criteria for reporting qualitative research (COREQ): a 32-item checklist for interviews and focus groups Int. J. Qual. Health Care 19 2007 349 357 17872937 Wernham B.G. Trumpatori B. Hash J. Lipsett J. Davidson G. Wackerow P. Thomson A. Lascelles B.D. Dose reduction of meloxicam in dogs with osteoarthritis-associated pain and impaired mobility J. Vet. Intern. Med. 25 2011 1298 1305 22092620 Wiseman-Orr M.L. Nolan A.M. Reid J. Scott E.M. Development of a questionnaire to measure the effects of chronic pain on health-related quality of life in dogs Am. J. Vet. Res. 65 2004 1077 1084 15334841 Wroe A.L. Bhan A. Salkovskis P. Bedford H. Feeling bad about immunising our children Vaccine 23 2005 1428 1433 15670876 Zhang W. Moskowitz R.W. Nuki G. Abramson S. Altman R.D. Arden N. Bierma-Zeinstra S. Brandt K.D. Croft P. Doherty M. Dougados M. Hochberg M. Hunter D.J. Kwoh K. Lohmander L.S. Tugwell P. OARSI recommendations for the management of hip and knee osteoarthritis, Part II: OARSI evidence-based, expert consensus guidelines Osteoarthritis Cartilage 16 2008 137 162 18279766 Ziebland S. McPherson A. Making sense of qualitative data analysis: an introduction with illustrations from DIPEx (personal experiences of health and illness) Med. Educ. 40 2006 405 414 16635119 Ziebland S. Chapple A. Dumelow C. Evans J. Prinjha S. Rozmovits L. How the internet affects patients' experience of cancer: a qualitative study Br. Med. J. 328 2004 564 15001506 Zoetis Inc., 2013. Available at: https://www.zoetisus.com/products/pages/rimadyldvm/docs/compliance.pdf (Accessed 23rd September 2015)

PMC005xxxxxx/PMC5001197.txt

==== Front Bioresour TechnolBioresour. TechnolBioresource Technology0960-85241873-2976Elsevier Applied Science ;, Elsevier Science Pub. Co S0960-8524(16)30922-110.1016/j.biortech.2016.06.105ArticleSupercapacitive microbial fuel cell: Characterization and analysis for improved charge storage/delivery performance Houghton Jeremiah aSantoro Carlo aSoavi Francesca bSerov Alexey aIeropoulos Ioannis cdArbizzani Catia bAtanassov Plamen plamen@unm.edua⁎a Department of Chemical & Biological Engineering, Center for Micro-Engineered Materials (CMEM), University of New Mexico, Albuquerque, NM 87131, USAb Department of Chemistry “Giacomo Ciamician”, Alma Mater Studiorum – Università di Bologna, Via Selmi, 2, 40126 Bologna, Italyc Bristol BioEnergy Centre, Bristol Robotics Laboratory, Block T, UWE, Coldharbour Lane, Bristol BS16 1QY, UKd Biological, Biomedical and Analytical Sciences, UWE, Coldharbour Lane, Bristol BS16 1QY, UK⁎ Corresponding author at: Center for Micro-Engineered Materials (CMEM), Department of Chemical & Biological Engineering, University of New Mexico, Albuquerque, NM 87131, USA.Center for Micro-Engineered Materials (CMEM)Department of Chemical & Biological EngineeringUniversity of New MexicoAlbuquerqueNM87131USA plamen@unm.edu1 10 2016 10 2016 218 552 560 3 5 2016 23 6 2016 25 6 2016 © 2016 The Author(s)2016This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).Highlights • Supercapacitive MFCs with various anode and cathode dimensions are investigated. • Cathode is limiting bottle supercapacitive MFC performance. • Increase in cathode area led to decrease in ohmic resistances and increase in capacitance. • The performance of a hypothetical cylindrical MFC is linearly modelled. • A 21 cm3 cylindrical MFC can deliver a peak power of 25  mW at 70  mA and 1300 W m−3. Supercapacitive microbial fuel cells with various anode and cathode dimensions were investigated in order to determine the effect on cell capacitance and delivered power quality. The cathode size was shown to be the limiting component of the system in contrast to anode size. By doubling the cathode area, the peak power output was improved by roughly 120% for a 10 ms pulse discharge and internal resistance of the cell was decreased by ∼47%. A model was constructed in order to predict the performance of a hypothetical cylindrical MFC design with larger relative cathode size. It was found that a small device based on conventional materials with a volume of approximately 21 cm3 would be capable of delivering a peak power output of approximately 25 mW at 70 mA, corresponding to ∼1300 W m−3. Keywords Microbial fuel cell (MFC)Supercapacitor (SC)Electrode areaLinear modelPower performance ==== Body 1 Introduction Microbial fuel cell (MFC) technology has been an area of interest over the past few decades as a potential source for sustainable alternative energy generation and simultaneous wastewater treatment (Pandey et al., 2016). Although power production of MFCs has increased greatly since the late 90’s, MFCs still deliver current/power densities that are approximately three orders of magnitude lower than those of methanol or hydrogen based fuel cells (Logan, 2009). These low current densities make it difficult to employ MFCs to directly power devices which require high energy output. Improving the power quality delivered by MFCs is a key challenge in the development of this technology. Microbial fuel cells utilize innate bacterial respiratory processes to convert organic materials to usable energy through the process of extracellular electron transfer. Electro-active bacteria oxidize organic substrates and electrons are conducted through the bacterial membrane to an extracellular electron acceptor using specialized proteins (c-type cytochromes) and appendages (nanowires) that are present on the bacterial surface (Busalmen et al., 2008, Gorby et al., 2006, Logan, 2009). In MFCs, a conductive anode serves as the final electron acceptor in the bacterial respiratory process. Conductive 3-D carbonaceous (Chen et al., 2012, Wei et al., 2011, Liu et al., 2012) or metallic (Dumas et al., 2007, Guerrini et al., 2014, Baudler et al., 2015) materials have been used as anode electrodes. The electrons flow through a circuit that is terminated with the reduction of oxygen at the cathode, creating a gradient in electrical potential and generating current. Oxygen is commonly used as the oxidant at the cathode due to its high electrochemical potential and its environmental availability. At near neutral pH, the cathodic reaction has a large overpotential, so a catalyst is necessary to complete the reaction (Zhao et al., 2006, Erable et al., 2012). Typically, this is accomplished through the use of platinum or platinum group metals (PGMs), enzymes, bacteria, high surface area carbon materials, or high surface area carbon materials with PGM-free catalyst. While platinum is one of the most effective materials currently known for the electrocatalytic oxygen reduction, it is very cost prohibitive resulting in 47% of the capital cost of the device (Rozendal et al., 2008). Furthermore, platinum catalysts are also subject to poisoning in the conditions present in MFC environments, leading to reduced efficiency over time (Santoro et al., 2016a). In this work, we utilize a platinum group metal-free organic catalyst (Fe-AAPyr) to catalyze the oxygen reduction reaction (ORR). Fe-AAPyr is competitive with platinum-based catalysts with the advantage of being much more cost effective, sustainable and less prone to catalyst poisoning (Santoro et al., 2016a). The maximum theoretical cell voltage, Vmax,theoretical, of an MFC can be calculated by considering the equilibrium potentials of the anode and cathode reactions (Ecathode and Eanode in V vs. SHE): (1) Vmax,theoretical=Ecathode-Eanode=0.805V-0.300V=1.105V The above equation assumes that acetate is used as a fuel source (16.9 mM) for the anode at pH = 7 and oxygen, at a partial pressure of 0.2 atm, is the oxidant for the cathode (Fradler et al., 2014). During operation, losses occur as a result of ohmic, activation, and mass-transport limitations, resulting in lower cell potentials of around 0.3–0.5 V. These voltage levels are insufficient to operate low-power consuming devices such as microprocessors (270 uA, 2.2 V), LEDs (10–20 mA, 2 V), or photodiodes (10 mA, 3.3 V) (Fradler et al., 2014). Various approaches have been explored for improving cell potential and power output in MFCs including: stacking of individual MFCs with series and parallel connections (Ieropoulos et al., 2008, Ledezma et al., 2013), maximum power point tracking (MPPT) techniques (Park and Ren, 2012), and the use of external capacitors with DC/DC converters (Dewan et al., 2009, Rozendal et al., 2008, Wang et al., 2015). Electrochemical supercapacitors (SCs) are an attractive energy storage technology that is capable of storing and delivering energy at high current and power densities with little variation in performance over the course of millions of charge/discharge cycles (Conway, 1999). In addition, SCs offer the advantage of duty cycles more compatible with BES technologies, whereby the charge/discharge cycles can be within minutes, rather than hours, days or months, which is the case for conventional batteries. SCs differ from conventional capacitors in that they do not make use of a solid dielectric material. Instead, they rely on the principles of electric double-layer capacitance and/or pseudocapacitance as the charge storage mechanisms (Conway, 1999). As stated above, external SCs have been utilized as an energy storage system to harvest the low power produced by MFCs and to deliver higher current pulses in order to power small electronic devices (Dewan et al., 2009, Ieropoulos et al., 2010, Ieropoulos et al., 2013, Wang et al., 2015). Furthermore, it has been demonstrated that more energy can be harvested by operating MFCs intermittently rather than continuously. Dewan et al. showed a 111% increase in power by intermittent operation of the MFC connected to a SC when compared to continuous operation (Dewan et al., 2009, Ieropoulos et al., 2016, Papaharalabos et al., 2013). Another approach to improve power quality is the utilization of the inherent capacitive features of MFC electrodes. MFCs and SCs both utilize high surface area carbon as their electrode material. Recently, efforts have been made to integrate capacitive materials with MFC electrodes in order to improve power quality and charge storage capabilities (Deeke et al., 2015). In 2005, Ieropoulos et al. first demonstrated that biofilms in MFCs were capable of storing electrons when the device was left in open circuit for an extended period of time, providing higher power upon reconnection of the circuit (Ieropoulos et al., 2005). It has been shown that cytochromes present within MFC biofilms exhibit pseudocapacitive behavior and can act as electron sinks (Esteve-Núñez et al., 2008, Schrott et al., 2011, Uría et al., 2011). Formation of a Helmholtz layer by electrolyte ion adsorption at the MFC/electrode interfaces further contributes to the observed capacitance of the cell (Fradler et al., 2014). Fradler et al. showed that double layer capacitance contributed approximately ten times the capacitance of the biofilm in a tubular MFC which was shown to achieve charge storage capacities comparable to SCs with minimal current leakage (Fradler et al., 2014). An integrated self-charging supercapacitive MFC has been constructed by integrating an additional high surface area carbon brush short-circuited with the cathode and operating an MFC in a controlled manner (Santoro et al., 2016b). The additional electrode (AdE) confers increased surface area available for formation of a Helmholtz double layer, thus increasing the device’s capacitance. The AdE also leads to lower observed ohmic resistance during galvanostatic (GLV) discharge of the microbial supercapacitor. This design significantly improves recharge times of the system when compared to designs that incorporate external capacitors, allowing for more frequent use of the accumulated energy (Santoro et al., 2016b). It was previously shown that the increase in cathode area affected positively on the performance output of the MFCs (Cheng and Logan, 2011, Kim et al., 2015). In the present study, we investigate a supercapacitive MFC (SC-MFC) system and the effect of relative anode and cathode size on the overall performance of the system. We use the experimental data from these experiments to construct a simple predictive linear model for a hypothetical SC-MFC with a cylindrical design in order to forecast performance of a larger scale device. We demonstrate that the performance of a SC-MFC based on conventional materials can be improved to levels suitable for powering practical electronic devices by optimizing design parameters. 2 Materials and method 2.1 MFC configuration A single chamber glass bottle microbial fuel cell design with a volume of 125 mL was used to investigate the effect of relative anode and cathode geometric area on supercapacitive MFC (SC-MFC) performance (Fig. S1). The cell consisted of a Pyrex glass bottle modified with two lateral glass tubes to serve as attachment sites for cathode electrodes. The MFC was operated in a membraneless configuration with the anode fully immersed in the solution with air-breathing cathodes. One face of the cathode was exposed to the electrolyte and the other was exposed to the air. The effect of changing the relative area of the anode and cathode of the MSC was investigated. 2.2 Anode construction Carbon brush electrodes (Millirose, USA) were employed as the anode material for all experiments. The carbon brushes used had a diameter of 3 cm and a length of 3 cm, giving a projected surface area of 9 cm2. Prior to our experiments, all anodes were pre-colonized with electro-active bacterial biofilms by incubation in a mixture (by volume) of 50% activated sludge (obtained from Albuquerque Southeast Water Reclamation facility, Albuquerque, NM) and 50% buffer solution composed of 0.1 M KCl and 0.1 M potassium phosphate buffer (pH = 7.5). The same anodes have been used for previous experiments (Santoro et al., 2016b, Soavi et al., 2016). Additional carbon brushes were added to the cell to investigate the effects of increased anode surface area. All carbon brushes were colonized with electroactive bacteria as described above (Santoro et al., 2016a, Santoro et al., 2016b). 2.3 Cathode construction Air-breathing gas diffusion electrodes were used for the cathodes. The electrode consisted of a hydrophobic-hydrophilic gradient of carbon infused with iron-aminoantipyrine (Fe-AAPyr), a PGM-free catalyst (Serov et al., 2014). Cathodes were constructed by pressing carbon-based materials onto stainless steel mesh used as current collector. First, carbon black teflonized with 50 wt% of PTFE (XC50) with a loading of 30 ± 1 mg cm−2 was mixed using a blade-type coffee grinder and pressed in a circular pellet die at 2 metric tons (mT) for 5 min. A secondary layer was added consisting of 20 ± 1 mg cm−2 carbon black teflonized with 35 wt% of PTFE (XC35) mixed with 2 ± 0.1 mg cm−2 Fe-AAPyr and pressed in the pellet die at 2 mT for 5 min at room temperature. The preparation of Fe-AAPyr was based on the sacrificial support method (SSM) that has been previously described (Serov et al., 2014). Cathodes were attached to the MFC using stainless steel screw clamps and rubber gaskets. The cathode area was modified using rubber gaskets with circular holes with diameters of 1.2 cm and 1.8 cm. The area of the circular holes was 1.13 cm2 (d = 1.2 cm) and 2.54 cm2 (d = 1.8 cm) respectively. The different cathode areas used in the experiments were: i) 2.54 cm2 (single cathode), ii) 3.67 cm2 (two cathodes with one 2.54 cm2 gasket and one 1.13 cm2 gasket), iii) 5.09 cm2 (two cathodes each with a 2.54 cm2 gasket). 2.4 Electrochemical measurements Electrochemical measurements were carried out using a BioLogic SP-50 potentiostat using a three-electrode setup with an Ag/AgCl (3 M KCl, +210 mV vs. SHE) reference electrode. The cell was left in open circuit until a steady state potential was attained. Then, a sequence following the order: rest – galvanostatic discharge – rest, was repeated. The galvanostatic discharge was run at different current levels (ipulse) (Fig. 1) with pulse times of 2 s and 10 ms while monitoring the anode and cathode potentials by the use of a reference electrode (Ag/AgCl 3 M KCl). Following each pulse, the SC-MFC was allowed to rest (no current applied, the circuit is opened) until the potential returned to the original open circuit voltage, (Vmax,OC). During this time, the electrode potentials are restored to their equilibrium values exhibited before the pulse, recharging the SC-MFC independently of an external power source (Santoro et al., 2016b). During the GLV discharge, an initial drop in cell voltage (from Vmax,OC to a lower value, Vmax) is observed This initial drop in potential (Vmax,OC–Vmax = ΔVohmic) is directly related to the equivalent series resistance (ESR) of the cell. The ΔVohmic includes contributions from the electrolyte as well as the electrodes. The relationship between ESR and ΔVohmic is demonstrated by Eq. (2): (2) ESR=ΔVohmicipulse The ESR of the cell can be further analyzed to investigate the individual contributions of the anode and cathode by examining each electrode profile under the GLV discharges. The ohmic losses observed at each electrode can be used to estimate the anode (RA) and cathode resistances (RC). Specifically, RA and RC are obtained by dividing the electrode ohmic losses per ipulse. The reference electrode is centered between the anode and the cathode and the bulk electrolyte resistance is assumed to be negligible. The cell ESR is related to RA and RC by Eq. (3): (3) ESR=RA+RC The capacitance of the SC-MFC influences the rate at which the cell voltage (ΔVcapacitive) decreases during the GLV discharge, following the initial ohmic drop. The slope of the GLV discharge curve over time (dV/dt) is inversely related to the capacitance of the cell. Capacitance (C) was calculated using Eq. (4): (4) Ccell=ipulsedVdt Anode capacitance (CA) and cathode capacitance (CC) were calculated by analyzing the slopes of the corresponding electrode potentials over time. The total cell capacitance (Ccell), is related to CA and CC by Eq. (5): (5) Ccell=1CA+1CC-1 Maximum power output (Pmax) was calculated for each SC-MFC configuration by multiplying the maximum cell voltage after the pulse (Vmax) by the pulse current: (6) Pmax=ipulse×Vmax Since this calculation does not account for the capacitive decrease of cell voltage (ΔVcapacitive) observed during discharge of the SC-MFC, the Pmax value is higher than the actual power delivered by the device over the pulse duration. This pulse power (Ppulse) is calculated on the basis of the energy delivered during the pulse (Epulse), which in turn is calculated by Eq. (7): (7) Epulse=ipulse∫0tVdt where t is the discharge time. The pulse power is obtained by Eq. (8): (8) Ppulse=Epulset 3 Results and discussion 3.1 Effect of cathode geometric area on SC-MFC performance Fig. 2 reports the results of the GLV discharge of SC-MFCs featuring cathodes of various areas. The results are summarized in Table 1. Cell voltage and electrode potential profiles for 2 s discharges at 3 mA are shown in Fig. 2a and b. Cell voltage and electrode potential profiles for 10 ms discharges at 3 mA are shown in Fig. S2. As cathode area doubled from 2.54 cm2 to 5.09 cm2, ΔVohmic decreased by approximately 47%. The values for the overall ΔVohmic were measured to be 176 ± 1.5 mV for the 2.54 cm2 cathode, 115 ± 3.5 mV for the 3.67 cm2 cathode, and 91 ± 3 mV for the 5.09 cm2 cathode area. The ESR for each cell was calculated to be 58.6 ± 0.3 Ω, 38.1 ± 0.9 Ω, and 30.5 ± 0.9 Ω respectively. Cell capacitance (Ccell) also increased with increasing cathode geometric area, with measured values of 24 ± 2 mF, 27 ± 0.1 mF and 30 ± 1.4 mF respectively (Fig. 2a and Table 1). Fig. 2b shows that the cathode is the main contributor to ΔVohmic and ESR. Cathode resistances (RC) of 57.1 ± 2.6 Ω (2.54 cm2), 35.7 ± 1.4 Ω (3.67 cm2), and 27.9 ± 4.2 Ω (5.09 cm2) were observed. These values correspond to a cathode resistance normalized to electrode geometric area (RC′) of approximately 140 Ω cm2. Cathode capacitance (CC) increased with increasing cathode area, with recorded values of 51.3 ± 1.9, 61.5 ± 0.5, and 73.2 ± 1.3 mF. This translates to a cathode capacitance normalized to electrode geometric area (areal capacitance density, CC′) of ∼17 mF cm−2 (Table 1). Anode resistance and capacitance remained constant at approximately 0.5 Ω and 48 mF (Table 1). Fig. 2c shows the Pmax values of the three cells as calculated by Eq. (4) for various discharge currents. Pmax increased significantly with increasing cathode geometric area. Doubling the cathode’s geometric area resulted in a 113% increase in Pmax indicating quasi-linear positive dependence between cathode area and Pmax. Recorded Pmax values were 2.65 ± 0.05 mW (i = 6 mA) for the SC-MFC with a 2.54 cm2 cathode area, 4.08 ± 0.1 mW (i = 10 mA) for the SC-MFC with a 3.67 cm2 cathode area, and 5.58 ± 0.08 mW (i = 14 mA) for the SC-MFC with a cathode area of 5.09 cm2 (Fig. 2d). These values correspond to volumetric power densities of 21.2 ± 0.4 W m−3, 32.64 ± 0.8 W m−3, and 44.64 ± 0.64 W m−3 respectively (based on 125 mL volume) (Fig. 2c). Fig. 2d reports the Pmax vs ipulse curves with power and current normalized to the cathode geometric area, i.e. in terms of areal power and current densities. It was found that the areal Pmax density was similar among the different SC-MFCs. Values ranged from 10.4 ± 0.2 W m−2 (at 23.6 A m−2) for the 2.54 cm2 cathode area to 11.0 ± 0.16 W m−2 (at 27.5 A m−2) for the 5.09 cm2 cathode area. Similar measured power density values indicated a roughly linear positive relationship between the power generated and the cathode area. Therefore, an increase in cathode area led to a roughly linear increase in power performance. Fig. 2e and f report the pulse power (Ppulse) delivered over 10 ms and 2 s pulses. As expected, the longer pulse time led to smaller power produced due to the capacitive decrease of cell voltage (ΔVcapactive). In agreement with Pmax data (Fig. 2c), the increase in cathode area led to higher power (Fig. 2e). For pulse durations of 10 ms, the SC-MFC with smaller cathode area delivered 2.3 ± 0.13 mW at 6 mA (19 ± 1 W m−3). Doubling the cathode’s geometric area resulted in ∼118% higher power (5.1 ± 0.21 mW, 41 ± 1.7 W m−3) at 13 mA for 10 ms discharge pulse durations. A similar trend was observed for 2 s pulse durations. The SC-MFC with cathode area of 2.54 cm−2 generated 1.38 ± 0.07 mW at 3 mA (11 ± 0.56 W m−3), the SC-MFC with 5.09 cm−2 cathode area increased the power output (2.5 ± 0.25 mW) under the same conditions. Volumetric power increased by approximately 80% (20 ± 2 W m−3) by doubling the cathode area (Fig. 2e). 4 Effect of anode geometric area on SC-MFC performance Fig. 3a and b show discharge profiles for cell voltage and individual electrode potentials at discharge times of 2 s at a pulse current of 3 mA for SC-MFCs with different anode geometric areas (number of anode brushes). Cell voltage and cell potentials at discharge time of 10 ms (ipulse 3 mA) are reported on the Fig. S3. The results of the GLV curve analyses are reported in Table 1. In these experiments, the cathode area was maintained at 5.09 cm2, as it showed the highest performance from previous experiments. As mentioned in the description of materials, each anode was considered to have a projected area of 9 cm2 per brush. The ESR of the cell was measured as 30.5 ± 0.9 Ω (one brush), 29.4 ± 0.5 Ω (two brushes), and 26.8 ± 0.5 Ω (three brushes) respectively (Fig. 3a and Table 1). The anode electrode contribution to ohmic losses (RA of 0.64 ± 0.2 Ω) is very low compared to the ohmic losses contributed by the cathode electrode (Fig. 3b). Anode resistance remained relatively constant under all three experimental conditions, suggesting that the bulk electrolyte resistance, which remained constant in all three cells, is the main contributor to anode resistance. Total cell capacitance (Ccell) increased when the number of brushes was increased (Fig. 3a), with measured values of 30 ± 1.4, 50 ± 3.7 and 63 ± 1.9 mF (Table 1). This trend is directly related to CA, which increased from 50 ± 4 mF to 121 ± 14 mF and 194 ± 18 mF (Fig. 3b), in agreement with a capacitance of ∼53 mF per brush (Table 1). The Pmax vs. I plots are shown in Fig. 3c. Pmax was relatively consistent between all three cells (Fig. 3c) with a measured value of ∼5.6–6.0 mW (45–48 Wm−3). A slightly higher value was obtained for the cell with three brushes and is likely due to the lower observed ESR of the cell. When Pmax was represented as areal power density (Fig. 3d), the SC-MFC with three brushes showed the lowest value (2.2 ± 0.1 W m−2) followed by the SC-MFC with two anode brushes (3.16 ± 0.05 W m−2) and the highest areal power density was observed with just one anode brush (6.2 ± 0.19 W m−2). This trend shows that increasing anode area does not improve the areal power performance of the system, and further demonstrates that the cathode is the limiting component of the system. The pulse power for 10 ms and 2 s pulse durations is reported in (Fig. 3e and f). Unlike the trend observed in the Pmax plots (Fig. 3c), the Ppulse curves expressed in mW and as volumetric power (W m−3) show a clear increasing trend with the anode size (Fig. 3e). During the 10 ms discharge, the peak power achieved for SC-MFC with three anode brushes was 6.03 ± 0.16 mW (48 ± 1.3 W m−3) followed by SC-MFC with two anode brushes (5.6 ± 0.11 mW, 44.5 ± 0.88 W m−3) and the one with just single brush (5.12 ± 0.22 mW, 41 ± 1.68 W m−3) (Fig. 3e). The 2 s discharge pulses show a more pronounced difference between the three cells. The difference can be attributed to the increase in Ccell. The maximum Ppulse values (tpulse of 2 s) for each SC-MFC configuration were as follows: 3.53 ± 0.09 mW (28.2 ± 0.72 W m−3) for the cell with three anode brushes, 2.9 ± 0.15 mW (23 ± 1.2 W m−3) for the cell with two anode brushes, and 2.51 ± 0.25 mW (20 ± 3 W m−3) for the cell with a single anode brush (Fig. 3e). Areal power density (W m−2) decreased with increasing anode area, indicating that increasing anode geometric area did not improve performance, and that cathode is the limiting factor of the system in terms of performance (Fig. 3f). 4.1 Simple linear predictive model The experimental data described above suggest that the SC-MFC performance can be significantly improved by properly balancing the anode and cathode geometric areas. Cathode is the limiting component of system performance as demonstrated by the linear relationship between cathode area and power generated. We have shown that increasing cathode geometric area is a viable strategy for improving SC-MFC power output by increasing the capacitance and decreasing the ESR. The most efficient design that minimizes SC-MFC volume and maximizes cathode area is a cylinder with the air cathode comprising the cylinder wall surrounding the anode brush and adequately spaced in order to avoid short circuit (Fig. 4). We devised a simple linear model in order to predict the performance of the cylindrical SC-MFCs with various radial diameters using experimental GLV data obtained at 3 mA and 2 s (Table 1). The minimum cell volume is limited by the size of the anode brush, which has a radius (rbrush) of 1.5 cm and a height (h) of 3 cm, resulting in a volume of 21.3 cm3 (V = π rbrush2 h). The cathode area for a cylinder of these dimensions is 28.4 cm2 (cathode area = 2 π r h) (Table 2). Since the height of the brush (h) is constant, the cathode area scales with the cylinder radius by 2 π r, reaching a maximum value of 94.2 cm2 at r = 5 cm. We calculated the Ccell and ESR of hypothetical cylindrical cells with radii (r) ranging between 1.5 and 5 cm as described in Table 2 and by the following Eqs. (9), (10), (11), (12). Since all cells in the model utilize a single anode brush, anode capacitance (CA) remained at a value of 53 mF (Table 1). Cathode capacitance (CC) was calculated by utilizing a cathode areal capacitance density (CC′) of 17 mF cm−2 using Eq. (9): (9) Cc=CC′×2πrh CC values ranged from 480 mF for r = 1.5 cm to 1600 mF for r = 5 cm. The calculated CC values are more than one order of magnitude higher than CA, therefore total cell capacitance (Ccell) is roughly unaffected by cathode geometric area. Indeed, substituting Eq. (9) into Eq. (5), Ccell is represented by Eq. (10): (10) Ccell=1CA+1CC′×2πrh-1=153+1320r-1 Ccell is ∼50 mF for all cells and it can be improved by increasing the height of the anode brush. The ESR of the cell is directly related to the cylinder radius and cathode area. ESR was calculated as the sum of the anode resistance (RA) and cathode resistance (RC) according to Eq. (3). The contribution of the electrolyte to ESR was considered to be negligible. RA was set at a constant value of 0.5 Ω (Table 1). RC was obtained for each cylindrical radius using the RC′ value of 140 Ω cm2 (Table 1) using Eq. (11): (11) RC=RC′2πrh Substituting Eq. (10) into Eq. (3), ESR is represented by Eq. (12): (12) ESR=RA+RC′2πrh Table 2 shows that RC decreases from a value of 5 Ω to 1.5 Ω when r increases from 1.5 cm to 5 cm, resulting in ESR values of 5.5 Ω to 2 Ω respectively. Calculated Ccell and ESR values were used to simulate GLV discharge curves for cylindrical SCMFCs with different radii at various currents (i) using Eqs. (13), (14): (13) Vcell=Vmax-ΔVohmic-ΔVcapacitive where Vmax=0.8V,ΔVohmic=i×ESRandΔVcapacitive=i×tCcell, therefore, (14) Vcell=0.8-i×ESR-i×tCcell Fig. 5 shows the simulated discharge curves for two cells with r = 1.5 cm and 5 cm with discharge current i = 25, 50 and 100 mA. It can be seen that r has little effect on capacitance (slope), but significantly affects ΔVohmic, which is related to the cell’s ESR (see Table 1). The Vcell vs. time profiles were analyzed to calculate the maximum power, Pmax by Eq. (15), and the energy E and power P delivered during a full discharge at different i by the Eqs. (6), (7): (15) Pmax=i×(Vmax-ΔVohmic)=i×(Vmax-i×ESR) Fig. 6 shows the Pmax vs I plots calculated for cells of various radii. Fig. 6a shows that an increase in the diameter of the cylindrical SC-MFC improves power performance. The cell with a radius of 5 cm is capable of producing up to 80 mW at 200 mA discharges current. However, the increased cylinder diameter leads to a larger cell volume, and therefore has a negative impact on the volumetric Pmax values. This is evident in Fig. 6b, where the highest power density of 1300 W m−3 is obtained at 70 mA with the smallest cell. It is worth noting, that despite a lower maximum power density of 320 W m−3, the biggest cell permits operation at higher discharge currents. These projected performance levels have been calculated on the basis of 2 s discharges at 3 mA. The model could be further implemented by the use of parameterization data that refer to specific operative conditions of the SC-MFC, and which take into account the effects of discharge current and time on the capacitive response of the cell, which in turn is expected to increase at lower currents (from mA to tens of microA) and for longer periods (from seconds to minutes) (Conway, 1999). In previous cases, single MFCs of 6.25 mL volume, produced 0.1 mW at 0.45 mA and 220 mV, but when two such units were connected together, a digital wristwatch was powered via an ultra-low power boost converter (Papaharalabos et al., 2013). In the second instance, individual 100 mL MFCs were generating between 1 and 2 mW continuously, and a 36-MFC module produced 40–60 mW continuously, which was part of a stack powering LED modules, (via a voltage regulator and an external supercapacitor), consuming approximately 1.5 W (Ieropoulos et al., 2016). The amount of absolute power generated by the larger cell (r = 5 cm, 80 mW) in this study is calculated to be at a cell voltage of 0.4 V. Although this is transient (i.e. not continuous) since it is generated as a pulse, it is still higher compared with the output levels of MFC modules previously reported to power practical applications (Ieropoulos et al., 2016, Papaharalabos et al., 2013). Hence, the data generated by this simple linear predictive model, suggest that with intermittent operation, the SC-MFC could easily power practical applications such as LEDs or other low-power dc applications. The Ragone plots in Fig. 6c and d show the calculated values for energy and power for complete discharges, from 0.8 V to 0 V, at different currents. Fig. 6c shows the highest energy of ∼4 μWh (15 mJ) is delivered at the lowest currents for all the cells. This is due to the fact that at the lower currents, the cell voltage profile over time is mainly affected by the capacitance of the cell (Ccell), which is nearly identical for all the cylinder diameters. At higher currents, the ohmic drop is larger, leading to lower quantities of delivered energy. This phenomenon is less prominent in the larger cells that exhibit lower ESRs. The highest power is 30 mW and is delivered by the SC-MFC with r = 5 cm under a 150 ms pulse at 150 mA. Fig. 6d confirms that the smallest cell exhibits the highest volumetric energy and power densities of 700 J m−3 (195 mW h m−3) and 600 W m−3. 5 Conclusions Cathode geometric area is a critical design component towards the improvement of SC-MFCs power performance. By increasing cathode area, the internal resistance decreased substantially and the peak power of the device scaled roughly linearly. A simple linear model was developed to predict the performance of a cylindrical SC-MFC. The model demonstrates that a SC-MC design with a greater relative cathode area should greatly improve the system performance. Due to the increased cathode surface area imparted by this design, volumetric power output is forecast to improve by more than two orders of magnitude, with an anticipated maximum value of ∼1300 W m−3. Appendix A Supplementary data Supplementary data Acknowledgements JH, CS, AS and PA would like to thank the Bill & Melinda Gates Foundation grant: “Efficient Microbial Bio-electrochemical Systems” (OPP1139954). FS and CA would like to acknowledge Alma Mater Studiorum – Università di Bologna (RFO, Ricerca Fondamentale Orientata). Appendix A Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.biortech.2016.06.105. Fig. 1 Schematic representation of the rest-galvanostatic discharge – rest sequence. No current is applied during the rest period. Fig. 2 Cell voltage (a) and electrode potential (b) profiles under 2 s pulses at 3 mA of SC-MFCs after 5 s rest. Pmax (c and d) and Ppulse (e and f) vs. I plots for SC-MFC with different cathodes. Cathode area is 2.54 cm−2 (blue), 3.67 cm−2 (green) and 5.09 cm−2 (red). Volumetric power densities (c and e) are normalized to the cell volume (125 mL). Areal power and current densities (d and f) are normalized to the cathode geometric area. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Fig. 3 Cell voltage (a) and electrode potential (b) profiles under 2 s pulses at 3 mA after 5 s rest. Pmax (c and d) and Ppulse (e and f) vs. I for SC-MFC with different number of brush anodes. The projected anode areas are 9 cm−2 (yellow), 18 cm−2 (black) and 27 cm−2 (grey). Volumetric power densities (c and e) are normalized to the cell volume (125 mL). Areal power and current densities (d and f) are normalized to the projected anode area (9 cm2 brush−1). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Fig. 4 Schematic of the cylindrical SC-MFC used for predictive model. Fig. 5 Cell voltage profiles of two cylindrical SC-MFCs with r = 1.5 cm (a) and 5 cm (b) calculated at i = 25, 50 and 100 mA by Eqs. (14), (15) on the basis of the values of ESR and Ccell reported in Table 2. Fig. 6 Projected Pmax vs current plots (a and b) and Ragone plots (c and d) of cylindrical SC-MFCs with different radius (r). Table 1 ESR and capacitance of the SC-MFC and electrode resistances and capacitances evaluated from the GLV discharge curves at 3 mA reported in Figs. 3a and b and 4a and b. RA′, RC′, CA′ and CC′ are the anode and cathode resistances and capacitances normalized to the electrode geometric areas. n. anode brush Anode brush area Cathode area ΔVohmic ESR Rc Rc′ RA RA′ cm2 cm2 mV Ω Ω Ω cm−2 Ω Ω brush−1 1 9 2.54 176 ± 1.5 58.6 ± 0.3 57 ± 2.6 145 0.8 ± 0.9 0.8 1 9 3.67 115 ± 3.5 38.1 ± 0.9 36 ± 1.4 131 0.8 ± 0.7 0.8 1 9 5.09 91 ± 3 30.5 ± 0.9 28 ± 4.2 142 0.35 ± 0.3 0.35 1 9 5.09 91 ± 3 30.5 ± 0.9 28 ± 4.2 142 0.35 ± 0.3 0.35 2 18 5.09 88 ± 1 29.4 ± 0.5 29 ± 0.7 148 0.4 ± 0.1 0.20 3 27 5.09 81 ± 1.5 26.8 ± 0.5 26.4 ± 0.6 134 0.6 ± 0.3 0.20 Average 140 0.5 n. anode brush Anode brush area Cathode area ΔVcapac. Ccell Cc Cc′ CA CA′ cm2 cm2 mV mF mF mF cm−2 mF mF cm−2 1 9 2.54 250 ± 20 24 ± 2 51 ± 1.9 20. 46 ± 8.3 46 1 9 3.67 226.5 ± 0.7 26.5 ± 0.1 61.5 ± 0.5 17 46 ± 2.8 49 1 9 5.09 203 ± 9.2 30 ± 1.4 73 ± 1.3 14 50 ± 4 50 1 9 5.09 203 ± 9.2 30 ± 1.4 73 ± 1.3 14 50 ± 4 50 2 18 5.09 122 ± 9.2 50 ± 3.7 86 ± 18 17 121 ± 13.7 60 3 27 5.09 95 ± 2.9 63 ± 1.9 95 ± 1 19 194 ± 17.7 65 Average 17 53 Table 2 The figures of merit of cylindrical SC-MFCs with increasing radius (r). Size Eq. r 1.5 cm 3 cm 4 cm 5 cm Anode rbrush = 1.5 cm h = 3 cm Cathode h = 3 cm area = 2 π r h, with r > rbrush lowest area = 28.3 cm2 28.3 cm2 56.5 cm2 75.5 cm2 94.2 cm2 Cell h = 3 cm volume = π r2 h, with r > rbrush lowest volume = 21.2 cm3 21.2 cm3 85 cm3 150 cm3 235 cm3 Capacitance Anode, CA 53 mF 53 mF 53 mF 53 mF 53 mF Cathode, CC CC = CC′ × 2 π r h with CC′ = 17 mF cm−2 (9) 480 mF 960 mF 1280 mF 1600 mF Cell, Ccell 1CA+1CC′×2πrh-1==153+1320r-1 (10) 48 mF 50 mF 51 mF 51 mF ESR Anode resistance, RA 0.5 Ω 0.5 Ω 0.5 Ω 0.5 Ω 0.5 Ω Cathode resistance, RC RC=RC′2πrh, with RC′ = 140 Ω cm2 (11) 5 2.5 1.9 1.5 Cell, ESR ESR = RA + RC = RA + RC′2πrh==0.5+8r (12) 5.5 3 2.4 2 ==== Refs References Baudler A. Schmidt I. Langner M. Greiner A. Schröder U. Does it have to be carbon? Metal anodes in microbial fuel cells and related bioelectrochemical systems Energy Environ. Sci. 8 2015 2048 Busalmen J.P. Esteve-Núñez A. Berná A. Feliu J.M. C-type cytochromes wire electricity-producing bacteria to electrodes Angew. Chem. 120 26 2008 Chen S. He G. Hu X. Xie M. Wang S. Zeng D. Hou H. Schröder U. A three-dimensionally ordered macroporous carbon derived from a natural resource as anode for microbial bioelectrochemical systems ChemSusChem 5 6 2012 1059 1063 22467379 Cheng S. Logan B.E. High hydrogen production rate of microbial electrolysis cell (MEC) with reduced electrode spacing Bioresour. Technol. 102 2011 3571 3574 21036036 Conway B.E. Electrochemical Supercapacitors: Scientific Fundamentals and Technological Applications 1999 Springer Science and Business Media Deeke A. Sleutels T.H.J.A. Donkers T.F.W. Hamelers H.V.M. Buisman C.J.N. Ter Heijne A. Fluidized capacitive bioanode as a novel reactor concept for the microbial fuel cell Environ. Sci. Technol. 49 3 2015 1929 1935 25514015 Dewan A. Beyenal H. Lewandowski Z. Intermittent energy harvesting improves the performance of microbial fuel cells Environ. Sci. Technol. 43 12 2009 4600 4605 19603683 Dumas C. Mollica A. Féron D. Basséguy R. Etcheverry L. Bergel A. Marine microbial fuel cell: use of stainless steel electrodes as anode and cathode materials Electrochim. Acta 53 2 2007 468 473 Erable B. Féron D. Bergel A. Microbial catalysis of the oxygen reduction reaction for microbial fuel cells: a review ChemSusChem 5 6 2012 975 987 22615123 Esteve-Núñez A. Sosnik J. Visconti P. Lovley D.R. Fluorescent properties of c-type cytochromes reveal their potential role as an extracytoplasmic electron sink in Geobacter sulfurreducens Environ. Microbiol. 10 2 2008 497 505 18093163 Fradler K.R. Kim J.R. Boghani H.C. Dinsdale R.M. Guwy A.J. Premier G.C. The effect of internal capacitance on power quality and energy efficiency in a tubular microbial fuel cell Process Biochem. 49 6 2014 973 980 Gorby Y.A. Yanina S. McLean J.S. Rosso K.M. Moyles D. Dohnalkova A. Beveridge T.J. Chang I.S. Kim B.H. Kim K.S. Electrically conductive bacterial nanowires produced by Shewanella oneidensis strain MR-1 and other microorganisms Proc. Natl. Acad. Sci. U.S.A. 103 30 2006 11358 11363 16849424 Guerrini E. Cristiani P. Grattieri M. Santoro C. Li B. Trasatti S. Electrochemical behavior of stainless steel anodes in membraneless microbial fuel cells J. Electrochem. Soc. 161 3 2014 H62 H67 Ieropoulos I. Greenman J. Melhuish C. Hart J. Energy accumulation and improved performance in microbial fuel cells J. Power Sources 145 2 2005 253 256 Ieropoulos I. Greenman J. Melhuish C. Microbial fuel cells based on carbon veil electrodes: stack configuration and scalability Int. J. Energy Res. 32 13 2008 1228 1240 Ieropoulos I. Greenman J. Melhuish C. Horsfield I. EcoBot-III-A robot with guts ALIFE 2010 733 740 Ieropoulos I.A. Ledezma P. Stinchcombe A. Papaharalabos G. Melhuish C. Greenman J. Waste to real energy: the first MFC powered mobile phone Phys. Chem. Chem. Phys. 15 37 2013 15312 15316 23939246 Ieropoulos I.A. Stinchcombe A. Gajda I. Forbes S. Merino-Jimenez I. Pasternak G. Sanchez-Herranz D. Greenman J. Pee power urinal – microbial fuel cell technology field trials in the context of sanitation Environ. Sci. Water Res. Technol. 2 2016 336 343 Kim K.-Y. Yang W. Logan B.E. Impact of electrode configurations on retention time and domestic wastewater treatment efficiency using microbial fuel cells Water Res. 80 2015 41 46 25996751 Ledezma P. Stinchcombe A. Greenman J. Ieropoulos I. The first self-sustainable microbial fuel cell stack Phys. Chem. Chem. Phys. 15 7 2013 2278 2281 23321677 Liu J. Qiao Y. Guo C.X. Lim S. Song H. Li C.M. Graphene/carbon cloth anode for high-performance mediatorless microbial fuel cells Bioresour. Technol. 114 2012 275 280 22483349 Logan B.E. Exoelectrogenic bacteria that power microbial fuel cells Nat. Rev. Microbiol. 7 5 2009 375 381 19330018 Pandey P. Shinde V.N. Deopurkar R.L. Kale S.P. Patil S.A. Pant D. Recent advances in the use of different substrates in microbial fuel cells toward wastewater treatment and simultaneous energy recovery Appl. Energy 168 2016 706 723 Papaharalabos G. Greenman J. Melhuish C. Santoro C. Cristiani P. Li B. Ieropoulos I. Increased power output from micro porous layer (MPL) cathode microbial fuel cells (MFC) Int. J. Hydrogen Energy 38 26 2013 11552 11558 Park J.-D. Ren Z.J. Hysteresis controller based maximum power point tracking energy harvesting system for microbial fuel cells J. Power Sources 205 2012 151 156 Rozendal R.A. Hamelers H.V. Rabaey K. Keller J. Buisman C.J. Towards practical implementation of bioelectrochemical wastewater treatment Trends Biotechnol. 26 8 2008 450 459 18585807 Santoro C. Serov A. Stariha L. Kodali M. Gordon J. Babanova S. Bretschger O. Artyushkova A.J. Atanassov P. Iron based catalysts from novel low-cost organic precursors for enhanced oxygen reduction reaction in neutral media microbial fuel cells Energy Environ. Sci. 2016 16596 Santoro C. Soavi F. Serov A. Arbizzani C. Atanassov P. Self-powered supercapacitive microbial fuel cell: the ultimate way of boosting and harvesting power Biosens. Bioelectron. 78 2016 229 235 26615513 Schrott G.D. Bonanni P.S. Robuschi L. Esteve-Nuñez A. Busalmen J.P. Electrochemical insight into the mechanism of electron transport in biofilms of Geobacter sulfurreducens Electrochim. Acta 56 28 2011 10791 10795 Serov A. Tylus U. Artyushkova K. Mukerjee S. Atanassov P. Mechanistic studies of oxygen reduction on Fe-PEI derived non-PGM electrocatalysts Appl. Catal. B 150 2014 179 186 Soavi F. Bettini L.G. Piseri P. Milani P. Santoro C. Atanassov P. Arbizzani C. Miniaturized supercapacitors: key materials and structures towards autonomous and sustainable devices and systems J. Power Sources 2016 Uría N. Muñoz Berbel X. Sánchez O. Muñoz F.X. Mas J. Transient storage of electrical charge in biofilms of Shewanella oneidensis MR-1 growing in a microbial fuel cell Environ. Sci. Technol. 45 23 2011 10250 10256 21981730 Wang H. Park J.-D. Ren Z.J. Practical energy harvesting for microbial fuel cells: a review Environ. Sci. Technol. 49 6 2015 3267 3277 25670167 Wei J. Liang P. Huang X. Recent progress in electrodes for microbial fuel cells Bioresour. Technol. 102 20 2011 9335 9344 21855328 Zhao F. Harnisch F. Schröder U. Scholz F. Bogdanoff P. Herrmann I. Challenges and constraints of using oxygen cathodes in microbial fuel cells Environ. Sci. Technol. 40 17 2006 5193 5199 16999088

PMC005xxxxxx/PMC5001198.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557119290010.1186/s12864-016-2900-2ResearchWhole genomic sequence analysis of Bacillus infantis: defining the genetic blueprint of strain NRRL B-14911, an emerging cardiopathogenic microbe Massilamany Chandirasegaran cmassilamany@unl.edu 1Mohammed Akram amohammed3@unl.edu 2Loy John Dustin jdloy@unl.edu 1Purvis Tanya tpurvis@vet.k-state.edu 3Krishnan Bharathi bharathipuji@gmail.com 1Basavalingappa Rakesh H. rakesh@unl.edu 1Kelley Christy M. Christy.Kelly@ARS.USDA.GOV 4Guda Chittibabu babu.guda@unmc.edu 2Barletta Raúl G. rbarletta@unl.edu 1Moriyama Etsuko N. emoriyama2@unl.edu 5Smith Timothy P. L. Tim.Smith@ARS.USDA.GOV 4Reddy Jay nreddy2@unl.edu 11 School of Veterinary Medicine and Biomedical Sciences, University of Nebraska-Lincoln, Lincoln, NE 68583 USA 2 University of Nebraska Medical Center, Omaha, NE 68198 USA 3 Kansas State Veterinary Diagnostic Laboratory, Manhattan, KS 66506 USA 4 Genetics, Breeding and Animal Health Unit, U.S. Meat Animal Research Center, Clay Center, NE 68933 USA 5 School of Biological Sciences and Center for Plant Science Innovation, University of Nebraska-Lincoln, Lincoln, NE 68588 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.511© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background We recently reported the identification of Bacillus sp. NRRL B-14911 that induces heart autoimmunity by generating cardiac-reactive T cells through molecular mimicry. This marine bacterium was originally isolated from the Gulf of Mexico, but no associations with human diseases were reported. Therefore, to characterize its biological and medical significance, we sought to determine and analyze the complete genome sequence of Bacillus sp. NRRL B-14911. Results Based on the phylogenetic analysis of 16S ribosomal RNA (rRNA) genes, sequence analysis of the 16S-23S rDNA intergenic transcribed spacers, phenotypic microarray, and matrix-assisted laser desorption ionization time-of-flight mass spectrometry, we propose that this organism belongs to the species Bacillus infantis, previously shown to be associated with sepsis in a newborn child. Analysis of the complete genome of Bacillus sp. NRRL B-14911 revealed several virulence factors including adhesins, invasins, colonization factors, siderophores and transporters. Likewise, the bacterial genome encodes a wide range of methyl transferases, transporters, enzymatic and biochemical pathways, and insertion sequence elements that are distinct from other closely related bacilli. Conclusions The complete genome sequence of Bacillus sp. NRRL B-14911 provided in this study may facilitate genetic manipulations to assess gene functions associated with bacterial survival and virulence. Additionally, this bacterium may serve as a useful tool to establish a disease model that permits systematic analysis of autoimmune events in various susceptible rodent strains. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2900-2) contains supplementary material, which is available to authorized users. Keywords Bacillus sp. NRRL B-14911GenomeHeartThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Heart failure (HF), a condition in which the heart is unable to adequately pump blood to rest of the body, is a leading cause of death worldwide. Estimates indicate that the current prevalence rate of HF is 2.8 %, and 825,000 new cases are reported annually in the United States alone [1]. While HF tends to be more prevalent in men than women in the age group of 40 to 79 years (1.73 to 2.1 %), women 80 years or older are more prone to the disease than men in that age group (1.4 %) [1]. Furthermore, the prevalence of HF is projected to increase from 2.8 % in 2010 to 3.3 % in 2025, and the economic loss resulting from HF is expected to double (~$34.1 billion in 2010 to ~$70 billion in 2025), in spite of continued efforts to contain the disease’s occurrence in the general population [2]. Various cardiovascular disease conditions have been implicated in the development of HF. These include pericardial and valvular diseases, atherosclerosis, hypertension, chronic ischemia, arrhythmia, diabetes, and myocarditis. Among the infectious causes, myocarditis has been identified as one important cause of HF in children and young adults. While most individuals affected with myocarditis may remain asymptomatic, 10 % can develop clinical heart disease. A proportion of these chronically affected individuals may develop dilated cardiomyopathy (DCM), and approximately half of them may undergo heart transplantation due to the lack of effective treatment options. Furthermore, it is estimated that approximately 2 million Americans appear to have inflammatory heart infiltrates, raising the possibility that a subset of people may have an ongoing silent myocarditis [3]. Two observations support this notion: (1) apparently healthy individuals like athletes can die from sudden death syndrome, and their autopsies suggest the presence of inflammatory infiltrates [4], and (2) a necropsy study involving more than 12,000 victims of accidental deaths not related to any cardiovascular diseases showed evidence of myocarditis in 1.05 % of cases [5]. Thus, identifying the triggers of myocarditis may provide opportunities to treat affected individuals in a timely fashion. The molecular mimicry hypothesis has been proposed as one major mechanism for the occurrence of autoimmune diseases including myocarditis, whereby the structural similarities between self and foreign antigens lead to recognition of self-antigens by generating cross-reactive immune responses [6, 7]. Numerous examples exist to support this theory in various disease conditions, such as experimental autoimmune encephalomyelitis/multiple sclerosis, experimental autoimmune uveoretinitis/uveitis and experimental autoimmune myocarditis/heart autoimmunity [7–13]. The importance of the molecular mimicry hypothesis can be summarized with two major predictable outcomes. (1) As the genomes of more microbes are sequenced, the search for mimicry sequences in the microbial databases has become relatively easy. These searches may result in the identification of microbes that are either natural pathogens of humans or are otherwise innocuous environmental isolates, but with the potential for them to trigger autoimmune diseases in those exposed. (2) Exposure to microbes carrying the mimicry sequences may result in the generation of cross-reactive immune responses leading to the induction of organ-specific autoimmunity. In our investigations, we identified a mimicry sequence for cardiac myosin heavy chain (Myhc)-α 334-352 contained in Bacillus sp. NRRL B-14911; the epitope, termed BAC 25-40, induces myocarditis by active immunization in A/J mice [12]. Because the biological and medical significance of this bacterium was not known, particularly regarding pathogenicity, we sought to analyze the complete genome of the organism to determine its phenotypic and virulence characteristics. By adopting phylogenetic analysis of 16S ribosomal RNA (rRNA) genes, sequence analysis of the 16S-23S rDNA intergenic transcribed spacers (ITS), phenotypic microarray (PM), and matrix-assisted laser desorption ionization time-of-flight (MALDI-TOF) mass spectrometry (MS), we propose the species and strain of Bacillus sp. NRRL B-14911 to be Bacillus infantis NRRL B-14911. The availability of the complete genome sequence for this bacterium may facilitate genetic manipulations to assess gene functions associated with bacterial survival and virulence. Additionally, this bacterium can serve as a useful tool to establish a novel disease model for autoimmune myocardial damage of infectious origin. Methods Bacterial strain, culture conditions and isolation of genomic DNA Bacillus sp. NRRL B-14911 was procured as a kind gift from the Agricultural Research Service (ARS) Culture Collection, United States Department of Agriculture (Washington, DC). Bacillus infantis sp. nov. (type strain SMC 4352-1 T = KCCM 90025 T = JCM 13438 T) was procured form Japan collection of microorganisms (Koyadai, Japan). For isolation of genomic DNA, bacteria were grown in Luria Bertani (LB) broth (volume to flask ratio of 1:10) at 37 °C with gentle shaking at 200 rpm for 36 h. The genomic DNA was isolated using Qiagen genomic-tip 100 as recommended by the manufacturer (Qiagen, Valencia, CA). Genome sequencing, analyses and annotation The genomic DNA of Bacillus sp. NRRL B-14911 was prepared for sequencing on the Pacific Biosciences RSII instrument as detailed in the procedure provided by the manufacturer (Procedure and Checklist - Greater Than 10 kb Template Preparation and Sequencing, Dec. 2012 version). Briefly, DNA was sheared using a Covaris G-tube (Covaris Inc., Woburn MA) to achieve fragments in the 5000-15,000 base pair range. A library was prepared from the DNA with size selection performed only by precipitation of the DNA onto AMPure PB beads, using the DNA Template Prep Kit 2.0 (Pacific Biosciences, Menlo Park CA). Sequencing was performed using XL/C2 chemistry in two SMRT cells, producing 176,090 reads with average read length 5185 bases, and 240,232 subreads (912 Mb total sequence) with average subread length 2972 bases (N50 = 4481 bases). Assembly was performed using PacBioToCA for error correction and Celera Assembler v7 for assembly as described [14]. Two contigs, one representing the bacterial chromosome and one representing a plasmid, were produced. The ends of the chromosomal contig were examined for overlap using nucleotide Basic Local Alignment Search Tool (BLASTN), which identified the most likely position at which the chromosome could be circularized. After removing redundant sequence, the origin was estimated by GenSkew (http://genskew.csb.univie.ac.at) analysis, and the linear contig was reset so that the estimated origin was base 1. The assembly was improved by polishing with Quiver from the SMRTportal software package (Pacific Biosciences), which fixed the insertions common in initial assemblies, and also confirmed the correct positioning during the circularization step. The finished assembly was annotated by NCBI, and annotation anomalies identified were curated and revised in Geneious (Biomatters ltd., New Zealand). Phylogenetic analysis Phylogenetic analysis was done using the 16S ribosomal DNA sequences from 28 selected species (23 from the genus Bacillus and 5 from its related genera Halobacillus, Oceanobacillus, Geobacillus and Paenibacillus). All sequences were obtained from the GenBank database at the National Center for Biotechnology Information (NCBI). The sequences were aligned using MAFFT (v7.130b) with the L-INS-i algorithm [15]. The maximum likelihood phylogeny was reconstructed using PhyML (version 3.0) [16] with the GTR substitution model, the proportion of invariable sites and gamma shape parameters both estimated, and the option to choose the best tree from the nearest-neighbor interchange tree-rearrangement and subtree-pruning/regrafting. Non-parametric bootstrap analysis was done with 1000 pseudoreplicates. Comparison of conserved ITS sequences between Bacillus sp. NRRL B-14911 and B. infantis based on genomic DNA PCR analysis Because that 16S-23S rDNA ITS sequences are hypervariable, but conserved within the same species, their sequence analyses have been successfully used for speciation of the genus Bacillus [17–19]. Briefly, 16S-23S rDNA ITS regions were amplified from the genomic DNA extracted from Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T using the primers specific to Bacillus genus as described previously [17, 18]. The primer sequences used were: 5'-GTCGTAACAAGGTAGCCGTA-3'/5'-CAAGGCATCCACCGT-3'; 5'-CCTTGTACACACCGCCCGT-3'/5'-AAAATAGCTTTTTGGTGGAG-3' ; and 5'-AAATTTGTATGGGCCTATAG- 3'/5'-GTGGGTTTCCCCATTCGG-3', and the amplifications were performed using the following conditions: 94o C for 4 min followed by 32 cycles, each consisting of 94o C for 1 min, 54o C for 1 min, 72o C for 2 min with a final extension at 72o C for 10 min. After resolving the PCR products in 1 % agarose gel, the PCR amplicons were excised, purified using gel extraction kit (Qiagen, San Jose, CA) and subjected for sequencing. After excluding the 16S and 23S rDNA sequences from the amplicons, the nucleotide sequences representing only the ITS regions were recovered, and their percent identities were analyzed using William Pearson’s lalign program (http://www.ch.embnet.org/software/LALIGN_form.html). Phenotypic and biochemical characterization Phenotypic analysis was performed using the Biolog GEN III microplate using Omnilog Data collection software (Biolog, Inc., Hayward, CA) [20, 21]. The components in the wells of the 96-well plates were comprised of sources for carbon (C), nitrogen (N), phosphorous (P), sulfur (S) and amino acids. The tests included: utilization of sugars, amino acids and organic acids; tolerance to NaCl; and susceptibility to antibiotics. To perform a comparative analysis, Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T were subcultured twice in isolation medium (trypticase soy agar with 5 % sheep blood, Remel, Thermofisher Scientific, KS) and inoculated individually in the wells of microplate. Protocol A was used and analysis was performed at 10 h postincubation, as per the manufacturer’s recommendation for members of the genus Bacillus. In addition, testing for Gram-staining, oxidase and catalase activities, and endospore formation were performed by standard microbiological procedures. Carbohydrate fermentation testing results for sorbitol, inulin, and lactose were confirmed using rapid fermentation tablets (Wee-Tabs, Key Scientific, Stamford Texas). Antimicrobial susceptibility testing for vancomycin was performed using Kirby-Bauer disk diffusion susceptibility testing following clinical and laboratories standards institute (CLSI) guidelines. As no interpretive criteria for assessing disk diffusion breakpoints for Bacillus sp. have been determined, interpretive criteria from CLSI M100-S22 for Staphylococcus sp. was applied to determine in vitro susceptibility breakpoints. Spore staining Bacterial smears prepared on glass slides were fixed by methanol and air-dried. The smears were then stained with malachite green solution (5 min) under steam, washed and counterstained with safranin (30 s). After washing, the slides were air-dried and examined under the microscope with an oil immersion lens. MALDI-TOF MS analysis The Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T cells grown in LB broth at 37o C were plated onto tryptic soy agar with 5 % sheep blood agar plates. Following overnight incubation, individual colonies were picked and spotted onto the MALDI-TOF target. The spots were overlaid with 1 μl of α-cyano-4-hydroxycinnamic acid (HCCA) matrix (Bruker), and the mass spectra were acquired using MALDI-TOF MS, Microflex LT system in a linear positive mode (Bruker Daltonik, Billerica, MA). Instrument calibration was performed using standard reference BTS Escherichia coli (Bruker). For bacterial identification, MALDI Biotyper 3.0, Reference Library 1.0 Version 3.1.2 was used [22, 23]. The cut-off scores used for species identification were: 2.300 to 3.000–highly probable species identification; 2.000 to 2.299–secure genus identification and probable species identification; 1.700 to 1.999–probable genus identification; and 0.000 to 1.699–not reliable for species identification. Sequence analysis of Allantoate amidohydrolase gene Allantoate amidohydrolase (AAH) gene was amplified from the genomic DNA obtained from Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T using sequence specific primers (5'-GCTGGCTTGAAAAAAATC-3'/5'-GGAGGCAAATTCATCTGG-3'). The PCR products were resolved by 1 % agarose gel electrophoresis, and the amplified product was purified using gel extraction kit and subjected for sequencing. Analysis of Bacillus sp. NRRL B-14911 genome for virulence factors The Virulence Factor Database (VFDB; http://www.mgc.ac.cn/VFs/) is a database constructed by the virulence-guided classification system. The core dataset of VFDB (VFs.faa) consisting of 502 virulence factors (VFs) from 2505 VF-related genes representing 25 genera of pathogenic microbes [24] was downloaded and used as the database for protein sequence similarity search. Using this information, we performed three types of analyses: First, we identified potential virulence factors from the proteins encoded in the Bacillus sp. NRRL B-14911 genome using protein Basic Local Alignment Search Tool (BLASTP) (version 2.2.25+) [25]; second, we blasted the proteomes of three pathogenic bacilli [B. anthracis str. Ames, GenBank Accn# NC_003997.3; B. cereus ATCC 14579, GenBank Accn# NC_004722.1; B. licheniformis ATCC 14580, GenBank Accn# NC_006270.3]; and three non-pathogenic bacilli [B. pseudofirmus OF4, GenBank Accn# NC_013791.2; B. amyloliquefaciens subsp. plantarum UCMB5033, GenBank Accn# NC_022075.1; B. subtilis subsp. subtilis str. OH 131.1, GenBank Accn# NZ_CP007409.1] against the protein sequences in the VFDB database to identify the potential virulence factors present in the respective groups. These were then compared with those of Bacillus sp. NRRL B-14911 to identify the virulence factors unique to this bacterium; and third, we compared the virulence factors of the pathogenic bacilli as described above with those of Bacillus sp. NRRL B-14911 to identify those that are common to both. The thresholds used were E-value of 1×10-10 and bit score of 40. CGView server was used to draw the circular map to show the location of potential virulence factor genes in Bacillus sp. NRRL B-14911 [26]. Comparative analysis of the Bacillus genomes For comparative genomic analysis of Bacillus sp. NRRL B-14911 with other species within the genus Bacillus, we downloaded the complete genome annotations of B. subtilis strain 168 [GenBank Accn# 225184640], B. megaterium DSM 319 [GenBank Accn# CP001982.1], B. thuringiensis serovar kurstaki strain HD73 [GenBank Accn# CP003889.1] and B. cereus ATCC 14579 [GenBank Accn# AE016877.1] from the NCBI database. The presence or absence of genes encoding methyltransferases and transporters as well as insertion sequence (IS) elements was determined based on the “product” assignments in each genome annotation. The comparative analysis for enzymes and biochemical pathways was performed using pathway mapping for each genome in the Kyoto Encyclopedia of Genes and Genomes (KEGG) database [27]. Results and discussion We report here the complete genome sequence analysis of the bacterium Bacillus sp. NRRL B-14911 that has a potential to induce heart autoimmunity by molecular mimicry. Bacillus sp. NRRL B-14911 was originally isolated from ocean water at a depth of 10 m in a sea expedition seeking to study the marine microflora in the Gulf of Mexico and around the Bimini Islands. However, the possible significance of this bacterium as a pathogen was unknown [28]. Based on our discovery that Bacillus sp. NRRL B-14911 contains a disease-inducing mimicry epitope for cardiac myosin, we sought to determine the biological significance of this organism to humans. To this end, we decided to sequence the complete genome of Bacillus sp. NRRL B-14911 and characterize its phenotypic and biochemical features with the expectation that identification of its species may create opportunities to establish a new disease model to study the autoimmune events of bacterial myocarditis in experimental settings. Genome sequencing Morphologically, Bacillus sp. NRRL B-14911 was found to be a Gram-positive, rod-shaped, sub-terminal endospore-forming, aerobic bacterium with rounded ends as observed by light microscopy (Additional file 1: Figure S1). For complete genome sequencing, we isolated the genomic DNA and performed sequencing in long-reads by PacBio RS SMRT sequencing technology [29, 30]. The sequencing produced >900 Mb of post-filter sequences, consisting 176 K reads of average >5 100 bases. Initial assembly of the genome produced a circularizable contig of 4,884,884 bases. The assembly was further refined in Quiver to generate a final assembly of 4,884,713 bases; wherein, 84 % of subreads mapped back to the assembly, resulting in consensus calling with an average base coverage of 114X. A GC skew analysis indicated that the origin of replication occurred at the 3,496,945th position. This position was then renumbered as position 1 to indicate the origin of replication. The overall GC content of the genome was estimated to be 46 %, which is relatively higher than the GC content of genomes from B. subtilis strain 168, B. megaterium DSM 319, B. thuringiensis serovar kurstaki strain HD73 and B. cereus ATCC 14579 (Table 1). These bacteria were chosen for comparison because their complete genome sequences were available in the NCBI database. Nonetheless, the GC content correlated positively with the percent coding region, but no other major differences were noted, except that the number of tRNAs and rRNAs were relatively low in Bacillus sp. NRRL B-14911 (Table 1). The genomic sequence was annotated and submitted to the NCBI [GenBank Accn# CP006643].Table 1 Comparative analysis of the genome of Bacillus sp. NRRL B-14911 with the genomes of other Bacillus species Parameter Bacillus sp. NRRL B-14911 B. subtilis strain 168 B. megaterium DSM 319 B. thuringiensis serovar kurstaki strain HD73 B. cereus ATCC 14579 Genome size (bp) 4884713 4215606 5097447 5646799 5411809 GC content (%) 46 43.5 38.1 35.2 35.3 Protein coding (bases) 4161286 3694614 4174023 4718007 4364840 Protein coding (%) 85.19 87.65 81.89 83.56 80.66 Number of genes 5179 4421 5272 6334 5501 Gene density (bp per gene) 943 953 967 892 987 tRNAs 85 86 115 104 108 rRNAs 27 30 33 36 39 Plasmids 1 - - 7 1 Accession numbers: CP006643.1, Bacillus sp. NRRL B-14911; 225184640, B. subtilis strain 168; CP001982.1, B. megaterium DSM 319; CP003889.1, B. thuringiensis serovar kurstaki strain HD73; AE016877.1, B. cereus ATCC 14579 Analysis of the genomic sequence of Bacillus sp. NRRL B-14911 revealed several noteworthy features. (1) The bacterium was found to contain one large plasmid with a size of 144,911 bases [Table 1; KF831061, awaiting confirmation from the GenBank Accn#]. The plasmid encodes several proteins, including component of the type IV secretion system (a conserved large VirB4 domain protein; Additional file 1: Figure S2). (2) Base-modification analysis of the genome revealed two different motifs, one with methylation of adenine at the 6th position to yield N6-methyladenine (m6A) on both the DNA strands, and the other on only one strand. The respective motifs are: CACNNNNNCTNG/CNAGNNNNNGTG (786/815 occurrences = 96.4 %; mean modification QV = 147.7) and GGAGT (4926/5958 occurrences = 82.7 %; mean modification QV = 134.7). We further scanned the chromosome and plasmid sequences using PacBio software to identify the restriction and modification systems that could be responsible for methylation at the specific motifs. While the genome did not reveal any of the above, the plasmid was found to contain a type I restriction-modification system specific for m6A modification. Although such modifications were suspected to have a role in pathway regulation [31], the role of the type I restriction-modification system in Bacillus sp. NRRL B-14911 has not been determined. (3) The genome encoded two vancomycin and one kanamycin resistance genes; however, the vancomycin resistance genes were found to be non-functional due to frameshifting insertions. Previous efforts to assemble the Bacillus sp. NRRL B-14911 genome using the 454 sequencing technology [Bioproject of Siefert, et al. 2000; GenBank accn# NZ_AAOX00000000] differed from our long-read sequencing approach [28]. The BioProject had sequences for a total of 74 contigs, consisting of 18 large and 56 small contigs (Fig. 1; and Additional file 1: Table S1). The 18 large contigs had coverage of 4,002,278 bases, accounting for 82 % of the total genomic length (Fig. 1). Eight of them, however, showed sequencing errors/ambiguity as denoted by ‘Ns’ in most places. Five of the 56 small contigs did not match our assemblies of either chromosome or plasmid (Additional file 1: Table S1). One of these (4281 bp; NZ_AAOX01000070) failed to match with any of the sequences in the non-redundant nucleotide database at NCBI (using BLASTN similarity search). Eight out of the remaining 51 small contigs matched only partially with our plasmid sequence, with coverage of 8 % (11,415 bases; Additional file 1: Table S1). Thus, the plasmid did not seem to be identified in their data set. Finally, among the remaining 43 small contigs, seven (<1.5 kb) matched our chromosome sequence, and the matching of another contig (NZ_AAOX01000091) was nearly perfect but fragmented in nine locations (Additional file 1: Table S1). Sequences of the remaining 35 small contigs matched with our genomic sequence. Comparison of our genome with the scaffold sequences derived in the bioproject also revealed several alterations in the orientation of the sequences. Overall, the total size of the genomic sequence submitted by Siefert et al. 2000 [28] was estimated to be 5,086,957 bases, 202,244 more bases than in our assembly (4,884,713 bases). This inflated genome size may be caused by many ‘Ns’ included in their sequences. The elimination of these Ns makes the relative coverage of their sequences (4,844,207 bases) to our genome to be 99.2 % (Fig. 1; Additional file 1: Figure S3), suggesting that their sequencing may possibly be near completion, but the sequences were fragmented and not assembled. Taken together, our sequencing approach using PacBio SMRT led to complete genome assembly with no errors, as also reported by others [29, 30]. However, heterogeneity within and across colonies generated from the same samples cannot be ruled out for the differences observed between the two approaches.Fig. 1 Comparison of assemblies of the Bacillus sp. NRRL B-14911 genome based on sequencing long-reads followed by de novo assembly as opposed to sequencing short-reads with subsequent scaffold building. Coverage maps depict short-read contigs assembled by scaffolding (top panel) and short-read contigs prior to scaffolding (middle panel) from a previous BioProject assembly of the Bacillus sp. NRRL-14911 genome, as aligned against our de novo assembly using long-reads. Alignment of short-read scaffolds and contigs is shown in the bottom panel. Red denotes 1X coverage by scaffolds, and blue denotes 1X coverage by the remaining unscaffolded contigs from the BioProject final assembly. Arrows point to areas with 2X coverage by sequences in their final contig list (visible as green sections in the online version with zoom) Identification of the species for Bacillus sp. NRRL B-14911 as B. infantis To identify the species of Bacillus sp. NRRL B-14911, we adopted four approaches: phylogenetic analyses of the 16S rRNA gene sequences, analysis of the 16S-23S rDNA ITS sequences, biochemical, and MALDI-TOF analyses. Phylogenetic analyses of 16S rRNA gene sequences To determine the species-identity of Bacillus sp. NRRL B-14911, we performed phylogenetic analysis of the 16S rRNA gene sequences, a system that has been routinely used for speciating various bacteria [32–34]. We compared the 16S ribosomal DNA sequences of Bacillus sp. NRRL B-14911 with those from 28 selected species including Bacillus and four other related genera (Halobacillus, Oceanobacillus, Geobacillus and Paenibacillus) (Fig. 2). The phylogenetic analysis revealed that Bacillus sp. NRRL B-14911 formed a clade distinct from the soil-dwelling bacilli (Fig. 2). As expected, the marine inhabitants from the genera Halobacillus, Oceanobacillus, Geobacillus and Paenibacillus formed separate clades, further validating the reliability of using 16S rRNA gene sequence analysis for species identification (Fig. 2). Within the marine bacilli, Bacillus sp. NRRL B-14911 formed a well-supported cluster with two strains of B. infantis (B. infantis 9 and B. infantis SMC 4352-1) as well as B. mangrovensis (88 % bootstrap supporting value) indicating their close relationships (Fig. 2). Bacillus sp. NRRL B-14911 was particularly close to the two strains of B. infantis suggesting a possibility that Bacillus sp. NRRL B-14911 is likely to be B. infantis.Fig. 2 Construction of phylogenetic tree representing the 16S rRNA gene sequences derived from various Bacillus species. Bootstrap supporting values (%) are shown at the nodes only when higher than 70 %. Box indicates clustering of Bacillus sp. NRRL B-14911, B. infantis strain 9 and B. infantis strain SMC 43521-1 together in one clade Sequence analysis of ITS regions Because that phylogenetic analysis suggested the probable species of Bacillus sp. NRRL B-14911 to be B. infantis JCM 13438 T, we sought to analyze the sequences of their ITS regions for further analysis. The 16S-23S rDNA ITS regions are considered to be the most variable regions of the ribosomal operon [17]. The sequences of these regions have been proposed to be species/strain-specific for the genus, Bacillus [17, 18]. Therefore, we performed a comparative evaluation of the ITS regions of both Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T to determine the similarities between the two. Using three sets of primers as previously described [17–19], we performed PCR analysis of genomic DNA from both the microorganisms. We named the resulting six amplicons as ITS1, ITS2, ITS3, ITS4, ITS5 and ITS6, and sequenced them. This analysis yielded the following information: (i) The sizes and patterns of PCR amplicons obtained from both Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T were similar (Fig. 3); (ii) By excluding the sequences of 16S and 23S rDNAs, we were able to determine the identities of five ITS-amplicons, ITS1, ITS2, ITS4, ITS5 and ITS6, and their sizes ranged from 151 to 268 bp (Table 2). These sequences also matched with the ITS regions of Bacillus sp. NRRL B-14911 [GenBank Accn# CP006643]. However, one amplicon, ITS3 did not yield ITS-specific information; and (iii) Comparisons of sequences of ITS1, ITS2, ITS4, ITS5 and ITS6 between Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T revealed their identities ranged from 96.7 to 100 % (Table 2). As the cut-off value for species identification based on ITS sequences has been suggested to be at least 92 % and the identities of both Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T meet this criterion, it is likely that both strains belong to the same species.Fig. 3 Comparative analysis of 16S-23S rDNA ITS of Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T. The ITS regions of Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T were amplified by PCR using the genomic DNA as a template as described in the Methods section. The PCR products were resolved in 1 % agarose gel electrophoresis and stained with ethidium bromide Table 2 Comparison of the sequence identities of the different ITS region of Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T Patterns Length of ITS (bp) % Similarity ITS-1 151 96.7 ITS-2 268 98.9 ITS-4 193 100 ITS-5 167 100 ITS-6 230 100 Biochemical analyses To phenotype and compare the biochemical characteristics between Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T, we used Biolog PM (Biolog Omnilog) to analyze parameters, such as utilization of sugars and amino acids and other carbon sources, ability to grow at high salt concentrations, and growth in the presence of acids and antibiotics (Biolog Inc.). Biolog PM involves the reduction of tetrazolium compounds due to the utilization of a specific substrate in query under minimally defined nutrient conditions [20, 21]. PM analysis predicted the probable species of B. infantis JCM 13438 T and Bacillus sp. NRRL B-14911 to be the same using inoculation protocol A, the recommended protocol for Bacillus. The Omnilog software determined a final identification of B. infantis with a similarity index value (SIM) of 0.314 and 0.271 for B. infantis JCM 13438 T and Bacillus sp. NRRL B-14911 respectively. The next closest matches in the Omnilog database were Brevibacterium otitidis for Bacillus sp. NRRL B-14911; and Staphylococcus equorum subsp. equorum for B. infantis JCM 13438 T with SIM values of 0.314 and 0.126 respectively, which may not be the reliable identification for the genus Bacillus. It is to be noted that the SIM values are lower because of the recommend truncated incubation time used for Bacillus, where the automated ID software would normally incubate a protocol A 20 h or more. Further, by comparing the biochemical characteristics of Bacillus sp. NRRL B-14911 with B. infantis JCM 13438 T [35], we noted that most of the parameters of Bacillus sp. NRRL B-14911 complemented those of B. infantis JCM 13438 T (Table 3). Likewise, by examining metabolism in the GENIII plate containing various antimicrobials, we noted that both the bacteria were likely susceptible to vancomycin, troleandomycin, lincomycin, and a few tetrazolium compounds. Sensitivity to vancomycin was further confirmed with disc diffusion testing and it supports the finding of the genomic sequence analysis which also found the corresponding gene to be non-functional as described above.Table 3 Comparison of biochemical characteristics of Bacillus sp. NRRL B-14911 with B. infantis JCM 13438 T Biochemical Test Bacillus sp. NRRL B-14911 B. infantis JCM 13438 Ta α-D-Glucose + + D-Mannose +/- +/- D-Mannitol + + D-Maltose + + D-Melbiose + + D-Fructose + + D-Trehalose + + D-Galactose + + Sorbitol - - Inulin - +/- Esculin +/- + Glycerol +/- +/- Cellobiose +/- + Gentiobiose + + Sucrose + + Oxidase - - Catalase + + D-Raffinose +/- +/- Gelatin - - Pectin + + p-hydroxy-phenylacetic acid - - Tween 40 - +/- Dextrin + + α-D-Lactose +/- + Glylcyl-L-Proline - - Methyl Pyruvate - - γ-amino-butyric acid - - D- Arabitol - - L-Alanine - - D-Lactic Acid Methyl Ester - - α-Hydroxy Butyric Acid - +/- β-Methyl-D-Glucoside + +/- myo-Inositol - - L-Arginine - - D-Glucononic Acid + + β-hyrdoxy D, L, Butyric Acid - - D-Salicin +/- +/- L-Aspartic Acid - - D-Glucuronic Acid + + Citric Acid - - α-Keto-Butyric Acid - - D-Fucose +/- + D-Glucose-6-PO4 + + Glucuronamide + + α-Keto Glutaric Acid - - Acetoacetic Acid + +/- N-Acetyl-β-D Mannosamine - - L-Fucose + + D-Fructose-6-PO4 + + L-Histidine - - Mucic Acid +/- +/- D-Malic Acid - - Propionic Acid +/- - D-Turanose +/- + N-Acetyl-D-Galactosamine - - L-Rhamnose +/- + D-Aspartic Acid - - L-Pyroglutamic Acid - - Quinic Acid - - L-Malic Acid - - Acetic Acid +/- +/- N-Aceytl Neuraminic Acid - - Inosine - +/- D-serine - - L-Serine - - D-saccharic acid - - Bromo-Succinic Acid - - Formic Acid - - 1 % NaCl + + 1 % Sodium Lactate + + Troleandomycin - - Lincomycin - - Vancomycin - - Nalidixic Acid +/- - Aztreonam + + ph 6.0 + + 4 % NaCl + + Fusidic Acid - - Rifamycin SV - - Guanidine HCL +/- +/- Tetrazolium Violet +/- +/- Lithium Chloride + + Sodium Butyrate + + pH 5.0 - - 8 % NaCl +/- +/- Minocycline - - Niaproof 4 - - Tetrazolium Blue +/- +/- Potassium Tellurite + + Sodium Bromate - - +, present; -, absent; a,reference organism MALDI-TOF analysis MALDI-TOF analysis has been widely used to discriminate bacteria at genus, species, subspecies and strain levels [23, 36–38]. MALDI-TOF analysis was conducted on the Bacillus sp. NRRL B-14911 proteome using Bruker Daltonik MALDI Biotyper. The analysis predicted the bacterium to be B. infantis after four independent analysis with scores 2.137, 2.139, 2.228, 2.097. The second closest matches in the database were B. nealsonii and B. firmus. To further validate this finding, we repeated the MALDI-TOF analysis using proteomes from both Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T. These analyses predicted both the bacteria to be B. infantis with scores ranging from 1.883 to 2.065 for Bacillus sp. NRRL B-14911, and 1.809 to 2.024 for B. infantis JCM 13438 T suggesting that their proteomic profiles are similar. Little is known about the diversity within the species of B. infantis as only a very small number of isolates have been described. As the data generated with both biochemical and MALDI-TOF analyses agreed with the phylogenetic analysis, but the sequence analyses of ITS regions revealed identities in the range of 96.7 % to 100 %, we believe that Bacillus sp. NRRL B-14911 may represent strain variation within the species B. infantis. Thus, we suggest the species and strain of Bacillus sp. NRRL B-14911 as B. infantis NRRL B-14911. Analysis of virulence factors The ability of bacteria to cause disease in susceptible hosts is determined largely by their virulence factors. We attempted to identify the genes from B. infantis NRRL B-14911 that encode for various virulence factors based on the sequence similarities with virulence factor proteins found in the VFDB database. A total of 623 proteins from Bacillus sp. NRRL B-14911 were identified to be potential virulence factors. A list of these proteins and their gene locations are shown in Additional file 1: Table S2 and Fig. 4. Among these, 18 proteins were found to be unique to Bacillus sp. NRRL B-14911 when compared with those from pathogenic (B. anthracis str. Ames, B. cereus ATCC 14579, and B. licheniformis ATCC 14580) and non-pathogenic bacilli (B. pseudofirmus OF4, B. amyloliquefaciens subsp. plantarum UCMB5033, B. subtilis subsp. subtilis str. OH 131.1) (Additional file 1: Table S3 and Additional file 1: Figure S4). Similarly, by comparing with the pathogenic bacilli alone, we noted that 225 genes to be common for both Bacillus sp. NRRL B-14911 and the pathogenic bacilli (Additional file 1: Table S4 and Additional file 1: Figure S5). The notable virulence factor genes include intercellular adhesion protein, invasion-associated protein, accessory colonization factor, laminin-binding surface protein, toxin co-regulated pilus biosynthesis protein, transporters, and regulatory proteins PhoP/PhoQ.Fig. 4 Circular genome map of Bacillus sp. NRRL B-14911 showing the location of genes for virulence factors. The map shows the locations of 623 genes of Bacillus sp. NRRL B-14911 potentially encoding virulence factors. The two outer blue circles represent the genes for virulence factors shown in forward and reverse directions. The innermost circle represents the GC skewness, and the second innermost circle represents the GC content of the Bacillus sp. NRRL B-14911 genome. The coordinates of each gene are listed in Additional file 1: Table S2. The start and end positions in Additional file 1: Table S2 match with the location of genes in Fig. 4 Other pathways We evaluated the genome of Bacillus sp. NRRL B-14911 for the presence or absence of methyltransferases, transporters, enzymatic and biochemical pathways and IS elements and compared the results with those from four other bacilli: B. subtilis strain 168, B. megaterium DSM 319, B. thuringiensis serovar kurstaki strain HD73 and B. cereus ATCC 14579. Methyl transferases and transporters In prokaryotes, DNA-methylation controls a number of physiological processes, including transcription, DNA mismatch repair and initiation of replication. Three classes of methyltransferases have been identified in bacteria: the first two classes transfer a methyl group from S-adenosyl-L-methionine (SAM) to adenine and cytosine to yield m6A and N4-methylcytosine (m4C), respectively; and the third class transfers a methyl group from SAM to cytosine to generate 5-methylcytosine (m5C) [39]. Five methyl transferases were found to be present uniquely in B. infantis NRRL B-14911. These include 50S rRNA methyltransferase, lysine methyltransferase, N5-glutamine SAM-dependent methyltransferase, protein-L-isoaspartate O-methyltransferase, and SAM-dependent methyltransferase (Additional file 1: Table S5). It is reported that rRNA methyltransferases confer antibiotic resistance to the bacteria by adding methyl groups specifically to the 23S rRNA, and prevent binding of drugs/antibiotics to the large subunit of the ribosome [40]. Thus, bacteria like B. infantis NRRL B-14911 that possess 50S rRNA methytransferases may have a survival advantage under antibiotic selection pressure. Similarly, lysine methyltransferases are known to mediate methylation of lysine residues in ribosomal and flagellar proteins and have a role in the posttranslational modification processes [41]. Two main superfamilies of transporters have been identified in bacteria. These include ion-coupled transporters and the ABC solute ATPases, which maintain in- and out-flow of nutrients and wastes. We noted a number of transporters present in Bacillus sp. NRRL B-14911 (Additional file 1: Table S6). A few unique transporters include (1) antibiotic ABC transporter ATP-binding protein and arsenic transporter ATPase/arsenite efflux transporter, which determine resistance to antibiotics and arsenic by extrusion [42, 43]; (2) C4-dicarboxylate ABC transporter, a tripartite ATP-independent periplasmic transporter that transports organic acids like succinate, malate fumarate, keto-acids and N-acetyl neuraminic acid [44]; (3) corrinoid ABC transporter that facilitates the intake of complex cyclic tetrapyrrole molecules such as hemes, chlorophylls and coenzyme F430 [45], (4) macrolide transporter, an efflux transporter of macrolide drugs like erythromycin and azithromycin, which determines resistance to antibiotics [46], (5) nicotinamide riboside transporter that aids in the uptake of nicotinamide riboside into the cytoplasm [47, 48]; (6) nitrate ABC transporter that mediates uptake of nitrate into the cell [49]; (7) peptide ABC transporter, which is often present in firmicutes, that determines resistance to antimicrobial peptides by substrate extrusion from the cell [50]; and (8) riboflavin transporter fmnp, which is involved in the uptake of riboflavin into the cell and a frequently occurring transporter in firmicutes [51]. Enzymes and biochemical pathways As described above, Bacillus sp. NRRL B-14911 contains a mimicry epitope (BAC 25-40; EGFTRLSFTAEEKAAH) for cardiac myosin peptide (Myhc-α 334-352; DSAFDVLSFTAEEKAGVYK) (identical residues are bolded), with allantoate amidohydrolase (AAH) as the source protein [12]. As expected, the annotated gene sequence of AAH also contains the exact amino acid sequence of the mimicry epitope as indicated above [GenBank protein ID: AGX06322]. To further confirm whether B. infantis JCM 13438 T also contains the gene for AAH, and if so, whether the sequence for mimicry epitope BAC 25-40 is conserved, we amplified the AAH gene from both Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T using the genomic DNA as a template, and sequenced the PCR products. These analyses revealed the presence of AAH gene in B. infantis JCM 13438 T (Fig. 5), and the amino acid sequence of the mimicry epitope, BAC 25-40 was also conserved except one silent mutation (GAG in place of GAA for glutamic acid, E; Additional file 1: Table S7). Previously, we had reported the conservation of the mimicry epitope, BAC 25-40 in various other Bacillus species [12]. Functionally, the biochemical reaction carried out by AAH is a two-step conversion of allantoate to ureidoglycolate and ammonia [52], and AAH functionality has been detected in both plants and bacteria. It is possible that the AAH gene may have been laterally transferred between plants and bacteria for recycling nitrogen [52, 53]. We also noted that B. infantis NRRL B-14911 is capable of biosynthesizing LPS and steroids (Additional file 1: Table S8). Additionally, as reported by others [54], our sequence analysis revealed the presence of a novel class of extracellular poly (3-hydroxybutyrate) (PHB) depolymerase. This enzyme is required for degradation of PHB to produce 3-hydroxybutyrate as an intracellular carbon and energy source under conditions of limited or unbalanced nutrient-availability [54]. We speculate that the PHB depolymerase may be critical for bacterial survival in the environment.Fig. 5 Conservation of AAH gene in Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T. AAH gene was amplified from Bacillus sp. NRRL B-14911 and B. infantis JCM 13438 T using genomic DNA as a template by PCR. The PCR amplicons were resolved in 1 % agarose gel electrophoresis and stained with ethidium bromide IS elements The IS elements are recombinationally active, mobile, genetic segments of bacterial DNA (600 to 3000 bp) that move from one position to another within the same chromosome or to a different chromosome. One consequence could be inactivation of gene expression if the insertion of IS elements takes place within the coding sequence [55, 56]. We noted that IS1, IS1380/IS942, IS5/IS903 and Tn3 are uniquely present in the genome of B. infantis NRRL B-14911 (Additional file 1: Table S9) compared to other bacilli. Previously, it was demonstrated that the transposon Tn3 (4957 bp) carries the enzyme β-lactamase, in addition to transposase and resolvase, and confers resistance to β-lactam antibiotics [57, 58]. Whether the Tn3 present in B. infantis NRRL B-14911 perform similar functions requires additional studies. Conclusions In summary, we have described the complete genome sequence analysis of a marine microbe called Bacillus sp. NRRL B-14911. The bacterial genome sequence analysis allowed us to identify genes for a wide range of virulence factors and enzymatic and biochemical pathways, including IS elements that are distinct from other closely related bacilli. The availability of the complete genomic sequence of Bacillus sp. NRRL B-14911 may thus provide opportunities to genetically manipulate its genome to study the genes in bacterial survival and virulence. Furthermore, phylogenetic and 16S-23S rDNA ITS sequence analyses including biochemical and phenotypic characterizations suggested a close-association with B. infantis, and as such, we have proposed the species and strain of Bacillus NRRL B-14911 as B. infantis NRRL B-14911. Of note, B. infantis was previously identified as one of the six bacterial isolates from a newborn child with sepsis, but its pathological significance was unknown [35]. Similarly, a variety of pathogens have been implicated in the causation of heart autoimmunity, but their direct causal links remain tenous clinically. Thus, it becomes difficult to explain the persistent inflammation in the absence of detectable infectious particles. In these circumstances, autoimmunity is suspected with a challenge being able to prove the cause and effect relationship. Mechanistically, break in self-tolerance as a result of exposure to microbes carrying mimicry epitopes for self-antigens like cardiac myosin can lead to heart autoimmunity through the generation of cross-reactive T cells as we have demonstrated for BAC 25-40 present in Bacillus spp. NRRL B-14911 ([12], Additional file 1: Table S7 and Fig. 5). This bacterium, may thus serve as a useful tool to establish a disease model that permits systematic analysis of autoimmune events with respect to the appearance, disappearance, persistence, and/or reappearance of cross-reactive T cells and their functionalities experimentally in susceptible rodent strains. Abbreviations AAH, allantoate amidohydrolase; BLASTN, Basic Local Alignment Search Tool; BLASTP, Basic Local Alignment Search Tool; DCM, dilated cardiomyopathy; HF, heart failure; IS, insertion sequence; ITS, intergenic transcribed spacers; m4C, N4-methylcytosine; m5C, 5-methylcytosine; MALDI-TOF, matrix-assisted laser desorption ionization time-of-flight; MS, mass spectrometry; Myhc, cardiac myosin heavy chain; PHB, poly (3-hydroxybutyrate); PM, phenotypic microarray; rRNA, ribosomal RNA; SAM, S-adenosyl-L-methionine; VFDB, Virulence Factor Database Additional file Additional file 1: Figure S1. Identification of Bacillus sp. NRRL B-14911 spores. Bacterial smear was stained with malachite green and safranin as described in the Methods section and examined under oil immersion microscope. Arrows indicate round and symmetrical endospores present both within and outside the bacteria. Original magnification: 100x. Figure S2. The genes encoded by the plasmid of Bacillus sp. NRRL B-14911. The plasmid sequence was generated from the sequence data obtained by sequencing of the genomic DNA from Bacillus sp. NRRL B-14911 in long-reads. The genes encoded by the plasmid were annotated as described for the bacterial chromosome (see methods).Figure S3. Alignment of contigs previously reported for Bacillus sp. NRRL B-14911 to the new long-read-based finished assembly. The inner ring (blue) represents contigs assembled from short-reads without any scaffolding (middle panel in Figure 1) as aligned to our de novo assembly based on long-reads. The outer ring (pink and black) represents alignment of contigs after scaffolding (top and bottom panels in Fig. 1). Note, white lines and blocks show large areas without any coverage in the prior assembly. Designations for each scaffold and contig are derived from GenBank accession numbers, which are abbreviated for convenience in display. Scaffolds are shown in pink and have full accessions with the format CH6723XX, where XX is the number shown on the figure following “CH”. Black- and blue-shaded regions represent un-scaffolded contigs, which have full accessions with the format AAOX01000XYY, where X is either 0 (for YY between 23 and 99) or 1 (for YY between 00 and 09). Only contigs with ≥ 99.900 % identity are shown. The scale in the middle of the circle is based on the finished de novo assembly, made using long-read sequencing. Figure S4. Circular genome map of Bacillus sp. NRRL B-14911 showing the location of genes for virulence factors that are unique to this bacterium in relation to other Bacillus. The circular map of Bacillus sp. NRRL B-14911 shows the locations of 18 virulence factor genes that are unique to this bacterium as determined by comparing the genomes of three pathogenic, and three non-pathogenic bacteria as described in the ‘Methods’ section. The two outer blue circles represent the genes for virulence factors present in forward and reverse directions. The innermost circle represents the GC skewness, and the second innermost circle represents the GC content of the Bacillus sp. NRRL B-14911 genome. The coordinates of each gene are listed in Additional file 1: Table S3. The start and end positions in Additional file 1: Table S3 match with the location of genes in Additional file 1: Figure S4. Figure S5. Circular genome map showing the location of virulence factor genes common to both Bacillus sp. NRRL B-14911 and other pathogenic bacilli. The circular map shows the locations of 225 virulence factor genes that are common to both Bacillus sp. NRRL B-14911 and three other pathogenic bacilli. Their genome comparisons were made as described in the ‘Methods’ section. The two outer blue circles represent the genes for virulence factors shown in forward and reverse directions. The innermost circle represents the GC skewness, and the second innermost circle represents the GC content of the Bacillus sp. NRRL B-14911 genome. The coordinates of each gene are listed in Additional file 1: Table S4. The start and end positions in Additional file 1: Table S4 match with the location of genes in Additional file 1: Figure S5. Table S1. Comparison of the issues with sequencing by scaffolds in the bioproject in relation to our method of sequencing by long-reads. Table S2. List of genes for virulence factor encoded in the genome of Bacillus sp. NRRL B-14911. Table S3. List of unique genes for virulence factor encoded in the genome of Bacillus sp. NRRL B-14911. Table S4. List of genes for virulence factor encoded in the genome of Bacillus sp. NRRL B-14911 that are in common with pathogenic Bacilli. Table S5. Comparative analysis of the presence or absence of common methyltransferases between Bacillussp. NRRL B-14911 and other Bacillus species. Table S6. Comparative analysis of the presence or absence of transporters between Bacillus sp. NRRL B-14911 and other Bacillus species. Table S7. Comparison of the nucleotide sequences and amino acid sequences corresponding to the epitope BAC 25-40 from Bacillus sp. NRRL B-14911and B. infantis JCM 13438 T. Table S8. Comparative analysis of the presence or absence of enzymes and biochemical pathways of Bacillus sp. NRRL B-14911 in relation to other Bacillus species. Table S9. Comparison of insertion sequence elements between Bacillus sp. NRRL B-14911 and other Bacillus species. (PDF 2089 kb) Acknowledgements This work was in part supported by the National Institutes of Health (HL114669). The use of product and company names is necessary to accurately report the methods and results; however, the United States Department of Agriculture (USDA) neither guarantees nor warrants the standard of the products, and the use of names by the USDA implies no approval of the product to the exclusion of others that may also be suitable. The USDA is an equal opportunity provider and employer. Declarations Publication charges for this article have been funded in part by the University of Nebraska-Lincoln and the National Institutes of Health. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The complete genome sequence of B. infantis NRRL B-14911 used in this paper can be found at GenBank under the Accn# CP006643. The plasmid sequence described in this manuscript (KF831061) is awaiting confirmation for the GenBank Accn#. Datasets supporting the results of this article are also included in the additional files. Authors’ contributions CM designed and performed the experiments, analyzed the data, and wrote the paper. AM analyzed the data. JDL designed and performed the experiments. TP performed MALDI analysis. BK, performed the experiments. RHB performed the experiments. CMK involved in the whole genome sequencing and analyzed the data. CG analyzed the data. RGB designed the experiments. ENM performed the phylogenetic analyses. TPLS sequenced the whole genome and analyzed the data. JR conceived and designed the study, coordinated the workflow, and wrote the paper. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Go AS Mozaffarian D Roger VL Benjamin EJ Berry JD Blaha MJ Dai S Ford ES Fox CS Franco S Heart disease and stroke statistics--2014 update: a report from the American Heart Association Circulation 2014 129 3 e28 e292 10.1161/01.cir.0000441139.02102.80 24352519 2. Heidenreich PA Trogdon JG Khavjou OA Butler J Dracup K Ezekowitz MD Finkelstein EA Hong Y Johnston SC Khera A Forecasting the future of cardiovascular disease in the United States: a policy statement from the American Heart Association Circulation 2011 123 8 933 944 10.1161/CIR.0b013e31820a55f5 21262990 3. Whitton JL Feuer R Myocarditis, microbes and autoimmunity Autoimmunity 2004 37 5 375 386 10.1080/08916930410001713089 15621561 4. Maron BJ Shirani J Poliac LC Mathenge R Roberts WC Mueller FO Sudden death in young competitive athletes. Clinical, demographic, and pathological profiles JAMA 1996 276 3 199 204 10.1001/jama.1996.03540030033028 8667563 5. Gravanis MB Sternby NH Incidence of myocarditis. A 10-year autopsy study from Malmo, Sweden Arch Pathol Lab Med 1991 115 4 390 392 2012501 6. Fujinami RS von Herrath MG Christen U Whitton JL Molecular mimicry, bystander activation, or viral persistence: infections and autoimmune disease Clin Microbiol Rev 2006 19 1 80 94 10.1128/CMR.19.1.80-94.2006 16418524 7. Massilamany C Huber SA Cunningham MW Reddy J Relevance of molecular mimicry in the mediation of infectious myocarditis J Cardiovasc Transl Res 2014 7 2 165 171 10.1007/s12265-013-9519-3 24263348 8. Fujinami RS Oldstone MB Amino acid homology between the encephalitogenic site of myelin basic protein and virus: mechanism for autoimmunity Science 1985 230 4729 1043 1045 10.1126/science.2414848 2414848 9. Gangaplara A Massilamany C Steffen D Reddy J Mimicry epitope from Ehrlichia canis for interphotoreceptor retinoid-binding protein 201-216 prevents autoimmune uveoretinitis by acting as altered peptide ligand J Neuroimmunol 2013 263 1-2 98 107 10.1016/j.jneuroim.2013.08.006 24029580 10. Guilherme L, Kalil J, Cunningham M. Molecular mimicry in the autoimmune pathogenesis of rheumatic heart disease. Autoimmunity. 2006;39(1):31–9. 11. Massilamany C Asojo OA Gangaplara A Steffen D Reddy J Identification of a second mimicry epitope from Acanthamoeba castellanii that induces CNS autoimmunity by generating cross-reactive T cells for MBP 89-101 in SJL mice Int Immunol 2011 23 12 729 739 10.1093/intimm/dxr084 22058327 12. Massilamany C Gangaplara A Steffen D Reddy J Identification of novel mimicry epitopes for cardiac myosin heavy chain-alpha that induce autoimmune myocarditis in A/J mice Cell Immunol 2011 271 2 438 449 10.1016/j.cellimm.2011.08.013 21939961 13. Massilamany C Steffen D Reddy J An epitope from Acanthamoeba castellanii that cross-react with proteolipid protein 139-151-reactive T cells induces autoimmune encephalomyelitis in SJL mice J Neuroimmunol 2010 219 1-2 17 24 10.1016/j.jneuroim.2009.11.006 20005578 14. Koren S Schatz MC Walenz BP Martin J Howard JT Ganapathy G Wang Z Rasko DA McCombie WR Jarvis ED Hybrid error correction and de novo assembly of single-molecule sequencing reads Nat Biotechnol 2012 30 7 693 700 10.1038/nbt.2280 22750884 15. Katoh K Standley DM MAFFT multiple sequence alignment software version 7: improvements in performance and usability Mol Biol Evol 2013 30 4 772 780 10.1093/molbev/mst010 23329690 16. Guindon S Dufayard JF Lefort V Anisimova M Hordijk W Gascuel O New algorithms and methods to estimate maximum-likelihood phylogenies: assessing the performance of PhyML 3.0. Syst Biol 2010 59 3 307 321 10.1093/sysbio/syq010 20525638 17. Cherif A Borin S Rizzi A Ouzari H Boudabous A Daffonchio D Bacillus anthracis diverges from related clades of the Bacillus cereus group in 16S-23S ribosomal DNA intergenic transcribed spacers containing tRNA genes Appl Environ Microbiol 2003 69 1 33 40 10.1128/AEM.69.1.33-40.2003 12513974 18. Daffonchio D Borin S Frova G Manachini PL Sorlini C PCR fingerprinting of whole genomes: the spacers between the 16S and 23S rRNA genes and of intergenic tRNA gene regions reveal a different intraspecific genomic variability of Bacillus cereus and Bacillus licheniformis [corrected] Int J Syst Bacteriol 1998 48 Pt 1 107 116 10.1099/00207713-48-1-107 9542081 19. Xu D Cote JC Phylogenetic relationships between Bacillus species and related genera inferred from comparison of 3' end 16S rDNA and 5' end 16S-23S ITS nucleotide sequences Int J Syst Evol Microbiol 2003 53 Pt 3 695 704 10.1099/ijs.0.02346-0 12807189 20. Bochner BR New technologies to assess genotype-phenotype relationships Nat Rev Genet 2003 4 4 309 314 10.1038/nrg1046 12671661 21. Bochner BR Gadzinski P Panomitros E Phenotype microarrays for high-throughput phenotypic testing and assay of gene function Genome Res 2001 11 7 1246 1255 10.1101/gr.186501 11435407 22. Mellmann A Bimet F Bizet C Borovskaya AD Drake RR Eigner U Fahr AM He Y Ilina EN Kostrzewa M High interlaboratory reproducibility of matrix-assisted laser desorption ionization-time of flight mass spectrometry-based species identification of nonfermenting bacteria J Clin Microbiol 2009 47 11 3732 3734 10.1128/JCM.00921-09 19776231 23. Mellmann A Cloud J Maier T Keckevoet U Ramminger I Iwen P Dunn J Hall G Wilson D Lasala P Evaluation of matrix-assisted laser desorption ionization-time-of-flight mass spectrometry in comparison to 16S rRNA gene sequencing for species identification of nonfermenting bacteria J Clin Microbiol 2008 46 6 1946 1954 10.1128/JCM.00157-08 18400920 24. Chen L Xiong Z Sun L Yang J Jin Q VFDB 2012 update: toward the genetic diversity and molecular evolution of bacterial virulence factors Nucleic Acids Res 2012 40 Database issue D641 645 10.1093/nar/gkr989 22067448 25. Camacho C Coulouris G Avagyan V Ma N Papadopoulos J Bealer K Madden TL BLAST+: architecture and applications BMC Bioinformatics 2009 10 421 10.1186/1471-2105-10-421 20003500 26. Grant JR Stothard P The CGView Server: a comparative genomics tool for circular genomes Nucleic Acids Res 2008 36 Web Server issue W181 184 10.1093/nar/gkn179 18411202 27. Kanehisa M Goto S KEGG: kyoto encyclopedia of genes and genomes Nucleic Acids Res 2000 28 1 27 30 10.1093/nar/28.1.27 10592173 28. Siefert JL Larios-Sanz M Nakamura LK Slepecky RA Paul JH Moore ER Fox GE Jurtshuk P Jr Phylogeny of marine Bacillus isolates from the Gulf of Mexico Curr Microbiol 2000 41 2 84 88 10.1007/s002840010098 10856371 29. Sharon D Tilgner H Grubert F Snyder M A single-molecule long-read survey of the human transcriptome Nat Biotechnol 2013 31 11 1009 1014 10.1038/nbt.2705 24108091 30. Ummat A Bashir A Resolving complex tandem repeats with long reads Bioinformatics 2014 30 24 3491 3498 10.1093/bioinformatics/btu437 25028725 31. Fuks F DNA methylation and histone modifications: teaming up to silence genes Curr Opin Genet Dev 2005 15 5 490 495 10.1016/j.gde.2005.08.002 16098738 32. Blackwood KS Turenne CY Harmsen D Kabani AM Reassessment of sequence-based targets for identification of bacillus species J Clin Microbiol 2004 42 4 1626 1630 10.1128/JCM.42.4.1626-1630.2004 15071016 33. Goto K Omura T Hara Y Sadaie Y Application of the partial 16S rDNA sequence as an index for rapid identification of species in the genus Bacillus J Gen Appl Microbiol 2000 46 1 1 8 10.2323/jgam.46.1 12483598 34. Ki JS Zhang W Qian PY Discovery of marine Bacillus species by 16S rRNA and rpoB comparisons and their usefulness for species identification J Microbiol Methods 2009 77 1 48 57 10.1016/j.mimet.2009.01.003 19166882 35. Ko KS Oh WS Lee MY Lee JH Lee H Peck KR Lee NY Song JH Bacillus infantis sp. nov. and Bacillus idriensis sp. nov., isolated from a patient with neonatal sepsis Int J Syst Evol Microbiol 2006 56 Pt 11 2541 2544 10.1099/ijs.0.64213-0 17082387 36. Fenselau C Demirev PA Characterization of intact microorganisms by MALDI mass spectrometry Mass Spectrom Rev 2001 20 4 157 171 10.1002/mas.10004 11835304 37. Lay JO Jr MALDI-TOF mass spectrometry of bacteria Mass Spectrom Rev 2001 20 4 172 194 10.1002/mas.10003 11835305 38. Welham KJ Domin MA Scannell DE Cohen E Ashton DS The characterization of micro-organisms by matrix-assisted laser desorption/ionization time-of-flight mass spectrometry Rapid Commun Mass Spectrom 1998 12 4 176 180 10.1002/(SICI)1097-0231(19980227)12:4<176::AID-RCM132>3.0.CO;2-T 9493412 39. Clark TA Murray IA Morgan RD Kislyuk AO Spittle KE Boitano M Fomenkov A Roberts RJ Korlach J Characterization of DNA methyltransferase specificities using single-molecule, real-time DNA sequencing Nucleic Acids Res 2012 40 4 e29 10.1093/nar/gkr1146 22156058 40. Vester B Long KS Antibiotic Resistance in Bacteria Caused by Modified Nucleosides in 23S Ribosomal RNA NCBI Bookshelf 2000 41. Botting CH, Talbot P, Paytubi S, White MF. Extensive lysine methylation in hyperthermophilic crenarchaea: potential implications for protein stability and recombinant enzymes. Archaea. 2010, 2010. doi:10.1155/2010/106341. 42. Higgins CF ABC transporters: physiology, structure and mechanism--an overview Res Microbiol 2001 152 3-4 205 210 10.1016/S0923-2508(01)01193-7 11421269 43. Yang HC Fu HL Lin YF Rosen BP Pathways of arsenic uptake and efflux Curr Top Membr 2012 69 325 358 10.1016/B978-0-12-394390-3.00012-4 23046656 44. Kelly DJ Thomas GH The tripartite ATP-independent periplasmic (TRAP) transporters of bacteria and archaea FEMS Microbiol Rev 2001 25 4 405 424 10.1111/j.1574-6976.2001.tb00584.x 11524131 45. Woodson JD Reynolds AA Escalante-Semerena JC ABC transporter for corrinoids in Halobacterium sp. strain NRC-1 J Bacteriol 2005 187 17 5901 5909 10.1128/JB.187.17.5901-5909.2005 16109931 46. Lu S Zgurskaya HI MacA, a periplasmic membrane fusion protein of the macrolide transporter MacAB-TolC, binds lipopolysaccharide core specifically and with high affinity J Bacteriol 2013 195 21 4865 4872 10.1128/JB.00756-13 23974027 47. Belenky PA Moga TG Brenner C Saccharomyces cerevisiae YOR071C encodes the high affinity nicotinamide riboside transporter Nrt1 J Biol Chem 2008 283 13 8075 8079 10.1074/jbc.C800021200 18258590 48. Sauer E Merdanovic M Mortimer AP Bringmann G Reidl J PnuC and the utilization of the nicotinamide riboside analog 3-aminopyridine in Haemophilus influenzae Antimicrob Agents Chemother 2004 48 12 4532 4541 10.1128/AAC.48.12.4532-4541.2004 15561822 49. Yan H Huang W Yan C Gong X Jiang S Zhao Y Wang J Shi Y Structure and mechanism of a nitrate transporter Cell Rep 2013 3 3 716 723 10.1016/j.celrep.2013.03.007 23523348 50. Gebhard S ABC transporters of antimicrobial peptides in Firmicutes bacteria - phylogeny, function and regulation Mol Microbiol 2012 86 6 1295 1317 10.1111/mmi.12078 23106164 51. Vogl C Grill S Schilling O Stulke J Mack M Stolz J Characterization of riboflavin (vitamin B2) transport proteins from Bacillus subtilis and Corynebacterium glutamicum J Bacteriol 2007 189 20 7367 7375 10.1128/JB.00590-07 17693491 52. Serventi F Ramazzina I Lamberto I Puggioni V Gatti R Percudani R Chemical basis of nitrogen recovery through the ureide pathway: formation and hydrolysis of S-ureidoglycine in plants and bacteria ACS Chem Biol 2010 5 2 203 214 10.1021/cb900248n 20038185 53. Yue J Hu X Sun H Yang Y Huang J Widespread impact of horizontal gene transfer on plant colonization of land Nat Commun 2012 3 1152 10.1038/ncomms2148 23093189 54. Ma WT Lin JH Chen HJ Chen SY Shaw GC Identification and characterization of a novel class of extracellular poly(3-hydroxybutyrate) depolymerase from Bacillus sp. strain NRRL B-14911 Appl Environ Microbiol 2011 77 22 7924 7932 10.1128/AEM.06069-11 21948827 55. Mahillon J Chandler M Insertion sequences Microbiol Mol Biol Rev 1998 62 3 725 774 9729608 56. Wagner A Lewis C Bichsel M A survey of bacterial insertion sequences using IScan Nucleic Acids Res 2007 35 16 5284 5293 10.1093/nar/gkm597 17686783 57. Heffron F Bedinger P Champoux JJ Falkow S Deletions affecting the transposition of an antibiotic resistance gene Proc Natl Acad Sci U S A 1977 74 2 702 706 10.1073/pnas.74.2.702 322140 58. Kretschmer PJ Cohen SN Selected translocation of plasmid genes: frequency and regional specificity of translocation of the Tn3 element J Bacteriol 1977 130 2 888 899 233726

PMC005xxxxxx/PMC5001199.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557137290110.1186/s12864-016-2901-1ResearchWhole blood transcriptional profiling comparison between different milk yield of Chinese Holstein cows using RNA-seq data Bai Xue baixue0716@163.com 1Zheng Zhuqing zzq1207@126.com 12Liu Bin liu207702@163.com 123Ji Xiaoyang jxyang0716@126.com 12Bai Yongsheng Yongsheng.Bai@indstate.edu 4Zhang Wenguang atcgnmbi@aliyun.com 151 College of Animal Science, Inner Mongolia Agricultural University, Hohhot, 010018 China 2 Institute of ATCG Nei Mongol Bio-Information, Hohhot, 010020 China 3 Nei Mongol BioNew Technology Co.Ltd, Hohhot, 010018 China 4 Department of Biology, Indiana State University, Terre Haute, IN 47809 U.S.A. 5 Kunming Institute of Zoology, Chinese Academy of Sciences, Kunming, China 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.512© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background The objective of this research was to investigate the variation of gene expression in the blood transcriptome profile of Chinese Holstein cows associated to the milk yield traits. Results We used RNA-seq to generate the bovine transcriptome from the blood of 23 lactating Chinese Holstein cows with extremely high and low milk yield. A total of 100 differentially expressed genes (DEGs) (p < 0.05, FDR < 0.05) were revealed between the high and low groups. Gene ontology (GO) analysis demonstrated that the 100 DEGs were enriched in specific biological processes with regard to defense response, immune response, inflammatory response, icosanoid metabolic process, and fatty acid metabolic process (p < 0.05). The KEGG pathway analysis with 100 DEGs revealed that the most statistically-significant metabolic pathway was related with Toll-like receptor signaling pathway (p < 0.05). The expression level of four selected DEGs was analyzed by qRT-PCR, and the results indicated that the expression patterns were consistent with the deep sequencing results by RNA-Seq. Furthermore, alternative splicing analysis of 100 DEGs demonstrated that there were different splicing pattern between high and low yielders. The alternative 3’ splicing site was the major splicing pattern detected in high yielders. However, in low yielders the major type was exon skipping. Conclusion This study provides a non-invasive method to identify the DEGs in cattle blood using RNA-seq for milk yield. The revealed 100 DEGs between Holstein cows with extremely high and low milk yield, and immunological pathway are likely involved in milk yield trait. Finally, this study allowed us to explore associations between immune traits and production traits related to milk production. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2901-1) contains supplementary material, which is available to authorized users. Keywords Whole bloodImmune responseMilk yieldDifferentially expressed genesDairy cattleThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Milk yield and milk composition of lactating cows are the most important economic traits in dairy cattle. In the past century, genetic selection on milk yield has improved milk production in cattle. With the development of quantitative trait loci (QTLs), genome-wide association studies (GWAS) and RNA sequencing (RNA-seq) technologies, a large number of candidate genes and SNPs associated with milk performance traits have been identified, such as DGAT1 and GHR genes [1–6]. Previous studies described the associations between DGAT1 K232A polymorphism and milk production traits [7, 8]. Blott et al. identified a significant SNP (BFGL-NGS-118998) inside the GHR gene that has an important role related to milk traits [9]. Mammary gland is an important organ to synthesize and secrete milk. The mammary epithelial cell has a remarkable ability to convert blood circulating nutrients into milk components [10]. Thus, almost all the studies related to milk performance traits are based on mammary gland. For example, Cui et al. collected mammary gland samples from four lactating cows after slaughter. They used RNA-seq to generate the bovine mammary transcriptome with high and low milk fat and protein percentage [11]. Finucane et al. compared bovine mammary expression profiles before and after parturition using microarray [12]. However, there are still limitations in analysis of milk performance traits using bovine mammary tissue, such as sampling difficultly, high cost of sampling and tremendous damage to the lactating cows, which resulted in small sample size. Milk is composed of fat, protein, lactose, minerals, vitamins and water, and all these nutrients derive from blood [10, 13–15]. Some data shows that synthesizing 1L milk requires 400-500L blood flow though the mammary gland. So plenty of blood is essential for synthesizing milk. Previous studies associated with milk traits were only focused on blood physiological and biochemical indexes, and there are very few studies associating gene expression with milk traits in cattle blood. A gene expression profile from blood provides new opportunities to clarify the basic mechanisms of complex traits in cattle milk. Besides, whole blood is a complex mixture of cells and can accurately reflect the physiological condition and health level of cows. Many studies have used blood to diagnose disease in dairy cattle, such as mastitis [16, 17]. Most importantly, blood is easier to sample in comparison with other tissues and involves limited handling of animals. Furthermore, the lactation process requires multiple tissues and organs to complete, such as mammary gland, liver and adipose tissue. Hence, blood has the potential ability to represent milk performance traits more directly and comprehensively. Sandri et al. (2015) analyzed the gene expression profile in the blood related to the gene merit for milk productive traits using microarray [18]. Marcel et al. analyzed porcine peripheral blood mononuclear cells transcription profile of pigs with divergent humoral immune response and lean growth performance [19]. To complete understand the impact of blood transcriptome profiles on milk yield, comprehensive cataloguing of gene expression change within high yielders and low yielders is required. The aim of this study is to compare gene expression profiles in bovine whole blood of high and low milk yield cows, and to investigate potential molecular biomarkers in the blood transcriptome that relate to the productive potential of lactating cows using RNA-seq techniques. Results Analysis of expressed transcripts between high and low yielders From 23 samples, we obtained total 74.6 Gb RNA-Seq data files (26,763,546 to 51,313,614 paired-end reads per sample). Nearly 68 % of the reads were mapped to the bovine genome UMD3.1.66 and approximately 62 % of the reads in every individual were uniquely mapped to the bovine genome. The alignment information for each sample is presented in Table 1. Of these, 16,314 and 16,151 expressed transcripts were revealed in high yielders and low yielders, respectively (Additional file 1: Table S1).Table 1 Summary of the mapping information for each sample Sample name Total reads Total mapped Multiple mapped Uniquely mapped L1 51313614 37633474 (73.34 %) 1891840 (3.69 %) 35741634 (69.65 %) L2 38408586 28132542 (73.25 %) 1626290 (4.24 %) 26506252 69.01 %) L3 38156116 23779230 (62.32 %) 1254082 (3.29 %) 22525148 (59.03 %) L4 28642682 18996768 (66.32 %) 873594 (3.05 %) 18123174 (63.27 %) L5 29417132 18996768 (64.58 %) 2319892 (7.89 %) 16676876 (56.69 %) L6 34296982 24799284 (72.31 %) 1048878 3.06%) 23750406 (69.25 %) L7 28755318 16604970 (57.75 %) 896548 (3.12 %) 15708422 (54.63 %) L8 31712370 21846718 (68.89 %) 980318 (3.09 %) 20866400 (65.80 %) L9 28387782 17806510 (62.73 %) 917208 (3.24 %) 16889302 (59.49 %) L10 32217694 20532020 (63.73 %) 896152 (2.78 %) 19635868 (60.95 %) H1 27038432 19212408 (71.06 %) 944260 (3.50 %) 18268148 (67.56 %) H2 32458530 23913728 (73.67 %) 999786 (3.08 %) 22913942 (70.59 %) H3 31355676 23459842 (74.82 %) 933980 (2.98 %) 22525862 (71.84 %) H4 32236792 22698066 (70.41 %) 949798 (2.95 %) 21748268 (67.46 %) H5 31144146 20791688 (66.76 %) 1004534 (3.23 %) 19787154 63.53 %) H6 32072196 22910946 (71.44 %) 1293062 (4.04 %) 21617884 (67.40 %) H7 38188984 25973682 (68.01 %) 1154850 (3.02 %) 24818832 64.99 %) H8 32100360 23224306 (72.35 %) 945914 (2.95 %) 22278392 (69.40 %) H9 30227962 20288464 (67.12 %) 997090 (3.30 %) 19291374 (63.82 %) H10 29321524 19085218 (65.09 %) 1073550 (3.66 %) 18011668 (61.43 %) H11 32631714 21820588 (66.87 %) 1021428 (3.13 %) 20799160 (63.74 %) H12 26763546 14871754 (55.57 %) 748546 (2.80 %) 14123208 (52.77 %) H13 28755192 16809674 (58.46 %) 870238 (3.03 %) 15939436 (55.43 %) DEGs and splice events between high and low yielders To provide a better understanding of the biological mechanism of milk yield, it is essential to identify the DEGs between high and low milk yield cows. Based on the Cuffdiff analysis, a total of 100 DEGs (p<0.05, FDR< 0.05) were examined between the high and low yielders (Fig. 1). All the DEGs were located in chromosomes randomly, but there were no DEGs identified in chromosome 14, 20, 27, and 28. The expression level of 100 DEGs was from 2 to 1063 FPKM in high group, and 0.4 to 1794 FPKM in low group. In addition, 43 of the 100 DEGs were highly expressed in the high yielders; whereas, the other 57 DEGs showed lower expression in low yielders. The expression level of 100 DEGs is shown in Fig. 2 and the detail information is presented in Additional file 2: Table S2.Fig. 1 Volcano plot displaying DEGs between the whole blood of 10 low yielders and 13 high yielders. The y-axis corresponds to the mean expression value of log10 (q-value), and the x-axis displays the log2 fold change value. The red dots circled in the frame represent the significantly differentially expressed transcripts (p value < 0.05 and FDR < 0.05) between high and low milk yield cows; the blue dots represent the transcripts whose expression levels did not reach statistical significance between the two groups Fig. 2 Expression of the 100 DEGs in bovine blood. The x-axis shows the gene expression level value of log10 (FPKM); y-axis shows the gene names. The left shows the 43 genes expressed higher in high yielders, and the right shows the rest 57 genes expressed higher in low yielders Splice events are thought to contribute to phenotypic complexity during the mammalian evolution [20]. In total, we obtained 44,572 and 36,467 splice events in high and low yielders, respectively, compared to the annotated bovine genome UMD3.1.66. Of these, 214 (in high yielders) and 202 (in low yielders) differentially expressed splice events were identified. Further analysis showed that the 214 and 202 splice events involved 59 DEGs and 57 DEGs, respectively. Major splicing events such as, exon skipping (ES), intron retention (IR), alternative 5’ splicing site (A5SS), alternative 3’ splicing site (A3SS) and mutually exclusive exon (MXE) were detected in our studied bovine blood transcriptomes. The A3SS was the major splicing pattern observed in high group; but in low group, the major type was ES (Fig. 3). This suggests that high milk yield cows are more inclined to take the A3SS pattern. In addition, more splice sites were found in chromosome 26 in both groups.Fig. 3 Statistics of mainly alternative splicing events. The first column shows the patterns of alternative splicing events; the second column shows the intron-exon structure, third and fourth column shows the number of AS events in high and low milk yield cows, respectively Functional classification of DEGs The DAVID tool [21] was used to annotate the function of the 100 DEGs with the particular categories focusing on the gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway. A total of 55 clusters (p<0.05) were significantly annotated with GO terms within three major function groups: biological process (BP), cellular component (CC), and molecular function (MF). The most significant GO categories observed were defense response, immune response, inflammatory response, icosanoid metabolic process, and fatty acid metabolic process (p<0.05) (Fig. 4). Only one KEGG pathway was enriched, which was the Toll-like receptor signaling pathway. On the other hand, based on the enrichment analysis of DEGs containing alternative splice sites revealed that the PPAR signaling pathway was only detected in the high yielded group. Detailed information of the DEG functional annotations are showed in Table 2.Fig. 4 Diagram showed DEGs from the top 10 GO functional annotations in blood samples. Genes circled in red were also enriched in KEGG pathway analysis Table 2 Gene Ontology and KEGG pathway annotation of DEGs between two groups Category GO ID GO term P value No.of DEGs DEGs BP GO:0006952 defense response 0.000001 15 KYNU, CARD9, OLR1, RSAD2, CCL5, TLR7, TNFRSF4, CXCL10, MIF, CCR3, HRH4, PLA2G7, THBS1, MX2, RAB27A BP GO:0009611 response to wounding 0.000008 13 NRP1, OLR1, CCL5, TLR7, TPM1, TNFRSF4, CXCL10, MIF, CCR3, HRH4, PLA2G7, THBS1, RAB27A BP GO:0006955 immune response 0.000024 14 KYNU, CARD9, OLR1, IFITM3, RSAD2, CCL5, TLR7, TNFRSF4, CXCL10, MIF, IGHE, TREM2, THBS1, RAB27A BP GO:0006954 inflammatory response 0.000027 10 OLR1, CCR3, HRH4, PLA2G7, CCL5, THBS1, TNFRSF4, TLR7, MIF, CXCL10 BP GO:0006690 icosanoid metabolic process 0.001543 4 GGT5, CYP2J2, ALOX12, MIF BP GO:0033559 unsaturated fatty acid metabolic process 0.001954 4 GGT5, CYP2J2, ALOX12, MIF BP GO:0009615 response to virus 0.001958 5 CARD9, RSAD2, CCL5, MX2, TLR7 BP GO:0006631 fatty acid metabolic process 0.002770 6 LPL, GGT5, CYP2J2, CPT1A, ALOX12, MIF BP GO:0032940 secretion by cell 0.003353 6 STX1A, LMF1, USE1, CCL5, RAB27A, CXCL10 BP GO:0051607 defense response to virus 0.004209 3 CARD9, RSAD2, TLR7 BP GO:0006633 fatty acid biosynthetic process 0.006735 4 LPL, GGT5, ALOX12, MIF BP GO:0046394 carboxylic acid biosynthetic process 0.006891 5 LPL, GGT5, KYNU, ALOX12, MIF BP GO:0016053 organic acid biosynthetic process 0.006891 5 LPL, GGT5, KYNU, ALOX12, MIF BP GO:0046456 icosanoid biosynthetic process 0.009951 3 GGT5, ALOX12, MIF BP GO:0006636 unsaturated fatty acid biosynthetic process 0.011894 3 GGT5, ALOX12, MIF BP GO:0008015 blood circulation 0.012860 5 TCAP, OLR1, MYOF, TPM1, CXCL10 BP GO:0003013 circulatory system process 0.012860 5 TCAP, OLR1, MYOF, TPM1, CXCL10 BP GO:0051270 regulation of cell motion 0.014554 5 COL18A1, NRP1, THBS1, TPM1, CXCL10 BP GO:0046903 secretion 0.015412 6 STX1A, LMF1, USE1, CCL5, RAB27A, CXCL10 BP GO:0016192 vesicle-mediated transport 0.021730 8 STX1A, RABEP2, HIP1R, USE1, TRAPPC5, CCL5, THBS1, RAB27A BP GO:0002252 immune effector process 0.027651 4 CARD9, RSAD2, TLR7, RAB27A BP GO:0002684 positive regulation of immune system process 0.028807 5 CARD9, IL13RA1, THBS1, TNFRSF4, TLR7 BP GO:0002544 chronic inflammatory response 0.028923 2 CCL5, THBS1 BP GO:0045087 innate immune response 0.029812 4 KYNU, TLR7, RAB27A, MIF BP GO:0051240 positive regulation of multicellular organismal process 0.031168 5 CARD9, CCL5, THBS1, TPM1, TLR7 BP GO:0045785 positive regulation of cell adhesion 0.034610 3 THBS1, TPM1, ALOX12 BP GO:0006968 cellular defense response 0.035673 3 CCR3, CCL5, TNFRSF4 BP GO:0033273 response to vitamin 0.041171 3 ALPL, KYNU, PMF1 BP GO:0002694 regulation of leukocyte activation 0.047354 4 IL13RA1, THBS1, TNFRSF4, MIF BP GO:0031349 positive regulation of defense response 0.049365 3 CARD9, CCL5, TLR7 BP GO:0030334 regulation of cell migration 0.049479 4 COL18A1, THBS1, TPM1, CXCL10 BP GO:0000267 cell fraction 0.002965 13 STX1A, KYNU, CYB5R2, CYP2J2, OLR1, HIP1R, SLC6A4, CCL5, CPT1A, PYGM, HRH4, ABCC4, CA4 CC GO:0005886 plasma membrane 0.009603 27 ALPL, STEAP4, NRP1, IFITM3, SLC6A4, SLC16A12, RSAD2, TNFRSF4, TPM1, HRH4, TGM3, THBS1, IL13RA1, MYOF, RAB27A, LPL, STX1A, OLR1, ZP3, FADD, GGT5, TARP, CCR3, CA4, ABCC4, TREM2, ALOX12 CC GO:0044433 cytoplasmic vesicle part 0.010180 5 STX1A, HIP1R, ABCC4, THBS1, RAB27A CC GO:0030659 cytoplasmic vesicle membrane 0.025142 4 STX1A, HIP1R, ABCC4, RAB27A CC GO:0044421 extracellular region part 0.027958 10 COL18A1, LPL, OGN, ZP3, PLA2G7, PMF1, CCL5, THBS1, MIF, CXCL10 CC GO:0005624 membrane fraction 0.028767 9 STX1A, CYP2J2, OLR1, HIP1R, HRH4, SLC6A4, ABCC4, CA4, CPT1A CC GO:0012506 vesicle membrane 0.031088 4 STX1A, HIP1R, ABCC4, RAB27A CC GO:0005615 extracellular space 0.034524 8 COL18A1, LPL, PLA2G7, PMF1, CCL5, THBS1, MIF, CXCL10 CC GO:0005626 insoluble fraction 0.034688 9 STX1A, CYP2J2, OLR1, HIP1R, HRH4, SLC6A4, ABCC4, CA4, CPT1A CC GO:0030141 secretory granule 0.048228 4 STX1A, ABCC4, THBS1, RAB27A MF GO:0019955 cytokine binding 0.013930 4 CCR3, IL13RA1, THBS1, TNFRSF4 MF GO:0008144 drug binding 0.028489 3 PYGM, SLC6A4, TLR7 KEGG PATHWAY hsa04620 Toll-like receptor signaling pathway 0.029157 4 FADD, CCL5, TLR7, CXCL10 qRT-PCR validation of DEGs in high and low milk yielders Furthermore, we randomly selected 4 DEGs identified from the RNA-seq data, LGALSL, IL-8, FAM213B and CCL5, to validate their expression patterns using qRT-PCR. The results from the qRT-PCR confirmed that the DEGs had the same expression pattern observed with the RNA-seq (Fig. 5). This indicates that the gene expression observed in blood transcriptome between high yielders and low yielders was highly credible.Fig. 5 The expression level of DEGs validation by qRT-PCR. LGALSL and IL-8 were highly expressed in high yielders. FAM213B and CCL5 were highly expressed in low yielders detected by RNA-seq Discussion The object of this study is to investigate potential molecular biomarkers in whole blood related to milk production traits in lactating dairy cows, with the aim of putting forward a non-invasive method that identifies the DEGs for milk performance traits. In this study, whole blood genome transcriptome profiles of high and low yield milk cows were investigated using RNA-seq technique. RNA-seq has many advantages over traditional cDNA microarray technologies and it can easily detect low-abundance genes [22, 23]. Marioni et al. demonstrated that RNA-seq and qRT-PCR have a high correlation, and that the Pearson correlation could reach 0.929 [23], which means RNA-Seq is accurate and reproducible. Among these 100 DEGs, many genes were also detected by other studies associated with milk yield, such as BOLA-DQA1. BOLA-DQA1 is one of the primary histocompatibility complex (MHC) class II molecules that plays an important role in the immune system. It is thought that MHC genes indirectly affect milk production traits by increasing the disease resistance of the cows. In this present study, the expression level of BOLA-DQA1 was higher in high milk yielders than in low milk yielders. Other research has reported that the BOLA-DQA1 gene was associated with resistance to mastitis progression [24]. Splicing event analysis showed that DEGs in high and low milk yield cows were different. In low yielders, cows are more inclined to take the ES pattern but in high yielders A3’SS was more likely. Oxidized low-density lipoprotein receptor (OLR1), which can degrade the oxidized form of low-density lipoprotein and plays important roles in fatty acid transport, was reported as a potential gene for milk-fat percentage and milk fat yield [25]. The splice site analysis in our study also found that AS in OLR1 was significant. There were 5 splice sites of OLR1 in high milk yield including one ES, one A5’SS and three other events. But in low yielders, there was only one splice site on OLR1, which was A5’SS. Also, the expression level of OLR1 was higher in high yielders than in low yielders. Splice events analysis of DEGs can reveal differences between high and low milk yield cows. Among those DEGs containing splice sites, gene ontology enrichment indicated that the PPAR signaling pathway was significantly different between high and low groups. Study showed that PPARA is expressed in heart, liver, adipose tissue, and muscle tissue and is involved in fatty acid catabolism [26]; PPARs play important roles in the regulation of metabolic and inflammatory signaling pathways [26, 27]. It is also reported that PPAR-γ (PPARG) is over-expressed in adipose tissue and macrophages and primarily regulates adipogenesis [27, 28]. PPAR-γ has been reported to significantly increase its expression during lactation in bovine mammary gland [28]. Also, anti-inflammatory properties were observed between PPARA and PPARG [29]. In our study, we found that this PPAR pathway is also in the blood transcriptome, and it was only detected in high yielders. Moreover, these three PPAR pathway DEGs: LPL, OLR1, and CPT1A, were highly expressed in high yielders. Also, the Toll-like receptor signaling pathway existed only in high yielders. The Toll-like receptor pathway is also involved in innate immunity [29]. These results indicate that the regulation of metabolic and immune function is more active in high milk yield cows. The immune system plays a key role in health maintenance, pathogenesis, diseases resistance, and production performance. Further research is required to explore the relation between immune response and milk performance traits in cattle blood transcriptomes. Conclusion The present study provides a non-invasive method to identify the DEGs in cattle blood using RNA-seq for milk yield traits. The study revealed 100 DEGs between high yielders and low yielders, and discovered different alternative splicing patterns between the two groups. The enrichment analysis also revealed that specific metabolic and immunological pathways are related to cattle milk yield traits, and could be considered a signature of blood biomarkers selection in dairy cattle. These results provide the valuable resources of biological research in Chinese Holstein cows milk production, but also offer some potential guidelines to understand the relationship between milk production and the immune function. Methods Blood samples collection Twenty three Chinese Holstein cows in their second or third lactation were selected based on their current lactation and previous lactation records from BingZhouHai Dairy Farm. All the cows are reared under the same standard. 13 high yielders: ~28kg/day/cow and 10 low yielders: ~18kg/day/cow were selected for this study. 5ml of whole blood was drawn from the jugular vein of each cow. The blood samples were frozen in liquid nitrogen then transferred to −80 °C for further RNA extraction. All experimental methods used in this study were approved by the Inner Mongolia Agricultural University (Hohhot, China) Institutional Animal Care and Use Committee. RNA extraction and sequencing We used TRIzol reagent (TaKaRa, USA) to extract total RNA from blood samples following the manufacturer’s instructions. RNA degradation and purity were monitored on 1% agarose gels. Illumina TruSeq Stranded mRNA Sample Preparation Kit (Illumina, San Diego, CA, USA) was used to generate cDNA libraries according to the manufacturer’s recommendations. A total of 1μg of high quality RNA from each individual was used to prepare the sequencing library. The poly-T oligo-attached magnetic beads selection procedure was used to obtained mRNAs from the total RNA. After fragmentation, random oligonucleotides and SuperScript II was used to synthesize first-strand cDNA. DNA polymerase I and RNase H was used to synthesis Second-strand cDNA. After adenylating of 3’ ends of DNA fragments, Hybridization was initiated by ligating Illumina PE adapter and index. cDNA fragments (200 bp) were generated by the AMPure XP system in purifying the library. DNA fragments were selectively enriched to construct the final sequencing library using Illumina PCR Primer Cocktail [11]. An Illumina Hiseq 2000 platform sued to sequence the libraries. Differentially expressed genes and splice events analysis between high and low yielders Clean reads were acquired by removing low quality reads with adapter and poly-N from raw reads and were used for subsequent analysis. We downloaded the cattle genome (UMD3.1.66) and annotation files from the ensembl database. Bowtie v0.12.8 was used to build the index of cattle reference genome [22] and TopHat v2.1.0 [23] was employed to align clean reads for each sample against the cattle reference genome. DEGs between the high and low yielders were detected using Cuffdiff [30]. P value <0.05 and FDR<0.05 were as the threshold in this study to select differentially expressed genes between high and low yielders. The Alternative Splicing Transcriptional Landscape Visualization Tool (Astalavista) web server (http://genome.crg.es/astalavista/) extracts and displays splice events from genomic annotation of exon-intron gene coordinates. Astalavista v3.0 [20] was employed to identify the alternative splice events for all available transcripts and to study the five basic splice events of 100 DEGs for 23 samples. Enrichment analysis of DEGs To further investigate the function of 100 DEGs, we performed the gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis using DAVID bioinformatics resource tool [21]. GO and KEGG terms with p<0.05 were considered significantly enriched. Validation of DEGs by qRT-PCR DEGs identified by the RNA-seq method were validated using qRT-PCR. GAPDH was used as the internal quality control. RT-PCR experiments were performed with 2×SYBR Green master mix technology (Takara) on the Mx3000P Real-Time PCR System (Agilent, USA). The reaction was performed using the following program: 95 °C for 10 min; 40 cycles of 95 °C for 15s and 61 °C for 60s; 95 °C for 30s, 55 °C for 30s and 95 °C for 30s. Primer sequences can be found in Additional file 3: Table S3. Abbreviations A3SS, alternative 3’ splicing site; A5SS, alternative 5’ splicing site; AS, Alternative splicing; BP, biological process; CC, cellular component; DEG, differentially expressed gene; ES, exon skipping; EST, expressed sequence tag; FDR, false discovery rate corrected p values; FPKM, Fragments per kilobase of transcript per million mapped fragments; GO, Gene ontology; IR, intron retention; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; MXE, mutually exclusive exon; RNA-Seq, RNA-Sequencing Additional files Additional file 1: Table S1. The total transcripts analyzed in investigated samples. (XLSX 2461 kb) Additional file 2: Table S2. The detail information of 100 DEGs between high and low milk yield cows. (XLSX 19 kb) Additional file 3: Table S3. Primer sequences used for qRT-PCR. (DOCX 14 kb) This work was supported by National Natural Science Foundation of China (No.30960246), and “National Science and Technology Support Plan” of China (No.2011BAD28B05). Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The sequence data of 23 Chinese Holstein cows used in this paper can be used by public. Datasets supporting the results of this article are included in the additional files. Authors’ Contributions WZ and YB conceived and designed the experiments. XB and XJ performed the experiments. XB, ZZ and BL analyzed the data. XB, YB and WZ wrote the paper. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate All of the animals were handled in strict, accordance with good animal practices as defined by the relevant national and local animal welfare bodies. The experimental procedure was approved by the Animal Care and Use Committee of Inner Mongolia Agricultural University, China and was performed in accordance with the animal welfare and ethics guidelines. ==== Refs References 1. Sun D Jia J Ma Y Zhang Y Wang Y Yu Y Zhang Y Effects of DGAT1 and GHR on milk yield and milk composition in the Chinese dairy population Anim Genet 2009 40 6 997 1000 10.1111/j.1365-2052.2009.01945.x 19781040 2. Grisart B Coppieters W Farnir F Karim L Ford C Berzi P Cambisano N Mni M Reid S Simon P Positional candidate cloning of a QTL in dairy cattle: identification of a missense mutation in the bovine DGAT1 gene with major effect on milk yield and composition Genome Res 2002 12 2 222 231 10.1101/gr.224202 11827942 3. Boichard D Grohs C Bourgeois F Cerqueira F Faugeras R Neau A Rupp R Amigues Y Boscher YM Levéziel H Detection of genes influencing economic traits in three French dairy cattle breeds Genet Sel Evol 2003 35 1 77 101 10.1186/1297-9686-35-1-77 12605852 4. Daetwyler HD Schenkel FS Sargolzaei M Robinson JAB A genome scan to detect quantitative trait loci for economically important traits in Holstein cattle using two methods and a dense single nucleotide polymorphism map J Chengdu Sport Univ 2008 91 8 3225 3236 5. Kolbehdari D Wang Z Grant JR Murdoch B Prasad A Xiu Z Marques E Stothard P Moore SS A whole genome scan to map QTL for milk production traits and somatic cell score in Canadian Holstein bulls J Anim Breed Genet 2009 126 3 216 227 10.1111/j.1439-0388.2008.00793.x 19646150 6. Grisart B Farnir F Karim L Cambisano N Kim J-J Kvasz A Mni M Simon P Frère J-M Coppieters W Genetic and functional confirmation of the causality of the DGAT1 K232A quantitative trait nucleotide in affecting milk yield and composition Proc Natl Acad Sci U S A 2004 101 8 2398 2403 10.1073/pnas.0308518100 14983021 7. Kaupe B Brandt H E-M P Erhardt G Joint analysis of the influence of CYP11B1 and DGAT1 genetic variation on milk production, somatic cell score, conformation, reproduction, and productive lifespan in German Holstein cattle J Anim Sci 2007 85 1 11 21 10.2527/jas.2005-753 17179535 8. Bovenhuis H, Visker M, Poulsen N, Sehested J, van Valenberg H, van Arendonk J, Larsen L, Buitenhuis A. Effects of the diacylglycerol o-acyltransferase 1 (DGAT1) K232A polymorphism on fatty acid, protein, and mineral composition of dairy cattle milk. J Dairy Sci. 2016. 9. Sarah B Jong-Joo K Sirja M Anne SK Anne C Paulette B Nadine C Christine F Bernard G Dave J Molecular dissection of a quantitative trait locus: a phenylalanine-to-tyrosine substitution in the transmembrane domain of the bovine growth hormone receptor is associated with a major effect on milk yield and composition Genetics 2003 163 1 253 266 12586713 10. Bauman DE Mather IH Wall RJ Lock AL Major advances associated with the biosynthesis of milk J Dairy Sci 2006 89 4 1235 1243 10.3168/jds.S0022-0302(06)72192-0 16537956 11. Cui X Hou Y Yang S Xie Y Zhang S Zhang Y Zhang Q Lu X Liu GE Sun D Transcriptional profiling of mammary gland in Holstein cows with extremely different milk protein and fat percentage using RNA sequencing BMC Genomics 2014 15 1 1 15 10.1186/1471-2164-15-226 24382143 12. Finucane KA Mcfadden TB Bond JP Kennelly JJ Zhao FQ Onset of lactation in the bovine mammary gland: gene expression profiling indicates a strong inhibition of gene expression in cell proliferation Funct Integr Genomics 2008 8 3 251 264 10.1007/s10142-008-0074-y 18259788 13. Akers RM: Lactation and the mammary gland: Iowa State Press; 2002 14. Mcmanaman JL Neville MC Mammary physiology and milk secretion Adv Drug Deliv Rev 2003 55 5 629 641 10.1016/S0169-409X(03)00033-4 12706546 15. Baldi A Cheli F Pinotti L Pecorini C Nutrition in mammary gland health and lactation: advances over Eight Biology of Lactation in Farm Animals meetings J Anim Sci 2008 86 13_suppl 3 9 10.2527/jas.2007-0286 17644785 16. Zeng R Bequette BJ Vinyard BT Bannerman DD Determination of milk and blood concentrations of lipopolysaccharide-binding protein in cows with naturally acquired subclinical and clinical mastitis J Dairy Sci 2009 92 3 980 989 10.3168/jds.2008-1636 19233791 17. Jhambh R Dimri U Gupta VK Rathore R Blood antioxidant profile and lipid peroxides in dairy cows with clinical mastitis Veterinary World 2013 6 6 271 273 10.5455/vetworld.2013.271-273 18. Sandri M Stefanon B Loor JJ Transcriptome profiles of whole blood in Italian Holstein and Italian Simmental lactating cows diverging for genetic merit for milk protein J Dairy Sci 2015 98 9 6119 6127 10.3168/jds.2014-9049 26142863 19. Marcel A Eduard M Siriluck P Klaus W PBMC transcription profiles of pigs with divergent humoral immune responses and lean growth performance Int J Biol Sci 2013 9 9 907 916 10.7150/ijbs.6769 24155665 20. Kang X Liu G Liu Y Xu Q Zhang M Fang M Transcriptome profile at different physiological stages reveals potential mode for curly fleece in Chinese tan sheep Plos One 2013 8 8 e71763 10.1371/journal.pone.0071763 23990983 21. Dennis G Sherman BT Hosack DA Yang J Wei G Lane HC Lempicki RA DAVID: Database for Annotation, visualization, and Integrated Discovery Genome Biol 2003 2003 3 P3 10.1186/gb-2003-4-5-p3 22. Croucher NJ Thomson NR Studying bacterial transcriptomes using RNA-seq Curr Opin Microbiol 2010 13 5 619 624 10.1016/j.mib.2010.09.009 20888288 23. Marioni JC Mason CE Mane SM Stephens M Gilad Y RNA-seq: an assessment of technical reproducibility and comparison with gene expression arrays Genome Res 2008 18 9 1509 1517 10.1101/gr.079558.108 18550803 24. Miyasaka T Takeshima SN Matsumoto Y Kobayashi N Matsuhashi T Miyazaki Y Tanabe Y Ishibashi K Sentsui H Aida Y The diversity of bovine MHC class II DRB3 and DQA1 alleles in different herds of Japanese Black and Holstein cattle in Japan Gene 2010 472 1–2 42 49 20965236 25. Schennink A Bovenhuis H Léon-Kloosterziel KM Arendonk JAMV Visker MHPW Effect of polymorphisms in the FASN, OLR1, PPARGC1A, PRL and STAT5A genes on bovine milk-fat composition Anim Genet 2009 40 40 909 916 10.1111/j.1365-2052.2009.01940.x 19719788 26. Chinetti G Fruchart JC Staels B Peroxisome proliferator-activated receptors (PPARs): nuclear receptors at the crossroads between lipid metabolism and inflammation Inflamm Res 2000 49 10 497 505 10.1007/s000110050622 11089900 27. Cynthia H Peter T Coordination of inflammation and metabolism by PPAR and LXR nuclear receptors Curr Opin Genet Dev 2008 18 5 461 467 10.1016/j.gde.2008.07.016 18782619 28. Bionaz M Loor JJ Gene networks driving bovine milk fat synthesis during the lactation cycle BMC Genomics 2008 9 1 366 10.1186/1471-2164-9-366 18671863 29. Szanto A Nagy L The many faces of PPARgamma: anti-inflammatory by any means? Immunobiology 2008 213 9–10 789 803 10.1016/j.imbio.2008.07.015 18926294 30. Trapnell C Williams BA Pertea G Mortazavi A Kwan G van Baren MJ Salzberg SL Wold BJ Pachter L Transcript assembly and quantification by RNA-Seq reveals unannotated transcripts and isoform switching during cell differentiation Nat Biotechnol 2010 28 28 511 515 10.1038/nbt.1621 20436464

PMC005xxxxxx/PMC5001200.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557330291010.1186/s12864-016-2910-0ResearchA genomics-based systems approach towards drug repositioning for rheumatoid arthritis Xu Rong rxx@case.edu 1Wang QuanQiu qwang@thintek.com 21 Department of Epidemiology and Biostatistics, Institute of Computational Biology, School of Medicine, Case Western Reserve University, 2103 Cornell Road, Cleveland, 44106 OH USA 2 ThinTek LLC, Palo Alto, 94306 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.518© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Rheumatoid arthritis (RA) is a chronic autoimmune disease characterized by inflammation and destruction of synovial joints. RA affects up to 1 % of the population worldwide. Currently, there are no drugs that can cure RA or achieve sustained remission. The unknown cause of the disease represents a significant challenge in the drug development. In this study, we address this challenge by proposing an alternative drug discovery approach that integrates and reasons over genetic interrelationships between RA and other genetic diseases as well as a large amount of higher-level drug treatment data. We first constructed a genetic disease network using disease genetics data from Genome-Wide Association Studies (GWAS). We developed a network-based ranking algorithm to prioritize diseases genetically-related to RA (RA-related diseases). We then developed a drug prioritization algorithm to reposition drugs from RA-related diseases to treat RA. Results Our algorithm found 74 of the 80 FDA-approved RA drugs and ranked them highly (recall: 0.925, median ranking: 8.93 %), demonstrating the validity of our strategy. When compared to a study that used GWAS data to directly connect RA-associated genes to drug targets (“direct genetics-based” approach), our algorithm (“indirect genetics-based”) achieved a comparable overall performance, but complementary precision and recall in retrospective validation (precision: 0.22, recall: 0.36; F1: 0.27 vs. precision: 0.74, recall: 0.16; F1: 0.28). Our approach performed significantly better in novel predictions when evaluated using 165 not-yet-FDA-approved RA drugs (precision: 0.46, recall: 0.50; F1: 0.47 vs. precision: 0.40, recall: 0.006; F1: 0.01). Conclusions In summary, although the fundamental pathophysiological mechanisms remain uncharacterized, our proposed computation-based drug discovery approach to analyzing genetic and treatment interrelationships among thousands of diseases and drugs can facilitate the discovery of innovative drugs for treating RA. Keywords Systems biologyNetwork medicineDrug discoveryRheumatoid arthritisDisease genomicsThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background RA is a chronic inheritable autoimmune disease characterized by inflammation and destruction of synovial joints. RA affects up to 1 % of the population worldwide [1]. The cause of RA remains unknown, with multiple genetic and environmental factors involved [2]. Currently, there is no drug that can cure RA or achieve sustained remission. Disease genetics can lead to the identification of novel drug treatments. Over the past decade, genome-wide association studies (GWAS) have robustly identified genetic risk loci linked to many complex diseases, including 100 risk loci for RA [3], and have provided valuable biological insights into many common diseases [4]. Several recent studies have indicated that disease genetics identified by GWAS may lead to translational opportunities for drug discovery [5–11]. For example, a recent study showed that RA risk loci identified through meta-analysis of GWAS data provided therapeutic opportunities for the repositioning of existing drugs for the treatment of RA [7]. To capitalize on complex human disease genetics identified through GWAS, the National Center for Advancing Translational Science (NCATS) was established to use genomic information to determine whether drugs approved to treat one disease could be effective in treating others [8]. However, significant challenges exist in directly translating disease-associated genetic variants identified by GWAS into novel therapeutics [4]. Here, we propose an alternative drug discovery approach by combining lower-level disease genetics identified by GWAS with higher-level drug treatment databases that we recently constructed [12–14]. We hypothesize that the genetic overlap among diseases often reflects pathophysiological overlap. Though the majority of such shared pathophysiological features may remain unknown (i.e. RA biology), treatment insights from one disease may be used to inform our knowledge of others and potentiate their treatments. Instead of directly inferring drug targets from disease genetics as previous approaches [5–7, 15], we use diseases that share high degrees of genetic commonalities with RA (RA-related diseases) as the starting point for discovering new drug treatments for RA (Fig. 1). For example, if a drug treats many RA-related diseases, then this drug is more likely to be a promising candidate to treat RA than a drug that treats few RA-related diseases. Compared to the traditional disease genetics-based or target-based drug discovery, we can put such inference into practice without precise knowledge of disease biology or drug mechanisms. Fig. 1 Indirect disease genetics-based drug repositioning approach Computational drug repositioning approaches can be classified as either drug-based or disease-based [16, 17]. Drug-based approaches leverage upon known drug molecular structures or functions such as chemical structure and properties, molecular docking, gene expression, drug treatments, and drug side effects [18–22]. It was recognized that drug screening based on existing drugs (“drug-based”) might fail to identify new therapeutic mechanisms [23]. On the other hand, disease-based approaches put less emphasis on existing drugs and focus more on disease mechanisms or interrelationships among diseases, therefore have potential in discovering truly innovative drugs. Disease-based approaches have used disease-related data ranging from genome [19, 20] to phenome [24–27]. While existing drug repositioning systems often used well-established computational or statistical methods, including regression/classification, machine learning, network analysis, and text mining [17], they differ in the datasets included in the systems and how heterogeneous data are integrated. The keys to our drug repositioning system include both the unique datasets included in the system as well as innovative approaches to integrating various disease- and drug-related data. One of the key components of our system is four large-scale drug-disease treatment relationship knowledge bases (TreatKBs) that we recently constructed from multiple heterogeneous and complementary data resources using advanced computational techniques including natural language processing, text mining and data mining [12–14]. The four TreatKB include 9,216 drug-disease treatment pairs extracted from FDA drug labels, 111,862 pairs extracted from the FDA Adverse Event Reporting System (FAERS), a database supporting the FDA’s post-marketing drug safety surveillance, 34,306 pairs extracted from 22 million published biomedical literature abstracts, and 69,724 pairs extracted from 171,805 clinical trials. All together, TreatKB contains 208,330 drug-disease treatment pairs for 2484 drugs and 24,511 diseases. In addition, we applied a novel signal prioritization algorithm that we recently developed [25], which first identifies diseases that are genetically related to RA and then prioritizes drugs based on the relevance of their associated diseases to RA. Methods The experiment framework is depicted in Fig. 2 and consists of four steps: (1) we constructed a genetic disease network (GDN) using disease-gene associations from GWAS. We developed a network-based ranking algorithm to find diseases that shared high degrees of genetic commonality with RA; (2) we analyzed disease classes that were highly associated with RA in order to evaluate the disease-ranking algorithm and to gain insight into RA-related diseases; (3) we developed a drug prioritization algorithm to systematically reposition drugs from RA-related diseases to treat RA. We retrospectively validated the algorithm using 80 FDA-approved RA drugs. We evaluated our algorithm in novel predictions using 165 not-yet-approved RA drugs. We compared our approach (“indirect disease genetics-based”) to Okada’s study [7] (“direct disease genetics-based”) in both retrospective validation and novel predictions; and (4) we examined drug classes that were highly enriched among top-ranked drug candidates. Fig. 2 The overall experiment flow chart Find and analyze RA-related diseases Construct genetic disease network (GDN) We constructed a GDN using disease-gene associations from the Catalog of Published Genome-Wide Association Studies from the US National Human Genome Research Institute (NHGRI), which is an exhaustive source containing the descriptions of disease- and trait-associated single nucleotide polymorphisms (SNPs) from published GWAS data [28]. We obtained a total of 22,470 disease/trait-gene pairs, representing 881 diseases/traits and 8,689 genes. On GDN, two diseases were connected if their associated genes overlapped. The edge weights were determined by the cosine similarity coefficients of disease-associated genes [29]. We also experimented other similarity measures such as Jaccard similarity coefficient and overlap. Since some diseases do not share genes directly but their associated genes may interact or participate in the same pathways, we also investigated alternative approaches to connect diseases on the networks. We connected two diseases if their associated genes (proteins) interact or participate in the same pathways using data from the STRING database [30]. Experimental results showed that connecting diseases based on cosine similarity of their associated genes performed best. GDN comprised of 882 disease nodes and 200,758 edges. Apply network-based ranking algorithm to find RA-related diseases from GDN We have recently developed network-based ranking algorithms to prioritize genes for a given disease [26, 31] or to prioritize diseases for a given microbial metabolite [32]. The iterative network-based ranking algorithm is defined as: pt+1=(1−γ)Mpt+γp0, wherein M is the column-normalized adjacency matrix of GDN, γ is a preset probability of restarting from the initial seed node (γ=0.1 in this study), and pt is a vector in which the ith element holds the normalized ranking score of disease i at tth iteration. The initial probability vector p0 contains only RA, with a probability of 1.0. The iterative ranking algorithm was run until it converges, meaning that the change between pt+1 and pt is less than 10−6. Diseases are ranked according to values in the steady-state probability vector pt. Analyze RA-related diseases To better understand top-ranked diseases as well as to test the network construction and ranking algorithms, we examined the distributions of disease classes among RA-related diseases at different ranking cutoffs. We classified diseases based on the 10th revision of the International Statistical Classification of Diseases and Related Health Problems (ICD10) [33]. The ICD10 includes 22 highest-level disease classes (or chapters) such as “Neoplasms” and “Diseases of the nervous system.” During our experiments, we found that only six disease classes were well represented in the GWAS catalog: immune diseases, autoimmune diseases, cardiovascular diseases,metabolic disorders, mental or psychiatric disorders, and neoplasms. The two disease classes: immune diseases and autoimmune diseases, served as positive controls because RA is an autoimmune disease (and immune disease) and is expected to be related to other immune diseases. We retrieved a ranked list of diseases from GDN using RA. We calculated the percentage of these six disease classes among ranked diseases at 10 different ranking cutoffs (top 10 %, 20 %,... 100 %). Reposition drugs based on RA-related diseases Drug repositioning algorithm We have recently developed a drug prioritization approach to systematically reposition drugs that treat RA-related diseases to treat RA [25]. The algorithm is based on the assumption that if a drug treats many top-ranked RA-related diseases, it will rank higher than another drug that treats one or two lower-ranked diseases. We first ranked drugs based on the number of RA-related diseases that they could treat as well as the ranking scores of these diseases. The drug prioritization algorithm is defined as: Rdrug=∑i=1nR_disease_i, wherein n is the number of RA-related diseases that a drug can treat and R_disease_i is the disease ranking score (output from the network-based disease ranking algorithm). De novo validation using known RA drugs and comparison We evaluated our algorithm using a total of 80 FDA-approved RA drugs. This list of drugs were manually compiled from FDA drug label and several authoritative medical-related websites such as webMD.com and drugs.com. Our evaluation was a de-novo validation scheme, since RA and its associated drug treatment pairs were removed from the inputs to the repositioning algorithm. The standard precision, recall, and F1 measures were used. We compared our study to Okada’s study [7] in predicting both FDA-approved RA drugs as well as novel drugs (discussed in the next section). In Okada’s study, the authors first prioritized 98 RA risk genes, from which a total of 27 drug target genes of approved RA drugs were identified. These 27 drug target genes were then connected to 19 drugs. Among these 19 drugs, 14 were FDA-approved RA drugs and 5 were novel predictions. We could not compare our approach to one of the state-of-art drug repositioning systems, PREDICT [24], since it does not include RA. PREDICT used disease-disease and drug-drug similarities from multiple data resources to construct a classifier to determine treatment associations between 593 drugs and 313 diseases, the majority of which are Mendelian diseases. One of PREDICT’s limitations is that it can only infer new connections between the 593 drugs and 313 diseases included in the system. Evaluate Novel predications The fact that a drug repositioning algorithm has worked well in ranking known RA drugs (retrospective validation) does not imply that it will work equally well in novel drug prediction; therefore, evaluation of novel prediction capability is critical for any drug repositioning algorithm. Instead of manually searching literature or clinical trials for evidence supporting novel predictions, we automated this process using the drug-disease treatment knowledge bases that we constructed from over 22 million biomedical literature records and from 172,888 clinical trials. We extracted a total of 162 RA drugs from published biomedical literature, including 49 FDA-approved RA drugs and 113 novel drugs. We extracted a total of 103 RA drugs from clinical trials, including 37 FDA-approved RA drugs and 66 novel drugs. Combining RA drugs from these two resources, we obtained a total of 165 novel RA drugs after FDA-approved drugs were removed. We evaluated the performance of our algorithm in novel prediction using these 165 drugs. We calculated precisions, recalls, and F1 measures at different ranking cutoffs. We compared our study to Okada’s study in novel predictions using the same evaluation dataset. Analyze repositioned drug candidates To better understand top-ranked repositioned drug candidates, we examined which classes of drugs (as defined by the Anatomical Therapeutic Chemical (ATC) classification system [34]) were enriched. Drug classes enriched among top-ranked drug candidates could provide insights into the underlying mechanisms of action of drug candidates within those classes. The ATC system consists of 13 first-level codes, 94 second-level codes, 267 third-level codes, 882 fourth-level codes, and 4580 fifth-level codes, which are individual drugs. We experimented the drug classification using different levels of ATC codes and found that third level ATC codes gave sufficient granularity. We calculated percentages of drug classes associated with the top 10 % of ranked drugs and compared them to those for all drugs. We identified drug classes that showed at least a 2-fold enrichment. Results Evaluate network construction and disease ranking algorithms and analyze RA-related diseases We retrieved a ranked list of 842 diseases from the GDN using RA as the input. Two classes were highly enriched among top-ranked diseases: autoimmune diseases and immune diseases (Fig. 3). Among top 10 % of retrieved diseases, 18.6 % were autoimmune diseases, representing a significant 5.59-fold enrichment as compared to the 3.33 % among all retrieved diseases. We also observed a significant enrichment for immune diseases. Among the top 10 % of diseases, 37.21 % were immune diseases, representing a significant 2.45-fold enrichment as compared to 15.20 % among all retrieved diseases. These results were expected because RA itself is an immune disease as well as an autoimmune disease. These results demonstrate the validity of both network construction and disease ranking algorithms. The other four disease classes were not significantly enriched. Fig. 3 Disease class distribution among RA-related diseases at 10 ranking cutoffs Retrospective validation with 80 FDA-approved RA drugs Drug repositioning using the combined TreatKB has better performance than individual TreatKBs We validated our drug repositioning algorithm using 80 FDA approved RA drugs. Since the drug treatment knowledge bases (TreatKBs) were constructed from different data resources using different computational methods, we evaluated which TreatKBs performs better in drug repositioning for RA. We calculated recalls, mean, and median rankings of these FDA-approved RA drugs when the four different TreatKBs were used separately or combined (Table 1). When the TreatKB derived from FDA drug labeling (‘FDA-approved’) was used, we achieved a recall of 0.825, an average ranking of 36.58 %, and a median ranking of 34.46 %. We achieved significantly better rankings when the two TreatKBs that were constructed from post-marketing FAERS (recall: 0.775, mean ranking: 19.02 %, and median ranking: 8.53 %) and from the biomedical literature (recall: 0.663, mean ranking: 29.69 %, and median ranking: 19.17 %) were used, respectively. Significantly, when the combined TreatKB was used, we achieved a recall of 0.925 (74 out of 80 RA drugs), a mean ranking of 20.0 %, and a median ranking of 8.93 %. The significantly improved recall and rankings demonstrate the critical importance of a comprehensive drug treatment knowledge base in drug repositioning tasks. In comparison, such high rankings were not evident in our analysis for randomly selected FDA-approved drugs (49.76 % for mean ranking and 44.65 % for median ranking). As shown in Table 1, there is a significant difference between the median ranking of 8.93 % and the mean ranking of 20.05 %, demonstrating a skewed ranking distribution of these RA drugs. Fig. 4 shows that the rankings of RA drugs varied greatly from 0.04 % (prednisone) to 97.83 % (salicylamide). These results indicate that not all RA drugs can be discovered based on disease genetics. Fig. 4 Rankings of (74 out of 80) FDA-approved RA drugs among 2484 drugs. The combined TreatKB was used Table 1 Recalls, mean, and median rankings of 80 FDA-approved RA drugs when four TreatKBs were used separately and in tandem TreatKB Recall Mean ranking Median ranking FDA-approved 0.825 36.58 % 34.46 % Post-market 0.775 19.02 % 8.53 % Clinical trials 0.750 31.73 % 29.24 % Literature 0.663 29.69 % 19.17 % Combined 0.925 20.06 % 8.93 % The best performance was achieved when four TreatKBs were combined (bold data) Our approach has comparable overall performance but complementary precision and recall as compared to a “direct genetics-based” approach As shown in Fig. 5, our drug repositioning algorithm proved effective in ranking FDA-approved RA drugs at the top: the precision was 0.46 for top 25 drugs (top 1 %), which represents a significant 14.3-fold enrichment as compared to the 0.03 for all 2484 drugs (top 100 %). The best overall performance was achieved at a cutoff of 5 % (top 124 drugs): a precision of 0.22, a recall of 0.35, and an F1 of 0.27 were achieved. Fig. 5 The performance (precisions, recalls, and F1s) of our algorithm at six different cutoffs (top 1 %, 5 %, 10 %, 20 %, 50 % and 100 % [all 2484 drugs]) when evaluated using 80 FDA-approved RA drugs. For comparison, Okada’s study is shown in blue We compared our study to Okada’s study in prioritizing known RA drugs. When evaluated using the 80 FDA-approved RA drugs, Okada’s study achieved a recall of 0.175, a precision of 0.736 and an F1 of 0.28. Our algorithm, at a 5 % cutoff (top 124 drugs), achieved a precision of 0.22, a recall of 0.35, and an F1 of 0.27 (Fig. 5). While Okada’s study achieved a higher precision, our study achieved a higher recall, indicating that these two approaches are largely complementary. Evaluation of novel predictions using 165 novel RA drugs We evaluated our algorithm in novel prediction. As shown in Fig. 6, our drug repositioning algorithm proved effective in prioritizing novel RA drugs, achieving a precision of 0.89 for the top 25 drugs (top 1 %), which represents a significant 8.9-fold enrichment as compared to the 0.1 for all 2484 drugs. The best overall performance in novel prediction was achieved at a cutoff of 10 % (top 248 drugs): a precision of 0.46, a recall of 0.50, and an F1 of 0.47. Fig. 6 The performance (precisions, recalls and F1s) of our algorithm at six different cutoffs (top 1 %, 5 %, 10 %, 20 %, 50 % and 100 % [all 2484 drugs]) when evaluated using 165 novel RA drugs. For comparison, Okada’s study is shown in blue Okada’s study made a total of five novel predictions (auraofin, certolizumab pegol, lguratimod, tacrolimus, and temsirolimus). Among these five drugs, two drugs (certolizumab pegol, and tacrolimus) appeared in the evaluation dataset: both certolizumab pegol and tacrolimus currently are in active clinical trials. Therefore, Okada’s study has a precision of 0.40, a recall of 0.006 (2 out of 165 novel RA drugs), and an overall F1 of 0.01 when evaluated using the set of 165 drugs (Fig. 6). The best performance for our system was achieved at a cutoff of 10 % (top 248 drugs): a precision of 0.46, a recall of 0.50, and an F1 of 0.47, representing a 47-fold increase in F1 as compared to the F1 of 0.01 in Okada’s study. The drug tacrolimus from Okada’s study was ranked at top 0.76 % position (top 19 among all 2484 drugs) by our algorithm. In summary, we show that our repositioning strategy has comparable overall performance (yet complementary precision and recall) to Okada’s study in retrospective validation using the FDA-approved RA drugs. However, our algorithm has performed significantly better in finding novel RA drugs, and therefore has greater potential in the task of discovering innovative drug treatments for RA. Drug categories for top-ranked drug candidates offer insight into the underlying mechanisms of drug actions Figure 7 shows 15 third-level ATC codes that showed at least 100 % enrichment for the top 10 % ranked drug candidates as compared all drugs. As shown in the Fig. 7, 7 out of these 15 ATC codes are related to immune reaction modulation and inflammation, including Immunosuppressants, corticosteroids, and anti-inflammatory agents. In the disease class enrichment analysis, we also showed that both immune and autoimmune diseases were highly enriched among top-ranked RA-related diseases. Both analyses are consistent with the fact that immune deregulation is implicated in the development of RA. This result also demonstrates that common pathophysiological mechanisms are shared among RA-related diseases. Intriguingly, anticancer drugs are the most highly enriched drug class (Fig. 7). This is consistent with findings from Okada’s study indicating that SNPs in RA risk genes overlapped with somatic mutation genes in cancers. In particular, some genes involved in the development of hematological cancer were implicated. This is also consistent with the fact that several approved RA drugs (for example, rituximab) were initially developed for cancer treatment and subsequently repurposed for the treatment of RA. It will be interesting to further investigate top-ranked anticancer drugs in the treatment of RA experimentally. Fig. 7 Top 15 third-level ATC codes for top 10 % repositioned drug candidates that show at least 100 % enrichment as compared to percentages of the same ATC codes for all drugs. Both immune- and inflammation-related ATC codes were highlighted as red Discussion and conclusions In this study, we prioritized a total of 2484 drugs in terms of their relevance in the potential treatment of RA. While previous studies demonstrated that directly linking disease-associated genes from GWAS data to drug targets can lead to novel drug discovery, our study provides an alternative strategy to capitalize on complex human 5 genetics and comprehensive drug treatment data for other diseases for the discovery of innovative drug treatments for RA. Our algorithm retrieved 74 out of 80 FDA-approved RA drugs and ranked those drugs highly, demonstrating the validity of our approach. In addition, our algorithm proved effective in predicting innovative drugs for RA. Nonetheless, our study can be significantly improved upon in the future. Our current study was restricted by the limited number of diseases (881 diseases) in the GWAS catalog, even though TreatKB includes 24,511 diseases. With new studies being continually added to the GWAS catalog, as well as new disease-gene associations increasingly being revealed by next-generation sequencing studies, additional drug repositioning opportunities will arise through human genetic analysis. Additionally, disease genetics from rare Mendelian disorders represents another valuable source of novel drug targets and may lead to surprising and novel drug discovery opportunities [15, 35]. Another rich resource for knowledge of human disease genetics is computation-based candidate disease gene prediction. Computational disease gene prediction aims to find new disease-gene associations through integrative computational analysis of known data of diseases, genes, functional protein interactions, gene expression, and the biomedical literature, among many others [36]. We recently showed that computationally-predicted disease genetics can lead to novel drug discovery [26, 31]. In the future, we will combine comprehensive disease genetics data from all three of the above-described sources with a novel computational strategy to find new drug treatments for RA. This study focuses on disease genetics-based drug repositioning. Additional invaluable resources such as other disease-related data (i.e. disease phenotypic data or gene expression data) and drug-related data (i.e. drug side effects, drug chemical structure, and gene expression) can be incorporated into the currently proposed algorithm to further improve performance. Integrating and reasoning over such complex biological data poses a significant challenge that bears future investigation. Acknowledgements RX is funded by Case Western Reserve University/Cleveland Clinic CTSA Grant (UL1 RR024989), the Eunice Kennedy Shriver National Institute Of Child Health & Human Development of the National Institutes of Health under Award Number DP2HD084068, the Training grant in Computational Genomic Epidemiology of Cancer (CoGE) (R25 CA094186-06), and Grant #IRG-91-022-18 to the Case Comprehensive Cancer Center from the American Cancer Society. QW was partially supported by ThinTek LLC. Declarations Publication charges for this article have been funded by the Eunice Kennedy Shriver National Institute Of Child Health & Human Development of the National Institutes of Health under Award Number DP2HD084068. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials Data is available by contacting Rong Xu at rxx@case.edu. Authors’ contributions RX and QW jointly designed, implemented, and performed the experiments, and wrote the paper. Both authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. ==== Refs References 1 CDC: Rheumatoid Arthritis. http://www.cdc.gov/arthritis/basics/rheumatoid.htm. Accessed May 2015. 2 McInnes IB Schett G The pathogenesis of rheumatoid arthritis N Engl J Med 2011 365 23 2205 219 10.1056/NEJMra1004965 22150039 3 Diogo D Okada Y Plenge RM Genome-wide association studies to advance our understanding of critical cell types and pathways in rheumatoid arthritis: recent findings and challenges Curr Opin Rheumatol 2014 26 85 92 10.1097/BOR.0000000000000012 24276088 4 Hindorff LA Sethupathy P Junkins HA Ramos EM Mehta JP Collins FS Manolio TA Potential etiologic and functional implications of genome-wide association loci for human diseases and traits Proc Natl Acad Sci 2009 106 23 9362 367 10.1073/pnas.0903103106 19474294 5 Nelson MR Tipney H Painter JL Shen J Nicoletti P Shen Y Sanseau P The support of human genetic evidence for approved drug indications Nature genetics 2015 47 8 856 6 10.1038/ng.3314 26121088 6 Sanseau P Agarwal P Barnes MR Pastinen T Richards JB Cardon LR Mooser V Use of genome-wide association studies for drug repositioning Nature biotechnol 2012 30 4 317 20 10.1038/nbt.2151 22491277 7 Okada Y Wu D Trynka G Raj T Terao C Ikari K Toes RE Genetics of rheumatoid arthritis contributes to biology and drug discovery Nature 2014 506 7488 376 81 10.1038/nature12873 24390342 8 Collins FS. Reengineering translational science: the time is right. Science translational medicine. 2011;3(90). 9 Plenge RM Scolnick EM Altshuler D Validating therapeutic targets through human genetics Nat Rev Drug Discov 2013 12 8 581 94 10.1038/nrd4051 23868113 10 Franke A McGovern DP Barrett JC Wang K Radford-Smith GL Ahmad T Van Gossum A Genome-wide meta-analysis increases to 71 the number of confirmed crohn’s disease susceptibility loci Nature genetics 2010 42 12 1118 1125 10.1038/ng.717 21102463 11 Richards JB Zheng HF Spector TD Genetics of osteoporosis from genome-wide association studies: advances and challenges Nat Rev Genet 2012 13 8 576 88 10.1038/nrg3228 22805710 12 Xu R Wang Q Large-scale extraction of drug-disease treatment pairs from biomedical literature for drug repurposing BMC Bioinformatics 2013 14 1 181 90 10.1186/1471-2105-14-181 23742147 13 Xu R Li L Wang Q Towards building a disease-phenotype relationship knowledge base: large scale extraction of disease-manifestation relationship from literature Bioinformatics 2013 29 17 2186 194 10.1093/bioinformatics/btt359 23828786 14 Xu R Wang Q Automatic signal prioritizing and filtering approaches in detecting post-marketing cardiovascular events associated with targeted cancer drugs from the fda adverse event reporting system (faers) J Biomed Informa 2014 47 171 7 10.1016/j.jbi.2013.10.008 15 Wang ZY Zhang HY Rational drug repositioning by medical genetics Nature biotechnol 2013 31 12 1080 1082 10.1038/nbt.2758 24316641 16 Hurle MR Yang L Xie Q Rajpal DK Sanseau P Agarwal P Computational drug repositioning: From data to therapeutics Clin Pharmacol Ther 2013 93 4 335 41 10.1038/clpt.2013.1 23443757 17 Li J Zheng S Chen B Butte AJ Swamidass SJ Lu Z A survey of current trends in computational drug repositioning Brief Bioinform. 2016 17 1 2 12 10.1093/bib/bbv020 25832646 18 Lamb J Crawford ED Peck D Modell JW Blat IC Wrobel MJ Golub TR The connectivity map: using gene-expression signatures to connect small molecules, genes, and disease Science 2006 313 5795 1929 1935 10.1126/science.1132939 17008526 19 Sirota M Dudley JT Kim J Chiang AP Morgan AA Sweet-Cordero A Butte AJ Discovery and preclinical validation of drug indications using compendia of public gene expression data Sci Transl Med 2011 3 96 96 779677 10.1126/scitranslmed.3001318 20 Dudley JT Sirota M Shenoy M Pai RK Roedder S Chiang AP Butte AJ Computational repositioning of the anticonvulsant topiramate for inflammatory bowel disease Sci Transl Med 2011 3 96 96 769676 10.1126/scitranslmed.3002648 21 Chiang AP Butte AJ Systematic evaluation of drug–disease relationships to identify leads for novel drug uses Clin Pharmacol Ther 2009 86 5 507 10 10.1038/clpt.2009.103 19571805 22 Duran-Frigola M, Aloy P. Recycling side-effects into clinical markers for drug repositioning. Genome Med. 2012;4(3). 23 Nestler EJ Hyman SE Animal models of neuropsychiatric disorders Nature Neurosci 2010 13 10 1161 1169 10.1038/nn.2647 20877280 24 Gottlieb A, Stein GY, E R, Sharan R. Predict: a method for inferring novel drug indications with application to personalized medicine. Mole syst biol. 2011;7(1). 25 Xu R Wang Q Phenopredict: a disease phenome-wide drug repositioning approach towards schizophrenia drug discovery Journal of biomedical informatics 2015 56 348 355 10.1016/j.jbi.2015.06.027 26151312 26 Chen Y Xu R Phenome-driven disease genetics prediction towards drug discovery Bioinformatics 2015 31 12 i276 i283 10.1093/bioinformatics/btv245 26072493 27 Chen Y Cai X Xu R Combing Human Disease Genetics and Mouse Model Phenotypes Towards Drug Repositioning for Parkinson’s Disease AMIA Annu Symp Proc. 2015 2015 1851 1860 26958284 28 Welter D MacArthur J Morales J Burdett T Hall P Junkins H Parkinson H The nhgri gwas catalog, a curated resource of snp-trait associations Nucleic acids Res 2014 42 D1 1001 1006 10.1093/nar/gkt1229 29 Han J, Kamber M, Pei J. Data Mining: Concepts and Techniques: Concepts and Techniques: Elsevier; 2011. 30 Franceschini A Szklarczyk D Frankild S Kuhn M Simonovic M String v9. 1: protein-protein interaction networks, with increased coverage and integration Nucleic Acids Res 2013 41(D1) 808 15 10.1093/nar/gks1094 31 Chen Y Xu R Network-based gene prediction for plasmodium falciparum malaria towards genetics-based drug discovery BMC Genomics 2015 16 Suppl 7 9 10.1186/1471-2164-16-S7-S9 25612924 32 Xu R Wang Q Li L A genome-wide systems analysis reveals strong link between colorectal cancer and trimethylamine n-oxide (tmao), a gut microbial metabolite of dietary meat and fat BMC Genomics 2015 16 Suppl 5 6 10.1186/1471-2164-16-S5-S6 25608745 33 World Health Organization: International Statistical Classification of Diseases and Related Health Problems (Vol. 1): World Health Organization; 2004. Geneva. 34 http://www.who.int/classifications/atcddd/en/. 35 Brinkman RR Dubé MP Rouleau GA Orr AC Samuels ME Human monogenic disorders—a source of novel drug targets Nat Rev Genet 2006 7 4 249 60 10.1038/nrg1828 16534513 36 Moreau Y Tranchevent LC Computational tools for prioritizing candidate genes: boosting disease gene discovery Nat Rev Genet 2012 13 8 523 36 10.1038/nrg3253 22751426

PMC005xxxxxx/PMC5001201.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556637279210.1186/s12864-016-2792-1ResearchEpiTracer - an algorithm for identifying epicenters in condition-specific biological networks Sambaturu Narmada narmada.sambaturu@biochem.iisc.ernet.in 1Mishra Madhulika madhulika@biochem.iisc.ernet.in 2Chandra Nagasuma nchandra@biochem.iisc.ernet.in 121 IISc Mathematics Initiative, Indian Institute of Science, Bangalore, 560012 India 2 Department of Biochemistry, Indian Institute of Science, Bangalore, 560012 India 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.543© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background In biological systems, diseases are caused by small perturbations in a complex network of interactions between proteins. Perturbations typically affect only a small number of proteins, which go on to disturb a larger part of the network. To counteract this, a stress-response is launched, resulting in a complex pattern of variations in the cell. Identifying the key players involved in either spreading the perturbation or responding to it can give us important insights. Results We develop an algorithm, EpiTracer, which identifies the key proteins, or epicenters, from which a large number of changes in the protein-protein interaction (PPI) network ripple out. We propose a new centrality measure, ripple centrality, which measures how effectively a change at a particular node can ripple across the network by identifying highest activity paths specific to the condition of interest, obtained by mapping gene expression profiles to the PPI network. We demonstrate the algorithm using an overexpression study and a knockdown study. In the overexpression study, the gene that was overexpressed (PARK2) was highlighted as the most important epicenter specific to the perturbation. The other top-ranked epicenters were involved in either supporting the activity of PARK2, or counteracting it. Also, 5 of the identified epicenters showed no significant differential expression, showing that our method can find information which simple differential expression analysis cannot. In the second dataset (SP1 knockdown), alternative regulators of SP1 targets were highlighted as epicenters. Also, the gene that was knocked down (SP1) was picked up as an epicenter specific to the control condition. Sensitivity analysis showed that the genes identified as epicenters remain largely unaffected by small changes. Conclusions We develop an algorithm, EpiTracer, to find epicenters in condition-specific biological networks, given the PPI network and gene expression levels. EpiTracer includes programs which can extract the immediate influence zone of epicenters and provide a summary of dysregulated genes, facilitating quick biological analysis. We demonstrate its efficacy on two datasets with differing characteristics, highlighting its general applicability. We also show that EpiTracer is not sensitive to minor changes in the network. The source code for EpiTracer is provided at Github (https://github.com/narmada26/EpiTracer). Keywords Network miningInfluential nodesRipple centralityPerturbation analysisCondition-specific networkIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background A biological system consists of a large number of proteins involved in a series of intricate and tightly orchestrated interactions. Representing this complex system as a network allows us to harness network-mining methodologies to analyse the system as a whole. Diseases typically affect only a small number of proteins [1, 2]. The immediate interacting partners of these proteins can be expected to show a change in expression levels or behavior. In addition, the inter-connected nature of the system causes cascade effects, altering the levels of proteins far removed from the original source. At the same time, the system may attempt to restore its equilibrium by launching a stress-response [3]. It would be interesting and useful to identify the key players in this tug-of-war, which are most influential in either spreading or curtailing the perturbation. These key proteins are referred to as epicenters specific to that condition. A vast amount of data is generated by microarray experiments, which provide a snapshot of the active and inactive players of the system. These datasets are available on public databases such as Omnibus [4]. Most studies in biology focus on only a few proteins or pathways, and work with a restricted field of view. Through algorithms such as EpiTracer, we hope to enable the analysis of large scale and detailed models, giving a picture which reflects the intricate workings of living systems more closely. In this paper, we work with a dataset consisting of nearly half the complement of human genes. In this paper, we develop an algorithm called EpiTracer, which identifies the epicenters from which either the perturbation or the reaction to it ripples out. This is done using a protein-protein interaction (PPI) network into which gene expression levels before and after the perturbation are integrated. To the best of our knowledge, no method exists currently which can identify epicenters with this type of data. Other methods that provide insights into influential nodes require a causal network as input, where each edge depicts a causal relationship, and is directed from the cause to the effect [5, 6]. However, clear-cut causal dependencies have been established for only a small set of proteins, making it impossible to analyse large networks. Network motifs have also been used to highlight important proteins in directed biological networks [7]. However these methods do not make use of information about changes in expression levels of genes, thus losing out on a rich source of information. Methods also exist which highlight the nodes which, when intentionally perturbed, spread the perturbation the fastest [8]. This is not the same as identifying the epicenter of a naturally occurring perturbation, which is a more complicated and biologically relevant scenario. The EpiTracer algorithm is based on the observation that an epicentric protein would have to be highly active in order to exert its influence, and also have good connectivity in order for its influence to spread. We define a new centrality measure called ripple centrality, which gives a combined measure of a node ′s activity as well as its connectivity, thus allowing us to rank proteins on their ability to be an effective epicenter. The top-ranked proteins qualify to be epicenters. The algorithm combines the PPI network and gene expression levels in such a way as to ease the computation of active paths. The sub-network with high activity paths only in the perturbed condition is extracted, thus reducing the search space for the next step. The nodes in this sub-network are then ranked on the basis of their ripple centrality score, with the top 10 nodes considered as epicenters. The efficacy of the algorithm is demonstrated through two case studies. The first case study analyses human glioma cell line (U251) upon overexpression of the gene PARK2 (GSE61973) [9]. The algorithm was able to identify PARK2 as the most important epicenter without any prior knowledge of the perturbation. Functional enrichment analysis showed that most of the top 10 epicenters play a role in enabling or countering the activity of PARK2. Also, 5 of the top 10 epicenters showed no significant fold change, proving that our method is capable of identifying more than simple differential expression analysis. The EpiTracer pipeline includes a program for extracting the immediate influence zone of the epicenters. Analysis of the immediate influence zone of the top-ranked epicenter (PARK2) showed that it was enriched in genes involved in cell-cycle regulation. The second case study attempts to identify the target genes regulated by transcription factor SP1 by knocking down the expression of SP1 in HeLa cells (GSE37935) [10]. In this study, EpiTracer was able to identify SP1 among the top ranked epicenters. Sensitivity analysis was carried out by increasing the gene expression levels of all nodes by upto 5 % (100 independent experiments), and decreasing the gene expression levels of all nodes by upto 5 % (100 independent experiments). It was found that irrespective of the direction or extent of perturbation, 9 nodes always appear in the top 10 ranks, and 16 nodes always appear in the top 20 ranks of epicenters. This shows that the nodes ranked as epicenters remain largely unaffected even when every gene in the system is subjected to a minor change. Methods A high-density protein-protein interaction network was reconstructed for use in this work. Condition-specific gene expression profiles were obtained from published literature. The inputs as well as the algorithm are explained below. Protein-protein interaction network A base network containing known and predicted protein-protein interactions, genetic interactions and regulatory interactions with directions was taken from Khurana et. al., 2013 [11]. Metabolic interactions from KEGG [12] were added to this, resulting in a directed network with 10,306 nodes and 74,404 edges. Gene expression profiles Two gene expression datasets were obtained from the GEO database [13]. In the first dataset GSE61973 [9], PARK2 gene was overexpressed in human glioma cell line (U251). In the second study GSE37935 [10], SP1 gene was knocked down using siRNA in HeLa cells. These two case studies were selected to demonstrate the general applicability of the EpiTracer algorithm. GeneSpringX 12.6.1, with Robust Multichip Averaging (RMA) [14] was used for microarray data normalization. A 1.5 fold cut-off was applied for differential gene expression analysis (P-value ≤ 0.05 by T-test with Benjamini-Hochberg false discovery rate correction). Combining the inputs The gene expression profile of each condition was mapped onto the PPI network, to create one weighted network per condition (Fig. 1(a)). The nodes (proteins) were given a weight equal to the normalized signal intensity for the corresponding gene in that condition. wix=SIx where wix is the weight of node i in condition x, and SIx is the normalized signal intensity in condition x. This formulation stems from the assumption that the expression level of a gene gives a reasonably good approximation of the abundance of the protein in the system. The cost of an edge (protein-protein interaction) was taken as a function of the abundance of the participating proteins, as cix=1wux∗wvx where cix is the cost of edge i in condition x, and wux, wvx are the weights of the nodes comprising the edge. This follows from the assumption used in mass-action kinetics, that the activity of a reaction is directly proportional to the concentration of the participants. Taking the inverse makes sure that a highly active interaction has a very low edge cost. Fig. 1 The EpiTracer workflow. a Gene expression profiles of each condition are mapped onto the base PPI network. b Highest activity paths are calculated for each condition, and common paths are discarded, giving condition-specific highest activity paths (CSHAPs). c The network induced by the CSHAPs form the condition-specific highest activity networks (CSHANs). d Nodes in the perturbed highest activity network are ranked according to ripple centrality. e The ranked list of nodes is split into two lists based on overlaps with the control highest activity network. Top 10 nodes in the list unique to the perturbed condition form the epicenters specific to the perturbation Given a path with n edges, the sum of costs of the edges involved in the path gives the cost of the path. pathcost=∑i=1ncix where cix is the edge cost for each edge in the path, and n is the length of the path. A shortest path algorithm will preferentially choose edges with the least cost for a given source and destination, which in our formulation translates to identifying the highest activity path. EpiTracer algorithm – rationale In order to be effective, an epicenter should be highly active and participate in high activity paths only in the perturbed condition. To capture this, we calculate highest activity paths in each condition and discard common paths. The common paths correspond to the paths which remain highly active and unchanged irrespective of the perturbation. Such paths add no information about the perturbation (Fig. 1(b)). The edges involved in these CSHAPs induce a sub-network of the original network, referred to as the condition-specific highest activity networks (CSHANs) (Fig. 1(c)). An epicenter should also be able to reach many nodes in the network in order to exert its influence, and the paths from the epicenter to these nodes must also be highly active. This is captured by the new centrality measure proposed here termed ripple centrality, and is explained below. Closeness centrality Closeness centrality [15] of a node u is defined as the reciprocal of the sum of shortest path costs from u to every reachable node v C(u)=1∑vσ(u,v) where σ(u,v) is the cost of the shortest path from u to v. Because of the way edge costs are formulated, a node u with highly active paths to a set of nodes v will have high closeness centrality. This is depicted by node Acl in Fig. 2a. Here a thicker edge corresponds to a highly active reaction. Fig. 2 Illustration of ripple centrality. a Node Acl is the source of highly active paths, and has high closeness centrality. However it can only reach 4 nodes, and is not a good epicenter. b Node Aor can reach 14 nodes, but paths originating at Aor have low activity. Thus it is not a good epicenter. c Node Arc is the source of highly active paths and can reach a large number of nodes (7), making it the best candidate for an epicenter. The hexagon represents candidate epicenters Outward reachability Given a node u, the number of nodes reachable from u is termed its outward reachability [16]. Rout(u)=nodesreachablefromu where Rout(u) denotes outward reachability of u. Ripple centrality In Fig. 2a, the node Acl represents nodes which have very high activity paths, but to only a small number of nodes. Such a node would have high closeness centrality [15], but would not be a good candidate for an epicenter as any perturbation arising at this point could not spread to a large number of nodes. On the other hand, node Aor (Fig. 2b) represents nodes which have very good connectivity, but participate in relatively low activity paths. These types of nodes would have high outward reachability, but are poor candidates for epicenters. Thus neither closeness centrality nor outward reachability are sufficient on their own. Node Arc in Fig. 2c has highly active paths to a large number of nodes, and is the best candidate for an epicenter. We formulate a new measure, ripple centrality, which serves as a logical AND between closeness centrality and outward reachability. Ripplecentrality(u)=C(u)∗Rout(u) In the calculation, both closeness centrality and outward reachability are normalized. Ripple centrality is calculated for the nodes in the perturbed CSHAN, resulting in the proteins being ranked on the basis of their effectiveness as potential epicenters (Fig. 1(d)). The ranked list is then split into two lists – (a) nodes occurring only in the perturbed CSHAN, and (b) nodes common to both CSHANs (Fig. 1(e)). Common nodes work as global epicenters, playing key roles both before and after the perturbation. Since identical paths have already been discarded (Fig. 1(b)), these proteins are those which have undergone re-wiring, and participate in a different pathway upon perturbation. The nodes occurring only in the perturbed CSHAN are epicenters specific to the perturbation, involved in either the spread of the perturbation or the reaction to it. EpiTracer algorithm The EpiTracer algorithm consists of three modules (1) highest_activity_paths extracts the paths with cost inside a user-defined percentile threshold, (2) condition_specific_han uses highest_activity_paths to identify the highest activity network specific to each condition, and (3) the main module, get_epicenters, uses the above two modules to identify the top 10 epicenters in the perturbed condition, as well as the top 10 epicenters common to both conditions. The pseudocode for each module is provided in Algorithms 1, 2 and 3. The symbols GA and GB refer to the graph for condition A and the graph for condition B, respectively. Biological analysis The proteins identified as epicenters, as well as the proteins surrounding them were subjected to biological and functional analysis. Immediate influence zone The nodes that occur within two hops upstream or downstream from an epicenter are designated the immediate influence zone of that epicenter. For the top-ranked epicenter, the immediate influence zone was identified manually and was restricted to the perturbed highest activity network. Downregulated genes which occur within two hops of the epicenter were picked from the full network and added to the influence zone. Since manually examining the full network for dysregulated genes in the vicinity of every epicenter is a time consuming and laborious task, an automated script was developed to facilitate the quick extraction of the influence zone. This can be done on the full network or on the highest activity network. This allows for easy identification of nodes with significant dysregulation, and can be used for further analysis. This script uses a default fold change cut-off of 2.0. Both the number of hops and the fold change can be varied by the user if necessary. Functional enrichment Gene set enrichment was performed against the KEGG [12] database using WebGestalt [17]. A hypergeometric test with P-value of 0.05 with FDR correction was used for statistical analysis. Network visualization was carried out with Cytoscape, and the Cytoscape plugin ClueGO [18] was used for GO module enrichment. Sensitivity analysis Two separate sensitivity analyses were carried out, one by increasing the expression levels of all genes by a randomly chosen value between 0 and 5 %, and the other by decreasing the expression levels of all genes similarly. This reflects measurement errors that can be introduced in the microarray data due to variability in the sensitivity of the detector. All numbers reported are an average of 100 independent experiments. Results The algorithm was implemented in Python 2.7, and uses the functions provided by Networkx 1.7 for computing all the centrality measures. Dijkstra ′s algorithm [19] was used for finding shortest paths. The EpiTracer algorithm was able to analyse a dataset consisting of 10,306 nodes and 74,404 edges on a 16 core Xeon server in less than 30 minutes. The results of the first case study are provided in detail in the next section, followed by a summary of the second case study. Case study 1 Microarray data for the overexpression of PARK2 in human glioma cell line (U251) and control (GFP) were taken from (E–GEOD–61973) [9]. PARK2 (PARKIN) is an E3 ubiquitin ligase whose dysfunction has been associated with Parkinsonism. The authors of this data, in their study [9], show that PARK2 is frequently deleted or downregulated in human glioma, and demonstrate that overexpression of PARK2 can significantly inhibit glioma cell growth. Through the EpiTracer algorithm, we uncover the global reprogramming of gene expression resulting from this perturbation, and highlight the epicenters of this process. We also provide a ranked list of influential players in this perturbation. System description The gene expression profiles were normalized and filtered, and the list of differentially expressed genes was extracted using a fold change cut-off of 1.5. It was found that 605 genes were downregulated and 1,089 genes were upregulated as a result of the overexpression of PARK2. In general, genes associated with cell cycle, ubiquitin mediated proteolysis, ErbB signaling pathway, MAPK, JAK-STAT signaling, WNT signaling, Hedgehog signaling pathway and pathways related to lipid metabolism were differentially expressed. A summary of network properties is shown in Fig. 3a. Fig. 3 Case study 1 (PARK2 overexpression in human glioma cell line). Data corresponding to overexpression of PARK2 in human glioma cell line (U251). a Human PPI network comprising of 10,306 nodes and 74,404 edges. Nodes colored red are upregulated upon perturbation, and nodes colored green are downregulated. (a1) Table of network properties. b Perturbation-specific *HAN (highest activity network), with network properties in table b1. c The 5 epicenters which were differentially expressed, along with their immediate neighbors d List of epicenters specific to the perturbation as well as global epicenters Highest activity paths (HAPs) All-pairs-shortest paths were calculated for the control network as well as the perturbed network. Paths with length ≥ 2 were sorted in the ascending order of path cost. It was found that the number of paths retained at a percentile cut-off of 0.2 was twice that retained when a cut-off of 0.1 was used. Thus the conservative threshold of 0.1 percentile was chosen, resulting in 67,728 paths being retained as highest activity paths (HAPs) in the perturbed network and 58,570 HAPs in the control network. Condition-specific highest activity network (CSHAN) Highest activity paths common to both conditions correspond to the paths which are highly active all the time, and are unaffected by the perturbation. Such paths were removed, giving us 9,621 HAPs specific to the control condition, and 18,779 HAPs specific to the perturbed condition. The edges involved in these paths correspond to the condition-specific highest activity networks (CSHANs). Interestingly, the CSHANs were themselves well-connected networks (Fig. 3b). Of the 1,756 genes in the perturbed CSHAN, 75 genes were found to be downregulated, and 130 were found to be upregulated. These belonged to the functional categories of cell cycle, MAPK, ErbB, p53 and mTOR signaling pathway, ubiquitin mediated proteolysis, regulation of actin cytoskeleton and oocyte meiosis. Tracing the epicenter The nodes in the perturbed CSHAN were ranked in descending order of their ripple centrality. This ranked list was then split into two - nodes occurring only in the perturbed CSHAN, and nodes common to both CSHANs (global epicenters). Since common paths have already been removed, nodes common to both CSHANs correspond to the nodes which participate in a different pathway after the perturbation. Nodes occurring only in the perturbed CSHAN are those which have become active and influential after the perturbation. The top 10 nodes from each list were considered as epicenters, and are listed in Fig. 3d. PARK2 was identified as the highest ranked epicenter among the nodes unique to the perturbed CSHAN, in spite of the fact that the algorithm was given no prior knowledge of the perturbation. Only 5 out of the 10 epicenters specific to the perturbed condition were found to have significant differential expression. This shows that EpiTracer is able to capture information that simple differential expression analysis cannot. The 5 epicenters which were differentially expressed, along with their immediate neighbors, have been depicted in Fig. 3c. Biological interpretation Top global epicenters were found to correspond to highly conserved and ubiquitously expressed proteins such as TUBB, GAPDH, VCL, ACTG1, DYNLL1 and ANXA2. RAC1 is known to promote cell migration and invasion in glioma cells. APP is associated with axonogenesis, neurite growth and neuronal adhesion [20]. PRDX1 is involved in redox regulation of the cell. B2M is associated with MHC Class I antigen presentation. Further, the top epicenters specific to the perturbed (PARK2 overexpression) condition were examined. It was found that 5 out of the 10 genes being examined showed significant differential expression, namely PARK2, RGS2, EPHA2, DNAJC1 and FGF2 (Fig. 3c). PARK2 was highlighted as the most important epicenter specific to the PARK2 overexpression condition. PARK2 negatively regulates cell cycle by degrading Cyclin E and D through its activity as an E3 ubiquitin ligase. RGS2 is involved in G0 to G1 transition [20]. Inhibition of EPHA2 leads to stalling of cells in G0/G1 phase [21]. In the PARK2 overexpression condition, EPHA2 was found to be upregulated. FGF2 blocks cell proliferation and causes a G2/M arrest [22]. When considered together, our analysis revealed that most of the top ranked genes were associated with cell cycle regulation. Immediate influence zone of the top-ranked epicenter In order to understand the cellular response to the top-ranked epicenter specific to the perturbed condition (PARK2 in this case), the influence zone around it was analysed. The subgraph induced by considering nodes upto two hops up/downstream of PARK2 in the perturbed CSHAN were considered to be in the PARK2 influence zone. Any downregulated nodes within 2 hops of PARK2 in the complete network were also added (Fig. 4a). GO enrichment was carried out specifically for cell cycle regulation as PARK2 is known to be a cell cycle regulator. Interestingly, it was found that the PARK2 influence zone was highly enriched for cell cycle regulation (Fig. 4b), including G2/M transition and G1/S transition of mitotic cell cycle, mitotic cell cycle, positive and negative regulation of cell cycle. Fig. 4 PARK2 influence zone. Detailed biological interpretation of PARK2 influence zone. a The PARK2 influence zone consists of 118 nodes and 119 edges. Red colored nodes correspond to upregulated genes, and green corresponds to downregulated genes. The epicenter is depicted using a hexagon. b GO enrichement of genes in PARK2 influence zone shows that most genes are involved in cell-cycle regulation. c Nodes downstream of PARK2 d Mechanistic insights into cell-cycle dysregulation upon PARK2 overexpression The influence exerted by PARK2 was studied by focusing on the nodes downstream of PARK2 (Fig. 4c). It was found that many downstream genes such as MDM2, CHEK1, SQSTM1 and DUSP1 were involved in cell cycle regulation. Since overexpression of PARK2 inhibits the progression of cell cycle, the expected response from the cell would be to modify other regulatory mechanisms of cell cycle progression to counteract this arrest. Examination of the nodes downstream of the top-ranked epicenter (PARK2) showed that this was indeed the case (Fig. 4d). Major remodeling can be inferred from the G0/G1 and G1/S transition. SQSTM1 (P63) is involved in exiting of the cell from the M phase in the cell cycle. CD44, EPHA2, RGS2 and ARL6IP1 are positive regulators for G0/G1 transition. MDM2 is an activator of G1/S transition as it inhibits P53 and Rb proteins. However, CHEK1 and DUSP1 are repressors of G1/S phase transition. CHEK1 acts as a Cyclin E repressor by inhibiting Cdc at the DNA-repair check-point. DUSP1 is a repressor of the MAPK pathway [23]. FGF2 and NEK6 are repressors of G2/M phase transition [24]. Since creation of such influence zones for every highly ranked gene is a tedious task, an automated script was developed to output the influence zone as well as to summarize the details of differentially expressed genes in an easy-to-read table. The table thus generated for the second highest ranked epicenter, CD44, is shown in Table 1. Table 1 Automated summary for CD44 (case study 1). CD44 was the 2 nd ranked epicenter specific to the perturbed condition. The table shows the nodes in the immediate influence zone (up to 2 hops up/downstream) of CD44 which showed significant differential expression (2-fold). The first row corresponds to the input node, CD44. In the subsequent rows, the first column shows the differentially expressed gene (DEG). If the DEG is more than 1 hop away from CD44, the intermediate nodes on the unweighted shortest path are described in columns 6 onwards Node Direction Num_hops Fold_change Which_network Intermediate_node_1 Fold_change Significant_fc? Which_network CD44 input_node 0 1.0776225906 unique to perturbed CSHAN L1CAM down_CD44 2 4.3995043553 unique to perturbed CSHAN EZR 1.4077313131 False common to both CSHANs CBLB down_CD44 2 2.7450193138 unique to perturbed CSHAN EGFR 0.9625043727 False common to both CSHANs TNNT1 down_CD44 2 0.3519677136 unique to control CSHAN FYN 1.047836048 False common to both CSHANs RPS6KA2 down_CD44 2 3.7937197295 unique to perturbed CSHAN EGFR 0.9625043727 False common to both CSHANs NEDD9 down_CD44 2 0.4699818669 not in any CSHAN FYN 1.047836048 False common to both CSHANs TFPI down_CD44 2 2.7743880699 not in any CSHAN MMP7 0.1323159054 True not in any CSHAN MBNL3 down_CD44 2 4.0574758277 not in any CSHAN LCK 0.967172212 False not in any CSHAN IVNS1ABP down_CD44 2 2.0134951539 common to both CSHANs ARHGEF1 1.0053092034 False not in any CSHAN ITGB3 down_CD44 2 0.2450269938 not in any CSHAN COL1A2 1.6105271705 False common to both CSHANs PLA2G4A down_CD44 2 7.496384226 not in any CSHAN COL1A2 1.6105271705 False common to both CSHANs FN1 down_CD44 2 2.2964207553 unique to perturbed CSHAN COL1A2 1.6105271705 False common to both CSHANs MEF2C down_CD44 2 2.2667831761 unique to perturbed CSHAN CD74 0.9449408306 False not in any CSHAN PTK2 down_CD44 2 0.2821339725 not in any CSHAN EGFR 0.9625043727 False common to both CSHANs CHN1 down_CD44 2 3.1958019401 unique to perturbed CSHAN TGFBR1 0.7284950123 False common to both CSHANs EGR1 down_CD44 2 2.1681746909 common to both CSHANs ARHGEF1 1.0053092034 False not in any CSHAN SRGN down_CD44 1 0.4947740703 not in any CSHAN OCLN down_CD44 2 2.9381790192 not in any CSHAN TGFBR1 0.7284950123 False common to both CSHANs ADAM12 down_CD44 2 2.3217184015 unique to perturbed CSHAN IGFBP3 1.9432883668 False common to both CSHANs TIMP1 down_CD44 2 0.4692561311 unique to control CSHAN MMP1 0.9852575467 False not in any CSHAN MMP7 down_CD44 1 0.1323159054 not in any CSHAN L1CAM up_CD44 2 4.3995043553 unique to perturbed CSHAN ANK1 0.7930398677 False not in any CSHAN ITGB3 up_CD44 2 0.2450269938 not in any CSHAN COL1A2 1.6105271705 False common to both CSHANs MMP7 up_CD44 1 0.1323159054 not in any CSHAN Sensitivity analysis The gene expression levels of all the genes were either increased or decreased as indicated in the Methods section. The results of the 200 independent runs were then analysed to check how the top ranked epicenters fared. It was found that 9 nodes were always present in the top 10 ranked epicenters specific to the perturbed condition irrespective of the direction or extent of perturbation. When the top 20 ranks were considered, 16 nodes were common to all 200 experiments. Also, PARK2 was ranked the 9.6th most important epicenter specific to the perturbed condition on average out of 10,306 possible candidates. This shows that even when every single node in the network was perturbed, the nodes ranked as epicenters remained largely unaffected. Case study 2 Microarray data for the knockdown of SP1 gene in HeLa cells were taken from GSE37935 [10]. The knockdown was carried out by treating HeLa cells with an siRNA directed against the SP1 mRNA. SP1 is a global transcription factor, and regulates various important biological processes such as proliferation, cell differentiation and oncogenesis. Since the knockdown of a transcription factor can lead to downregulation of its target genes which are positively regulated, these genes will have higher activity in the control condition. Hence in this scenario, we analyse epicenters specific to the perturbed as well as the control condition. Biological interpretation In the perturbed (SP1 knockdown) condition, the top 10 ranked epicenters consist of 14 genes. 5 genes are assigned the same rank due to similar activity and connectivity. Out of the 14 epicenters, 5 genes, namely GPRC5A, EBF1, PTPN4, FAS and ADCK2 were differentially expressed. GPRC5A, EBF1 and PTPN4 genes play important roles in development, cellular growth, and differentiation [20]. FAS is involved in physiological regulation of programmed cell death. The function of ADCK2 is not yet clear. In the control condition, top 10 epicenters include 50 genes, with 30 genes being ranked 7th and 9 genes being ranked 2 nd due to similar activity. In this case, SP1 appeared as the 10th ranked epicenter. Immediate influence zone The immediate influence zone of the top 10 epicenters was constructed as a combined network. The targets of SP1 and their first interactors were added to this network, and the entire network was pruned to retain only epicenters, targets of SP1, differentially expressed genes, and genes which were essential for the connectivity of the graph. This pruned graph contains 142 nodes and 228 edges, and is shown in Fig. 5a. Analysis of the graph showed that epicenters were generally indirect regulators of the targets of SP1. This could indicate that alternative methods of regulating SP1 targets gained importance due to the knockdown of SP1. Many targets of SP1 were found in the highest activity paths which trace back to the epicenters. For example, MYC and TP53 were highlighted as important genes regulated by SP1, and 10 regulators of MYC were ranked as epicenters, with 5 of them being assigned the same rank due to similar activity. The paths tracing back to the epicenters clearly illustrate the cascade of influence of the epicenters to the targets of SP1, involving mediator genes. The most prominent mediator genes in the SP1 knockdown condition are EEF1A1 and HSPA8. EEF1A1 is regulated by 7 epicenters, of which FAS, EBF1 and PTPN4 are differentially expressed. EEF1A1 in turn regulates 5 targets of SP1, of which EP300 is differentially expressed. EP300 is a transcriptional co-activator protein, and is important in the processes of cell proliferation and differentiation. Fig. 5 Case study 2 (SP1 knockdown in HeLa cell line). Influence zone of the top 10 epicenters was constructed from the condition-specific highest activity network and enriched with the targets of SP1 and their immediate neighbors. This network was pruned to retain only epicenters, SP1 targets, differentially expressed genes, and the genes connecting them. Nodes with a hexagonal shape represent epicenters, a golden border around the node indicates SP1 target, and a pink border around the node indicates mediator gene. The rank of each epicenter is written next to it in red. (a) SP1 knockdown condition. 14 genes occur in the list of top 10 epicenters (5 genes correspond to rank 5). (b) Control condition. 50 genes correspond to top 10 epicenters. 30 genes correspond to rank 7, and regulate MYC, a target of SP1. Similarly, 9 genes correspond to rank 2, and regulate CEBPB Since the data being analysed is of knockdown of a transcription factor (SP1), we investigate the targets regulated by SP1 by focusing on the highest activity network specific to the control condition. The influence zone for the epicenters with top 10 ranks was constructed and pruned as in the perturbed condition. This graph contains 125 nodes and 168 edges, and is shown in Fig. 5b. MYC and CEBPB emerged as important genes in this condition. MYC is a direct target of SP1, and also regulates other targets of SP1. CEBPB is an important mediator gene, which regulates 4 targets of SP1, and is regulated by 9 epicenters, all of which were ranked 2 nd. In conjunction, the analysis of the two conditions revealed that the effect of SP1 knockdown spreads through 3 important hubs - MYC, CEBPB and TP53. Regulators of these important hubs were ranked as epicenters by our algorithm. MYC has 11 target genes which are differentially expressed. ZNFX1, TAF1D, NFX1, TFIIF and NFX1 are involved in tanscriptional and post transcriptional regulation [20]. CDKN2A activity leads to cell cycle arrest. ODC1 is an enzyme of polyamine metabolism and PFAS participates in purine metabolism [12]. Both metabolic pathways are necessary for DNA replication and transcription. NFX1 is mainly involved in inflammatory response. TP53 also regulates 11 genes, of which three genes, namely PTTG1, COPS5 and CDKN2A, are differentally expressed. COPS5 is one of the members of the COP9 signalosome which regulates mutiple signaling pathways [20]. PTTG1 is involved in cell cycle regulation. CEBPB regulates 10 gene in the control condition, of which two were differentally expressed - SP1 and INHBE. Discussion EpiTracer identifies nodes at which highly active paths originate and which are able to reach a large fraction of the active network. When annotated with the condition in which they are active, these nodes correspond to the most influential players in that specific condition and are termed epicenters. It is important to note that the epicenter does not necessarily correspond to the source of the perturbation. EpiTracer can be expected to have wide applicability, demonstrated here by two entirely different datasets studied in this work. Since the algorithm focuses on active nodes and edges, the network on which the analysis is carried out must be chosen based on the context. As demonstrated in case study 1, analysing the perturbed highest activity network is preferable when the perturbation is expected to be an upregulation event. If the perturbation is expected to be a downregulation event, analysing the control highest activity network will yield the set of nodes which were influential before the knockdown (case study 2). An analysis of the perturbed highest activity network is also useful since it can yield a list of epicenters that are activated in the perturbed condition upon removal of the knocked-out regulator. If the nature of the perturbation is unknown, both highest activity networks should be analysed. A limitation of the algorithm is that the source of the perturbation may not appear in the highest activity networks if its expression level remains low both before and after the perturbation. In such cases EpiTracer will be able to highlight the highly active nodes close to the source of the perturbation, but not the source itself. It was observed during the course of this work that the largest strongly connected component (LSCC) plays an important role in spreading a perturbation through the network. The largest strongly connected component is the largest subgraph in which there exists a path from every node to every other node. It was found that the epicenter was a part of the LSCC in the highest activity network under study. If the LSCC comprises a big enough percentage of the graph, we believe it might be possible to speed up the algorithm by restricting the search only to the nodes in the LSCC. Conclusion We propose a new algorithm, EpiTracer, to trace the epicenter of perturbations in a condition-specific biological network. The algorithm is capable of extracting the highest activity network specific to each condition under study and ranking the nodes in these highest activity networks with a ripple centrality score, which reflects how well any influence from that node can ripple out into the rest of the network. The algorithm has been demonstrated on two case studies, one where a gene was overexpressed, and another where a gene was knocked down. In the case of overexpression, EpiTracer was able to identify the overexpressed gene as the most important epicenter. Biological analysis of the top-ranked epicenters showed that all of them had functions relevant to cell cycle progression, and highlighted a scenario where the most important epicenters were involved in either spreading the influence of PARK2 or working to counteract its effect. Also, 5 of the top 10 epicenters showed no significant change in expression level, and yet were found to be biologically meaningful epicenters. This shows that our algorithm is able to highlight more than simple differential expression. The immediate influence zone of PARK2 generated by the EpiTracer pipeline, and the dysregulated genes in this were also found to be enriched in genes involved in cell cycle regulation. In the knockdown case study, alternative regulators of the knocked-down gene’s targets were highlighted as epicenters. Also, the gene that was knocked down was picked up as an epicenter in the control condition. This demonstrates the general applicability of the algorithm. Sensitivity analysis has been carried out to show that the epicenters identified by EpiTracer are largely unaffected by small changes in the network. The EpiTracer algorithm identifies the epicenters which either spread a perturbation or respond to it. The paths along which the influence ripples out of the epicenters is highlighted by the condition-specific highest activity network. This gives a system-wide, unbiased view of a disease phenotype, and how the organism responds to it. Acknowledgements We thank Dr. Arun Konagurthu, Monash University, Australia for the useful discussions. Declarations Publication of this article was partly funded by the Department of Biotechnology (DBT, India). This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4. Authors’ contributions NS developed and implemented the algorithm. MM analysed the networks and biological significance. NC generated the idea and supervised the whole project. All authors wrote and approve the manuscript. Competing interests The authors declare that they have no competing interests. ==== Refs References 1 Goh KI Cusick ME Valle D Childs B Vidal M Barabasi AL The human disease network Proc Natl Acad Sci 2007 104 21 8685 90 10.1073/pnas.0701361104 17502601 2 Padiadpu J Vashisht R Chandra N Protein–protein interaction networks suggest different targets have different propensities for triggering drug resistance Syst Synth Biol 2010 4 4 311 22 10.1007/s11693-011-9076-5 22132058 3 Rowland MA Fontana W Deeds EJ Crosstalk and competition in signaling networks Biophys J 2012 103 11 2389 98 10.1016/j.bpj.2012.10.006 23283238 4 Edgar R Domrachev M Lash AE Gene expression omnibus: Ncbi gene expression and hybridization array data repository Nucleic Acids Res 2002 30 1 207 10 10.1093/nar/30.1.207 11752295 5 Martin F Sewer A Talikka M Xiang Y Hoeng J Peitsch MC Quantification of biological network perturbations for mechanistic insight and diagnostics using two-layer causal models BMC Bioinformatics 2014 15 1 238 10.1186/1471-2105-15-238 25015298 6 Krämer A Green J Pollard J Tugendreich S Causal analysis approaches in ingenuity pathway analysis (ipa) Bioinformatics 2013 30 4 523 30 10.1093/bioinformatics/btt703 24336805 7 Wang P Lü J Yu X Identification of important nodes in directed biological networks: A network motif approach PLoS ONE 2014 9 8 106132 10.1371/journal.pone.0106132 8 Kitsak M Gallos LK Havlin S Liljeros F Muchnik L Stanley HE Makse HA Identification of influential spreaders in complex networks Nat Phys 2010 6 11 888 93 10.1038/nphys1746 9 Lin DC Xu L Chen Y Yan H Hazawa M Doan N Said JW Ding LW Liu LZ Yang H Genomic and functional analysis of the e3 ligase park2 in glioma Cancer Res 2015 75 9 1815 27 10.1158/0008-5472.CAN-14-1433 25877876 10 Oleaga C Welten S Belloc A Solé A Rodriguez L Mencia N Selga E Tapias A Noé V Ciudad CJ Identification of novel sp1 targets involved in proliferation and cancer by functional genomics Biochem Pharmacol 2012 84 12 1581 91 10.1016/j.bcp.2012.09.014 23018034 11 Khurana E Fu Y Chen J Gerstein M Interpretation of genomic variants using a unified biological network approach PLoS Comput Biol 2013 9 3 1002886 10.1371/journal.pcbi.1002886 12 Kanehisa M Goto S Sato Y Kawashima M Furumichi M Tanabe M Data, information, knowledge and principle: back to metabolism in kegg Nucleic Acids Res 2014 42 D1 199 205 10.1093/nar/gkt1076 13 Barrett T Wilhite SE Ledoux P Evangelista C Kim IF Tomashevsky M Marshall KA Phillippy KH Sherman PM Holko M Ncbi geo: archive for functional genomics data sets-update Nucleic Acids Res 2013 41 D1 991 5 10.1093/nar/gks1193 14 Irizarry RA Hobbs B Collin F Beazer-Barclay YD Antonellis KJ Scherf U Speed TP Exploration, normalization, and summaries of high density oligonucleotide array probe level data Biostatistics 2003 4 2 249 64 10.1093/biostatistics/4.2.249 12925520 15 Sabidussi G The centrality index of a graph Psychometrika 1966 31 4 581 603 10.1007/BF02289527 5232444 16 Bondy JA Murty USR Graph theory with applications 1976 London Macmillan 17 Wang J Duncan D Shi Z Zhang B Web-based gene set analysis toolkit (webgestalt): Update 2013 Nucleic Acids Res 2013 41 W1 77 83 10.1093/nar/gkt439 18 Bindea G Mlecnik B Hackl H Charoentong P Tosolini M Kirilovsky A Fridman WH Pagès F Trajanoski Z Galon J Cluego: a cytoscape plug-in to decipher functionally grouped gene ontology and pathway annotation networks Bioinformatics 2009 25 8 1091 3 10.1093/bioinformatics/btp101 19237447 19 Dijkstra EW A note on two problems in connexion with graphs Numer Math 1959 1 1 269 71 10.1007/BF01386390 20 Safran M Dalah I Alexander J Rosen N Stein TI Shmoish M Nativ N Bahir I Doniger T Krug H Genecards version 3: the human gene integrator Database 2010 2010 020 10.1093/database/baq020 21 Yuan W Chen Z Wu S Guo J Ge J Yang P Huang J Silencing of epha2 inhibits invasion of human gastric cancer sgc-7901 cells in vitro and in vivo Neoplasma 2011 59 1 105 13 10.4149/neo_2012_014 22103904 22 Salotti J Dias MH Koga MM Armelin HA Fibroblast growth factor 2 causes g2/m cell cycle arrest in ras-driven tumor cells through a src-dependent pathway PloS ONE 2013 8 8 72582 10.1371/journal.pone.0072582 23 Shah S King EM Chandrasekhar A Newton R Roles for the mitogen-activated protein kinase (mapk) phosphatase, dusp1, in feedback control of inflammatory gene expression and repression by dexamethasone J Biol Chem 2014 289 19 13667 79 10.1074/jbc.M113.540799 24692548 24 Lee MY Kim HJ Kim MA Jee HJ Kim AJ Bae YS Park JI Chung JH Yun J Nek6 is involved in g2/m phase cell cycle arrest through dna damage-induced phosphorylation Cell Cycle 2008 7 17 2705 9 10.4161/cc.7.17.6551 18728393

PMC005xxxxxx/PMC5001202.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556159279010.1186/s12864-016-2790-3ResearchStructure-guided selection of specificity determining positions in the human Kinome Moll Mark mmoll@rice.edu 1Finn Paul W. 2Kavraki Lydia E. 11 Department of Computer Science, Rice University, PO Box 1892, Houston, 77251 TX USA 2 University of Buckingham, Hunter St, Buckingham, UK 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.431© Moll et al. 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background The human kinome contains many important drug targets. It is well-known that inhibitors of protein kinases bind with very different selectivity profiles. This is also the case for inhibitors of many other protein families. The increased availability of protein 3D structures has provided much information on the structural variation within a given protein family. However, the relationship between structural variations and binding specificity is complex and incompletely understood. We have developed a structural bioinformatics approach which provides an analysis of key determinants of binding selectivity as a tool to enhance the rational design of drugs with a specific selectivity profile. Results We propose a greedy algorithm that computes a subset of residue positions in a multiple sequence alignment such that structural and chemical variation in those positions helps explain known binding affinities. By providing this information, the main purpose of the algorithm is to provide experimentalists with possible insights into how the selectivity profile of certain inhibitors is achieved, which is useful for lead optimization. In addition, the algorithm can also be used to predict binding affinities for structures whose affinity for a given inhibitor is unknown. The algorithm’s performance is demonstrated using an extensive dataset for the human kinome. Conclusion We show that the binding affinity of 38 different kinase inhibitors can be explained with consistently high precision and accuracy using the variation of at most six residue positions in the kinome binding site. We show for several inhibitors that we are able to identify residues that are known to be functionally important. Keywords Protein kinasesSpecificity determining positionsBinding affinityIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Predicting affinity profiles remains a challenging task for computational and medicinal chemists. This is particularly true of the kinase family of enzymes because of their large number and structural similarity. Despite their structural similarity, the kinases exhibit large phylogenetic diversity. As a result, binding site sequence dissimilarity alone cannot explain the differences in binding affinity [1]. Selectivity patterns obtained by experimental screening in enzyme assays are often difficult to rationalize in structural terms. Additional tools are needed to improve our capabilities to design inhibitors that selectively bind to only a small subset of the kinases. The rapidly increasing number of kinase structures has made it possible to study how structural differences affect binding affinity. For instance, different inhibitors have been designed to target the inactive, DFG-out conformation and active, DFG-in conformation [2–5]. In general, determining exactly how functional changes relate to structural ones remains an important open challenge [6, 7]. This is caused in part by the fact that not all structural changes cause a functional change. Additionally, the available structures are non-uniformly distributed over the known kinase sequences: for many kinases there is no structural information, while other kinases are overrepresented, which can lead to overfitting. In previous work [1], we introduced the Combinatorial Clustering Of Residue Position Subsets (CCORPS) method and demonstrated that it could be used to predict binding affinity of kinases. CCORPS considers structural and chemical variation among all triplets of binding site residues and identifies patterns that are predictive for some externally provided labeling. The labeling can correspond to, e.g., binding affinity, Enzyme Commission classification, or Gene Ontology terms, and only needs to be defined for some of the structures. CCORPS corrects for the non-uniform distribution of structures. From the patterns CCORPS identifies, multiple predictions are combined into a single consensus prediction by training a Support Vector Machine. A limitation of this work is that it is difficult to identify the most important Specificity Determining Positions (SDPs). In this paper, we are not trying to construct a better predictor, but, rather, a better explanation for some labeling. The explanation is better in the sense that it provides a simple explanation of a labeling in terms of the dominant SDPs. Rather than using all patterns discovered by CCORPS, it uses a small number of patterns that involve only a small number of residues yet is able to accurately recover binding affinity. The main contribution of this paper is an algorithm that computes the Specificity Determining Positions that best explain binding affinity in terms of structural and chemical variation. More generally, the algorithm can identify a sparse pattern of structural and chemical variation that corresponds to an externally provided labeling of structures. This work extends our prior work on CCORPS, but shifts the focus from optimal predictions to concise, biologically meaningful, explanations of functional variation. There has been much work on the identification and characterization of functional sites. Most of the techniques are broadly applicable to many protein families, but we will focus in particular on their application to kinases, when possible. Much of the work on computing SDPs is based on evolutionary conservation in multiple sequence alignments (see, e.g., [8–10]). There has also been work on relating mutations to an externally provided functional classification in a phylogeny-independent way [11, 12]. This work is similar in spirit to what CCORPS does, but based on sequence alone. While sequence alignment techniques can reveal functionally important residues in kinases [13], structural information can provide additional insights. This is especially true for large, phylogenetically diverse families such as the kinases. The FEATURE framework [14, 15] represents a radically different way of identifying functional sites. Instead of alignment, FEATURE builds up a statistical model of the spatial distribution of physicochemical features around a site. Another approach to modeling functional sites has been the comparison of binding site cavities [3, 16]. In [17] a functional classification of kinase binding sites is proposed based on a combination of geometric hashing and clustering. This approach is similar in spirit to our prior work [1], but our work considers variations in a small sets of binding site residues, which makes it possible to separate non-functional structural changes from functional ones. In [18] a method called FLORA is proposed for analysis of structural conservation across whole domains (rather than binding sites). FLORA was shown to be able to identify functional subfamilies (defined by Enzyme Commission classifications) within large protein superfamilies. It relies on the construction of structural feature vectors, which shares some similarities with our approach. However, FLORA is completely unsupervised and it is not clear how it could be extended to explain patterns of kinase binding affinity. In [19] many of the ideas above are combined into one framework. Given sequences from a PFAM alignment [20] and some reference structures, homology models are constructed for all sequences. Next, cavities are extracted, aligned, and clustered. Unlike our work, the approach in [19] is completely unsupervised and does not aim to provide an explanation for an externally provided classification (such as kinase binding affinity). Methods CCORPS overview Our algorithm builds on the existing CCORPS framework [1]. CCORPS is a semi-supervised technique that takes as input a set of partially labeled structures and produces as output the predicted labels for the unlabeled structures. Of course, this is only possible if the labels can be related to variations in the structures. In previous work [1] we have shown this to be the case for labelings based on binding affinity and functional categorization (Enzyme Commission classification). CCORPS [1] consists of several steps. First, a one-to-one correspondence needs to be established between relevant residues (e.g., binding site residues) among all structures. This correspondence can be computed using a multiple sequence alignment or using sequence independent methods [21–24]. Second, we consider the structural and physicochemical variation among all structures and all triplets of residues. The triplets are not necessarily consecutive in the protein sequence and can be anywhere in the binding site. Each triplet of residues constitutes a substructure: a spatial arrangement of residues. For each triplet, we compute a distance matrix of all pairwise distances between substructures. The distance measure used is a combination of structural distance and chemical dissimilarity introduced in [22]. In particular, the distance between any two substructures s1 and s2 is defined as: d(s1,s2)=dside chain centroid(s1,s2)+dsize(s1,s2)+daliphaticity(s1,s2)+daromaticity(s1,s2)+dhydrophobicity(s1,s2)+dhbond acceptor(s1,s2)+dhbond donor(s1,s2). The dside chain centroid(s1,s2) term is the least root-mean-square deviation of the pairwise-aligned side chain centroids of the substructures. The remaining terms account for differences in the amino acid properties between the substructures s1 and s2 as quantified by the pharmacophore feature dissimilarity matrix as defined in [22]. Each row in the distance matrix can be thought of as a “feature vector” that describes how a structure differs from all others with respect to a particular substructure. The n×n distance matrix for n structures is highly redundant and we have shown that the same information can be preserved in a 2-dimensional embedding computed using Principal Component Analysis [25]. Each 2D point is then a reduced feature vector. The set of n 2-dimensional points is clustered using Gaussian Mixture Models in order to identify patterns of structural variation. Not all structural variation is relevant; we focus on patterns of structural variation that align with the classification provided by the labeling. The final stage of CCORPS is the prediction of labels for the unlabeled structures. Suppose a cluster for one of the residue triplets contains structures with only one type of label as well as some unlabeled structures. This would suggest that the predicted label for the unlabeled structures should be the same as for the other cluster members. We call such a cluster a Highly Predictive Cluster (HPC). These HPC are a critical component of the algorithm presented in the next section. There are many clusterings and each clustering can contain several HPCs (or none at all). For example, in the human kinome the binding site consists of 27 residues, leading to 273=2,925 clusterings. Typically, an unlabeled structure belongs to several HPCs and we thus obtain multiple predictions. These predictions might not agree with each other. In our prior work we trained a Support Vector Machine [26] to obtain the best consensus prediction from the multiple predictions. Structure-guided selection of specificity determining positions While CCORPS has been demonstrated to make accurate predictions, it has been difficult to interpret the structural basis for these predictions. This has motivated us to look at alternative ways to interpret the clusterings produced by CCORPS. Rather than trying to build a better predictor, we have developed an algorithm that constructs a concise structural explanation of a labeling. It determines a set of Specificity Determining Positions (SDPs). An algorithm that would predict that almost every residue position is important would not be very helpful. We therefore wish to enforce a sparsity constraint: for a set of labeled structures S we want to find the smallest possible number of HPCs that cover the largest possible subset of S and involve at most λ residues. The problem of finding SDPs can be formulated as a variant of the set cover problem. The set cover problem is defined as follows: given a set S and subsets Si⊆S,i=1,…,n, what is the smallest number of subsets such that their union covers S? This is a well-known NP-Complete problem, but the greedy algorithm that iteratively selects the subset that expands coverage the most can efficiently find a solution with an approximation factor of ln|S|. As mentioned above, in our case, S is the set of labeled structures. We keep track of the residues involved in the selected HPCs and mark them as SDPs. Solving this as a set cover problem would likely still select most residues. The intuition for this can be understood as follows. The number of clusterings each residue is involved in is quadratic in the number of residues in the alignment. Each of those clusterings could contain a HPC that covers at least one structure that is not covered yet by other HPCs. Even in completely random data some patterns will appear, which could in turn be classified as HPCs. We measure sparsity of the cover in terms of the number of residues and not the number of HPCs, since this facilitates an easier interpretation of the results shown later on. As noted before, there can be several HPCs per clustering. This means that once we have selected an HPC, we might as well include all other HPCs from that same clustering (we have already “paid” for using the corresponding residues). As an algorithmic refinement, we may also wish to limit the degree at which we are fitting the data to avoid overfitting and get a simpler description of the most significant residues positions whose variation can be used to explain the labeling. The algorithm for computing SDPs is shown in Algorithm 1. It is similar to the greedy set cover algorithm. The input to the algorithm consists of a list of labeled structures, a list of all 3-residue subsets of the binding site, and a list of sets of structures that belong to HPCs. After initializing the set of SDPs and the set of selected subset indices in S, the main loop performs the following steps. First, the indices of all subsets are computed that will not grow the set of SDPs beyond a size limit λ (line 5). Second, the subset index is computed that will increase the cover of the known labels with HPC structures the most (line 9). Next, the algorithm checks whether the increase is “large enough,” i.e., greater than or equal to δ (line 11). If so, the set of SDPs and the sets of not-yet-covered structures are updated (line 13–14). If not, the algorithm terminates and returns the set of SDPs. The final output of Algorithm 1 provides a concise explanation of which structural and chemical variations correlate highly with a given labeling. In the context of the kinases, it can identify triplets of residues whose combined structural and chemical variation give rise to patterns that allow one to separate binding from non-binding kinases. As we will see in the next section, often only a very small set of residues is sufficient to obtain HPCs that cover most of the structures with known binding affinity. Results In [27] a quantitative analysis is presented of 317 different kinases and 38 kinase inhibitors. For every combination of a kinase and an inhibitor, the binding affinity was experimentally determined. This dataset also formed the basis for the evaluation of CCORPS [1]. The kinase inhibitors vary widely in their selectivity. Inhibitors like Staurosporine bind to almost every kinase, while others like Lapatnib bind to a very specific subtree in the human kinase dendrogram. The structure dataset was obtained by selecting all structures from the Pkinase and Pkinase_Tyr PFAM alignments [20]. The binding site, as defined in [1], consists of 27 residues. After filtering out structures that had gaps in the binding site alignment, 1,958 structures remained. The binding affinity values were divided into two categories (i.e., labels): “binds” and “does not bind.” This gives rise to two different types of HPCs: clusters predictive for binding (which we call true-HPCs below) and clusters predictive for not binding (which we call false-HPCs below). All other structures corresponding to kinases that were not part of the Karaman et al. study [27] do not have a label. CCORPS was run on this dataset, consisting of all 1,958 structures along with the binding affinity data. This resulted in 273=2,925 clusterings, one for every triplet of residues. The median number of true-HPCs per inhibitor was 591, while the median number of false-HPCs per inhibitor was 13,632. In the next subsection we look in detail at results of our algorithm with one parameter setting to get a sense of what kind of output is produced. In the subsequent subsection we will describe different ways to measure coverage of the SDPs as well as their predictive potential. We then evaluate these measures on all inhibitors with different parameter settings. Specificity-determining positions While in our prior work [1] the emphasis was on predicting the affinity of kinases, here we are focused on creating a concise explanation of the affinity. Thus, here we are not performing cross validation experiments. We have run Algorithm 1 on the kinome dataset with λ=6 residues and δ=16 (statistics for different values of λ and δ are reported in the next subsection). With λ=6, the algorithm can select at most two non-overlapping triplets. We computed the SDPs for all inhibitors (see Fig. 1). With some additional bookkeeping we can keep track of which residue was involved in which selected subsets. The bar chart for each inhibitor can be interpreted as follows. Along the x-axis is the residue position in the multiple sequence alignment of the 27 binding site residues. The relative height of each bar indicates how often a residue position was part of a selected 3-residue subset. Blocks with the same color correspond to residues belonging to the same residue subset. This can provide important contextual information. It shows not only which residues are important to help explain binding affinity, but also the context in which its variation should be seen. It could, e.g., indicate that one residue’s variation relative to some other residue(s) is important. The contextual residues themselves may not always vary much and are perhaps not of as much functional importance in the traditional sense. As λ is increased, more bars would be added to each profile as long as they improve coverage by at least δ structures. Similarly, as δ is decreased, more bars would be added to each profile as long as no more than λ residues are involved. Fig. 1 The SDP profiles computed for every inhibitor in the kinome dataset. The x-axis represents the residue position in the 27-residue multiple sequence alignment of the binding site. Each row shows the SDPs for one inhibitor whose name is shown on the y-axis. For each inhibitor, blocks with the same color correspond to one of the 3-residue subsets. If there are multiple colors in a given position, then the same residue was part of several selected subsets. This means that the same residue in different structural contexts can help explain the binding affinity of different kinases Figure 2 shows some examples of the clusterings that have been selected by Algorithm 1. These clusterings contain a large number of structures belonging to HPCs. The distance between points represents how different the corresponding structures are, structurally and chemically. The examples show that we can identify very strong spatial cohesion among the structures that bind when looking at the right residues (i.e., the SDPs). Not all clusterings selected by Algorithm 1 show such a strong relationship between structure and function. Especially for inhibitors that bind more broadly to kinases this relationship is harder to untangle. Fig. 2 Examples of the kind of clusterings selected by our algorithm. The axes correspond to the 2D, PCA-reduced feature vector representation of the pairwise distances between structures as described in the Methods section. Each point represents one structure. Red: known to bind, black: known to not bind, gray: binding affinity unknown. Discs: structures belonging to HPCs, circles: all other structures There is significant variation among the SDP profiles. For a very selective inhibitor like SB-431542 the variation of only three positions is sufficient to explain the binding affinity (see also the next subsection), while for ABT-869 many combinations of 3 residues out of the 6 selected residues seem to be helpful in explaining the binding affinity. Figure 3 shows a visualization of the SDPs for the inhibitor Imatinib. Figure 3a shows the structural variation (or lack thereof) in the selected residue positions for all structures that bind Imatinib. In contrast, if the same positions in all structures that do not bind Imatinib are superimposed, the structural variation is very high as is shown in Fig. 3b. Fig. 3 Structural visualization of SDPs. P38 (PDB ID 3HEC) is shown in ribbon representation along with the superimposed (a) SDPs for all the structures that bind imatinib and (b) the same residue positions for all structures that do not bind to imatinib Coverage and predictive power of SDPs Based on the set of SDPs we can (a) try to “recover” the labels of labeled structures that were not part of the selected HPCs and (b) predict labels for the unlabeled structures. There are at least four simple strategies to do this: We could assume that the union of all true-hpcs contains all the structures that bind and that all others do not bind. We could assume that the union of all false-hpcs contains all the structures that do not bind and all others do bind. We could omit the false-hpcs altogether from the input H to Algorithm 1 and select residue subsets based on large true-hpcs only. The labels are then recovered as in (1). We could omit the true-hpcs altogether from the input H to Algorithm 1 and select residue subsets based on large false-hpcs only. The labels are then recovered as in (2). Note that the SDPs computed with Algorithm 1 are the same in the first two strategies, but will generally look different when using strategies 3 and 4. We have evaluated each of these strategies on all 38 ligands. For each we can evaluate the coverage: the percentage of known labels that are included in the HPCs. We can also count the number of unlabeled structures included in HPCs, which can be interpreted as the number of new binding affinities we can predict. For the first two strategies we get predictions for both binding and not-binding, while for the latter two we only get predictions for one type of affinity. Finally, we can calculate the usual statistical performance measures (sensitivity, specificity, precision, and accuracy) to measure how well the selected HPCs can predict binding affinity for all labeled structures. The results were computed with λ=6 and δ=16 and are summarized in Table 1. Note that specificity is equal to 1 in strategies 1 and 3 by construction. Similarly, sensitivity is equal to 1 in strategies 2 and 4 by construction. In general, assuming that the union of all true-HPCs contains all the structures that bind (as is done in strategies 1 and 3) results in poor sensitivity. Strategy 2 seems to strike a good balance between sensitivity and specificity as well as between precision and accuracy. Strategy 4 performs even better than strategy 2, but provides poorer coverage. Table 1 Coverage of labeled structures, number of predicted affinities for unlabeled structures, as well as sensitivity, specificity, precision, and accuracy for HPC-based prediction of binding affinity Strategy Cov. #pred. Sens. Spec. Prec. Acc. 1 83 % 215 0.486 1.000 0.921 0.904 2 83 % 520 1.000 0.887 0.783 0.929 3 15 % 1,084 0.617 1.000 0.921 0.932 4 71 % 364 1.000 0.900 0.806 0.937 Each row summarizes the average over all 38 ligands for the corresponding strategy The results in Table 2 show more detailed results for each ligand with strategy 2. While there is some variation among the inhibitors, the coverage is almost always very high. In cases where it is not, such as AST-487, JNJ-7706621 and Sunitinib, it is usually a inhibitor that binds to many different parts of the kinome tree (see kinome interaction maps in [27]). Finally, we analyzed the sensitivity to the parameter δ and λ. As is shown in Tables 3 and 4, performance varies significantly with both λ and δ (as is expected). However, even with very large values of δ, the algorithm is still able to cover the vast majority of known binding affinities. Even more surprisingly, even when restricting SDPs to only λ=3 residues (corresponding to a single clustering), over 60 % of the structures with known binding affinity are covered. Table 2 Coverage of labeled structures, number of predicted affinities for unlabeled structures, as well as specificity, precision, and accuracy for HPC-based prediction of binding affinity as recovered from SDPs computed using our algorithm (with λ=6 and δ=16). Sensitivity is equal to 1 in all cases Inhibitor Cov. #pred. Spec. Prec. Acc. ABT-869 86 % 557 0.922 0.633 0.931 AMG-706 83 % 558 0.928 0.707 0.938 AST-487 65 % 426 0.661 0.806 0.859 AZD-1152HQPA 85 % 568 0.914 0.668 0.927 BIRB-796 67 % 391 0.766 0.653 0.838 BMS-387032/SNS-032 96 % 670 0.984 0.959 0.988 CHIR-258/TKI-258 81 % 420 0.947 0.861 0.960 CHIR-265/RAF265 87 % 473 0.960 0.801 0.966 CI-1033 77 % 475 0.882 0.710 0.909 CP-690550 96 % 629 0.989 0.736 0.989 CP-724714 99 % 684 0.999 0.982 0.999 Dasatinib 83 % 500 0.897 0.837 0.933 EKB-569 70 % 474 0.876 0.688 0.902 Erlotinib 80 % 532 0.902 0.693 0.920 Flavopiridol 80 % 515 0.844 0.754 0.895 GW-2580 99 % 677 1.000 1.000 1.000 GW-786034 79 % 485 0.920 0.737 0.934 Gefitinib 81 % 470 0.906 0.561 0.916 Imatinib 86 % 587 0.936 0.590 0.941 JNJ-7706621 59 % 356 0.580 0.704 0.790 LY-333531 83 % 413 0.912 0.652 0.924 Lapatinib 99 % 684 0.999 0.982 0.999 MLN-518 94 % 659 0.989 0.808 0.989 MLN-8054 87 % 493 0.948 0.766 0.956 PI-103 99 % 654 0.999 0.988 0.999 PKC-412 54 % 217 0.621 0.687 0.793 PTK-787 97 % 664 0.999 0.974 0.999 Roscovitine/CYC202 98 % 650 1.000 1.000 1.000 SB-202190 84 % 500 0.929 0.815 0.946 SB-203580 69 % 349 0.792 0.641 0.849 SB-431542 100 % 670 1.000 1.000 1.000 SU-14813 71 % 343 0.761 0.667 0.838 Sorafenib 70 % 509 0.919 0.801 0.939 Staurosporine 91 % 646 0.681 0.956 0.959 Sunitinib 61 % 343 0.652 0.654 0.790 VX-680/MK-0457 78 % 410 0.844 0.767 0.897 VX-745 85 % 583 0.912 0.680 0.926 ZD-6474 87 % 511 0.939 0.823 0.952 average 83 % 520 0.887 0.783 0.929 The last row lists the average performance over all inhibitors Table 3 Sensitivity to the value of λ with δ=16 λ Cov. #pred. Spec. Prec. Acc. 3 62 % 312 0.669 0.493 0.778 4 73 % 419 0.781 0.661 0.864 5 79 % 482 0.844 0.729 0.907 6 83 % 520 0.887 0.783 0.929 7 86 % 537 0.909 0.810 0.943 8 88 % 554 0.921 0.838 0.951 9 89 % 565 0.930 0.858 0.958 Each row represents an average over all 38 inhibitors Table 4 Sensitivity to the value of δ with λ=6 δ Cov. #pred. Spec. Prec. Acc. 1 85 % 587 0.904 0.820 0.941 2 85 % 580 0.903 0.817 0.940 4 85 % 565 0.900 0.812 0.938 8 84 % 547 0.895 0.800 0.935 16 83 % 520 0.887 0.783 0.929 32 81 % 490 0.871 0.723 0.916 64 78 % 456 0.848 0.658 0.898 128 74 % 413 0.817 0.612 0.876 Each row represents an average over all 38 inhibitors Discussion Frequency analysis of SDP positions The 27 residues that make up the binding site (see Fig. 4) are not equally represented in the SDP profiles. For example, position 2 does not occur in any of the SDP profiles, whereas position 12 occurs in 31 out of the 38 (see Fig. 5). The residues occurring most frequently in SDP profiles are often residues that have been observed to be important for inhibitor selectivity. Fig. 4 The kinase binding site. Selected residues of P38 α are shown in complex with imatinib (PDB ID 3HEC) Fig. 5 Frequency of each residue position occurring in SDPs across all inhibitors. The x-axis represents the residue position in the 27-residue multiple sequence alignment of the binding site SDP position 8, which occurs in 22 of the SDP profiles, corresponds to the well-known “gatekeeper” residue [28]. The size of this residue controls access to the hydrophobic binding pocket accessed by Type II inhibitors. Most kinase inhibitors are ATP-competitive and mimic to a greater or lesser extent the hydrogen bonding interactions that the adenine aromatic moiety of ATP makes with the hinge region of the protein. The hinge region corresponds to positions 9–11 and each of these positions occurs frequently in the SDP profiles, particularly at positions 9 and 10. Note that the interactions of inhibitors with the hinge are through hydrogen bonds to the protein backbone and are thus, in this sense, not sequence specific. Also, position 10 is rarely involved in hydrogen bonding because the canonical orientation of the backbone orients the NH and CO backbone groups away from the binding site. A recent analysis has shown that the potency of kinase inhibitors is not correlated with the number of hinge hydrogen bonds, but that there is a trend, albeit not pronounced, for compounds that make more hydrogen bonds to be less selective [29]. Large conformational changes that alter the canonical binding pattern have been observed when the conformationally less constrained glycine residue occurs in hinge positions. The SDP analysis indicates that subtler alterations in geometry and sequence in this region play an important role in modulating selectivity. It is not inevitable that frequently observed interactions automatically translate into modulators of binding profile. Position 20 of the SDP profile corresponds to a conserved glutamic acid residue in the middle of the C-helix that forms a salt bridge with a conserved lysine and is often involved in hydrogen bonds to amides or ureas of Type II kinase inhibitors. However, this position occurs in only one SDP (SB-431542). The most frequently selected position in SDP profiles is number 12, occurring in 31 out of the 38 profiles. This residue occurs in the “selectivity surface”, a relatively solvent exposed region with significant structural variation. For many inhibitors, this position contributes information from multiple 3-residue subsets, enabling the geometric and sequence variability of this region of the protein relative to the rest of the structure to be captured. Positions 16 and 17 correspond to the Asp and Phe residues of the DFG motif. This motif occurs in “DFG-in” or “DFG-out” conformations, with DFG-in being the active conformation of the enzyme and DFG-out a catalytically inactive form that is stabilized by Type-II inhibitors such as imatinib. Despite this geometrical variability, these positions occur rarely in SDP profiles. Only a small percentage of kinases have been observed in the DFG-out state crystallographically. Interestingly, the ability of a kinase to adopt this inactive conformation has been postulated to be controlled by two other residues, the gatekeeper and the residue immediately N-terminal to the DFG sequence [30]. This later residue is at position 15 and occurs in the SDPs with moderate frequency. The number of 3-position subsets that contribute to the SDP profile is related to inhibitor selectivity. The histograms in Fig. 6 show the number of contributing 3-position subsets (x-axis) plotted against the various selectivity metrics calculated by Karaman et al. [27] (y-axis). The selectivity values are the average of the compound values with SDP profiles derived from that number of 3-position subsets. Note that the selectivity value can be zero. For all metrics other than the Kd ratio measure, the most selective inhibitors have SDP profiles derived from one to three 3-position subsets. The pattern is similar whether the kinases are considered as a whole (S(3 μM), S(100nM)) or the tyrosine kinases (STK(3 μM), STK(100nM)) or serine/threonine kinases (SSTK(3 μM), SSTK(100nM)) are considered separately. A very similar result is obtained by calculating S(10 μM) from the Karaman et al. data [27], in order to match the activity cutoff threshold used in the CCORPS analysis (data not shown). The Kd ratio measure differs from the others by focusing on off-targets with affinity within 10-fold of the primary target. Such compounds are considered active by the 10 μM IC50 cutoff value used to generate the SDPs and thus the lack of correlation with the Kd ratio measure is expected. A similar trend is observed in specificity of the SDP profiles. In Table 5 we see that SDP profiles derived from a small number of 3-position subsets tend to a higher specificity. Fig. 6 Different measures of selectivity as a function of the number of 3-position subsets that contribute to the SDP profile Table 5 Average specificity over all inhibitors as a function of the number of 3-position subsets that determine the SDPs # 3-pos. subsets 1 2 3 4 5 6 7 Specificity 1.00 0.99 0.97 0.83 0.86 0.86 0.83 Frequency 1 6 4 10 7 6 4 The last row shows the number of inhibitors whose sdps are determined by a given number of 3-position subsets Comments on specific compounds CP-690550 (Tofacitinib) Tofacitinib is a clinically used selective Janus Kinase inhibitor. An SDP Word Logo is shown in Fig. 7a. Fig. 7 Sequence logos (created by WebLogo [38]) for the SDPs of structures known to bind to different inhibitors as well as a logo for all structures There are PDB structures for 5 kinases, for each of which tofacitinib is a potent inhibitor (JAK1, JAK3, JAK3, TYK2 and PKN1). In the X-ray structure 3lxk (JAK3) elements 9, 10 and 19 are close to the inhibitor, but elements 24–26 are distant. Figure 1 shows that this arises from two 3-position subsets (9, 10, 26 and 19, 24, 25) [This being the case, I’m not sure why there is variability at positions 24 and 25]. The tofacitinib complexes with JAK1, JAK2, JAK3 and TYK2 are very similar to each other. The structure 4oti is the PKN1-tofacitinib complex, for which tofacitinib is a medium potency inhibitor. Superposition of the ligand between 3lxk and 4oti shows an essentially identical conformation. This aligns the residues of the N-lobe quite well, but the C-lobe is displaced. The 3-position subsets that span the N- and C-lobes could capture this range of possibilities in HPCs and thus enable the binding to PKN1 to be accounted for. Positions 24–26 occur quite frequently as SDPs, even for inhibitors that are not in contact with these residues. The p38 α structure (e.g., 3hec) is not inhibited by tofacitinib. The superposition (based on the 27 alpha-carbon positions of the binding site residues used by CCORPS) shows a broadly similar disposition of the N- and C-lobes. In this case there are sequence differences at five of the six SDP positions. The CDK8 structure (3rdf) has more subtle differences that are hard to distinguish from the active examples based on visual inspection. Weigert et al. [31] generated resistance mutants to JAK2. Of the three mutants identified, one E864K (JAK2 numbering) is not within our 27 residue active site definition. However, Y931C (Position 10 in the logo) conferred resistance to all of the JAK inhibitors studied, including tofacitinib, in agreement with the SDP result. G935R (Position 12 in the logo) conferred resistance to all inhibitors except tofacitinib, also in agreement with the SDP. Lapatinib Lapatinib is a selective inhibitor of ErbB2 and EGFR. An SDP Word Logo is shown in Fig. 7b. The general pattern is fairly typical, with the gatekeeper, hinge and selectivity surface represented. Kancha et al. [32] reports several mutations observed in ERBb2 in various solid tumors. Most of these are distant from the binding site, but one T862A corresponds to position 15 in the Logo and is associated with modest lapatinib resistance. An analogous mutation is also found in EGFR. Trowe et al. [33] report that T798 is the most frequently mutated ErbB2 residue in an in vitro screen using a randomly mutagenized ErbB2 expression library and shows the greatest lapatinib resistance. This corresponds to position 8 in the logo (gatekeeper). A less frequently observed mutation L726 is not found in the logo (position 1). Other mutated residues are not in the binding site set. The gatekeeper residue is also mutated in EGFR, but other EGFR resistance-inducing mutations do not map to the corresponding logo positions. Imatinib (Gleevec) Imatinib is an Abl/Kit/VEGFR inhibitor. An SDP Word Logo is shown in Fig. 7c. The profile is similar to that of lapatinib to the extent that gatekeeper, hinge and selectivity surface residues are represented. Mutation at positions 8 or 10 is a common cause of imatinib resistance. Note that the presence of the gatekeeper in the profile of a Type II inhibitor is not unexpected, but that not all Type-II logos have this. As noted above, Type II inhibitors such as imatinib bind to a DFG-out enzyme conformation, but these residues are not in the profile and thus do not provide the strongest selectivity signal. Position 19, which is in the hydrophobic pocket, is also of interest. Mutation at this position in BCR/ABL has been reported to confer moderate Imatinib resistance [34]. This position was also the most frequently mutated residue found in imatinib-resistant KIT mutants from analysis of tumor samples obtained from patients enrolled in a Phase II clinical study of imatinib [35]. The gatekeeper residue was also frequently mutated in this population. Sunitinib (Sutent™) is approved for the treatment of advanced GIST after failure of imatinib due to resistance or intolerance. It is effective against the imatinib-resistant V654A (position 19) mutant, a position which does not occur in the sunitinib SDPs. If false-HPCs are omitted (i.e., strategy 3 in the subsection Coverage and Predictive Power ofSDPs), the SDPs also include position 24. This position is frequently mutated in resistant tumors, with positions 10 and 24 together accounting for 14 % of BCR/ABL mutations. The SDPs of the more selective KIT/VEGFR inhibitor PTK-787 also includes position 24. The occurrence of other positions in the imatinib logo is harder to rationalize. In the structures, the side chain at position 4 points away from the inhibitor and is not in direct contact with it. This may point to an indirect role in modulating the conformation of the protein in this region. Position 4 is actually selected quite frequently (9 times). As part of the hydrophobic core of the N-lobe, it may act as a marker for the relative disposition of the two domains of the enzyme. Differential flexibility of the kinases is often discussed in the literature as playing a role in selectivity, see for example [36]. Conclusion We have described a general method for identifying Specificity Determining Positions in families of related proteins. The method was shown to be very effective in identifying SDPs within the human kinome that help explain the binding affinity of 38 different inhibitors. Consistent with prior studies, we were able to identify the gatekeeper residue and the hinge region as generally very important for the binding specificity of kinases. It has also highlighted the selectivity surface as a region that is key in determining selectivity profiles. An in-depth analysis of the SDPs for three specific kinase inhibitors provides further evidence that we can identify other residues that are known to be important in each case, including positions that are mutated in drug-resistant tumors. Of particularly interest are these that are not in direct contact with the inhibitor (some examples of which were discussed above) but which may be involved indirectly through, for example, influencing the conformation or flexibility of the protein. This would be a significant benefit, as such residues are difficult to identify by other means. Not only could this potentially provide a new insight into the structural biology of kinases, but such knowledge may be helpful in the design of inhibitors with novel, or improved, selectivity profiles. In this regard, it would be interesting to explore expanding the approach to include additional, non-binding site residues, that have been implicated in resistance through modulation of conformational plasticity and investigated by molecular dynamics. In prior work [37] we have demonstrated that the addition of homology models leads to an improvement in the prediction of binding affinity. Homology models can fill in gaps in structural coverage, thereby potentially eliminating “accidental” HPCs and create new ones. In future work we plan to investigate whether homology models can provide similar benefits in the identifications of SDPs. Abbreviations CCORPS, combinatorial clustering of residue position subsets; HPC, highly-predictive cluster; SDP, specificity-determining position. From IEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA. 9–12 November 2015 Declarations This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4; Funding Work on this paper by Mark Moll and Lydia E. Kavraki has been supported in part by NSF ABI 0960612, NSF CCF 1423304, and Rice University Funds. Authors’ contributions MM, PWF, and LEK conceptualized and conceived the project and its components. MM carried out the experiments. MM, PWF, and LEK analyzed the data. MM and PWF wrote the manuscript and LEK edited it. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. The authors wish to thank Drew Bryant. Without his work on creating the ccorps software infrastructure and preparing the human kinase dataset for processing with ccorps the results presented here would not have been possible. ==== Refs References 1 Bryant DH Moll M Finn PW Kavraki LE Combinatorial clustering of residue position subsets predicts inhibitor affinity across the human kinome PLOS Comput Biol 2013 9 6 1003087 10.1371/journal.pcbi.1003087 2 Liu Y Gray NS Rational design of inhibitors that bind to inactive kinase conformations Nat Chem Biol 2006 2 7 358 64 10.1038/nchembio799 16783341 3 Kuhn D Weskamp N Hüllermeier E Klebe G Functional classification of protein kinase binding sites using Cavbase ChemMedChem 2007 2 10 1432 47 10.1002/cmdc.200700075 17694525 4 Bikker JA Brooijmans N Wissner A Mansour TS Kinase domain mutations in cancer: implications for small molecule drug design strategies J Med Chem 2009 52 6 1493 509 10.1021/jm8010542 19239229 5 Milletti F Hermann JC Targeted kinase selectivity from kinase profiling data ACS Med Chem Lett 2012 3 5 383 6 10.1021/ml300012r 24900482 6 Gani OA, Thakkar B, Narayanan D, Alam KA, Kyomuhendo P, Rothweiler U, Tello-Franco V, Engh RA. Assessing protein kinase target similarity: Comparing sequence, structure, and cheminformatics approaches. Biochim Biophys Acta. 2015. doi:10.1016/j.bbapap.2015.05.004. 7 Finn PW Kavraki LE Computational approaches to drug design Algorithmica 1999 25 2 347 71 10.1007/PL00008282 8 Kalinina OV Mironov AA Gelfand MS Rakhmaninova AB Automated selection of positions determining functional specificity of proteins by comparative analysis of orthologous groups in protein families Protein Sci 2004 13 2 443 56 10.1110/ps.03191704 14739328 9 Chakrabarti S Lanczycki CJ Analysis and prediction of functionally important sites in proteins Protein Sci 2007 16 1 4 10.1110/ps.062506407 17192586 10 Capra JA Singh M Characterization and prediction of residues determining protein functional specificity Bioinformatics 2008 24 13 1473 80 10.1093/bioinformatics/btn214 18450811 11 Pazos F Rausell A Valencia A Phylogeny-independent detection of functional residues Bioinformatics 2006 22 12 1440 8 10.1093/bioinformatics/btl104 16551661 12 Rausell A Juan D Pazos F Valencia A Protein interactions and ligand binding: from protein subfamilies to functional specificity Proc Natl Acad Sci U S A 2010 107 5 1995 2000 10.1073/pnas.0908044107 20133844 13 Mok J Kim PM Lam HYK Piccirillo S Zhou X Jeschke GR Sheridan DL Parker SA Desai V Jwa M Cameroni E Niu H Good M Remenyi A Ma J-LN Sheu YJ Sassi HE Sopko R Chan CSM De Virgilio C Hollingsworth NM Lim WA Stern DF Stillman B Andrews BJ Gerstein MB Snyder M Turk BE Deciphering protein kinase specificity through large-scale analysis of yeast phosphorylation site motifs Sci Signal 2010 3 109 12 10.1126/scisignal.2000482 14 Halperin I Glazer DS Wu S Altman RB The FEATURE framework for protein function annotation: modeling new functions, improving performance, and extending to novel applications BMC Genomics 2008 9 Suppl 2 2 10.1186/1471-2164-9-S2-S2 18173843 15 Liu T Altman RB Using multiple microenvironments to find similar ligand-binding sites: application to kinase inhibitor binding PLoS Comput Biol 2011 7 12 1002326 10.1371/journal.pcbi.1002326 16 Chen BY Honig B VASP: a volumetric analysis of surface properties yields insights into protein-ligand binding specificity PLoS Comput Biol 2010 6 8 1000881 10.1371/journal.pcbi.1000881 17 Kinnings SL Jackson RM Binding site similarity analysis for the functional classification of the protein kinase family J Chem Inf Model 2009 49 2 318 29 10.1021/ci800289y 19434833 18 Redfern OC Dessailly BH Dallman TJ Sillitoe I Orengo CA FLORA: a novel method to predict protein function from structure in diverse superfamilies PLoS Comput Biol 2009 5 8 1000485 10.1371/journal.pcbi.1000485 19 de Melo-Minardi RC Bastard K Artiguenave F Identification of subfamily-specific sites based on active sites modeling and clustering Bioinformatics 2010 26 24 3075 082 10.1093/bioinformatics/btq595 20980272 20 Finn RD Tate J Mistry J Coggill PC Sammut SJ Hotz HR Ceric G Forslund K Eddy SR Sonnhammer ELL Bateman A The Pfam protein families database Nucleic Acids Res 2008 36 Database issue 281 8 21 Menke M Berger B Cowen L Matt: local flexibility aids protein multiple structure alignment PLoS Comput Biol 2008 4 1 10 10.1371/journal.pcbi.0040010 22 Schalon C Surgand JS Kellenberger E Rognan D A simple and fuzzy method to align and compare druggable ligand-binding sites Proteins 2008 71 4 1755 1778 10.1002/prot.21858 18175308 23 Xie L Bourne PE Detecting evolutionary relationships across existing fold space, using sequence order-independent profile-profile alignments Proc Natl Acad Sci U S A 2008 105 14 5441 446 10.1073/pnas.0704422105 18385384 24 Moll M Bryant DH Kavraki LE The LabelHash algorithm for substructure matching BMC Bioinforma 2010 11 555 1 15 25 Jolliffe IT Principal Components Analysis 1986 New York Springer 26 Steinwart I Christmann A Support Vector Machines 2008 Verlag New York Springer 27 Karaman MW Herrgard S Treiber DK Gallant P Atteridge CE Campbell BT Chan KW Ciceri P Davis MI Edeen PT Faraoni R Floyd M Hunt JP Lockhart DJ Milanov ZV Morrison MJ Pallares G Patel HK Pritchard S Wodicka LM Zarrinkar PP A quantitative analysis of kinase inhibitor selectivity Nat Biotechnol 2008 26 1 127 32 10.1038/nbt1358 18183025 28 Huang D Zhou T Lafleur K Nevado C Caflisch A Kinase selectivity potential for inhibitors targeting the ATP binding site: a network analysis Bioinformatics 2010 26 2 198 204 10.1093/bioinformatics/btp650 19942586 29 Xing L Klug-Mcleod J Rai B Lunney EA Kinase hinge binding scaffolds and their hydrogen bond patterns Bioorg Med Chem 2015 23 19 6520 527 10.1016/j.bmc.2015.08.006 26358279 30 Hari SB Merritt EA Maly DJ Sequence determinants of a specific inactive protein kinase conformation Chem Biol 2013 20 6 806 15 10.1016/j.chembiol.2013.05.005 23790491 31 Weigert O Lane AA Bird L Kopp N Chapuy B van Bodegom D Toms AV Marubayashi S Christie AL McKeown M Paranal RM Bradner JE Yoda A Gaul C Vangrevelinghe E Romanet V Murakami M Tiedt R Ebel N Evrot E De Pover A Régnier CH Erdmann D Hofmann F Eck MJ Sallan SE Levine RL Kung AL Baffert F Radimerski T Weinstock DM Genetic resistance to jak2 enzymatic inhibitors is overcome by hsp90 inhibition J Exp Med 2012 209 2 259 73 10.1084/jem.20111694 22271575 32 Kancha RK von Bubnoff N Bartosch N Peschel C Engh RA Duyster J Differential sensitivity of ERBB2 kinase domain mutations towards lapatinib PLoS ONE 2011 6 10 26760 10.1371/journal.pone.0026760 33 Trowe T Boukouvala S Calkins K Cutler Jr RE Fong R Funke R Gendreau SB Kim YD Miller N Woolfrey JR Vysotskaia V Yang JP Gerritsen ME Matthews DJ Lamb P Heuer TS EXEL-7647 inhibits mutant forms of ErbB2 associated with lapatinib resistance and neoplastic transformation Clin Cancer Res 2008 14 8 2465 75 10.1158/1078-0432.CCR-07-4367 18413839 34 Redaelli S Piazza R Rostagno R Magistroni V Perini P Marega M Gambacorti-Passerini C Boschelli F Activity of bosutinib, dasatinib, and nilotinib against 18 imatinib-resistant BCR/ABL mutants J Clin Oncol 2009 27 3 469 71 10.1200/JCO.2008.19.8853 19075254 35 Heinrich MC Corless CL Blanke CD Demetri GD Joensuu H Roberts PJ Eisenberg BL von Mehren M Fletcher CDM Sandau K McDougall K Ou W-b Chen CJ Fletcher JA Molecular correlates of imatinib resistance in gastrointestinal stromal tumors J Clin Oncol 2006 24 29 4764 74 10.1200/JCO.2006.06.2265 16954519 36 Blanc J Geney R Menet C Type II kinase inhibitors: an opportunity in cancer for rational design Anticancer Agents Med Chem 2013 13 5 731 47 10.2174/1871520611313050008 23094911 37 Chyan J, Moll M, Kavraki LE. Improving the prediction of kinase binding affinity using homology models. In: Computational Structural Bioinformatics Workshop at the ACM Conf. on Bioinf., Comp. Bio. and Biomedical Informatics. Washington: 2013. p. 741–8, doi:10.1145/2506583.2506704. 38 Crooks GE Hon G Chandonia JM Brenner SE WebLogo: a sequence logo generator Genome Res 2004 14 6 1188 1190 10.1101/gr.849004 15173120

PMC005xxxxxx/PMC5001203.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556923279410.1186/s12864-016-2794-zResearchMBDDiff: an R package designed specifically for processing MBDcap-seq datasets Liu Yuanhang 12Wilson Desiree 2Leach Robin J. 23Chen Yidong Cheny8@uthscsa.edu 141 Greehey Children’s Cancer Research Institute, University of Texas Health Science Center at San Antonio, San Antonio, TX USA 2 Department of Cellular and Structure Biology, University of Texas Health Science Center at San Antonio, San Antonio, TX USA 3 Department of Urology, University of Texas Health Science Center at San Antonio, San Antonio, TX USA 4 Department of Epidemiology & Biostatistics, University of Texas Health Science Center at San Antonio, San Antonio, TX USA 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.432© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Since its initial discovery in 1975, DNA methylation has been intensively studied and shown to be involved in various biological processes, such as development, aging and tumor progression. Many experimental techniques have been developed to measure the level of DNA methylation. Methyl-CpG binding domain-based capture followed by high-throughput sequencing (MBDCap-seq) is a widely used method for characterizing DNA methylation patterns in a genome-wide manner. However, current methods for processing MBDCap-seq datasets does not take into account of the region-specific genomic characteristics that might have an impact on the measurements of the amount of methylated DNA (signal) and background fluctuation (noise). Thus, specific software needs to be developed for MBDCap-seq experiments. Results A new differential methylation quantification algorithm for MBDCap-seq, MBDDiff, was implemented. To evaluate the performance of the MBDDiff algorithm, a set of simulated signal based on negative binomial and Poisson distribution with parameters estimated from real MBDCap-seq datasets accompanied with different background noises were generated, and then performed against a set of commonly used algorithms for MBDCap-seq data analysis in terms of area under the ROC curve (AUC), number of false discoveries and statistical power. In addition, we also demonstrated the effective of MBDDiff algorithm to a set of in-house prostate cancer samples, endometrial cancer samples published earlier, and a set of public-domain triple negative breast cancer samples to identify potential factors that contribute to cancer development and recurrence. Conclusions In this paper we developed an algorithm, MBDDiff, designed specifically for datasets derived from MBDCap-seq. MBDDiff contains three modules: quality assessment of datasets and quantification of DNA methylation; determination of differential methylation of promoter regions; and visualization functionalities. Simulation results suggest that MBDDiff performs better compared to MEDIPS and DESeq in terms of AUC and the number of false discoveries at different levels of background noise. MBDDiff outperforms MEDIPS with increased backgrounds noise, but comparable performance when noise level is lower. By applying MBDDiff to several MBDCap-seq datasets, we were able to identify potential targets that contribute to the corresponding biological processes. Taken together, MBDDiff provides user an accurate differential methylation analysis for data generated by MBDCap-seq, especially under noisy conditions. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2794-z) contains supplementary material, which is available to authorized users. Keywords MBDCap-seqDNA methylationDifferentially methylationDifferential methylated regionsMBDDiffXBSeqIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background The human genome is composed of billions of heritable non-static epigenetic arrangement of histone and DNA sequence that controls how genes are expressed [1] DNA methylation, along with some other covalent modifications of histone or DNA sequences, have regulatory control over gene expression. In 1975, two key publications have suggested that methylation of cytosine residues in the context of CpG dinucleotide could be an epigenetic marker of DNA sequences [2, 3]. A large majority of CpG islands of the vertebrate genome reside in or near the promoter regions [4]. Since then, DNA methylation has been intensively studied and, specifically, DNA methylation in promoter regions has been shown to be associated with cell development, tumor progression, and aging [5–7] Through years of efforts, many experimental methods have been developed to assay the methylation status of CpG in a genome-wide manner. Currently, the golden standard of genome-wide profiling of DNA methylation is the whole genome bisulphite sequencing, which involves treatment of sodium bisulphite followed by high throughput sequencing (BS-seq) [8]. However, there are two major disadvantages for BS-seq. Firstly it requires genome-wide deep sequencing in order to precisely identify modification at base-pair level, which currently is not cost effective. Secondly, datasets generated by BS-seq might also give rise to alignment difficulties due to C/T modification. On the other hand, affinity-based method, such as methylated DNA immunoprecipitation followed by high throughput sequencing (MeDIP-seq) [9] and methyl-CpG binding domain-based capture followed by high throughput sequencing (MBDCap-seq) [8], are established as alternatives to BS-seq for genome-wide DNA methylation profiling which are more cost effective. It has been shown that MeDIP-seq is more sensitive to highly methylated, high-CpG densities regions and MBDCap-seq is more sensitive to highly methylated, moderate-CpG densities [10]. MBDCap-seq approach uses methyl-CpG binding domain of the MBD2 protein to capture double-stranded DNA, combined with subsequent high throughput sequencing, to systematically identify methylated regions in the genome. There are some unique characteristics of MBDCap-seq. DNA methylation profiling by MBDCap-seq is biased by underlying CpG properties of the genome, more precisely,, methylated regions with high GC contents are more likely to be eluted than regions with low GC contents. In terms of quantification of DNA methylation levels and the determination of differential methylation (DM), current methods used for testing of differential methylated regions (DMRs) generally do not take the sample specific background noise during MBD capture, which is caused by non-specific pull-down behavior of methyl-CpG binding domain, into consideration. Last but not least, up till now, there is no software that is designed specifically for processing large MBDCap-seq datasets. Previously, we developed an algorithm called BIMMER for testing genome-wide differential methylation, where we constructed a two-layer hidden Markov model (HMM) to model the differential methylation status [10]. However, because of the complexity of the algorithm and the nature of Expectation-Maximization (EM) solution, BIMMER is relatively slow in speed and is not suited for analyzing MBDCap datasets in large scale. Our aim for the study presented here is to provide an efficient computational pipeline specifically designed for identification of DM genes by using MBDCap-seq protocol. Methods Genome-wide MBDCap sequencing for prostate cancer patient samples MBDCap-seq protocol was carried out to identify methylated regions across the genome for a set of prostate cancer samples as listed in Table 1. Total of 6 primary prostate tumors derived from patients that have different clinical outcome (3 Metastasized (METs) and 3 No Evidence of Disease (NEDs)) were processed and sequenced. Methylated fragments, bound to a methyl-CpG binding domain protein, were eluted for sequencing with the Illumina HiSeq 2000 sequencer with 50 bp single read (SR) sequencing module. Approximately 295 million sequence reads were generated and around 78 % reads were mapped to unique genome locations for all 6 samples. The MBDCap-seq analysis pipeline is:Table 1 6 prostate tumor samples profiled with MBDCap Sample ID Outcome Gleason Reads (millions) Mapped reads (millions) 1 MET 7 51.57 40 2 MET 7 50.43 40.54 3 MET 10 53.43 41.83 4 NED 6 45.71 34.63 5 NED 7 47.05 27.03 6 NED 46.49 44.85 Apply FastQC to short read sequences to examine sequencing QC and other characteristics of sequence reads. Extract the file fastqc_data.txt for GC enrichment analysis; Perform BWA aligner to align sequence reads to UCSC human genome build hg19 [11]; Remove sequence reads with equal and more than 2 bp mis-match and non-uniquely mapped to the genome; Sort, convert and index BAM file for each sample; Count number of reads in 100 bp bins tiling through entire genome by using BedCoverage/BedTools [12]; Count number of reads within 4kbp regions (+/- 2kbp around transcription start sites (TSS)) of each gene by using BedCoverage tool. Determination of GC enrichment Normal human genome has GC content percentage roughly around 40 % (see UCSC genome statistics at http://genome.ucsc.edu/goldenPath/stats.html#hg18). The FastQC algorithm generates a GC counting statistic from all reads in fastqc_data.txt. If a sample contains portion of DNA that are enriched in GC content, we expect to see a shifted distribution, as illustrated in the Results section Fig. 4, to the right side of the normal genomic DNA GC distribution. Thus, by assuming a mixture model of 2 Gaussian distributions (gray-dashed line in Fig. 4), or G = p1N(μ1, σ1) + p2N(μ2, σ2), where p1 + p2 = 1, we can determine GC enrichment score,Fig. 1 Workflow for MBDDiff and simulation procedure Fig. 2 ROC curves of MBDDiff, MEDIPS and DESeq for simulated datasets in different scenarios. ROC curves for simulated MBDcap-seq datasets with low, intermediate or high level of background noise with 3 number of replicates in each group, 10 % of DM promoters with 2 fold of difference (a); Different levels of background noise but only for highly methylated promoters above 75 % quantile of methylation levels (b); Different levels of background noise but only for lowly methylated promoters below 25 % quantile of methylation levels (c); ROC curves for simulated MBDcap-seq datasets with low, intermediate or high level of background noise with 6 number of replicates in each group, 10 % of DM promoters with 2 fold of difference (d); Simulation was carried out 100 times and the average results is used Fig. 3 False discovery curves of MBDDiff, MEDIPS and DESeq for simulated datasets in different scenarios. False discovery curves for simulated MBDcap-seq datasets with low, intermediate or high level of background noise with 3 number of replicates in each group, 10 % of DM promoters with 2 fold of difference (a); Different levels of background noise but only for highly methylated promoters above 75 % quantile of methylation levels (b); Different levels of background noise but only for lowly methylated promoters below 25 % quantile of methylation levels (c); False discovery curves for simulated MBDcap-seq datasets with low, intermediate or high level of background noise with 6 number of replicates in each group, 10 % of DM promoters with 2 fold of difference (d); Simulation was carried out 100 times and the average results is used Fig. 4 Statistical power of MBDDiff, MEDIPS and DESeq for simulated datasets in different scenarios. Bar plot of statistical power for simulated MBDcap-seq datasets with low, intermediate or high level of background noise with 3 number of replicates in each group, 10 % of DM promoters with 2 fold of difference (a); Different levels of background noise but only for highly methylated promoters above 75 % quantile of methylation levels (b); Different levels of background noise but only for lowly methylated promoters below 25 % quantile of methylation levels (c); Bar plot of statistical power for simulated MBDcap-seq datasets with low, intermediate or high level of background noise with 6 number of replicates in each group, 10 % of DM promoters with 2 fold of difference (d); Simulation was carried out 100 times and the average results is used 1 ES=p2μ2−μ1/p1×20 We initialized the model fitting with following parameters (for human genome), μ1 = 40, μ2 = 60, and p1 = p2 = 0.5. For our default setting (50 % reads are GC enriched), ES = 1.0. If only 20 % reads are enriched for GC content (p2/p1 = 20/80 = 1/4), we will have ES = 0.25, assuming other parameters stay the same. We select samples with ES > 0.2, otherwise, samples will be discarded without further data analysis. Construction of reference regions to measure background noise Gene annotation (refFlat table) of human genome hg19 build was downloaded using UCSC table browser (http://genome.ucsc.edu/). Promoter regions were defined as regions ranging from 2 Kb upstream of TSS to 2Kb downstream of TSS. We only kept one TSS for transcripts with same TSS. In total, we constructed annotation for 33,178 promoters. To identify regions that potentially contribute to background noise, we built a 100 bp tiling window across the whole genome. We then used BEDTools to count the number of mapped reads within each tiling window of the 6 prostate cancer samples. In order to identify the regions for measuring background noise, we applied following procedures:Filtering step: We exclude any 100 bp genome-wide tiling windows that reside in promoter regions, predicted CpG islands regions and any windows that contain ambiguous bases (gaps); Construction Step: We then select preliminary background regions based on GC content as follows: for each promoter region, identify 80 100 bp windows nearby that has low in GC content (<40 %) and also relatively proximal to the corresponding TSS; and Finalize background regions based on average transcript per million (TPM): for each promoter region, choose 40 out of 80 100 bp windows that are relatively low in TPM [13] as defined as TPM=rg×rl×106flg×∑Grg×rlflg, where rg is the number of reads mapped to each 100 bp window, rl is the read length, flg is fragment length, in our case, 100. Statistical framework and differential methylation testing of MBDDiff algorithm The read count for promoter region of gene i can be decomposed into two components, true signal Si, which is directly associated with real methylation level, and background noise Bi, which is attributed mainly to the random pull-down events from the wet-lab procedure of MBDCap-seq. Previously, we have developed an algorithm, XBSeq [14], for testing for differential expression for RNA-seq experiments. Here we applied similar statistical framework for MBDCap-seq experiments. Simply speaking, we assumed that the true signal (what we would like to estimate) Si possesses a negative binomial distribution and background noise Bi follows a Poisson distribution. Then the observed signal (what we typically measured) Xi is a convolution of Si and Bi, which is governed by a Delaporte distribution [15]. Xi=Si+Bi 2 Si~NBripi Bi~Poissonλi We further assumed that background noise Bi and true signal Si are independent. By default, we applied a non-parametric method for parameter estimation. The Poisson parameter λι for Bi can be estimated as: 3 λi=1m∑j=1mbij where bij denotes estimated background noise for promoter i across m replicates for each condition. Following the estimation of Poisson parameter, we will be able to infer mean μSi and standard deviation σSi for each promoter region: 4 μSi=ESi=EXi−EBi 5 σSi2=σXi2+σBi2−2ρσXiσBi, Then the parameters for negative binomial distribution can be estimated by 6 ri=μSi2/σSi'2−μSi 7 pi=μSi/σSi'2 where σSi' denotes adjusted variance for Si. This has proven to be useful when the sample size is small [16]. Details regarding non-parametric parameter estimation can be found in our previous publication of XBSeq [14]. When sample size is relatively large (>5), the maximum likelihood estimation (MLE) is applied to estimate parameters. The likelihood function is given by 8 Lθi=∏j=1mpXij|αi,βi,λi⋅∏j=1mpBij|λi=∏j=1m∑k=0XijΓαi+kβikλiXij−ke−λiΓαik!1+βiαi+kXij−k!⋅∏j=1mλiBije−λiBij! which has no closed form. Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm is used to estimate the parameters by iterative updating. αi and βi are parameters for gamma portion of Delaporte distribution which are related to negative binomial parameters by: 9 ri=αi 10 pi=1/βi+1 After successful estimation of all parameters, differential methylation testing of each promoter between two groups (with read count x and y) will be carried out by using moderated Fisher’s exact test: 11 p=∑pab≤pxypab∑allpab where a and b are constrained by a + b = x + y Simulation In order to evaluate the performance of our method, we generated a set of simulated datasets where we can control the differential methylation status of each promoter region. In this study, true signal S was simulated from a negative binomial distribution and background noise B was simulated from a Poisson distribution with parameters estimated from real MBDCap-seq datasets. We compared MBDDiff with MEDIPS [17], an R package designed for general purpose DNA fragments enrichment experiments, such as MeDIP-seq and MBDCap-seq for their ability to detect differntially methylated regions. We also compared MBDDiff with DESeq [16], an algorithm originally designed for differential expression analysis where the background noise is not considered for testing of expression difference between two conditions. We choose DESeq algorithm due to the fact that it has similar signal distribution assumption (Negative binomial) and differential test statistic (Fisher’s Exact test). We followed a similar simulation procedure described in XBSeq. Basically, to estimate model parameters from a given MBDCap-seq dataset, 5000 promoters were randomly selected with replacement after discarding promoters with relatively low mapped reads or larger dispersion (top 10 %). The true signal S was simulated from a negative binomial distribution based on the mean and variance estimated from the 5000 promoters. 10 % or 30 % of the promoters were selected to be differentially methylated with enrichment fold change either 2 or 3. We simulated experiments with either 3 or 6 replicate samples per group to examine the potential effect of the number of replicates. To simulate background noise B, we first simulated read counts for the selected 100 bp windows. Then for each promoter region, background noise B was the summation of the read counts from all the corresponding 100 bp windows. We generated background noise in three different scenarios, with different level of dispersion, to examine the performance of our method in normal and noisy conditions. Background noise with different dispersion levels were simulated for each 100 bp window from a hybrid model: 12 Binc~M*Normμσ where μ is from a Poisson distribution μ ~ Poisson(λ + NF). In our simulation, we set M = 10, σ = 3. The noise factor NF can be chosen from 0, 7, 20, each represents experiments with low background noise, intermediate background noise and high background noise. Simulations were repeated 100 times and statistical metrics were evaluated based on the average performance. We evaluated different algorithms (MBDDiff, MEDIPS and DESeq) for their ability to discriminate between differentially methylated and non-differentially methylated promoters in terms of the following metrics: area under the ROC curve, number of false discoveries, statistical power, false discovery rate at pre-selected p value cutoff, distribution of p values under null model where there are no differentially methylated promoters. We also examined the performance of different methods separately for lowly and highly methylated promoters to see whether the performance is affected by methylated level of the promoter. In order to investigate performance of different methods when the underlying model assumption can not be met. We simulated true signal from normal distribution with parameters estimated from a real MBDcap-seq dataset and background noise from either normal or uniform distribution to see whether the performance of different methods is affected or not. For instance, to simulate background noise from normal distribution. We also applied equation (12). The difference is that, the parameter μ is from a normal distribution with parameters estimated from a real MBDcap-seq dataset. Additional DNA methylation datasets for testing In addition to our in-house prostate cancer samples, we also applied our method MBDDiff to previously published MBDCap-seq datasets where we compared endometrial cancer patient with either recurrent or non-recurrent outcome. 3 patients in each group were selected and processed with different methods to identify potential factors that contribute to endometrial cancer recurrence. Details about the experiments can be found in GSE26592 [18]. Similarly, we also selected and processed three tumor and normal samples from public domain dataset GSE58020, where MBDCap-seq were carried out to investigate DNA methylation profile for tripe negative breast cancers (TNBCs) [19]. Comparison with other software for MBDCap-seq datasets We also compared our algorithm with some other methods for MBDCap-seq datasets, including MEDIPS (1.20.0), DESeq (1.20.0). All these evaluations were carried out under R version 3.2.0 and Bioconductor version 3.1. Details regarding simulation procedure and workflow of MBDDiff is illustrated in Fig. 1. Results Implementation of MBDDiff In order to use MBDDiff, we need to construct species specific background annotations that will be used to measure background noise for MBDCap-seq datasets. Several background annotation files have already been constructed and can be downloaded from github page (https://github.com/Liuy12/MBDDiff_files). Please follow instructions from MBDDiff package if you want to construct background annotation for your organism of choice. Currently MBDDiff uses BEDTools to count reads mapped to promoter and background regions. So users need to have BEDTools installed on their computer. After counting reads mapped to promoter and background regions, MBDDiff then applies quality control procedure to examine the quality of MBDCap-seq datasets. Differential methylation testing will be carried out by XBSeq algorithm. In addition, MBDDiff also provides a set of visualizing tools for MBDCap data analysis. Discrimination between DM and non-DM promoters In order to compare MBDDiff, MEDIPS and DESeq for processing MBDCap-seq datasets, we generated synthetic datasets when the differential methylation status and fold change of each promoter region can be controlled. We followed the simulation procedure described in the Methods section. Briefly, methylation levels from 5000 promoter regions and their corresponding background noise were simulated with model parameters estimated from a given prostate cancer MBDCap-seq dataset. Among 5000 promoter region simulated, 10 % or 30 % promoters were designated for enrichment fold-change at 2 or 3. Different background noise levels were simulated and all statistical metric calculation were reported as an average over 100 repeats. We first compared the three methods for their ability to discriminate between differentially methylated and non-differentially methylated promoters in terms of the area under the Receiver Operating Characteristic (ROC) curve (AUC). As shown in Fig. 2 and Additional file 1: Table S1 and S2, 6 sample per group generally performed better than 3 samples per group in terms of AUC under various conditions. Overall, MBDDiff performs better with a higher AUC compared to MEDIPS and DESeq with different levels of background noise. For instance, under the condition of 3 samples per group, 10 % of differentially methylated promoters and 2-fold difference between two groups, MBDDiff achieved AUC 0.84 when background noise is relatively low, while the AUCs for MEDIPS and DESeq are 0.81 and 0.80 respectively. Even though all three methods have decreased AUC when background noise is increased (MBDDiff drop 0.04 to 0.80, MEDIPS dropped 0.1 to 0.71, and DESeq also dropped 0.1 to 0.70), MBDDiff is relatively resistant to higher background noise and performs much better even when the background noise is at very high level. We then examined the performance of the three methods separately for highly methylated (>75 %) and weakly methylated (<25 %) promoters. MBDDiff performs only slightly better compared to MEDIPS and DESeq for highly methylated promoters (Fig. 2b and Additional file 1: Figure S1a). However, for weakly methylated promoters, MBDDiff performs much better than MEDIPS and DESeq which indicates that background noise estimation procedure is essential for accurate DM detection under weakly methylation condition (Fig. 2c and Additional file 1: Figure S1b). Control of false discoveries We then compared the performance of MBDDiff, MEDIPS and DESeq in terms of the false discoveries encountered among the top ranked differential methylated promoters based on p value. Overall, MBDDiff picked up the least number of false discoveries in various conditions (Fig. 3 and Additional file 1: Table S1 and S2). For example, under the condition of 3 samples per group, 10 % of differentially methylated promoters and 2-fold difference between two groups, MBDDiff identified 248 out of 500 number of false discoveries compared to MEDIPS (271) and DESeq (279) when the background noise has relatively low dispersion. When the background noise is increased, MBDDiff also picked up increased number of false discoveries (add 31 to 279), but is relatively resistant to the increase of background noise compared to MEDIPS (add 69 to 340) and DESeq (add 64 to 344). We also took a similar approach to examine the three methods separately for highly methylated (>75 %) and weakly methylated (<25 %) groups. The three methods picked up similar number of false discoveries for highly methylated promoters (Fig. 3b and Additional file 1: Figure S1c). However, for weakly methylated promoters, MBDDiff again performs the best with lowest number of false discoveries encountered (Fig. 3c and Additional file 1: Figure S1d). We also examined false discovery rate at pre-selected p value cutoff (p value = 0.05, Additional file 1: Figure S2c). MBDDiff has relatively low false discovery rate (around 0.4) compare to MEDIPS (around 0.5) and DESeq (around 0.5). Overall, MBDDiff is more robust against false discoveries compared to MEDIPS and DESeq especially for weakly methylated promoters even when the background noise is relatively high. Statistical power Finally, we compared MBDDiff, MEDIPS and DESeq in terms of the statistical power at selected cutoff (p value = 0.05). As shown in Fig. 4 and Additional file 1: Table S1&2, MEDIPS performs the best when the background noise is relatively low. However, when we increased the dispersion of background noise, MBDDiff became the best method with the largest statistical power. For instance, under the condition of 3 samples per group, 10 % of differentially methylated promoters and 2-fold difference between two groups, MBDDiff achieved statistical power of 0.41 compared to MEDIPS (0.45) and DESeq (0.34) when the background noise is relatively low. However, when the background noise is higher, MBDDiff performs better with statistical power of 0.34 compared to MEDIPS (0.27) and DESeq (0.20). Similarly, we also compared statistical power of the three algorithms separately for highly methylated (>75 %) and weakly methylated (<25 %) promoters. For highly methylated group, MEDIPS performs slightly better than MBDDiff as shown in Fig. 4b and Additional file 1: Figure S1e. In contrast, for weakly methylated groups, MBDDiff performs much better than MEDIPS when the background noise is relatively high (Fig. 1f and 4c). Overall, MBDDiff remains one of the best method in terms of statistical power in various conditions and is more robust against the increase of background noise. Apply MBDDiff to prostate cancer datasets We recently carried out MBDCap-seq to examine whole genome DNA methylation profile in prostate tumors in order to identify potential factors that are involved in the development of prostate cancer after treatment. We excluded one patient sample because of a relative different Gleason score from other 5 samples. Then we applied MBDDiff to 5 samples in order to identify differentially methylated promoters between MET group and NED group. As shown in Fig. 6(a), there is a clear enrichment of GC content for highly methylated regions, which indicates the effectiveness in MBD2 capture procedure for our prostate samples. The enrichment of GC content was further assessed in Fig. 5, where 2 samples (one for MET and one for NED) were selected and their GC enrichment scores are 0.25 an 0.33, respectively. Figure 6(b) showed the distribution of background noise and promoter regions. We also examined the relationship between tumor samples. As we can observe from Fig. 6(c), patients with NED outcome are more dispersed than patients with MET outcome which might suggests a common mechanism for prostate cancer metastasis. After performing differential methylation algorithm, we identified 57 differentially methylated promoters with absolute log2 fold change greater than 1, p-value smaller than 0.001 and averaged methylation levels greater than 15 (read count unit). ANO7, a gene hyper-methylated in prostate cancer identified by our study, has been reported that it may act as a target gene for antibody-based immunotherapy [20]. F13A1, which is also identified as differentially methylated, is associated with bone metastasis in prostate cancer [21]. To conclude, by using MBDDiff, we were able to identify 57 promoters that were differentially methylated between MET and NED. This set of genes might be of helpful for future studies of prostate cancer metastasis.Fig. 5 GC enrichment assessment. We perform GC enrichment test for all samples processed with MBDCap-seq protocol. a Enrichment score (ES) of 0.25 were detected for a MET sample. The dash lines are two Gaussian mixture models (light dash line for normal human genome GC distribution estimated from 50 bp short reads, and dark dash line for enriched 50 bp reads). ES score is evaluated by using Eq. 1. b ES of 0.33 for a NED sample. Both samples pass the threshold of > 0.20 requirement. In both (a) and (b), blue-line is the empirical density from the actual data, and red-line is the mixture model density. Both figure showed a tight estimation Fig. 6 Apply MBDDiff to prostate cancer datasets. (a) Density plot of GC content grouped by different levels of methylation for all 100 bp windows across the whole genome of one example prostate MET sample; (b) Distribution of background noise and promoter counts; (c) 3D PCA plot of sample relation. C indicates NED patients and T indicates MET patients; (d) Heatmap of selected promoters for 5 prostate cancer samples Apply MBDDiff to breast cancer datasets To demonstrate the functionalities of MBDDiff, we also applied MBDDiff to a public dataset derived from TNBCs (GSE58020). We selected and processed 3 samples in both tumor and normal conditions. Firstly, MBDDiff assessed the quality of the samples by examining the GC enrichment for all 100 bp windows with different levels of methylation. As shown in Fig. 7(a), there is a clear enrichment of GC content for regions with higher mapped reads compared to ones with lower mapped reads which potentially contribute to background noise. After quality assessment, MBDDiff continued to estimate the context-specific background noise for each sample based on the preliminary background annotation. As shown in Fig. 7(b), background noise clearly coincides with the left hump of promoter mapped reads, indicating that the background noise we estimated indeed reflects the nature of mixture model for our observation. MBDDiff also provides some visualization functionalities including 3D principal component analysis (PCA) to examine sample relationship based on DNA methylation levels of promoter regions, more versatile heatmap visualization of DNA methylation levels across samples, etc. Alternatively, you can also generate an html report that contains several dynamic visualizations. Details can be found at our github page: https://github.com/Liuy12/MBDDiff.Fig. 7 Functionalities for MBDDiff for processing TBNC datasets. a Density plot of GC content grouped by different levels of methylation for all 100 bp windows across the whole genome of one example TNBC sample; (b) Distribution of background noise and promoter counts; (c) 3D PCA plot of sample relation, T indicates tumor samples, N indicates normal samples; (d) Heatmap of selected promoters for 6 TNBC samples As illustrated in Fig. 7(c), there is a clear separation between TNBC tumor and normal samples. Figure 7(d) showed heatmap of methylation levels of selected promoters across all samples (around 1800). Finally, MBDDiff performed differential methylation statistical tests based on XBSeq. Very interestingly, the dispersion for promoters showed a ‘S’ shape curve caused by promoters with either very small or very large dispersions, which probably is due to the artifacts of methylation measure from the MBDCap-seq protocol (Additional file 1: Figure S3A). Differential methylated promoters are identified by selecting promoters with absolute log2 fold-change greater than 1, adjusted p-value less than 0.01 and averaged methylation levels greater than 30. Additional file 1: Figure S3B showed the MA plot after differential methylation tests. With this stringent selection criterion, 348 unique promoters with significant differential methylation were obtained (significantly more than what we obtained in prostate cancer data, since here we compared tumor vs normal samples, while for prostate application, we studied tumors with NED outcome vs tumors with MET outcome). Interestingly, by using DAVID [22], we found that the genes of these promoters are enriched in biological processes, including regulation of transcription (adjusted p-value < 0.01) and regulation of neuron differentiation (adjusted p-value < 0.01), which might indicate the involvement of neuronal stem cell regulators in TNBCs [23]. Apply MBDDiff to endometrial cancer datasets Finally, we applied MBDDiff to our previously published dataset where MBDCap-seq procedure was carried out to examine the global methylation patterns for a total of 232 primary samples in endometrial cohorts, breast cancer cohorts and breast cancer cell lines (GSE26592). For the purpose of testing our algorithm, we randomly selected three samples from endometrial cancer patients with either recurrent or non-recurrent outcome to identify potential factors that contribute to recurrence of endometrial cancer. After alignment procedure with bwa, the six samples were processed with MBDDiff. As shown in Fig. 8a, there is a clear enrichment of GC content for regions with relatively high methylation levels, which indicates that the dataset generated is of good quality. Figure 8b shows the distribution of background noise and true signal. Then MBDDiff assesses sample relationship based on methylation patterns of promoter regions. As shown in Fig. 8c, there is a clear separation between patients with recurrent and non-recurrent outcome. Figure 8d shows heatmap of promoter regions that are differentially methylated between the two groups. Finally, differential methylation test was carried out to identify potential factors that might cause recurrence of endometrial cancer. Totally we identified 66 differentially methylated promoters with fold change larger than 2, adjusted p value smaller than 0.1 and average mean methylation level bigger than 9. Compared to original paper we identified much less number of differentially methylated regions, the reason might be: 1) We only focus on differential methylation testing for promoter regions; 2) We only used 3 patient samples in each group. Among all the differential methylated genes, DKK1 has been shown to be a novel biomarker for endometrial carcinoma.Fig. 8 Apply MBDDiff to endometrial cancer datasets. a Density plot of GC content grouped by different levels of methylation for all 100 bp windows across the whole genome of one example endometrial sample with recurrence outcome; (b) Distribution of background noise and promoter counts for one example endometrial sample with recurrence outcome; (c) 3D PCA plot of sample relation, R indicates samples with recurrence outcome, NR indicates samples with non-recurrence outcome; (d) Heatmap of selected promoters for 6 samples Discussion In order to compare MBDDiff, MEDIPS and DESeq, we carried out simulation procedure to generate datasets with different levels of background noise. Then we compared the three methods in terms of various statistical metrics. As we showed in the Results section, MBDDiff generally outperformed MEDIPS and DESeq in terms of AUC (Fig. 2) and number of false discoveries (Fig. 3). In terms of statistical power, MEDIPS performed slightly better than the other two methods when the background noise is relatively low. However, with higher background noise, MBDDiff has a better statistical power compared to MEDIPS and DESeq. Taken together, MBDDiff is more robust against higher background noise for accurately identifying differentially methylated promoters. As designed in the simulation procedure, true signal was simulated from negative binomial distribution and background noise from Poisson distribution. While these are commonly consented models for data generated from NGS protocols, real data may deviate from NB and/or Poisson distributions. To investigate whether there is a potential bias with regards to the models we used for simulating true signal as well as background noise, we tested performance of the three methods with different models for true signal and background noise for simulation. Firstly, we simulated the true signal from a normal distribution and background noise from Poisson distribution (at normal level as estimated from the one selected prostate cancer dataset). As shown in Additional file 1: Figure S2a, irrespective of which models we used for simulation, MBDDiff performed better then the other two methods in terms of AUC (Additional file 1: Table S3). All these three methods tend to perform better when negative binomial is used for simulating true signal as we expected: the simulated data derived from models match the underlying models in these algorithms. Then we also tested the effect of using different models for simulating background noise (Additional file 1: Figure S2b). We used negative binomial model to simulate true signal but coupled with background noise derived from normal, uniform, or Poisson distribution. As shown in Fig S2b and Additional file 1: Table S4, MBDDiff outperformed MEDIPS and DESeq, regardless of which noise model you use for simulating background noise. In summary, MBDDiff is robust under different signal and background noise models. While for MEDIPS and DESeq, they work better when uniform and Poisson distributions (they virtually overlapping in Additional file 1: Figure S2b) rather than normal distribution for background noise. Potential bias can also be observed by examining the distribution of p values under a null hypothesis when no promoters are differentially methylated. As shown in Additional file 1: Figure S2d, p values generated by all three algorithms roughly follow the uniform distribution as we expected, with higher variation generated from MEDIPS. We also benchmarked the three algorithms for experiments with different numbers of replicates. As shown in Additional file 1: Figure S4, DESeq consumes the most amount of time, followed by MBDDiff and MEDIPS. Higher computing cost for DESeq and MBDDiff with more replicate samples in each condition is mostly attributed to Fisher’s Exact test for assessing statistical significance. Conclusions We developed an R package, MBDDiff, which aims specifically to process MBDCap-seq datasets. MBDDiff provide users the ability to assess quality of datasets, test for differential methylation of promoter regions and visualization functionalities. Abbreviations AUC, area under the curve; BS-seq, bisulphite sequencing; DM, differential methylation; MBDCap-seq, methyl-CpG binding domain-based capture followed by high throughput sequencing; MeDIP-seq, methylated DNA immunoprecipitation followed by high throughput sequencing; MET, metastasis samples; NB, negative binomial (distribution, model); NED, no evidence of disease samples; ROC, receiver operating characteristic. Additional file Additional file 1: Supplementary information. Supplementary figures and tables to provide additional analysis results. (PDF 1214 kb) Acknowledgements This research was supported in part by the Genome Sequencing Facility of the Greehey Children’s Cancer Research Institute, UTHSCSA, which provided MBDCap-seq service. Fundings for this research were provided partially by the National Institutes of Health Cancer Center Shared Resources (NIH-NCI P30CA54174) to YC and NIGMS (R01GM113245) to YC and YL. Declaration The publication costs for this article were funded by the aforementioned NIGMS grants to YC. This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4. Authors’ contributions All authors contribute to the manuscript. YL, DW, RL and YC conceived and designed the study. YL and DW contributed to the initial design of the algorithm. YL implemented the algorithm carried out the simulation procedure. DW and RL contributed MBDCap-seq datasets for prostate samples. All authors read and approved the final manuscript. Competing interests The authors declare that they no competing interests. ==== Refs References 1. Inbar-Feigenberg M Choufani S Butcher DT Roifman M Weksberg R Basic concepts of epigenetics Fertil Steril 2013 99 3 607 15 10.1016/j.fertnstert.2013.01.117 23357459 2. Holliday R Pugh JE DNA modification mechanisms and gene activity during development Science 1975 187 4173 226 32 10.1126/science.1111098 1111098 3. Riggs AD X inactivation, differentiation, and DNA methylation Cytogenet Cell Genet 1975 14 1 9 25 10.1159/000130315 1093816 4. Deaton AM Bird A CpG islands and the regulation of transcription Genes Dev 2011 25 10 1010 22 10.1101/gad.2037511 21576262 5. Heyn H Esteller M DNA methylation profiling in the clinic: applications and challenges Nat Rev Genet 2012 13 10 679 92 10.1038/nrg3270 22945394 6. Gentilini D Mari D Castaldi D Remondini D Ogliari G Ostan R Bucci L Sirchia SM Tabano S Cavagnini F Role of epigenetics in human aging and longevity: genome-wide DNA methylation profile in centenarians and centenarians’ offspring Age 2013 35 5 1961 73 10.1007/s11357-012-9463-1 22923132 7. Robinson MD Stirzaker C Statham AL Coolen MW Song JZ Nair SS Strbenac D Speed TP Clark SJ Evaluation of affinity-based genome-wide DNA methylation data: effects of CpG density, amplification bias, and copy number variation Genome Res 2010 20 12 1719 29 10.1101/gr.110601.110 21045081 8. Taiwo O Wilson GA Morris T Seisenberger S Reik W Pearce D Beck S Butcher LM Methylome analysis using MeDIP-seq with low DNA concentrations Nat Protoc 2012 7 4 617 36 10.1038/nprot.2012.012 22402632 9. Li N Ye M Li Y Yan Z Butcher LM Sun J Han X Chen Q Zhang X Wang J Whole genome DNA methylation analysis based on high throughput sequencing technology Methods 2010 52 3 203 12 10.1016/j.ymeth.2010.04.009 20430099 10. Mao Z Ma C Huang TH Chen Y Huang Y BIMMER: a novel algorithm for detecting differential DNA methylation regions from MBDCap-seq data BMC Bioinformatics 2014 15 Suppl 12 S6 10.1186/1471-2105-15-S12-S6 25474268 11. Li H Durbin R Fast and accurate short read alignment with Burrows-Wheeler transform Bioinformatics 2009 25 14 1754 60 10.1093/bioinformatics/btp324 19451168 12. Quinlan AR BEDTools: the swiss-army tool for genome feature analysis Curr Protoc Bioinformatics 2014 47 11 12 11 34 25199790 13. Baccarelli A Rienstra M Benjamin EJ Cardiovascular epigenetics: basic concepts and results from animal and human studies Circ Cardiovasc Genet 2010 3 6 567 73 10.1161/CIRCGENETICS.110.958744 21156932 14. Chen HI Liu Y Zou Y Lai Z Sarkar D Huang Y Chen Y Differential expression analysis of RNA sequencing data by incorporating non-exonic mapped reads BMC Genomics 2015 16 Suppl 7 S14 10.1186/1471-2164-16-S7-S14 26099631 15. Johnson NL Kemp AW Kotz S Univariate discrete distributions 2005 3 Hoboken Wiley 16. Anders S Huber W Differential expression analysis for sequence count data Genome Biol 2010 11 10 R106 10.1186/gb-2010-11-10-r106 20979621 17. Lienhard M Grimm C Morkel M Herwig R Chavez L MEDIPS: genome-wide differential coverage analysis of sequencing data derived from DNA enrichment experiments Bioinformatics 2014 30 2 284 6 10.1093/bioinformatics/btt650 24227674 18. Gu F Doderer MS Huang YW Roa JC Goodfellow PJ Kizer EL Huang TH Chen Y CMS: a web-based system for visualization and analysis of genome-wide methylation data of human cancers PLoS One 2013 8 4 e60980 10.1371/journal.pone.0060980 23630576 19. Stirzaker C Zotenko E Song JZ Qu W Nair SS Locke WJ Stone A Armstong NJ Robinson MD Dobrovic A Methylome sequencing in triple-negative breast cancer reveals distinct methylation clusters with prognostic value Nat Commun 2015 6 5899 10.1038/ncomms6899 25641231 20. Bera TK Das S Maeda H Beers R Wolfgang CD Kumar V Hahn Y Lee B Pastan I NGEP, a gene encoding a membrane protein detected only in prostate cancer and normal prostate Proc Natl Acad Sci U S A 2004 101 9 3059 64 10.1073/pnas.0308746101 14981236 21. Morrissey C True LD Roudier MP Coleman IM Hawley S Nelson PS Coleman R Wang YC Corey E Lange PH Differential expression of angiogenesis associated genes in prostate cancer bone, liver and lymph node metastases Clin Exp Metastasis 2008 25 4 377 88 10.1007/s10585-007-9116-4 17972146 22. da Huang W Sherman BT Lempicki RA Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources Nat Protoc 2009 4 1 44 57 10.1038/nprot.2008.211 19131956 23. Cimino-Mathews A Subhawong AP Elwood H Warzecha HN Sharma R Park BH Taube JM Illei PB Argani P Neural crest transcription factor Sox10 is preferentially expressed in triple-negative and metaplastic breast carcinomas Hum Pathol 2013 44 6 959 65 10.1016/j.humpath.2012.09.005 23260325

PMC005xxxxxx/PMC5001204.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557147290910.1186/s12864-016-2909-6ResearchIdentification of rifampin-regulated functional modules and related microRNAs in human hepatocytes based on the protein interaction network Li Jin lijin@hrbeu.edu.cn 1Wang Ying wangying0129@126.com 12Wang Lei wl@hrbeu.edu.cn 1Dai Xuefeng daixuefeng203@163.com 2Cong Wang cwcy9156@163.com 1Feng Weixing fengweixing@hrbeu.edu.cn 1Xu Chengzhen xu5115@163.com 1Deng Yulin deng@bit.edu.cn 1Wang Yue yuewang@iu.edu 3Skaar Todd C. tskaar@iu.edu 4Liang Hong lh@hrbeu.edu.cn 15Liu Yunlong yunliu@iupui.edu 1351 College of Automation, Harbin Engineering University, 145 Nantong Street, Nangang District, Harbin, Heilongjiang 150001 China 2 Network Information Center, Qiqihar University, No.42, Wenhua Street, Qiqihar, Heilongjiang 161006 China 3 Department of Medical and Molecular Genetics, Indiana University School of Medicine, Indianapolis, IN USA 4 Division of Clinical Pharmacology, Department of Medicine, Indiana University School of Medicine, Indianapolis, IN USA 5 Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indianapolis, IN USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.517© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background In combination with gene expression profiles, the protein interaction network (PIN) constructs a dynamic network that includes multiple functional modules. Previous studies have demonstrated that rifampin can influence drug metabolism by regulating drug-metabolizing enzymes, transporters, and microRNAs (miRNAs). Rifampin induces gene expression, at least in part, by activating the pregnane X receptor (PXR), which induces gene expression; however, the impact of rifampin on global gene regulation has not been examined under the molecular network frameworks. Methods In this study, we extracted rifampin-induced significant differentially expressed genes (SDG) based on the gene expression profile. By integrating the SDG and human protein interaction network (HPIN), we constructed the rifampin-regulated protein interaction network (RrPIN). Based on gene expression measurements, we extracted a subnetwork that showed enriched changes in molecular activity. Using the Kyoto Encyclopedia of Genes and Genomes (KEGG), we identified the crucial rifampin-regulated biological pathways and associated genes. In addition, genes targeted by miRNAs that were significantly differentially expressed in the miRNA expression profile were extracted based on the miRNA-gene prediction tools. The miRNA-regulated PIN was further constructed using associated genes and miRNAs. For each miRNA, we further evaluated the potential impact by the gene interaction network using pathway analysis. Results and Disccussion We extracted the functional modules, which included 84 genes and 89 interactions, from the RrPIN, and identified 19 key rifampin-response genes that are associated with seven function pathways that include drug response and metabolism, and cancer pathways; many of the pathways were supported by previous studies. In addition, we identified that a set of 6 genes (CAV1, CREBBP, SMAD3, TRAF2, KBKG, and THBS1) functioning as gene hubs in the subnetworks that are regulated by rifampin. It is also suggested that 12 differentially expressed miRNAs were associated with 6 biological pathways. Conclusions Our results suggest that rifampin contributes to changes in the expression of genes by regulating key molecules in the protein interaction networks. This study offers valuable insights into rifampin-induced biological mechanisms at the level of miRNAs, genes and proteins. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2909-6) contains supplementary material, which is available to authorized users. The International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Protein-protein interactions are intrinsic to most biological processes [1]. Expanded knowledge of the protein interaction network (PIN) may shed light on basic cellular mechanisms. An expression profile is a dynamic collection of data used to deduce a gene’s function, state, environment, etc. With the increasing availability of genome and proteome data, the PIN can be integrated with gene expression profiles to create conditional network modules within a specific biological state. This method has been used to explore cellular mechanisms associated with multiple diseases [2], including cancer. For instance, Zhang et al. [3] analysed the genes and crucial modules associated with coronary artery diseases (CAD), and suggested that two proteins were critical for the development of CAD. Lin et al. [2] studied dynamic functional modules and co-expressed protein interaction networks in cases of dilated cardiomyopathy. Previous studies suggest that the integrated analysis of PIN and gene expression profile information may contribute to the identification of the functional modules and key genes that are relevant to important biological pathways. Rifampin is a drug that is usually used to treat tuberculosis and inactive meningitis [4]. The molecular mechanisms and functions of rifampin-regulation have previously been identified. Our previous study has confirmed that rifampin altered expression level of miRNAs and many cytochrome P450 enzymes (CYPs), which are the major metabolic enzymes that control the metabolism of most clinically important drugs, and some of the changes exist in associated relationships that suggest that some of CYP mRNAs are targeted by some miRNAs [5–8]. Rifampin is also a typical ligand of the pregnane X receptor (PXR), which is a transcription factor and a key regulator of the CYPs and other genes involved in drug disposition [9, 10]. Furthermore, rifampin can rapidly downregulate hepatic angiogenesis- and mitogenesis-related genes. Therefore, it shows favorable antiproliferative effects on endothelial cell, which is make it potentially beneficial for targeting hepatobiliary cancer cells [11, 12]. Previous studies have demonstrated that the drug-metabolizing enzymes [6], transporters, and microRNAs (miRNAs) are regulated by rifampin [11, 12], and the mechanisms of the regulation of some of these genes are well-studied; however, little has been done to put the global gene expression effects of rifampin into biological pathways and interactive networks. Protein interaction network can depict and integrate information pertaining to domain architecture, post-translational modification, interaction networks and disease association for each protein in the human proteome [13]. Furthermore, by combining with mRNA expression profiles, they can be used to identify specific correlations of between the genes, and to identify the key genes and functional modules associated with critical biological pathways. In addition, the integration of the miRNA expression profiles can depict relationship between the altered expression of miRNAs and their targeted-mRNA. The implementation of an integrative method that incorporates protein interaction networks and gene expression profiles to reveal conditional network modules associated with the rifampin-regulated biological processes becomes increasingly important in clarifying the regulatory mechanisms responsible for the rifampin effects on gene expression. In this study, we focused on identifying the key genes, miRNAs, and the regulatory relationships among them. We also explored the rifampin-induced biological pathways by integrating the protein interaction networks and the miRNA and mRNA expression profiles. In this study, we propose a method which can be used to identify the rifampin-regulated functional modules in the protein interaction network of human hepatocytes, and can also be used to further analyze the rifampin-induced miRNAs and their functions. A schematic of the overall method is illustrated in Fig. 1. In this model, the gene expression profile and PIN are integrated to construct the rifampin-regulated protein interaction network (RrPIN). Then, in order to analyse the crucial biological pathways, we identify the functional modules that participate in a common biological function within the protein-protein interaction network. Next, the functional modules are extracted using BioNet and jActiveModules, and the rifampin-induced significant differentially expressed key genes are identified based on an analysis of Gene Ontology (GO) and the Kyoto Encyclopedia of Genes and Genomes (KEGG). Finally, the miRNA-regulated PINs are established using these key genes and gene-targeted miRNAs based on the miRNA expression profile and miRNA-target prediction databases, and the functions of the miRNAs are revealed based on GO and KEGG. The proposed analysis enables us to uncover rifampin-induced biological mechanisms in human hepatocytes.Fig. 1 The hierarchical chart of identification of rifampin-regulated functional modules and related miRNAs Methods Data The gene expression dataset and miRNA expression dataset were performed as our previous study in Ramamoorthy et al. [8]. In the current study, the miRNA and mRNA expression profiles were obtained from primary human hepatocyte cultures (obtained from CellzDirect) from 7 donors, each treated with rifampin or vehicle for a total of 14 datasets. Cultures from each subject were treated as biologic replicates (n = 7). The hepatocytes were treated with rifampin or vehicle for 24 h and the total RNAs were isolated using a miRNeasy kit. The mRNA expression profile included 12,780 genes. The miRNA expression profile, which included 334 miRNAs, was measured using the Taqman OpenArray Human miRNA Panel using a NT Cycler. The mRNAs expression was measured using a standard method including EZBead preparation, Next-Gene sequencing, read quality assessment, sequence alignment, and RNA-Seq differential expression analysis. Construction of RrPIN The PPI data was downloaded from the Human Protein Reference Database (HPRD) [13], which contains experimentally validated interactions within the human proteome. The human liver protein interaction network (HLPN) [14] contains proteome-scale protein interaction maps of the human liver. It is comprised of 3484 interactions among 2582 proteins and provides substantial new insights into systems biology, disease research, and drug discovery. To construct the human protein interaction network (HPIN), all the proteins and non-overlapped interactions in the HPLN and HPRD were merged as the nodes and interactions of HPIN. To construct the rifampin-regulated gene network, we integrated the gene expression profile and HPIN as follows: the SDGs which were included in the HPIN were used as RrPIN’s nodes, and the interactions of RrPIN’s nodes in the HPIN were used as the RrPIN’s interactions. Cytoscape version 3.0.2 software (http://chianti.ucsd.edu/cytoscape-3.2.0/) [15] was used to generate the network. Identification of the functional modules Particular interest of BioNet and jActiveModules were the identification of functional modules in the network in which the nodes have significant P-values by means of detecting differentially expressed regions in networks. This indicates a group of nodes which are densely connected and have significant differences in expression level, suggesting a module whose activity is influenced by the experimental context of the expression data. The functional modules tend to correspond to shared common cellular function beyond the scope of classical pathways [16–18]. The maximally scoring optimal module was identified using BioNet [17, 18]. And the jActiveModules plug-in of cytoscape was used to further identify multiple significant modules in the PPI network [16]. Enrichment analysis of functional modules The gene-annotation enrichment analysis was performed using the Database for Annotation, Visualization, and Integrated Discovery (DAVID), which provides a comprehensive set of functional annotation tools for biological interpretation of large gene lists. GO and KEGG are included in the set of functional annotation tools of DAVID. To study the rifampin-regulated biological process, we used DAVID’s GOTERM_BP_FAT (lower levels of biological process ontology), and KEGG pathway analysis to identify enriched biological themes, particularly GO terms [19]. Identification of miRNAs and analysis of their functions MiRNAs with p < 0.05 were regarded as significant differentially expressed miRNAs. We identified these miRNAs’ target genes using the R library RmiR.Hs.miRNA [20] which collects information from different miRNA target databases. In this study, Targetscan [21, 22], miRanda [23], PicTar [24] and miRTarBase [25, 26] were choosen. The BiomaRt [27] library, which provides a wide range of online queries from gene annotation to database mining, was used to convert gene IDs to gene symbols based on the hsapiens_gene_ensembl database. For each miRNA, the miRNA-targeted genes belonging to functional modules were considered as the nodes of the miRNA-regulated PIN. The interactions of these genes in PPI and of each miRNA with its target genes were considered as the interactions of the miRNA-regulated PIN. As a result, we obtained the miRNA-regulated PIN. For each miRNA, we analysed its potential functions by analysing the miRNA’s target genes based on GO and KEGG. Results SDGs and RrPIN The mRNA expression profile was obtained from RNA-seq data from primary hepatocytes from 7 subjects treated with rifampin or vehicle. In order to focus on cellular responses that are triggered by the rifampin treatment, we pre-selected the genes that are differentially expressed with a loose p-value threshold at p-value < 0.01 without multiple hypothesis correction. Our further analysis focuses on 1866 differentially expressed genes that pass the threshold. We mapped all the differentially expressed genes on the combined human protein interaction network, which consists of 10,210 proteins with 42,521 interactions. As shown in Fig. 2, the resultant rifampin interaction network includes 663 proteins with 1024 interactions.Fig. 2 The workflow of the RrPIN. Color is according to the fold change where red denotes upregulated and green denotes downregulated Identification and analysis of the functional modules The aforementioned network contains candidate differentially expressed genes with a flexible p-value cutoff. This is intentional since our network analysis will be further used to identify a cluster of interacting molecules that tend to be collectively differentially expressed, and therefore will reduce false positives. We used BioNet [17, 18], a bioconductor package for the functional analysis of biological networks, which uses the p-values obtained from differential expressed genes from RNA-seq data. The goal of this algorithm is to identify functional modules, or significantly differentially expressed subnetworks, within large networks [17]. This was achieved by computing a score for each node, reflected by its p-value, and used a network search algorithm to find the highest-scoring subgraph. In this study, the maximally functional module was identified by computing optimal scores based on the p-values from the RNA-seq data to evaluate how molecular activity changes were correlated with rifampin regulation. False discovery rate (FDR) is an adjustment parameter for controlling the resultant subnetwork size. Since FDR can be used for fine-tuning of the signal noise decomposition, we scan a large range of FDRs and evaluate the obtained modules according to true-positive rate and precision (ratio of true positives to all positively classify). As a result, a threshold value of >0.0001 was used, because others thresholds lead to either too small or too large size of the module. The derived module captures the characteristically differently expressed interactions associated with rifampin treatment. There were 84 genes and 89 interactions in the maximally functional module. P-values, fold-changes, and false discovery rates (FDR) for the genes of the maximally functional module are shown in Additional file 1. To avoid bias and to ensure generality of our results, besides the maximally functional module, we identified the multiple functional modules and key genes that also demonstrated the significant enrichment on differentially expressed genes; this analysis was done using the jActiveModules plug-in in cytoscape [15]. The maximum depth from the start node was set as 2 and the overlap threshold was set as 0. There were 31 nodes and 36 interactions within the five functional modules. P-values, Fold Changes, and false discovery rates (FDR) for the genes within the five functional modules are shown in Additional file 2. The maximally functional module and five functional modules are shown in Fig. 3.Fig. 3 The functional modules of RrPIN. a The maximally functional module of PPI network. Color is according to the fold change where red denotes upregulated and green denotes downregulated. The shape of the nodes depicts the aggregate score: circles indicate a negative score, rectangles denotes a positive score. b, c, d, e and f are the five functional modules of RrPIN. The regulatory relationships are denoted by colours in which red indicates upregulated genes, and green indicates downregulated. As well, the depth of the colour explains the size of fold change As expected, the results from BioNet are essentially in agreement with the results from the jActiveModules plug-in. The maximally functional module included all the nodes and interactions of the five functional modules. Enrichment analysis of functional modules To systematically determine the roles of genes in the functional modules, the online biological classification tool DAVID was used to carry out the functional classification based on GO and key signal pathways from KEGG. Since most of the genes within the five modules were included in the maximally functional module, we primarily focused on the analysis of the maximally functional module. Table of top 20 GO terms and top 10 KEGG terms for the genes of functional modules were shown in Table 1. Due to the redundant nature of the ontology analysis, functional annotation clustering was also derived from DAVID. A table of the top 20 functional annotations clustering for the genes of functional modules assessed by DAVID are provided in Additional file 3. Since the evidence suggests that rifampin have a broad spectrum of effect on enhancing drug metabolism, specifically, we focused on the GO terms “drug” and “metabolism,” and the top six listed KEGG pathways were extracted and the duplicates were eliminated. The p-value was used to evaluate the significance of the GO terms and KEGG pathways. Table 2 shows the results of the enrichment analysis of the maximally functional module in the RrPIN.Table 1 Table of top 20 GO terms and top 10 KEGG terms for the genes of functional modules Category Term Count Percent P-value Benjiamini GOTERM_BP_FAT regulation of apoptosis 19 22.6 5.10E-07 7.30E-04 GOTERM_BP_FAT regulation of programme cell death 19 22.6 5.90E-07 4.20E-04 GOTERM_BP_FAT regulation of cell death 19 22.6 6.20E-07 3.00E-04 GOTERM_BP_FAT negative regulation of apoptosis 13 15.5 7.90E-07 2.80E-04 GOTERM_BP_FAT negative regulation of programmed cell death 13 15.5 9.10E-07 2.60E-04 GOTERM_BP_FAT negative regulation of cell death 13 15.5 9.40E-07 2.20E-04 GOTERM_BP_FAT membrane organization 13 15.5 1.70E-06 3.50E-04 GOTERM_BP_FAT vesicle-mediated transport 15 17.9 4.50E-06 8.10E-04 GOTERM_BP_FAT membrane invagination 10 11.9 4.70E-06 7.40E-04 GOTERM_BP_FAT endocytosis 10 11.9 4.70E-06 7.40E-04 GOTERM_BP_FAT response to hypoxia 8 9.5 1.20E-05 1.70E-03 GOTERM_BP_FAT response to oxygen levels 8 9.5 1.60E-05 2.10E-03 GOTERM_BP_FAT response to inorganic substance 9 10.7 2.30E-05 2.70E-03 GOTERM_BP_FAT anti-apoptosis 9 10.7 2.40E-05 2.60E-03 GOTERM_BP_FAT positive regulation of multicellular organismal process 9 10.7 7.90E-05 8.00E-03 GOTERM_BP_FAT drug metabolic process 4 4.8 9.80E-05 9.20E-03 GOTERM_BP_FAT response to metal ion 7 8.3 9.80E-05 8.70E-03 GOTERM_BP_FAT phagocytosis 5 6 1.60E-04 1.40E-02 GOTERM_BP_FAT response to organic substance 14 16.7 2.20E-04 1.80E-02 GOTERM_BP_FAT regulation of tube size 5 6 2.40E-04 1.80E-02 KEGG_PATHWAY Metabolism of xenobiotics by cytochrome P450 6 7.1 2.00E-04 1.70E-02 KEGG_PATHWAY Retinol metabolism 5 6 1.40E-03 6.00E-02 KEGG_PATHWAY Drug metabolism 5 6 2.40E-03 6.70E-02 KEGG_PATHWAY Linoleic acid metabolism 3 3.6 2.70E-02 4.50E-01 KEGG_PATHWAY Pathways in cancer 8 9.5 2.90E-02 4.00E-01 KEGG_PATHWAY Focal adhesion 6 7.1 3.70E-02 4.10E-01 KEGG_PATHWAY Porphyrin and chlorophyll metabolism 3 3.6 3.70E-02 3.70E-01 KEGG_PATHWAY Small cell lung cancer 4 4.8 4.20E-02 3.70E-01 KEGG_PATHWAY ECM-receptor interaction 4 4.8 4.20E-02 3.70E-01 KEGG_PATHWAY TGF-beta signaling pathway 4 4.8 4.60E-02 3.60E-01 Table 2 Enrichment analysis of the maximally functional module in rifampin DAVID (Term) Genes P-value GO: Response to drug ABCB1,UGT1A4,CAV1,CAV2 3.6E-2 KEGG: Metabolism of xenobiotics by cytochrome P450 UGT1A4,ADH6,CYP1A1,CYP2C19,CYP2C9,CYP2E1 2.0E-4 KEGG: Retinol metabolism UGT1A4,ADH6,CYP1A1,CYP2C19,CYP2C9 1.4E-3 KEGG: Drug metabolism UGT1A4,ADH6, CYP2C19,CYP2C9,CYP2E1 2.4E-3 KEGG: Linoleic acid metabolism CYP2C19,CYP2C9,CYP2E1 2.7E-2 KEGG: Pathways in cancer CEBPA,CREBBP,SMAD3,TRAF2,BIRC3,EGLN2,FN1,IKBKG 2.9E-2 Focal adhesion BIRC3,CAV1,CAV2,FN1,ITGA1,THBS1 3.7E-2 The results show that the maximally functional module is relevant with seven functional enrichment terms: response to drug, metabolism of xenobiotics by cytochrome P450, retinol metabolism, drug metabolism, linoleic acid metabolism, cancer pathways, and focal adhesion. Among these terms, retinol metabolism, drug metabolism and linoleic acid metabolism contained many similarities in genes, since these three function terms were functionally correlated and clustered in functional annotation clustering in DAVID. In particular, the function pathways coincided with previously reported rifampin-induced biological functions. For example, rifampin affected the hepatic drug disposition and metabolism [28, 29] and it was a potent inducer of drug-metabolizing enzymes [6, 29–31]. Rifampin is also an inhibitor which rapidly downregulates angiogenesis and mitogenesis-related genes to target cancer cells [12, 32, 33]. In addition, we define the percentage of identified SDGs in each pathway relative to the total number of SDGs as pathway relative abundance. Assume that there are N SDGs-associated biological pathways, for i–th pathway, S(i) is number of identified SDGs, and N(i) is enriched number of the total SDGs. The pathway relative abundance E(x) is defined as: 1 Ei=SiNii∈1..N Figure 4 shows the pathway relative abundance of maximally functional module’s genes and all SDGs on the associated seven functional enrichment terms. It can be seen that the SDGs of maximally functional module are more enriched and representative on each terms than the total SDGs. This suggests that our strategy in integrating PPI network with the differential expression analysis helped us in capturing more biologically relevant signals.Fig. 4 The pathway relative abundance of maximally functional module’s genes and all SDGs on the associated seven functional enrichment terms To analyze each functional enrichment term, we focus on the analysis of their genes. There were 19 key genes associated with 7 functions that we extracted using DAVID. Then we extend the protein interaction network of 19 genes based on the RrPIN with one level interaction. The RrPIN extension network of 19 genes consists of 50 nodes and 53 interactions. The RrPIN extension network of 19 genes and associated 7 functions are shown as Fig. 5.Fig. 5 The RrPIN extension network of 19 genes and associated 7 functions It is worth noting that UGT1A4, ADH6, CYP1A1, CYP2C19, CYP2C9 and CYP2E1 are all associated with metabolism of xenobiotics, drug metabolism, retinol metabolism, and linoleic acid metabolism. BIRC3, CAV1, CAV2, FN1, ITGA1 and THBS1 were functionally enriched to focal adhesion, which contributes to antiangiogenic and anti-tumour effects. These results indicate that rifampin induced drug metabolism, partially, by regulating UGT1A4, ADH6, CYP1A1, CYP2C19, CYP2C9 and CYP2E1. These results also signify that rifampin can influence the anti-angiogenesis and anti-tumour effects of drugs by regulating BIRC3, CAV1, CAV2, FN1, ITGA1 and THBS1. Previous reports support these findings, stating that UGT1A4 CYP1A1, CYP2C19, CYP2C9 and CYP2E1 are drug-metabolizing enzymes [34, 35], and ADH6 modulates the risk for drug dependence [35]. BIRC3 contains anti-apoptotic genes, which can be suppressed to counteract cancerous activity [36]. CAV1 and CAV2 were correlated with tumour growth and metastasis [37–39], and FN1 was a potential biomarker for some cancers [40, 41], while ITGA1 and THBS1 were also associated with cancer risk [42, 43]. In addition, some of the 19 key genes were hub proteins that interacted with multiple proteins. For example, CAV1, CREBBP, SMAD3, TRAF2, KBKG and THBS1 had at least four interactions with other proteins. These results suggest that these six genes are important components in biological pathways regulated by rifampin. Joint analysis of key genes and associated miRNAs To identify miRNAs that may regulate these key genes within the functional modules, we correlated the alterations in the miRNA and gene expression. In this process, we extracted the significant differentially expressed miRNAs (p < 0.05), and identified 20 miRNAs. The significant differentially expressed miRNAs are shown in Table 3.Table 3 The significant differentially expressed miRNAs miRNA P-value miRNA P-value Upregulated Upregulated miR-886-3p 0.0002 miR-660 0.0297 miR-766 0.0075 miR-638 0.0302 miR-92a 0.0169 miR-25 0.0338 miR-107 0.0177 miR-616 0.0446 miR-30d# 0.0195 miR-576-3p 0.0453 miR-335 0.0241 miR-218 0.0499 Downregulated Downregulated miR-186 0.0018 miR-320 0.0376 miR-361 0.0111 miR-202 0.0396 miR-95 0.0219 miR-200b# 0.0426 miR-345 0.0239 let-7 g 0.0435 In order to identify the target genes of significant differentially expressed miRNAs, three databases (Targetscan, miRanda and PicTar) were used [20–23]. The miRNA-mRNA pairs were extracted for each significant differentially expressed miRNA. In order to include verified miRNA-mRNA pairs, we also extracted the miRNA-mRNA pairs from the miRTarBase [25, 26]. And the miRNA-mRNA pairs of which mRNA is a gene of maximally functional module were chosen. Then, we established the miRNA-regulated PIN, which showed a negative correlation between the miRNA and the mRNA. The miRNA-regulated PIN, which is constructed of the genes in the functional modules, is shown in Table 4 and Fig. 6.Table 4 The miRNA-regulated PIN which constructed by the genes of functional modules Gene list logFC miRNA P-value Fold change CYP2E1 −1.4341 miR-335 0.0242 1.3300 CAV1 −0.8518 miR-34b 0.1753 185.3764 miR-886-3p 0.0001 1.5645 miR-218 0.0499 1.9012 miR-576-3p 0.0453 2.1916 CAV2 −0.5386 miR-200c 0.0913 4.8313 miR-576-3p 0.0453 2.1916 CEBPA 0.5812 miR-186 0.0017 0.8356 CREBBP 0.3821 miR-186 0.0017 0.8356 miR-95 0.0216 0.6320 miR-769 0.1249 0.8388 EGLN2 0.4574 miR-202 0.0396 0.5988 let-7 g 0.0435 0.8402 ITGA1 −0.3754 miR-616 0.0446 1.3337 miR-660 0.0297 1.2642 miR-576-3p 0.0453 2.1916 miR-335 0.0242 1.3300 THBS1 −0.3951 miR-886-3p 0.0001 1.5645 miR-335 0.0242 1.3300 miR-616 0.0446 1.3337 miR-92a 0.0169 1.1319 Fig. 6 The miRNA-regulated PIN which is constructed by the genes of functional modules Eight genes and 14 miRNAs were identified to have significant differential expression changes. Each of these genes was regulated by multiple miRNAs. Due to the miRNAs control of gene expression, either by degradation of the target mRNAs or by inhibition of protein translation, miRNA-regulated PPI networks can uncover new rules of miRNA regulation or protein interaction. Thus, we predicted the potential functions of miRNAs based on the function of their target genes as shown in Table 5.Table 5 The potential functions of miRNAs DAVID (Term) miRNA GO: Response to drug miR-34b, miR-886-3p, miR-218, miR-576-3p, miR-200c KEGG: Metabolism of xenobiotics by cytochrome P450 miR-335 KEGG: Drug metabolism miR-335 KEGG: Linoleic acid metabolism miR-335 KEGG: Pathways in cancer miR-186, miR-95, miR-769 Focal adhesion miR-34b, miR-886-3p, miR-218, miR-576-3p, miR-200c, miR-616, miR-660, miR-335, miR-92a Twelve miRNAs were extracted which associated with 6 biological pathways including response to drug, metabolism of xenobiotics by cytochrome P450, drug metabolism, linoleic acid metabolism, cancer pathways, and focal adhesion through regulation of 8 target genes. The results suggest that miR-335 influences drug metabolism through negative regulation of CYP2E1, which is a drug metabolizing enzyme that is affected by rifampin treatment. Therefore, it is possible that rifampin may alter miRNA expression, which in turn affects the expression of the drug metabolizing enzyme gene CYP2E1. MiR-186 was found to regulate two genes (CEBPA, CREBBP), which were associated with cancer pathways. MiR-186, miR-769, miR-95, miR-202 and let-7 g were also relevant to cancer pathways, but did not serve other functions. Previous studies have demonstrated that rifampin also inhibited anti-angiogenesis by regulating the expression of multiple miRNAs (miR-34b, miR-886-3p, miR-218, miR-576-3p, miR-200c, miR-616, miR-660, miR-335, miR-92a), and further induced the gene expression of BIRC3, CAV1, CAV2, FN1, ITGA1 and THBS1. Conclusions In conclusion, a novel integrative network-based method was used to identify the functional modules and discover the potential functions of miRNAs based on human protein network, mRNA and miRNA expression profile in rifampin treated hepatocytes. Furthermore, this method identifies 19 genes and 7 crucial biological pathways. By analysing the miRNA-regulated PIN, we suggested that 12 miRNAs were associated with 6 biological pathways through regulation of 8 target genes. Our results suggest that rifampin contributes to changes in the expression of genes and miRNAs, and induces multiple biological pathways. This study not only provides an insight into functional modules that are associated with rifampin-treated human hepatocytes in human protein interaction network, it also shows that the integrated analysis of mRNA, miRNA expression profile and PIN can be used to study the molecular mechanism of rifampin-induced drug disposition. Additional files Additional file 1: P-values, Fold Change and false discovery rates (FDR) for the genes of the maximally functional module. (PDF 161 kb) Additional file 2: P-values, Fold Change and false discovery rates (FDR) for the genes of the five functional modules. (PDF 105 kb) Additional file 3: Top 10 GO terms and KEGG terms for the genes in functional modules from the DAVID were provided as Additional file 3. (PDF 264 kb) Declarations Publication charges for this article have been funded by the National Key Scientific Instrument and Equipment Development Projects of China (2012YQ04014001 and 2012YQ04014010), National Natural Science Foundation of China (61471139), Fundamental Research Funds for the Central Universities (HEUCF160412), Natural Science Fund of Heilongjiang Province (F201331, F201241). This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The complete RNA-seq data used in this paper can be downloaded from the GEO database with the accession number: GSE79933; The complete microRNA OpenArray data used in this paper can be downloaded from http://compbio.iupui.edu/group/6/pages/rifampin. Authors’ contributions JL and WY developed the programs and workflows, analysed the data, and wrote the manuscript. LW, XFD and YW contributed to the data analysis. WC provided some advice on analysis and contributed partly to writing of the manuscript. CZX and WXF contributed to the computational analyses. YLD, TCS was responsible for sample collection and processing for analysis. HL and YLL conceived and directed the project, arranged the sampling, provided advice on analysis, and contributed to writing of the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Koyuturk M Algorithmic and analytical methods in network biology Wiley Interdiscip Rev Syst Biol Med 2010 2 3 277 292 10.1002/wsbm.61 20836029 2. Lin CC Chen YJ Chen CY Oyang YJ Juan HF Huang HC Crosstalk between transcription factors and microRNAs in human protein interaction network BMC Syst Biol 2012 6 18 10.1186/1752-0509-6-18 22413876 3. Gennarino VA D’Angelo G Dharmalingam G Fernandez S Russolillo G Sanges R Mutarelli M Belcastro V Ballabio A Verde P Identification of microRNA-regulated gene networks by expression analysis of target genes Genome Res 2012 22 6 1163 1172 10.1101/gr.130435.111 22345618 4. Egelund EF Isaza R Brock AP Alsultan A An G Peloquin CA Population pharmacokinetics of rifampin in the treatment of Mycobacterium tuberculosis in Asian elephants J Vet Pharmacol Ther 2015 38 2 137 143 10.1111/jvp.12156 25236765 5. Liu CL Lim YP Hu ML Fucoxanthin attenuates rifampin-induced cytochrome P450 3A4 (CYP3A4) and multiple drug resistance 1 (MDR1) gene expression through pregnane X receptor (PXR)-mediated pathways in human hepatoma HepG2 and colon adenocarcinoma LS174T cells Marine drugs 2012 10 1 242 257 10.3390/md10010242 22363234 6. Mahatthanatrakul W Nontaput T Ridtitid W Wongnawa M Sunbhanich M Rifampin, a cytochrome P450 3A inducer, decreases plasma concentrations of antipsychotic risperidone in healthy volunteers J Clin Pharm Ther 2007 32 2 161 167 10.1111/j.1365-2710.2007.00811.x 17381666 7. Takahashi K Tatsumi N Fukami T Yokoi T Nakajima M Integrated Analysis of Rifampicin-induced MicroRNA and Gene Expression Changes in Human Hepatocytes Drug Metab Pharmacokinet 2014 29 4 333 340 10.2133/dmpk.DMPK-13-RG-114 24552687 8. Ramamoorthy A Liu Y Philips S Desta Z Lin H Goswami C Gaedigk A Li L Flockhart DA Skaar TC Regulation of microRNA expression by rifampin in human hepatocytes Drug Metab Dispos 2013 41 10 1763 1768 10.1124/dmd.113.052886 23935064 9. Kim B Moon JY Choi MH Yang HH Lee S Lim KS Yoon SH Yu KS Jang IJ Cho JY Global metabolomics and targeted steroid profiling reveal that rifampin, a strong human PXR activator, alters endogenous urinary steroid markers J Proteome Res 2013 12 3 1359 1368 10.1021/pr301021p 23320515 10. Kliewer SA Goodwin B Willson TM The nuclear pregnane X receptor: a key regulator of xenobiotic metabolism Endocr Rev 2002 23 5 687 702 10.1210/er.2001-0038 12372848 11. Benson EA Eadon MT Desta Z Liu Y Lin H Burgess KS Segar MW Gaedigk A Skaar TC Rifampin Regulation of Drug Transporters Gene Expression and the Association of MicroRNAs in Human Hepatocytes Front. Pharmacol. 2016 7 111 10.3389/fphar.2016.00111 27199754 12. Shichiri M Fukai N Kono Y Tanaka Y Rifampicin as an oral angiogenesis inhibitor targeting hepatic cancers Cancer Res 2009 69 11 4760 4768 10.1158/0008-5472.CAN-08-3417 19458074 13. Keshava Prasad TS Goel R Kandasamy K Keerthikumar S Kumar S Mathivanan S Telikicherla D Raju R Shafreen B Venugopal A Human Protein Reference Database--2009 update Nucleic Acids Res 2009 37 Database issue D767 D772 10.1093/nar/gkn892 18988627 14. Wang J Huo K Ma L Tang L Li D Huang X Yuan Y Li C Wang W Guan W Toward an understanding of the protein interaction network of the human liver Mol Syst Biol 2011 7 536 10.1038/msb.2011.67 21988832 15. Kohl M Wiese S Warscheid B Cytoscape: software for visualization and analysis of biological networks Methods Mol Biol 2011 696 291 303 10.1007/978-1-60761-987-1_18 21063955 16. Ideker T Ozier O Schwikowski B Siegel AF Discovering regulatory and signalling circuits in molecular interaction networks Bioinformatics 2002 18 Suppl 1 S233 S240 10.1093/bioinformatics/18.suppl_1.S233 12169552 17. Dittrich MT Klau GW Rosenwald A Dandekar T Muller T Identifying functional modules in protein-protein interaction networks: an integrated exact approach Bioinformatics 2008 24 13 i223 i231 10.1093/bioinformatics/btn161 18586718 18. Beisser D Klau GW Dandekar T Muller T Dittrich MT BioNet: an R-Package for the functional analysis of biological networks Bioinformatics 2010 26 8 1129 1130 10.1093/bioinformatics/btq089 20189939 19. Kanehisa M Goto S Sato Y Furumichi M Tanabe M KEGG for integration and interpretation of large-scale molecular data sets Nucleic Acids Res 2012 40 Database issue D109 D114 10.1093/nar/gkr988 22080510 20. Helvik SA Snove O Jr Saetrom P Reliable prediction of Drosha processing sites improves microRNA gene prediction Bioinformatics 2007 23 2 142 149 10.1093/bioinformatics/btl570 17105718 21. Friedman RC Farh KK Burge CB Bartel DP Most mammalian mRNAs are conserved targets of microRNAs Genome Res 2009 19 1 92 105 10.1101/gr.082701.108 18955434 22. Garcia DM Baek D Shin C Bell GW Grimson A Bartel DP Weak seed-pairing stability and high target-site abundance decrease the proficiency of lsy-6 and other microRNAs Nat Struct Mol Biol 2011 18 10 1139 1146 10.1038/nsmb.2115 21909094 23. Betel D Wilson M Gabow A Marks DS Sander C The microRNA.org resource: targets and expression Nucleic Acids Res 2008 36 Database issue D149 D153 18158296 24. Krek A Grun D Poy MN Wolf R Rosenberg L Epstein EJ MacMenamin P da Piedade I Gunsalus KC Stoffel M Combinatorial microRNA target predictions Nat Genet 2005 37 5 495 500 10.1038/ng1536 15806104 25. Hsu SD Lin FM Wu WY Liang C Huang WC Chan WL Tsai WT Chen GZ Lee CJ Chiu CM miRTarBase: a database curates experimentally validated microRNA-target interactions Nucleic Acids Res 2011 39 Database issue D163 D169 10.1093/nar/gkq1107 21071411 26. Hsu SD Tseng YT Shrestha S Lin YL Khaleel A Chou CH Chu CF Huang HY Lin CM Ho SY miRTarBase update 2014: an information resource for experimentally validated miRNA-target interactions Nucleic Acids Res 2014 42 Database issue D78 D85 10.1093/nar/gkt1266 24304892 27. Assefa SA Preston MD Campino S Ocholla H Sutherland CJ Clark TG estMOI: estimating multiplicity of infection using parasite deep sequencing data Bioinformatics 2014 30 9 1292 1294 10.1093/bioinformatics/btu005 24443379 28. Lam JL Shugarts SB Okochi H Benet LZ Elucidating the effect of final-day dosing of rifampin in induction studies on hepatic drug disposition and metabolism J Pharmacol Exp Ther 2006 319 2 864 870 10.1124/jpet.106.108282 16905688 29. Anderson MS Cote J Liu Y Stypinski D Auger P Hohnstein A Rasmussen S Johnson-Levonas AO Gutstein DE Effects of Rifampin, a potent inducer of drug-metabolizing enzymes and an inhibitor of OATP1B1/3 transport, on the single dose pharmacokinetics of anacetrapib J Clin Pharmacol 2013 53 7 746 752 10.1002/jcph.97 23670789 30. Hamilton M Wolf JL Drolet DW Fettner SH Rakhit AK Witt K Lum BL The effect of rifampicin, a prototypical CYP3A4 inducer, on erlotinib pharmacokinetics in healthy subjects Cancer Chemother Pharmacol 2014 73 3 613 621 10.1007/s00280-014-2390-3 24474302 31. Loboz KK Gross AS Williams KM Liauw WS Day RO Blievernicht JK Zanger UM McLachlan AJ Cytochrome P450 2B6 activity as measured by bupropion hydroxylation: effect of induction by rifampin and ethnicity Clin Pharmacol Ther 2006 80 1 75 84 10.1016/j.clpt.2006.03.010 16815319 32. Shichiri M Tanaka Y Inhibition of cancer progression by rifampicin: involvement of antiangiogenic and anti-tumor effects Cell Cycle 2010 9 1 64 68 10.4161/cc.9.1.10354 20016271 33. Fardel O Lecureur V Loyer P Guillouzo A Rifampicin enhances anti-cancer drug accumulation and activity in multidrug-resistant cells Biochem Pharmacol 1995 49 9 1255 1260 10.1016/0006-2952(95)00045-2 7763306 34. Cho YY Jeong HU Kim JH Lee HS Effect of honokiol on the induction of drug-metabolizing enzymes in human hepatocytes Drug Des Devel Ther 2014 8 2137 2145 25395831 35. Luo X Kranzler HR Zuo L Wang S Schork NJ Gelernter J Multiple ADH genes modulate risk for drug dependence in both African- and European-Americans Hum Mol Genet 2007 16 4 380 390 10.1093/hmg/ddl460 17185388 36. Kumar S Ingle H Mishra S Mahla RS Kumar A Kawai T Akira S Takaoka A Raut AA Kumar H IPS-1 differentially induces TRAIL, BCL2, BIRC3 and PRKCE in type I interferons-dependent and -independent anticancer activity Cell Death Dis 2015 6 e1758 10.1038/cddis.2015.122 25950488 37. Ando T Ishiguro H Kimura M Mitsui A Mori Y Sugito N Tomoda K Mori R Harada K Katada T The overexpression of caveolin-1 and caveolin-2 correlates with a poor prognosis and tumor progression in esophageal squamous cell carcinoma Oncol Rep 2007 18 3 601 609 17671707 38. Capozza F Trimmer C Castello-Cros R Katiyar S Whitaker-Menezes D Follenzi A Crosariol M Llaverias G Sotgia F Pestell RG Genetic ablation of Cav1 differentially affects melanoma tumor growth and metastasis in mice: role of Cav1 in Shh heterotypic signaling and transendothelial migration Cancer Res 2012 72 9 2262 2274 10.1158/0008-5472.CAN-11-2593 22396494 39. Elsheikh SE Green AR Rakha EA Samaka RM Ammar AA Powe D Reis-Filho JS Ellis IO Caveolin 1 and Caveolin 2 are associated with breast cancer basal-like and triple-negative immunophenotype Br J Cancer 2008 99 2 327 334 10.1038/sj.bjc.6604463 18612310 40. Amundson SA Smilenov LB Integration of biological knowledge and gene expression data for biomarker selection: FN1 as a potential predictor of radiation resistance in head and neck cancer Cancer Biol Ther 2010 10 12 1252 1255 10.4161/cbt.10.12.13731 20948301 41. Viana Lde S Affonso RJ Jr Silva SR Denadai MV Matos D Salinas de Souza C Waisberg J Relationship between the expression of the extracellular matrix genes SPARC, SPP1, FN1, ITGA5 and ITGAV and clinicopathological parameters of tumor progression and colorectal cancer dissemination Oncology 2013 84 2 81 91 10.1159/000343436 23128103 42. Yim DH Zhang YW Eom SY Moon SI Yun HY Song YJ Youn SJ Hyun T Park JS Kim BS ITGA1 polymorphisms and haplotypes are associated with gastric cancer risk in a Korean population World J Gastroenterol 2013 19 35 5870 5876 10.3748/wjg.v19.i35.5870 24124332 43. Lin XD Chen SQ Qi YL Zhu JW Tang Y Lin JY Polymorphism of THBS1 rs1478604 A > G in 5-untranslated region is associated with lymph node metastasis of gastric cancer in a Southeast Chinese population DNA Cell Biol 2012 31 4 511 519 10.1089/dna.2011.1344 22011138

PMC005xxxxxx/PMC5001205.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556924289710.1186/s12864-016-2897-6ResearchDetection of high variability in gene expression from single-cell RNA-seq profiling Chen Hung-I Harry uml679@my.utsa.edu 12Jin Yufang yufang.jin@utsa.edu 2Huang Yufei yufei.huang@utsa.edu 2Chen Yidong cheny8@uthscsa.edu 131 Greehey Children`s Cancer Research Institute, The University of Texas Health Science Center at San Antonio, San Antonio, TX 78229 USA 2 Department of Electrical and Computer Engineering, The University of Texas at San Antonio, San Antonio, TX 78249 USA 3 Department of Epidemiology and Biostatistics, The University of Texas Health Science Center at San Antonio, San Antonio, TX 78229 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.508© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background The advancement of the next-generation sequencing technology enables mapping gene expression at the single-cell level, capable of tracking cell heterogeneity and determination of cell subpopulations using single-cell RNA sequencing (scRNA-seq). Unlike the objectives of conventional RNA-seq where differential expression analysis is the integral component, the most important goal of scRNA-seq is to identify highly variable genes across a population of cells, to account for the discrete nature of single-cell gene expression and uniqueness of sequencing library preparation protocol for single-cell sequencing. However, there is lack of generic expression variation model for different scRNA-seq data sets. Hence, the objective of this study is to develop a gene expression variation model (GEVM), utilizing the relationship between coefficient of variation (CV) and average expression level to address the over-dispersion of single-cell data, and its corresponding statistical significance to quantify the variably expressed genes (VEGs). Results We have built a simulation framework that generated scRNA-seq data with different number of cells, model parameters, and variation levels. We implemented our GEVM and demonstrated the robustness by using a set of simulated scRNA-seq data under different conditions. We evaluated the regression robustness using root-mean-square error (RMSE) and assessed the parameter estimation process by varying initial model parameters that deviated from homogeneous cell population. We also applied the GEVM on real scRNA-seq data to test the performance under distinct cases. Conclusions In this paper, we proposed a gene expression variation model that can be used to determine significant variably expressed genes. Applying the model to the simulated single-cell data, we observed robust parameter estimation under different conditions with minimal root mean square errors. We also examined the model on two distinct scRNA-seq data sets using different single-cell protocols and determined the VEGs. Obtaining VEGs allowed us to observe possible subpopulations, providing further evidences of cell heterogeneity. With the GEVM, we can easily find out significant variably expressed genes in different scRNA-seq data sets. Keywords Single-cellSingle-cell RNA-SeqCell heterogeneityNegative binomial distributionGene expression variation modelVariably expressed genesThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Single-cell analysis has emerged a decade ago to understand the heterogeneity of a cell population, especially in biology contexts such as early embryonic development and tumor etiology [1]. Single-cell quantitative PCR (qPCR) [2–4] or single-molecule RNA fluorescence in situ hybridization (FISH) [5] have been widely used as low-throughput approaches to measure the expression of specific genes at a single-cell level. Although experiments using these methods can provide crucial information of cellular heterogeneity and the presence of distinct cell subpopulations, only a small number of genes can be monitored simultaneously. RNA sequencing (RNA-seq), a developed approach using next-generation sequencing (NGS) technology, can unbiasedly detect the genome-wide gene expression of a sample. Bulk RNA-seq experiments start with a large population of cells (> 105), and the gene expression levels are considered as the average expression across the population of a cell pool [6]. Bulk RNA-seq might be sufficient in many contexts such as revealing the aberration of mRNA expression between different treatments, conditions, or phenotypes. However, biological questions like diversity in early stage development embryonic cells, which each cell has distinct functions, can't be explained by bulk RNA-seq experiments. With recent introduction of Smart-seq protocol, the required volume of starting materials has been vastly reduced, making the single-cell RNA sequencing (scRNA-seq) achievable [7, 8]. There are already several protocols for sequencing of single cells, which allow researchers to assay high-throughput gene expression profiling at the single-cell level of a large number of cells. However, unlike the conventional RNA-seq where analysis tools are abundantly available, the lack of bioinformatics tools for single-cell RNA-seq limits its huge potential. Comparing with bulk RNA-seq measurements, single-cell RNA-seq data tend to have much lower read counts (~200,000 to 5 million reads per cell) [9], higher variability, and large number of outliers, and all these are poorly accommodated by conventional RNA-seq analysis methods [10]. Unlike the objectives of conventional RNA-seq where differential expression analysis and the detection of differentially expressed genes (DEGs) are integral components, the most important goal of scRNA-seq is to identify variably expressed genes (VEGs) across a population of cells to account for the discrete nature of single-cell gene expression and uniqueness of sequencing library preparation protocol for single-cell sequencing. As we observed, the transcriptional heterogeneity of the cell population can be assessed by the expression variation difference, whether they are lowly or highly expressed, which conventional RNA-seq analysis failed to identify due to the assumption of homogeneity within each cell subtype. In recent studies, gene expression variation models were proposed specifically for the corresponded scRNA-seq experiments in order to detect VEGs deviated from the Poisson model [11, 12]. However, different scRNA-seq data sets rendered different distributions and a common mathematical description is necessary. Hence, the purpose of this study is to provide a mathematical description of a gene expression variation model (GEVM) for scRNA-seq data. The model addresses the over-dispersion of single-cell data and the additional variability caused by different sources of variation. By exploiting existing statistical tools such as local regression and nonlinear least squares curve fitting, the parameters of gene variation model are estimated and statistical significant VEGs can be identified. To study the robustness of the model, we have built a simulation framework to generate single-cell RNA-seq data using different distributions in each step to imitate the dispersion of real data in different conditions. We demonstrated robustness of our method by applying it to the simulated data and test how precise we can estimate the parameters to the initial settings. Methods Modeling of single-cell RNA-seq data To develop a generic GEVM, we exploited the over-dispersion concept from edgeR [13]. Assuming each gene's expression follows a negative binomial (NB) distribution with parameter NB(ri, pi) for ith gene, we have 1 σi2=μi1−pi=μi+μi2ri, where the μ and σ2 are gene expression mean and variance, respectively. We further assume that in a given condition across a cell population, the model parameter r does not change (invariant to gene expression level), or 2 σ2=μ+αμ2, where α is defined as the dispersion, or α = 1/r. For simplicity, we omitted gene index from Eq. 2. Clearly, when α > 0, the data are from a NB distribution. If α = 0, the data can be represented by a Poisson distribution (or r → ∞), which follow the diagonal line with a slope of −12 in a log-log CV-mean plot where σ2 = μ in Eq. 2. However, there are many sources of technical variation that contribute to the variability of scRNA-seq data. For instance, single-molecule capture efficiency, 3』 end bias due to single-cell RNA library preparation protocol, and low expression of genes that are easily affected by noises [14]. In this respect, we assume σ2 = μ + bμ = βμ, where β = 1 + b, and bμ is an additive noise component (proportional to the mean signal strength). Thus, the data deviate from the original diagonal line, following a line of log10(CV)=−12log10(μ)+12log10(β). Consequently, we extended the relation between the mean and variance given in Eq. 2, by adding a model parameter β to represent the multiplicative effect of different sources of technical noises. 3 σ2=βμ+αμ2, where we also assumed β is invariant within each cell population. We further obtained, from Eq. 3, the relationships between the coefficient of variation (CV, defined as σ/μ) of each gene across the cell population and its average expression level as follows, 4 log10CV=12log10βμ+α. Therefore, by measuring the CV and mean abundance of gene expression μ from all genes, we can estimate the two parameters α and β and dissect the baseline of the cell population. Note that from Eq. 4 when the mean expression level μ becomes larger, CV→α, or a constant coefficient variation [15], and when μ ≪ 1 the σ2 → βμ, or equivalently to a Poisson parameter λ ' = βμ. Estimation of model parameters and selection of significant VEGs In order to identify genes whose variation of gene expression are larger than those defined by Eq. 4, we need to estimate model parameter α and β from a scRNA-seq data set derived from a given cell population. The estimation procedure is as follows (Fig. 1): firstly we calculate the mean and coefficient of variation of each gene across a set of cells; afterwards, we perform a robust local regression implemented in locfit (R package) for fitting a robust CV-mean relationship. The nonlinear curve starts at the point with enough neighboring points (>0.5 % of total genes) to prevent overemphasizing the low expression section due to the subsampling in the next step. In addition, we also terminate the nonlinear curve at the smallest CV point to constrain to a flat line. As a typical phenomenon in scRNA-seq, only a few genes with high expression levels, results in an inaccurate local fitting at the right-tail side. On the other hand, a large proportion of genes locates in the middle section, leading to a bias during least-squares fitting in the next step. To remedy this bias, we subsample the fitted data points in a fixed interval (0.01 in log10 scale) from the start to the terminal point. Then we employ nonlinear least-squares fitting implemented in nls (R package) to estimate the two model parameters (α and β) of the GEVM. Now we can get the CV difference Di, which is the shortest distance of gene i to the ideal model with parameter α and β as a measure of variability.Fig. 1 Workflow of identifying significantly variably expressed genes and the following analyses for single-cell RNA-seq data Determination of p-value of VEGs Instead of picking VEGs by the rank of CV difference Di, we hypothesize that under the assumption of a homogeneous cell population, the CV difference to the model curve (Eq. 4) possesses a normal distribution (around baseline). We further assume that majority of genes, in a heterogeneous cell population, do not deviate much. Therefore, we use the CV differences of the data points around the model curve (Eq. 4) to fit a normal distribution. Even though robust local regression is used to estimate the expression variation model, the model is still influenced by those outlier genes. Hence, we use kernel density to find the center of the normal distribution. Afterwards, we fit the normal distribution using the CV differences below the center point. We can calculate p-value of each data point from the normal distribution and determine the significance of VEGs comparing to initial homogenous cell population. The procedure of Benjamini and Hochberg [16] is also applied to obtain the false discovery rate (FDR). Fig. 1 shows the overall workflow for detecting VEGs in scRNA-seq data set. Simulation of scRNA-seq data from a homogeneous cell population In order to evaluate the robustness of our GEVM, we generated a set of simulated data where we could control the baseline parameters and the differential expression status for a set of genes in a random set of cells. First, we utilized exponential distribution (with 3 different mean values: 0.25, 1, and 10, respectively) to create a “master cell” and its genome-wide expression levels of a cell population. The two lower mean values were designed to reflect the nature of low expression events in scRNA-seq data. The master cell expression level Mi would be the base expression value of gene i for the other single cells in the population (children cells derived from a single master cell). Given the master cell expression level Mi, and the assigned parameters α, β, the children single cells xij were simulated with a negative binomial distribution, 5 xij~NBrij,pij where the two NB parameters rij and pij were further computed by, 6 rij=μij2σij2−μij=μijβ−1+αμij 7 pij=μijσij2=1β+αμij Equations 6 and 7 were obtained utilizing our model Eq. 3. The mean value of gene i in cell j, μij, was derived from the master cell expression level with a Gaussian distribution of μij = N(Mi, max(0.2, 0.2 * Mi)). Here we required standard deviation greater than 0.2 to avoid small or near 0 standard deviation. Simulation of scRNA-seq data from a heterogeneous cell population To generate a cell population with non-distinct grouping effects, we first select a percentage of cells to be deviated from its original homogeneous population governed by the master cell. To achieve that and with a set of selected cells, we determine a subset of genes (variable prct) whose expression levels to be altered, and we generate the log fold change of each selected gene from a normal distribution to simulate a gradual fold change, with majority of them with minimal alteration. The fold change of a selected gene k is generated as, 8 log2FCk~Normalmean=0,s=2 where the variation level can be controlled by modifying the standard deviation s of the normal distribution. To determine a subset of cells to be altered, the probability of each cell to be deviated is in a uniform distribution, uniform(0, 1) and a cell with probability larger than 0.9 is classified as a heterogeneous cell. By using different distributions for simulation, we are able to generate data close to real scRNA-seq data under different conditions by changing the assigned parameters. We also compare our model with the noise model (Eq. 9) from a previous study [12]. At last, we measure the root mean square error (RMSE) to test the robustness of both methods on the simulated data, where RMSE is evaluated against log10(CV) over μ at a fixed interval, between input and estimated models. 9 log10CV=log10μγ+δ Single-cell RNA-seq data set for testing Two mouse scRNA-seq data sets were obtained from Gene Expression Omnibus (GSE65525 and GSE60361) [11, 12]. GSE65525 is the mouse embryonic stem cells with 24,175 genes in 933 single cells, sequenced using CEL-seq protocol [17], and GSE60361 is the mouse cerebral cortex cells with 19,970 genes in 3,007 cells, sequenced using quantitative single-cell RNA-seq protocol [18]. Both data sets were counted using unique molecular identifiers (UMIs) to eliminate duplicated reads caused by library amplification. Following previous study [11], we also performed the same scaling normalization method on both UMI count data sets, 10 k^ij=kijK¯/Kj,whereKj=∑ikij where kij is the UMI count of gene i in cell j, Kj is the total UMI count of cell j and K¯ is the average UMI count among the cell population. Genes that expressed in less than 1 % of the cell population were removed before applying to the model. As we shown later, the two data sets distribute differently. Under these two distinct cases, we will test the performance of the proposed method under different conditions. Results and discussion Implementation of noise model on simulation data To understand the robustness and limitation of the noise model, simulated data sets with different parameters compositions were generated by using R and then proceeded to identify the significantly VEGs following the flow chart in Fig. 1. Simulation modules implemented were: 1) Master cell gene expression generation; 2) homogeneous cell population gene expression generation (with model parameter α and β); 3) heterogeneous cell population generation (with model parameter prct for number of genes deviated from homogeneous cell population, and s for gene expression variation, Eq. 8). The VEG analysis algorithm will first estimate model parameter α and β described in Eq. 4 by using a cascade of regression (local fit, subsampling, and nonlinear least-squares). For single-cell gene expression data, in the ideal condition all genes should obey CV = μ− 1/2 [11], following a Poisson distribution as depicted by a black diagonal line in log(μ) vs log(CV) plot shown in Fig. 2. In reality, the variance typically exceeds the sample mean, justifying the negative binomial distribution in many NGS applications (and in our simulation example, Eq. 5. The cyan curve in Fig. 2 is the likelihood model of robust local regression using the function locfit.robust in R where outliers were iteratively identified and down-weighted, which allowed us to accurately fit a baseline for the data. The red line in Fig. 2 is the fitted homogeneous variation model and the orange line is the noise model in Eq. 9. With the estimated model parameters α^andβ^, we will evaluate the regression robustness using RMSE. The parameter estimation process was evaluated by varying initial model parameters (α and β in Table 1, s and prct in Table 2, and then number of cells in Table 3) that deviated from master cell population.Fig. 2 CV-mean plot of data under different α and β. Other parameters were fixed as gene number = 15,000, cell number = 1,000 cells, prct = 10 %, and s = 2 Table 1 Estimation of model parameters α^andβ^ under different α and β with fixed number of cells, prct, s, and gene number = 15,000. Comparing with the noise model in Eq. 9, we have obtained fairly low RMSE in each condition Simulation parameters Regression results # of cells Prct (%) s α β α^ β^ RMSE RMSE (Eq. 9) 1,000 10 2 0 1 0.0003 ± 0.0002 1.0293 ± 0.0014 0.0074 ± 0.0006 0.044 ± 0.010 1.2 0.0004 ± 0.0003 1.2187 ± 0.0024 0.0047 ± 0.0006 0.066 ± 0.020 1.5 0.0007 ± 0.0004 1.5032 ± 0.0039 0.0028 ± 0.0007 0.091 ± 0.019 0.15 1 0.1557 ± 0.0005 1.0116 ± 0.0009 0.0047 ± 0.0003 0.049 ± 0.008 1.2 0.1562 ± 0.0007 1.1965 ± 0.0020 0.0038 ± 0.0004 0.030 ± 0.004 1.5 0.1569 ± 0.0007 1.4756 ± 0.0047 0.0043 ± 0.0006 0.017 ± 0.001 0.5 1 0.5146 ± 0.0013 1.0020 ± 0.0010 0.0038 ± 0.0004 0.060 ± 0.006 1.2 0.5161 ± 0.0011 1.1837 ± 0.0023 0.0039 ± 0.0003 0.047 ± 0.006 1.5 0.5187 ± 0.0016 1.4561 ± 0.0045 0.0054 ± 0.0005 0.030 ± 0.004 Table 2 Estimation of model parameters α^andβ^ under different prct and s with fixed number of cells, α, β, and gene number = 15,000 Simulation parameters Regression results # of cells α β s Prct (%) α^ β^ RMSE RMSE (Eq. 9) 1,000 0.15 1.2 1 10 0.1563 ± 0.0005 1.1965 ± 0.0018 0.0037 ± 0.0003 0.028 ± 0.002 30 0.1579 ± 0.0006 1.2017 ± 0.0019 0.0048 ± 0.0003 0.026 ± 0.001 50 0.1612 ± 0.0009 1.2076 ± 0.0023 0.0071 ± 0.0005 0.027 ± 0.001 2 10 0.1563 ± 0.0005 1.1961 ± 0.0019 0.0040 ± 0.0005 0.033 ± 0.007 30 0.1612 ± 0.0017 1.2015 ± 0.0024 0.0077 ± 0.0014 0.036 ± 0.001 50 0.1713 ± 0.0014 1.2080 ± 0.0024 0.0147 ± 0.0012 0.050 ± 0.002 3 10 0.1572 ± 0.0012 1.1963 ± 0.0026 0.0056 ± 0.0009 0.048 ± 0.008 30 0.1649 ± 0.0010 1.1997 ± 0.0027 0.0122 ± 0.0011 0.054 ± 0.002 50 0.1775 ± 0.0012 1.2078 ± 0.0030 0.0225 ± 0.0011 0.096 ± 0.003 Table 3 Estimation of model parameters α^andβ^ under number of cells with fixed α, β, prct, s, and gene number = 15,000 Simulation parameters Regression results Prct (%) s α β # of cells α^ β^ RMSE RMSE (Eq. 9) 10 2 0.15 1.2 50 0.1595 ± 0.0024 1.1078 ± 0.0040 0.0127 ± 0.0007 0.037 ± 0.007 100 0.1587 ± 0.0023 1.1416 ± 0.0044 0.0085 ± 0.0009 0.034 ± 0.006 500 0.1575 ± 0.0009 1.1836 ± 0.0036 0.0047 ± 0.0008 0.032 ± 0.008 Estimation of model parameters (α and β) We firstly fixed the data set size 15,000 genes and 1,000 cells with prct = 10 % and s = 2, only the model parameters α and β were changed, and the fit results of simulation data are shown in Fig. 2. When α = 0 and β = 1, we simply simulated the data in a Poisson distribution, following a diagonal line in the figure. When α became larger, the curve angled more prominent, which indicated data deviated from Poisson distribution at the larger expression level. The increase of β resulted in the entire data shifting away from the diagonal line, which might be associated with different sources of technical noises. We observed the robust parameter estimation as shown in Table 1 in all initial model parameters (with RMSE less than 0.01 for all these simulated cases). We noted that sometimes the current model failed to fit a straight line when α = 0, which we will investigate further for regression procedures at higher expression level specifically. When the input parameter β became larger, the two estimated model parameters were deviated from the input parameters. However, even in the extreme case where α = 0.5 and β = 1.5, the RMSE still very consistent in our model (0.0054 ± 0.0005, see Table 1). On the other hand, the orange line - the simple noise model fitting using Eq. 9, can hardly fit the baseline of the simulated data, which results in high RMSE (~0.05, 10x larger than our proposed method) in most conditions. We further examined the number of significant VEGs under each condition. The pale green points in the log(μ)-log(CV) plots in Fig. 2 were the selected as significant VEGs with FDR < 0.05. In the ideal condition where α = 0 and β = 1, there are in average 940 genes changed by at least two fold change and we have detected around 700 VEGs. Along with the increase α and β, the number of significant VEGs decreased. In the condition where α = 0.5 and β = 1.5, there are only around 250 VEGs detected, where around 950 genes are altered by at least two fold change. It is reasonable since the data are more disperse when α and β become larger. The dispersion affects the fitted normal distribution of CV difference while determining the p-value for VEGs, which results in worse FDR when the model parameters are large. Test estimation robustness with varying degree of heterogeneity of cell population Next we tested the performance of model under different percentage of genes affected by random log2 fold change values, which were generated by a normal distribution with zero mean and standard deviation s (Eq. 8). The data set size was still set as 15,000 genes and 1,000 cells, and we fixed the model parameters where α = 0.15 and β = 1.2. From the results in Table 2, we could observe that model parameter β is mostly identical and remained close to 1.2 under different levels and numbers of variable genes. However, the model parameter α became larger (from 0.156 to 0.178) with the increments of s and prct. This is unavoidable because α represents the dispersion of the data set. With more genes deviated from the homogeneous population, the dispersion increased and estimated α biased from the input model parameter value. Due to the deviation of α, RMSE also increased when s and prct became larger. We concluded that the scale and number of variable genes influence the estimation of model parameter α, which results in the increase of RMSE. Nevertheless, this issue is solved during the determination of the distribution of CV difference, where we use kernel density to adjust the center of the normal distribution. Test estimation robustness with varying number of cells At last, we would like to know if the model could be properly fit with limited number of cells. We reduced the population size to 50, 100, or 500 cells. To test under a moderate variation condition, we set prct = 10 %, and s = 2, with model parameters remained as α = 0.15 and β = 1.2. The results in Table 3 show that reducing the number of cells slightly affected the estimation of α: α is larger when the number of cells is smaller, in which CV of genes are more disperse. The estimation of β also deviated a bit with the decrease of the population size. Under 50 and 100 cells conditions, the scattering of the data points around the diagonal line resulted in the estimation error of β and a higher RMSE in lower number of cells. Moreover, the two factors that influenced the estimation of α and β also played a role in calling significant VEGs. Under the same number of genes, we determined only about 355 VEGs in 500 cells condition, whereas about 596 VEGs were called in 50 cells condition. With only a small number of cells, the normal distribution of homogeneous genes is difficult to estimate accurately, which might result in the increase or decrease of detected VEGs. Hence, a sufficient number of cells is necessary to accurately determine VEGs among a cell population. In conclusion, the major factors that influence the robustness of the noise model are how data distributes and the number of cells. Fitting errors arise from two situations, 1) the estimated parameters are unusually large (especially β) in the simulated cases, which is unlikely in real scRNA-seq data, 2) the data distribute close to the diagonal line in the CV-mean plot, but with many variable genes at higher expression level, which results in the failure of fitting a straight line. The cell population size is also a concern; however, in reality a single-cell experiment should be designed with a large number of cells. Hence, the population size may not be a major factor for most single cell applications. From the simulation results we could find out that a simple fitting method is not enough. By fitting the model in Eq. 9 straightforwardly, we got much larger RMSE in every condition. In contrast, our expression variation model design with multiple layers of estimates can be fitted properly for most of the experiment condition. However, in some cases the fitted model curve (red) lay under the local fit curve (cyan) at the middle mean abundance interval, which it might be a potential problem occasionally. Application on real data sets We have identified the VEGs for the two scRNA-seq data sets, and the respective CV-mean plots are shown in Fig. 3. From Fig. 3a, we can see that most genes in the first data set (GSE65525) distribute nearby the diagonal line, inferring that the data were only affected slightly by technical noises. Part of the fitted model overlaps with the Poisson distribution line until the mean abundance is larger than 1. Foreseeably, the two model parameters are close to the ideal case, we estimated that α = 0.044 and β = 1.260. In Fig. 3b, the cyan line is the kernel density estimation of CV difference to find the peak of the normal distribution of homogeneous genes. Using the left side of the peak, the red line is the fitted normal distribution and we identified 883 VEGs with FDR less than 0.001.Fig. 3 a CV-mean plot of data set GSE65525 and b the CV difference histogram The second data set, GSE60361 shown in Fig. 4a, is much more disperse and deviated away from the diagonal line. However, our method still fitted a reasonable noise model. Even though the local fit curve was terminated around μ = 10, the extension of the noise model at tail interval fitted well. The model parameters where α = 0.558 and β = 2.356 are much larger, and the histogram of CV difference is also widely distributed. Similar with the simulation case with high percentage of variable genes, the fitted model can't locate accurately on the center of the normal distribution of homogeneous genes. In Fig. 4b, we estimated the normal distribution where the peak is around −0.2. As a result, 3103 genes were defined as VEGs, which is a very large number. We found out that the average UMIs of each cell in the second data set is only around 14,000, which is far less than the first data set with around 29,500 UMIs. The small number of UMI counts results in large dispersion of data and detecting a large number of VEGs. Clearly, the total UMI reads per cell in this data is too small to obtain a precise estimation of model parameters. Additional simulation perhaps is needed to further evaluate the requirement of effect of number of UMIs for single cell study.Fig. 4 a CV-mean plot of data set GSE60361 and b the CV difference histogram Determination of single-cell subpopulations After the determination of VEGs, we can use different conventional bioinformatics tools to further study the heterogeneity and subpopulation of single-cell population. Principal component analysis (PCA) can be used to find out possible subpopulations among the entire single-cell population. Here we picked the first data set to demonstrate the subsequent scRNA-seq analysis. First, we used principal component analysis (PCA) on the log-transformed data of 883 selected genes to observe the heterogeneity among all cells, shown in Fig. 5. We could find some possible subpopulations at the left, top left, right, and bottom corners, which were labeled in different colors in Fig. 5 After we determined subpopulations from the PCA result, other methods can be applied to study the heterogeneity of the cell population: using the principal component (PC) loadings to classify the genes; or using Single-Cell Differential Expression (SCDE) [19] and/or DESeq [20] algorithms to identify differential expressed (DE) genes between different subpopulations. We can further perform functional annotation and pathway analyses on identified DE genes to understand the origins of cell heterogeneity.Fig. 5 3-D PCA plot of data set GSE65525 Even though the two scRNA-seq experiments obtained from GEO database used two different techniques to capture single cells with vastly different distributions in the CV-mean plots as shown in Figs. 3 and 4 , we could fit the expression variation models properly for both data. In the previous two studies [11, 12], it has been demonstrated that, using VEGs, cell heterogeneity has been detected along with associated biological functions of subpopulations. Clearly, finding the VEGs of a single-cell experiment is just the first step. The subsequent analyses that utilizing VEGs and their expression changes across the cell population are the key of single-cell RNA-seq analysis. Conclusion In this paper, we proposed a single cell gene expression variation model, and demonstrated the method to regress the model parameters for a single-cell RNA-seq experiment by exploiting the relationship between the coefficient of variation and mean transcript abundance of all genes in the genome. A single-cell data simulation was also designed and used to determine the robustness of the model parameter estimation. In most condition the model parameters were estimated precisely, and resistant to the influence of factors such as population size, and dispersion of genes. The results of testing on two real scRNA-seq data sets further confirmed our simulation, while additional modeling requirement due to lower total UMI count per cell warrants further investigation. Abbreviations CV, Coefficient of Variation; DE, Differential Expression; DEG, Differentially Expressed Gene; GEVM, Gene Expression Variation Model; NB, Negative Binomial (distribution, model); NGS, Next Generation Sequencing; PC, Principal Component; PCA, Principal Component Analysis; RMSE, Root-mean Square Error; scRNA-seq, single-cell RNA-seq; UMI, Unique Molecular Identifier; VEGs, Variably Expressed Genes Acknowledgement This research was supported in part by the Genome Sequencing Facility of the Greehey Children's Cancer Research Institute, UTHSCSA. Fundings for this research were provided partially by the NCI Cancer Center Shared Resources (NIH-NCI P30CA54174), NIH (CTSA 1UL1RR025767-01 to YC and CPRIT (RP120685-C2) grants to YC and HHC, and NIGMS (R01GM113245) to YH and YC. Declarations The publication costs for this article were funded by the CPRIT grants to YC mentioned below. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The R scripts of the algorithm and the UMI data for GSE65525 will be available from GitHub, https://github.com/hillas/scVEGs. Authors' contributions All authors contribute to the manuscript. HHC, YJ, YH and YC conceived and designed the study. HHC carried out the simulation procedure and implemented the algorithm in R. All authors read and approved the final manuscript. Competing interests Authors declare no competing interest in preparing the paper and developing the software associated to this paper. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Stegle O Teichmann SA Marioni JC Computational and analytical challenges in single-cell transcriptomics Nat Rev Genet 2015 16 3 133 145 10.1038/nrg3833 25628217 2. Taniguchi K Kajiyama T Kambara H Quantitative analysis of gene expression in a single cell by qPCR Nat Methods 2009 6 7 503 506 10.1038/nmeth.1338 19525960 3. Livak KJ, Wills QF, Tipping AJ, Datta K, Mittal R, Goldson AJ, Sexton DW, Holmes CC. Methods for qPCR gene expression profiling applied to 1440 lymphoblastoid single cells. Methods. 2013;59(1):71–9. 4. Sanchez-Freire V Ebert AD Kalisky T Quake SR Wu JC Microfluidic single-cell real-time PCR for comparative analysis of gene expression patterns Nat Protoc 2012 7 5 829 838 10.1038/nprot.2012.021 22481529 5. Raj A van den Bogaard P Rifkin SA van Oudenaarden A Tyagi S Imaging individual mRNA molecules using multiple singly labeled probes Nat Methods 2008 5 10 877 879 10.1038/nmeth.1253 18806792 6. Kim JK Marioni JC Inferring the kinetics of stochastic gene expression from single-cell RNA-sequencing data Genome Biol 2013 14 1 R7 10.1186/gb-2013-14-1-r7 23360624 7. Tang F, Barbacioru C, Wang Y, Nordman E, Lee C, Xu N, Wang X, Bodeau J, Tuch BB, Siddiqui A, et al. MRNA-Seq whole-transcriptome analysis of a single cell. Nat Methods. 2009;6(5):377–82. 8. Ramskold D, Luo S, Wang YC, Li R, Deng Q, Faridani OR, Daniels GA, Khrebtukova I, Loring JF, Laurent LC, et al. Full-length mRNA-Seq from single-cell levels of RNA and individual circulating tumor cells. Nat Biotechnol. 2012;30(8):777–82. 9. Streets AM Huang Y How deep is enough in single-cell RNA-seq? Nat Biotechnol 2014 32 10 1005 1006 10.1038/nbt.3039 25299920 10. Saliba AE Westermann AJ Gorski SA Vogel J Single-cell RNA-seq: advances and future challenges Nucleic Acids Res 2014 42 14 8845 8860 10.1093/nar/gku555 25053837 11. Klein AM, Mazutis L, Akartuna I, Tallapragada N, Veres A, Li V, Peshkin L, Weitz DA, Kirschner MW. Droplet barcoding for single-cell transcriptomics applied to embryonic stem cells. Cell. 2015;161(5):1187–201. 12. Zeisel A, Munoz-Manchado AB, Codeluppi S, Lonnerberg P, La Manno G, Jureus A, Marques S, Munguba H, He L, Betsholtz C, et al. Brain structure. Cell types in the mouse cortex and hippocampus revealed by single-cell RNA-seq. Science. 2015;347(6226):1138–42. 13. Robinson MD McCarthy DJ Smyth GK EdgeR: a bioconductor package for differential expression analysis of digital gene expression data Bioinformatics 2010 26 1 139 140 10.1093/bioinformatics/btp616 19910308 14. Marinov GK, Williams BA, McCue K, Schroth GP, Gertz J, Myers RM, Wold BJ. From single-cell to cell-pool transcriptomes: stochasticity in gene expression and RNA splicing. Genome Res. 2014;24(3):496–510. 15. Chen Y Kamat V Dougherty ER Bittner ML Meltzer PS Trent JM Ratio statistics of gene expression levels and applications to microarray data analysis Bioinformatics 2002 18 9 1207 1215 10.1093/bioinformatics/18.9.1207 12217912 16. Benjamini Y Hochberg Y Controlling the false discovery rate: a practical and powerful approach to multiple testing J R Stat Soc Ser B 1995 57 289 300 17. Hashimshony T Wagner F Sher N Yanai I CEL-Seq: single-cell RNA-Seq by multiplexed linear amplification Cell Rep 2012 2 3 666 673 10.1016/j.celrep.2012.08.003 22939981 18. Islam S, Zeisel A, Joost S, La Manno G, Zajac P, Kasper M, Lonnerberg P, Linnarsson S. Quantitative single-cell RNA-seq with unique molecular identifiers. Nat Methods. 2014;11(2):163–6. 19. Kharchenko PV Silberstein L Scadden DT Bayesian approach to single-cell differential expression analysis Nat Methods 2014 11 7 740 742 10.1038/nmeth.2967 24836921 20. Anders S Huber W Differential expression analysis for sequence count data Genome Biol 2010 11 10 R106 10.1186/gb-2010-11-10-r106 20979621

PMC005xxxxxx/PMC5001206.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556158291110.1186/s12864-016-2911-zResearchComprehensive comparison of molecular portraits between cell lines and tumors in breast cancer Jiang Guanglong ggjiang@iu.edu 12Zhang Shijun shijzhan@umail.iu.edu 12Yazdanparast Aida ayazdanp@iupui.edu 12Li Meng li487@iupui.edu 12Pawar Aniruddha Vikram apawar@umail.iu.edu 12Liu Yunlong yunliu@iupui.edu 12Inavolu Sai Mounika sinavolu@iupui.edu 12Cheng Lijun lijcheng@iupui.edu 121 Center for Computational Biology and Bioinformatics, School of Medicine, Indiana University, Indianapolis, IN 46202 USA 2 Department of Medical and Molecular Genetics, School of Medicine, Indiana University, Indianapolis, IN 46202 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.525© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Proper cell models for breast cancer primary tumors have long been the focal point in the cancer’s research. The genomic comparison between cell lines and tumors can investigate the similarity and dissimilarity and help to select right cell model to mimic tumor tissues to properly evaluate the drug reaction in vitro. In this paper, a comprehensive comparison in copy number variation (CNV), mutation, mRNA expression and protein expression between 68 breast cancer cell lines and 1375 primary breast tumors is conducted and presented. Results Using whole genome expression arrays, strong correlations were observed between cells and tumors. PAM50 gene expression differentiated them into four major breast cancer subtypes: Luminal A and B, HER2amp, and Basal-like in both cells and tumors partially. Genomic CNVs patterns were observed between tumors and cells across chromosomes in general. High C > T and C > G trans-version rates were observed in both cells and tumors, while the cells had slightly higher somatic mutation rates than tumors. Clustering analysis on protein expression data can reasonably recover the breast cancer subtypes in cell lines and tumors. Although the drug-targeted proteins ER/PR and interesting mTOR/GSK3/TS2/PDK1/ER_P118 cluster had shown the consistent patterns between cells and tumor, low protein-based correlations were observed between cells and tumors. The expression consistency of mRNA verse protein between cell line and tumors reaches 0.7076. These important drug targets in breast cancer, ESR1, PGR, HER2, EGFR and AR have a high similarity in mRNA and protein variation in both tumors and cell lines. GATA3 and RP56KB1 are two promising drug targets for breast cancer. A total score developed from the four correlations among four molecular profiles suggests that cell lines, BT483, T47D and MDAMB453 have the highest similarity with tumors. Conclusions The integrated data from across these multiple platforms demonstrates the existence of the similarity and dissimilarity of molecular features between breast cancer tumors and cell lines. The cell lines only mirror some but not all of the molecular properties of primary tumors. The study results add more evidence in selecting cell line models for breast cancer research. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2911-z) contains supplementary material, which is available to authorized users. Keywords HeterogeneousBreast cancerDNA mutationmRNA expressionCopy number alterationReverse-phase protein arrayMolecular portraitsCell linesThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background According to a recent World Health Organization report, breast cancer is the second most common type of cancer. Each year there are about 2300 new cases of breast cancer in men and 230,000 new cases in women in the U.S. [1]. While age and gender are two primary demographic risk factors in breast cancer, about 5–10 % of breast cancer risk is attributed to hereditary gene mutations in BRCA1, BRCA2 and TP53 [2]. Breast cancer is a complex disease. Its heterogeneous nature has been classified by its molecular characteristics. The protein expression status of estrogen receptor alpha (ER), progesterone receptor (PR), human epidermal growth factor receptor-2 (HER2) decide the group of breast cancers. It can be subtyped as Luminal A (ER+/PR+, HER2+), Luminal B (ER+/PR+, HER2-), HER2amp (HER2 positive) and Basal-like/triple negative (ER-,PR-, HER2-) [3, 4]. The Basal-like patients are correlated with biologically aggressive disease and often have a poor prognosis [3]. In Luminal A and Luminal B subtypes, ER was identified as the therapeutic target, and its targeted hormone therapies (such as tamoxifen and letrazole) have been well established. In HER2 amplification group, trasuszumab is the candidate drug. However, basal-like triple negative tumors still do not have recognizable therapies. The target identification and its subtype classification is an important aspect for therapy development in breast cancer [5, 6]. Cell lines, originated from human tumors, have historically acted as the primary experimental model to investigate the cancer biology and molecular pharmacology. Parallel massive drug screening on these cancer cells characterize the diverse cancer cell reactions to drugs by genomic features. As a salient example, the Cancer Cell Line Encyclopedia (CCLE) project conducts a detailed genetic characterization of a large panel of 997 human cancer cell lines in DNA copy number, mRNA expression and mutation [7]. Together with the drug screening data, CCLE becomes a powerful resource for the drug and target discovery researches. Breast cancer is heterogeneous in nature. Cell lines study is only an interpretation from a context of artifacts introduced by selection and establishment in vitro, and there exists large differences between cancer cell lines and tissue samples especially in its molecular genome [8, 9]. Selecting the right cells model to mimic tumor tissues helps to evaluate proper drug reactions in tumors in vitro [10, 11]. Gene-expression profiling has become an important tool to characterize both the similarity and dissimilarity between cell lines and tumors. A recent work by Ross DT [12] demonstrated the distinctive gene expression signature in breast cancer tissue: basal, luminal epithelial cell signature, as well as mesenchymal/stromal. Lacroix M [13] valuated some widely used breast cancer cell lines as breast tumor models by a comparative genetic expression features. Besides gene expression, CNV has gradually been recognized as important due to features in predicting cancer progression and recurrence. Jessica Kao et al. [14] compared the gene expression profiles and CNVs of breast cancer cells and tumor tissues to define relevant cell line models. Both Fridlyand et al. [10] and Richard M. et al. [15] conducted similar analyses, in which the similarity was further investigated within the breast cancer subtypes. Nevertheless, these researches provide important information for understanding a molecular mechanism from only one aspect of the breast cancer genome, such as mRNA or DNA or protein, but not both. No one has yet attempted to investigate the correlation between cell lines and tumor tissues from all CNV, mutation, gene expression and protein expression between and within breast cancer subtypes systematically. The Cancer Genome Atlas (TCGA) [15] aims to discover major cancer-causing genomic alterations. It publicly provides 1098 breast tumor samples with mRNA expression profiling, DNA exome parallel sequencing, CNV, and protein expression. Because of this valuable data, a number of important breast cancer genes and pathways were detected systematically during the past 3 years [16–18]. However, systematic comparisons between TCGA breast tumor samples and breast cell line data, such as Cancer Cell Line Encyclopedia (CCLE), have not yet been conducted. The primary innovation of this comparison is that, for the first time, four layers of genomic data: CNV, mutation, mRNA expression and protein expression, were investigated to seek the similarity or dissimilarity between breast cancer cells and tumors. Secondly, because of better sensitivity and broader dynamic range of sequencing technology comparing to the array platforms, genomic data was better captured in TCGA and CCLE by the platform data comparison. In this paper, a comprehensive comparison in CNV, mutation, mRNA expression and protein expression between CCLE breast cancer cell lines and TCGA primary breast tumors is presented separately. At the end, a total score that integrates four genomic features will be defined to investigate the overall similarity between breast cancer cell lines and its tumor tissues. Results Sixty-eight breast cancer cell lines were extracted from CCLE [7] and literature [19]. One thousand seven hundred five breast cancer tumor samples were obtained from TCGA and Gene Expression Omnibus (GEO). All of the datasets are listed in Table 1. Different subsets of samples were assayed on four different level platforms, including Affymetrix HU133 and Agilent G4502A_07_3 for mRNA expression microarrays irrespectively, Affymetrix 6.0 single nucleotide polymorphism (SNP) arrays for copy number variation, whole-exome sequencing in TCGA and hybrid capture sequencing 1651 genes in CCLE for mutation analysis. Reverse-phase protein lysate microarrays (RPPAs) are used to test basal phosphorylation and protein abundance in TCGA tumors and cell lines. Please note that not all samples were characterized on each platform. Different subsets of tumors and cell lines were analyzed in each platform (Additional file 1: Tables S1 and S2). Each one of the four platform data analyses focused on the overlapping genes between tumors and cell lines, and the overall similarity analysis by using all four platforms was conducted afterward. Figure 1 describes the overall analysis process between cell lines and tumors in breast cancer.Table 1 Four molecular profiles datasets for tumor and cell lines comparison in breast cancer Data types Sources Platforms Samples size Copy number variation TCGA;CCLE Affymatrix SNP 6.0 1033; 59 Mutation (Exome Sequencing) TCGA;CCLE Illumina GAIIx 967; 51 Gene expression TCGA; GEO; CCLE AgilentG4502A_07_3 (TCGA); Affymatrix HU133 Plus 2.0 (GEO; CCLE) 530; 279; 58 Protein TCGA; CCLE RPPA 197; 38 Fig. 1 The whole analysis process between cell lines and tumors in breast cancer using 4 genomic profiles. Sixty eight cell lines and 1375 tumors are compared in gene expression, copy number variation (CNV), mutation and protein across 10 aspects. A score that integrated four genomic features was used to evaluate the overall similarity of tumors and cell lines Gene expression profiles comparison between breast cancer cell lines and tumors PAM50 (Prediction Analysis for Microarrays) [20] is one of the most common genetic tests for breast cancer subtyping. The PAM50 was designed as a RT-qPCR 50-gene expression signature. It has been acknowledged as a prognostic gene signature assay by an authoritative organization, National Comprehensive Cancer Network (NCCN) (http://www.nccn.org/), in year 2015. Due to this, many breast tumor and cell line samples lacked of ER, PR, and HER2 status for breast cancer treatment classifications. As for the missing information of HER2 status, it has 182 in 1096 TCGA tumors and 15 in 68 CCLE cell lines. These samples are classified as subtypes of Luminal A, Luminal B, HER2amp, and Basal-like using the PAM50 signature. On the other hand, the RT-qPCR and mRNA-based PAM50 ER/PR/HER2 classification results are compared. Figure 2 displays the PAM50 gene expression signature predicted subtypes of cell lines and tumors in breast cancer, and the observed ER, PR, HER2 status. Eight hundred seventy five TCGA samples have information of ER/PR/HER2 status in 1096 tumors, while 53 cell lines in 68 CCLE samples have those. Figure 2a shows the PAM50 subtypes of 530 invasive breast cancer patients in TCGA using AgilentG4502A_07_3 array platform. Comparing to the standard ER, PR, and HER2 status for classification of breast carcinoma by using immunohistochemistry staining (Table 2), 341 tumors with PAM50 classification are in concordance with the standard classification in 514 tumors, where the normal-like patients (other) are excluded. The concordance rate is 66.3 %. Figure 2b shows the breast cancer subtype classification of 56 breast cancer cell lines in CCLE using 50 genes PAM analysis. Gene expression profile in CCLE was conducted in Affymetrix Hu133 Plus2.0 Array platform. Thirty-four cell lines with known classification are in concordance with PAM50 classification, and the concordance rate is 60.71 % (34/56). Some cell lines without ER/PR/HER2 status, such as KPL1, ZR751, HS742T, HS60T, HS281T, HS343T, HS274, received ER/PR/HER2 imputation from the PAM50 prediction. In the follow-up data analysis, we kept the known classification and imputed PAM50 for both cell lines and tumor samples. Additional file 1: Tables S1 and S2 list the classification results for cell lines and tumors based on PAM50 gene expression. Interestingly, we observed that the gene expression pattern of PAM50 between cell lines and tumors are similarity, but some genes in cell lines are not as highly expressed as in tumors, such as gene FOXA1 and ESR1.Fig. 2 Gene expression PAM50-breast cancer subtype classifications of cell lines and primary tumors for ER, PR, Her2 status. a The PAM50 subtype classification of 530 invasive breast cancer samples in TCGA, which uses AgilentG4502A_07_3 Array platform. b The PAM50 subtype classification of 56 breast cancer cell lines, which uses Affymetrix Human Genome U133 Plus 2.0 Array platform Table 2 Molecular classification of breast carcinoma Classification Immunoprofile Other characteristics Luminal A ER+/PR+/HER2-; ER+/PR-/HER2-; ER-/PR+/HER2-, Low tumor grade, Low expression of proliferation marker Ki67 Luminal B ER+/PR+/HER2+; ER+/PR-/HER2+; ER-/PR+/HER2+ High tumor grade, High expression of proliferation marker Ki67 HER2-enrichment ER-/PR-/HER2+; ER-/PR-/HER2+; ER-/PR-/HER2+ High tumor grade, High expression of proliferation marker Ki67 Basal-Like ER-/PR-/HER2- High tumor grade, High expression of proliferation marker Ki67 In order to compare the similarity of the whole genome expression profiles between primary breast cancer tumors and breast cancer cells (i.e. CCLE samples), the breast cancer tumors in Gene Expression Omnibus (GEO) GSE41998 (279 tumors) were selected because they shared the same Affymetrics gene expression platform (Additional file 1: Table S3). Figure 3 shows the correlation distributions of whole genome expression between breast cancer cell lines and primary tumors. The 56 box plots of the correlations illustrate the similarity between 56 cell lines and 279 tumors. The correlation coefficient is around 0.6–0.8 between cell lines and tumors. These results show that cell lines keep a high similarity to tumors in whole gene expression profile in breast cancer even though in different subtypes.Fig. 3 Whole genome expression correlation analysis between 56 breast cancer cell lines and 279 primary tumors. The x-axis indicates 56 cell lines, and y-axis is the correlation coefficient between tumors and cells. The cell line subtypes are denoted in different colors Copy number variations comparison between CCLE breast cancer cell lines and TCGA breast cancer tumors CNVs are compared between CCLE breast cancer cell lines and TCGA breast cancer primary tumors in various breast cancer subtypes. Figure 4 displays copy number distribution for both tumors and cell lines across 24 chromosomes. In Fig. 4a, chromosome 1 and 8 have the highest copy number amplification frequencies while chromosomes 13 and 16 have the most copy number deletion regions in both cell lines and tumor tissues. Figure 4b displays the significant genomic alterations in breast cancer tumors and cell lines. MYC, PVT1, RAD21 and TRPS1 are top four copy number amplified genes, while MAP2K4, ANKRD11, APRT, CSMD1 and ZFPM1 are top five genes with copy number deletions. Some important cancer genes, such as PIK3CA, BRCA1, BRCA2, and ERBB2, show a mixture of amplifications and deletions.Fig. 4 The CNV comparison between 1049 TCGA breast cancer samples and 59 CCLE breast cancer cell lines. a DNA copy number profiles across the whole chromosomes (right) and copy number amplification and deletion in highly mutated breast cancer driver genes. The red color indicates amplification, while the blue color indicates deletion. The top panel is TCGA, and the bottom panel is CCLE. b is the CNV frequency comparison for highly mutated genes between TCGA breast cancer samples and CCLE breast cancer cell lines The CNVs between cell lines and tumor samples of breast cancer are compared in sample segmentation mean and density calculation of copy number Fraction Genome Altered (FGA). Its calculation is presented in the method section. Figure 5a demonstrates that cell lines have more copy number deletions than tumors. In particular, HCC1599, MDA-MB-361, MDA-MB-157, and UACC893 are the top 4 CNV deletions cell lines. In Fig. 5b, it is evident that the frequency of copy number alteration are significantly higher in cell lines than in tumors. The mean cell line FGA is wider than that of tumor FGA. In order to evaluate the similarity between tumors and cell lines, the Pearson correlations for the top 10 % CNV in 2094 genes are calculated between 59 cell lines and 1049 tumors. Fig. 5c shows the CNV-based correlation coefficient distribution between cell lines and tumors in different breast cancer subtypes. We observe that cell lines HCC2218, MDA-MB-175-VII, ZR-75-30, BT-483, HCC1569 and MDA-MB-453 are more similar to tumors in CNV than the other cancer cells. Their correlation coefficients are larger than 0.55 (p < 10−18). On the other hand, HMEL, Hs 578 T, Hs 274.T, Hs 606.T, Hs 281.T, Hs 739.T, CAL-51, Hs 343.T and Hs 742.T had negative correlation coefficients with tumors samples (p < 10−2).Fig. 5 CNV similarity analysis between cell line and tumors in breast cancer. a is the average segmentation length comparison between cell lines and tumors samples, here the segmentation length is measured according to log 2(CN/2)). Positive number means amplification, and negative number means deletion; b shows the density distribution of copy number variation in breast tumors and cell lines; c shows CNV-based similarity between cell line and tumors in breast cancer according to Pearson correlation coefficient using the top 10 % genes (2094 genes) Mutation analysis in cell lines and tumors CCLE sequenced only 1347 cancer genes in breast cancer, while TCGA has whole exome sequencing. Our comparative analysis is only based on those 1347 overlapping genes and their somatic mutations. In CCLE, in order to remove background germline mutation, mutations reported in the 1000 Genome Project and dbSNP were filtered out using ANNOVAR tool, including the gene-based single nucleotide variants (SNVs) and insertions/deletions [21]. Figure 6 shows the comparisons of somatic mutations between cell lines and tumors across four aspects: somatic mutation frequency, somatic mutation density, average mutation sites distribution per million bases (Mb) in four subtypes, as well as mutation correlation variation between cell lines and tumors. Figure 6a illustrates the mutation frequency per Mb in TCGA and CCLE vs CNV fraction genome alteration. A subset of cell lines with hyper-mutated genes is revealed, such as MDAMB361, BT474, MDAMB453 and HCC1569. These cells of breast cancer show moderately higher mutation frequency than the tumors. Figure 6b shows the somatic mutation density. The median somatic mutational frequency for tumors in TCGA is around 13, while cell lines in CCLE is around 25. Figure 6c shows the somatic mutation distribution among four subtypes of breast cancer in TCGA and CCLE, where y-axis is the mutation rate per million bases and x-axis is mutation gene numbers. The wider the line is, the more the gene mutation number of samples is. It suggests that the gene mutation number in Luminal B subtype from TCGA is the largest. At the same time, its mutation rate is also higher than the other subtypes. Tumor and cell lines with Luminal A subtype have the lowest mutation numbers and mutation rate. Her2 subtype group in cell lines has a larger mutation number than the other subtypes. Figure 6d shows the 1347 somatic mutation genes-based correlation coefficient distributions between cell lines and tumors in different breast cancer subtypes. These genes were firstly denoted as 0 or 1 to illustrate non-mutation or mutation. The correlation is distributed in the range of [-0.1, 0.43]; Additional file 2: Table S7 shows the detail correlation coefficient between cell lines and tumors in four levels for gene expression, mutation, copy number variation and protein irrespectively. The top four cell lines that have the highest mutational correlation with tumors are: UACC893, JIMT1, EFM19 and HCC1954. The highest consistency coefficient is 0.4258.Fig. 6 DNA sequencing-based mutation comparison between CCLE 51 cell lines and TCGA 977 tumors. a The scatter plot of fraction genome altered and mutation per million bases for TCGA samples and CCLE samples; b The mutation densities in breast tumors and cell lines; c Mutation-based subtypes similarity of cell line and tumors in breast cancer; d Mutation correlation coefficient distributions between cell lines and tumors Thirty-one genes, reported in recent TCGA nature and science papers [16–18, 22–28], were selected as important driver mutation genes in the breast cancer. These genes were further investigated across 51 breast cancer cell lines. Figure 7 shows a landscape of these functional driver mutations in these cell lines of breast cancer. According to the mutation per megabyte base calculation, HCC1569, MDAMB361, and BT474 are hyper-mutated cell lines, while HS 281 T, HS 343 T, and ZR 751 are lowly mutated cell lines. The popular cell lines MCF7 and MDAMB231 have median mutation rates. The top mutated genes in breast cancer tumors are TP53 (31 %) and PIK3CA (33 %). TP53 has copy number deletion in almost all cell lines, and has mixed somatic mutation. CNV has a dominant role in PIK3CA across 19 cell lines with mixed somatic mutations. Genome integrity pathway genes, ATM, BAP1, BRCA2, TTN and TP53, almost all have strong gene copy number amplification in cell lines mixed with somatic mutation, except for TTN. Similar data has been observed in genes MAP2K4 and MAP3K1 on MAPK signaling pathway. Genes PRKCA, PTGS2 and ZNF217 have many copy number deletions. The important drug biomarkers BRAF and ERBB2 (HER2) are relatively conservative, which do not have much somatic mutations.Fig. 7 The landscape of functional driver mutations in cell lines of breast cancer. Upper rows show the gene mutation frequency and mutation rate per million bases (Mb) in 967 tumors. Left column shows the popularity of breast cancer cell lines denoted by the publication citation number in Pubmed and mutated rate per Mb in cell lines. Point mutations (germline mutation and somatic mutation) and copy number variation (CNV amplification is segment-mean > 0.3, CNV deletion is segment-mean < -0.3) are shaped into the horizontal bar and vertical bar with different color, respectively A comparison of mutation spectra across four subtypes (Fig. 8) reveals that the mutation transition rates of cell lines and tumors are similar within different subtypes. On the other hand, it can be observed that breast cancer contains larger C > T and C > G trans-versions in subtypes HER2amp and Luminal B. HER2amp has the highest C > T trans-version rate. Luminal A has the highest A > C trans-version. Figure 8b shows the correlation of six mutation categories in tumors and cell lines. It suggests that C > T and C > G trans-version possess the highest concordance between tumors and cell lines. Basal-Like subtypes between cells and tumor tissue are consistent in A > T and C > G transition, while only A > G trans-version showed the correlation between tumors and cell lines in subtypes of Luminal B.Fig. 8 Mutation spectra and contexts across 4 subtypes of breast cancer. a Mutation spectrum of six transition (Ti) and transversion (Tv) categories for each subtypes of breast cancer (Luminal A = LA, Luminal B = LB, HER2amp = HER2 and Basal-Like = BaL). b Hierarchically clustered mutation context (defined by the proportion of A, T, C and G nucleotides within + -2 bp of variant site) for six mutation categories. Colour denotes degree of correlation: red (r = 1), yellow (r =0.5), green (r = 0), blue (r = -1) Comparison analysis of proteins phosphorylation expression between cell lines and tumors in breast cancer Quantitative expression of 50 cancer-related proteins, phosphorylated-proteins by RPPA, were measured on 197 breast tumors and 38 cell lines. Pearson Correlation analysis and unsupervised hierarchical clustering analyses were conducted between cell lines and tumors (Fig. 9). The correlations in Fig. 9a suggest that all four cell line subtypes possess different correlation distributions with tumor samples. Luminal B cells have the highest correlations, while basal cells have the lowest correlations and also show the largest variations. Figure 9d illustrated hierarchy distance among cell lines. It suggests that the same subtype cell lines usually are closely clustered. Protein expressions for ER and PR have high concordance, and they are reversely correlated with Coveolin.1 in all subtypes, especially in the Basal-Like subtype. A similar variation phenomena was observed in several other groups’ of proteins in different subtypes: (EGFA, CCNB1), (4EBP1, MEK1), (mTOR, GSK3), and (GATA3, p70s6kp389, AKT). Correlations between cell lines and tumors are further illustrated in Fig. 9c. The correlation ranges from −0.61 to 0.84. Some cell lines, T47D, BT483, and AU565, are the top three cell lines that have closer correlations to tumors in protein level, while the most popular breast cancer cell line, MCF7, is somewhere in the middle. The exact correlations between cell lines and tumors are presented in Additional file 1: Table S6 based on 50 phosphor-proteins.Fig. 9 Protein-based comparison of cell lines and tumors in breast cancer according to 50 proteins in RPPA testing. a RPPA-based Pearson correlation distribution in different subtypes of cell lines in breast cancer. b RPPA-based cell lines correlation in breast cancer by Pearson correlation. The dot size expresses the correlation strong or weak, and the larger means it has a strong correlation. The color bar shows the positive or negative directions. c RPPA-based correlation distribution between cell lines and tumors in breast cancer by 50 phosphorylated proteins. A dot means a correlation coefficient between a cell and a tumor. The different colors represent the different breast cancer subtypes. The same subtype of tumors and cells are used to calculate their correlation coefficient. d RPPA-based cell lines hierarchy clusters in breast cancer. Rows are proteins while columns are cell line samples Figure 10 shows the hierarchical distance between cell lines and tumors based on the 50 phosphorylated-proteins. The cell lines and tumors are assembled together by these proteins. It clearly classifies these breast cancer samples into four distinctive subtypes. Interestingly, the Basal-like cell lines MDAMB436, SUM139PT and HCC2185 are similar to protein features of Luminal A subtypes in tumors. Another discovery is that the Basal-like cell line MDAMB453 is close to Luminal B tumors. All details of the result is referred to in Additional file 2: Table S7, protein RPPA correlation coefficient between cell lines and tumors.Fig. 10 50 protein RPPA-based hierarchical clustering between 197 tumors and 38 cell lines. Rows are different proteins and columns are tumors and cell lines samples. Two color bars represent subtypes of breast cancer and data types irrespectively Correlation analysis of gene expression verse phosphorylated protein expression between cell lines and tumors in breast cancer The correlations of the gene mRNA versus its phosphorylated protein was calculated in cell lines and tumors irrespectively. The average correlation coefficient (Fig. 11) of 38 genes’ mRNA with their 50 phosphorylated proteins concentration ranges from −0.3 to 0.9 both in cell lines and tumors. Nearly 60% of the genes had a positive correlation between mRNA and protein. ESR1 has the highest correlation coefficient r−0.89 in 173 TCGA tumors, and r = 0.68 in 29 CCLE cell lines of breast cancer between mRNA and protein. Drug-target genes, such as PGR, HER2, EGFR and AR, all have high correlation (r > 0.5, p < 0.01) between mRNA and protein both in TCGA tumors and cell lines. Two important oncogenes, GATA3 and RP56KB1, both have high mRNA- protein correlation. The correlation for GATA3 is 0.79 in cell lines and 0.81 in tumors, while the correlation for RP56KB1 is 0.92 in cell lines and 0.78 in tumors. The small figure in Fig. 11a shows the linear correlation of the gene-protein between cell lines and tumors, which the linear correlation coefficient is 0.7076 (p < 0.01). This strong signal indicates the consistency of gene expression and protein expression in both cell line and tumor. The potential discrepancy could be due to the stability of mRNA, the degradation of protein, the time dependent and site dependent nature of protein phosphorylation, and etc. The interesting result in the Fig. 11b illustrates the gene expression amount are irrelevant to the correlation of mRNA-protein. As a matter of fact, the highest expressed gene RP56 has a negative correlation with mRNA-protein correlations in both cell lines and tumors.Fig. 11 The comparison between 38 genes mRNA expression and their phosphorylated proteins expression in tumors and cell lines. a The correlation comparison of mRNA verse phosphorylated protein in cell lines and tumors. b The 38 gene expression average in 29 cell lines What kinds of cell lines are close to tumors? Gene expression profiles and proteins phosphorylation expressions of tumors and cell lines were compared to further corroborate our observations made on the CNV and mutation data. The correlations of four different molecular profiles of all cell line and tumor pairs were calculated (Fig. 12a). These four correlations differ greatly from each other. Gene expression-based correlation had the largest correlation, CNV correlation was the next highest, mutation and protein expression correlations were low. These four correlations were combined into a total score as formula (2). Figure 12b shows the ranked cell lines by their average total correlations with the tumors. BT483, T47D, MDAMB453 are the true top 3 cell lines in breast cancer research.Fig. 12 The cell lines correlation degree with tumors in 4 molecular levels to breast cancer. a The correlations in 4 different molecular datasets. (GE = genes expression, CNV = copy number variation, MU = DNA exome sequencing mutation, PRRA = proteins phosphorylation expression). b Whole score between cell lines and tumors according to 4 different molecular dataset’s correlation Discussion Breast cancer is a highly complex disease. The subsets of breast tumors show diverse patterns of gene expression, CNV, mutation, and protein expression. A considerable amount of knowledge on breast carcinomas have been derived from in vivo and in vitro studies performed on breast cancer cell lines. Whether breast cancer cells are representative of the tumors remains debatable. In this study, the comparisons between cell lines and primary tumors from molecular profiles: gene expression, CNV, mutation, and protein expression, show that the cell lines are similar to some but not all of the primary tumors. Among them, gene expressions have the highest while the mutation-based correlation was the lowest.In gene expression-based clustering analysis, cell lines possess similar clustering as with tumors using PAM50. At the same time, cell lines show stable genomic and expression patterns, as well as high correlation, with tumors in whole gene expression profile. From the mutation comparison between cell lines and tumors, some common features were found: the chromosome 1 and 8 regions show high frequency copy number amplification, and chromosome 13 and 16 display high frequency deletions. Some significant cancer-related genomic alterations: MYC, PVT1, RAD21, TRPS1, CDH1, RB1, PIK3CA, MAP2K4, and ANKRD11, are identified in both breast cancer tumors and cell lines. The results were verified partially in reference [10]. In the single point mutation comparison, the six trans-version distribution modes of mutation spectrum demonstrates the similarity between tumors and cell lines in four breast cancer subtypes. High frequent C > T and C > G transitions are observed in both tumors and cell lines, while few A > T happens; Basal-like tumors and cells show the high concordance. These results were confirmed by Philip J. et al. [22]. They suggested that the underlying mutation mechanism is related to transcription-coupled nucleotide excision repair (NER). NER removes bulky DNA adducts that distort the DNA double helix and introduces a strand bias for mutation. However, little is known about the trans-version processes of mutation. In analyzing the cancer landmark genes, gene PIK3CA and TP53 in cell lines are the top 2 mutated genes that tumors have [26]. In addition, Luminal A subtype in cell lines possess hyper mutations in three genes GATA3, PIK3CA, and MAP3KI. HER2 subtype cell lines have 72 % and 39 % mutation rates for TP53 and PIK3CA, respectively. In the recent report [26], similar results in tumors were reported, in which Luminal A is dominated with a high PIK3CA mutation frequency and Luminal B had high PIK3CA and TP53 mutation frequency. HER2 cell lines have a high PIK3CA and TP53 mutations frequency in company with HER2 amplification [26]. In addition, important drug biomarkers, such as BRAF, ERBB2 (HER2), KRAS, have very low somatic mutation. All these evidences suggest that the cancer cell lines have very similar CNVs and gene mutations patterns as tumors. On the other hand, cell lines have more genetic aberrations than primary tumors. Amplification, deletion and mutation are more frequent in the cell lines than in the tumors. This is consistent with a similar study in ovarian cancer [8]. One potential interpretation is that cell lines may have transformed numerous passages over the period of cell culture time or get contaminated with stromal cells [10]. Another interpretation could be that the cell line is derived predominantly from early-stage tumors or pleural effusions [10].c) In protein expression-based comparison, breast cancer subtype proteins ER, PR and HER2 have a high consistence in cell lines and in tumors. RPPA can identify breast cancer subtypes clearly and accurately not only in cell lines but also in tumors according to these protein statuses. RPPA is a sensitive and accurate technology to evaluate protein expression and activities. It helps the target identification, validation, and drug discovery [29, 30]. Some cell lines, T47D, BT483, and AU565, have much closer protein expression than the popular MCF7 cell does. On the other hand, protein expression correlation between cell lines and tumors in breast cancer vary greatly ranging from −0.1 to 0.4, it is also true in the same subtype cell lines and the variation is particularly high investigated in the basal-like subtype. The results were supported by Sorger et al. [31], who investigated the immediate-early signaling that regulates the AKT (AKT1/2/3) and ERK (MAPK1/3) pathways in different breast cancer cell types. They found that cell lines have diverse to ligand sensitivity and signaling biochemistry. In addition, they found that the basal-like cells have the largest variations in responding to growth factors while HER2amp cell lines have the least variations [31]. Basal-like breast cancer is a highly heterogeneous group without proper drug targets yet. Brian D. et al. investigated the subtypes for basal-like breast cancer and preclinical models for targeted therapy selection [5]. According to BRCA1, AR, PIK3CA and PTEN mutations, drugs are selected in cell lines to predict preclinical TNBC targeted therapies. d) There are many complicated post-transcriptional mechanisms in turning mRNAs into proteins. According to correlation analyses between gene expression and phosphorylated protein expression in both cell lines and tumors, significant results are found that important drug targets in breast cancer, such as ESR1, PGR, HER2, EGFR and AR show high correlated mRNA and protein levels. High mRNA-protein correlation. Two oncogenes GATA3 and RP56KB1 with high consistency correlation between mRNA and protein expression become a promising potential drug targets. On the other hand, the gene expression variation at the mRNA level is not necessarily consistent with its protein level, such as genes TP53, KDR, DECAM1, which has been well documented in the literature [32, 33]. Most interestingly, the mRNA-protein correlation patterns comparing cell lines with primary tumors show a great deal of consistency among 38 investigated genes. However, the gene expression amount is irrelevant to the translation processing from mRNA to protein directly. e) In the whole score overall comparison, cell lines and tumors show high gene expression-based correlations, but the correlations in mutation and protein expression level are low. The possible reason is that mutation data is discrete, and mutation rate is low. According to PubMed search builder (http://www.pubmed.org) in year 2015, the number of citations for all breast cancer cell lines at CCLE is sorted (see Fig. 7). The most commonly studied cell lines are MCF-7, MDA-MB-231, MDA-MB-468 and SK-BR-3. They each have more than 600 PubMed citations. However, the correlation between these cell lines and tumors lies in the middle according to a total score of four molecular profile analyses. On the other hand, less popular cell lines, such as BT483, T47D, MDAMB453, are in the top 3 for representing breast tumors.f) Breast cancer subtypes in tumors and cell lines. The breast cancer cell line classification provides a cell modeling system to primary tumors. Our study addresses the classification results for cell lines and tumors based on PAM50 (Additional file 1: Table S1 and S2). Although some classification results are not consistent with the known classification in cell lines and tumors, the whole subtype’s concordance reaches more than 60 %. Any cell line’s usage as a tumor’s model depends upon its subtype’s speculation. A hypothesis based on gene expression will lead to different cell selection versus another hypothesis based on mutation. Conclusion In this paper, a comprehensive comparison in CNV, mutation, mRNA expression and protein expression between CCLE breast cancer cell lines and TCGA primary breast tumors is conducted and presented. The following are our primary conclusion. (1) PAM50 gene expression differentiated four major breast cancer subtypes, such as Luminal A and B, HER2amp, and Basal-like, in both cells and tumors. Using whole genome expression arrays, strong correlations are observed between cells and tumors. (2) Consistent CNV patterns are observed between tumors and cells across the chromosome. High C > T and C > G trans-version rates are observed in both cell lines and tumors, while cells have slightly higher somatic mutation rates than tumors. (3) Although the ER/PR/HER2 show the consistent patterns between cells and tumors, the other proteins in the RPPA platforms do not. Clustering analysis on protein expression data can reasonably recover the breast cancer subtypes in both cells and tumors. However, low correlations were observed between cells and tumors in phosphorylated proteins. (4) Nearly 50 % gene expressions are not consistent with their protein levels both in tumors and cell lines. The high and low of gene expression is irrelevant to the translation processing from mRNA to protein directly. Nevertheless, important drug targets in breast cancer, such as ESR1, PGR, HER2, EGFR and AR possess highly correlated in mRNA-protein expression both in tumors and cell lines. (5) A total similarity score developed from the four correlations among four molecular profiles suggests that cell lines, BT483, T47D and MDAMB453 have the highest similarity with tumors. Methods Data collection Four levels of molecular profiles: mRNA gene expression, CNV, mutation, and protein expression, were retrieved from TCGA, CCLE and GEO (Table 1). The study cohort of breast cancer consists of 1375 patients and 68 cell lines. Tumors data and annotations were downloaded from TCGA data portal (https://gdc-portal.nci.nih.gov/) with tumor matched selections and level 3 data. DNA exome sequencing data was available from 967 tumors. mRNA expression by AgilentG4502A_07_3 platform test was collected for 530 samples, while copy number alteration was detected using Affymetrix 6.0 single nucleotide polymorphism array (SNP- array) in 1033 tumors, and protein expression by RPPA in 197 tumors was obtained. The total number of breast cancer cell lines in CCLE was 59 [7, 13]. DNA copy number data (59 cell lines), mutation data (51 cell lines), mRNA expression data (56 cell lines) and their annotations originate from CCLE websites (http://www.broadinstitute.org/ccle). According to reference [26], 38 cell lines of RPPA data was downloaded. ER, PR, and HER2 genes statuses in cell lines are found from references [5, 10, 34–36]. To compare the mRNA expression values between cell lines and tumors of breast cancer, the same platform datasets in tissue were downloaded from the GEO data set (GSE41998). It consisted of 279 tumor samples [37] with the entity histopathology information. Table 3 shows all of the cell lines samples annotation and classification information which used in this paper. Additional file 1: Tables S1–S3 lists all samples annotation of cell lines and patients in this paper.Table 3 Cell lines annotation of breast carcinoma Cell line name Gender Hist subtype1 Source ER PR Her2 PAM50 mRNA Our classification AU565 F ATCC - - + Her2amp Her2amp BT-20 F ductal_carcinoma ATCC - - - Basal-like Basal-like BT-474 F ductal_carcinoma ATCC + + + Luminal B Luminal B BT-483 F ductal_carcinoma ATCC + + - Luminal A Luminal A BT-549 F ductal_carcinoma ATCC - - - Basal-like Basal-like CAL-120 F DSMZ - - - Basal-like Basal-like CAL-148 F ductal_carcinoma DSMZ - - - Luminal B Basal-like CAL-51 F DSMZ - - - Basal-like Basal-like CAL-85-1 F DSMZ - - - Basal-like Basal-like CAMA-1 F ATCC + - - Luminal B Luminal A DU4475 F ATCC - - - Basal-like Basal-like EFM-19 F ductal_carcinoma DSMZ + + - Luminal B Luminal A EFM-192A F DSMZ + - + Her2amp Luminal B EVSA-T F DSMZ + - + NON Luminal B HCC1143 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC1187 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC1395 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC1419 F ductal_carcinoma ATCC - - + Her2amp Her2amp HCC1428 F ATCC + + - Luminal B Luminal A HCC1500 F ductal_carcinoma ATCC - - - Luminal A Basal-like HCC1569 F metaplastic_carcinoma ATCC - - + Basal-like Her2amp HCC1599 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC1806 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC1937 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC1954 F ductal_carcinoma ATCC - - + Her2amp Her2amp HCC202 F ductal_carcinoma ATCC - - + Her2amp Her2amp HCC2157 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC2218 F ductal_carcinoma ATCC - - + Luminal A Her2amp HCC38 F ductal_carcinoma ATCC - - - Basal-like Basal-like HCC70 F ductal_carcinoma ATCC - - - Basal-like Basal-like HDQ-P1 F ductal_carcinoma DSMZ - - - Basal-like Basal-like HMC-1-8 F HSRRB NON Luminal A Hs 274.T F ATCC Basal-like Luminal B Hs 281.T F ATCC Basal-like Her2amp Hs 343.T F ATCC Basal-like Her2amp Hs 578 T F ductal_carcinoma ATCC - - - Basal-like Basal-like Hs 606.T F ATCC Luminal A Luminal B Hs 739.T F ATCC Basal-like Basal-like Hs 742.T F ATCC Luminal A Luminal A JIMT-1 F ductal_carcinoma DSMZ - - Basal-like Her2amp KPL-1 F ductal_carcinoma DSMZ Luminal A Basal-like MCF7 F ATCC + + - Luminal A Luminal A MDA-MB-134-VI F ductal_carcinoma ATCC + - Luminal A Luminal A MDA-MB-157 F ductal_carcinoma ATCC - - - Basal-like Basal-like MDA-MB-175-VII F ductal_carcinoma ATCC + - Luminal B Luminal A MDA-MB-231 F ATCC - - Basal-like Basal-like MDA-MB-361 F ATCC + + + Luminal B Luminal B MDA-MB-415 F ATCC + - - Luminal B Luminal A MDA-MB-436 F ATCC - - - Basal-like Basal-like MDA-MB-453 F ATCC - - - Luminal B Her2amp MDA-MB-468 F ATCC - - - Basal-like Basal-like SK-BR-3 F ATCC - - + Her2amp Her2amp T-47D F ductal_carcinoma ATCC + + - Luminal B Luminal A UACC-812 F ductal_carcinoma ATCC + - + Her2amp Luminal B UACC-893 F ductal_carcinoma ATCC - - + Her2amp Her2amp YMB-1 F HSRRB + - - Luminal B Luminal A ZR-75-1 F ductal_carcinoma ATCC Luminal A Her2amp ZR-75-30 F ductal_carcinoma ATCC + + - Her2amp Luminal A HCC2185 - - - NON Basal-like HMEL NON Basal-like HCC3153 - - - NON Basal-like ZR75B + - - NON Luminal A 600MPE + - - NON Luminal A SUM1315MO2 - - NON Basal-like SUM149PT - - - NON Basal-like SUM159PT - - - NON Basal-like SUM225CWN - - + NON Her2amp LY2 + - - NON Luminal A Samples are classified as different subtypes Breast cancer classification, in clinic, is measured according to these features: histological type, tumor grade, lymph node status and markers, such as oestrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor 2 (HER2) [4, 6]. Breast cancer could be classified into at least four subtypes known as Luminal A, Luminal B, HER2-enriched and Basal-like (triple negative,TN), according to molecular characteristics which are summarized in Table 2. PAM50 (Prediction Analysis for Microarrays) test is a risk model to identify the intrinsic subtypes in recent 5 years according to 50 gene expressions, including gene ESR1(ERα), PGR(PR) and ERBB2(HER2) [4]. This technique is based on Nano-string counter technology [38, 39]. PAM50 analysis was performed in R following the instructions therein [40]. Here, a threshold of 4.0 was chosen based on the false discovery rate, resulted in the 50-gene classifier. For the sake of missing data imputation, the status of ER, PR, HER2 and the PAM50 subtype calls were regarded as the subtype’s classification reference of breast cancer in this paper. If the sample status of ER, PR, and HER2 is known, samples classification of breast carcinoma is referenced to Table 2. Otherwise its subtype is assigned by mRNA gene expression-based PAM50 prediction, Additional file 1: Tables S1 and S2 provide all the classification information. Data processing mRNA expression analysis and clustering between cell lines and tumors All raw files of microarray mRNA expression, in the form of ‘CEL’ files, were downloaded from GEO GSE36133 and GSE41998. These raw data were normalized by the Affymetrix Microarray Suite 5.0 (MAS5.0) algorithm in accordance with background adjustments, scaling, and aggregation to remove non-biological elements of the signal. Common 22,267 probe sets, corresponding to 14,970 genes, are used comparison analysis for cell line and tumors. All samples in cell lines and tumors are divided into four subtypes group based on ER, PR, HER2 status: luminal A, luminal B, HER2-enrichment and Basal-like as the description before had shown in Additional file 1: Tables S1 and S2. Mean correlation value was obtained for each cell line and tumor in R platform by Pearson correlation analysis. Hierarchy clustering is analyzed between cell lines and tumors of breast cancer in GENE-E software. DNA copy number data analysis A total copy number of changes of TCGA 1033 tumors and CCLE 59 cell lines was detected using Affymetrix 6.0 single nucleotide polymorphism array (SNP 6.0 array) across 28,918 genes. Copy number was measured by a probe corresponding to a segment. They were then inferred and normalized based upon specific linear calibration curves. The circular binary segmentation (CBS) algorithm was used to normalize the segmentations (generally, log2(CN/2)) for further analysis. These segmentations were used to identify focal amplification/deletions and arm-level gains. Fraction genome altered calculation CNVs correspond to relatively large regions of the genome that have been deleted and inserted. To quantitate the extent of the genomic instability in each sample, we calculated the Fraction of Genome Altered (FGA, the fraction of genome lost and gained) as formula (1). The equation represents that sum lengths of all segments (L(i)) whose copy number (CN) segment is above the set threshold (T) and divide by sum of lengths of all segments (L(i)) [8]. Hence, the length of a segment having value equal to or greater than a set threshold are added and are divided by the sum of length of all segments. 1 FGA=∑CNi>TLi/∑Li Here, the threshold T is set to 0.2 for tumor samples and 0.3 for CCLE cell line samples. The threshold values are based on the average distribution density after samples CNV analysis. Cell lines always keep a copy number hyper-mutation degree than tumors’. Copy number correlation calculation With the help of Bioconductor package called ‘CNTools’ [41], these segments are mapped to corresponding gene region across 28,918 genes for both TCGA data and CCLE data, segments file is converted into gene files,then is used for next step correlation analysis. In order to reduce data contamination, only select the top 10 % CNV in 2094 genes segments mean for cross-Pearson’s-correlations calculation between 58 cell lines and 1049 tumors. DNA exome mutation analysis The mutation data was obtained directly from DNA sequence mutation annotation format (.maf) files where Illumina GA platform is used to test. In TCGA, 997 breast invasive cancer Level 2 somatic data is bulk downloaded and hybrid capture 1650 genes in CCLE 59 samples are obtained. According to software ANNOVAR gene-based annotation [21], gene mutation function is reported according to the 1000 Genomes Project and dbSNP database, somatic and germline mutation are identified in CCLE. Mutations are limited to somatic mutations and functional mutations. Hence intronic, silent and other mutations were ignored and only exonic mutations were considered. Mutation frequency calculation Gene mutational frequency can be described as a ratio of total number of gene mutations in samples to total number of samples. Actually, it is the measure of gene mutations probability in the breast cancer population. Mutation rate calculation The mutation number of bases for TCGA are detected from the bed files. The bed file contains a number of bases covered for each chromosome, in form of start and end location. Subtracting end from start gives number of bases covered by the reads. All bases obtained for each sample are summed together to obtain a whole number of bases covered, it is the given sample mutations rate per million bases (Mb). Bed files derive from ‘Wig’ format file. ‘Wig’ provides the number of reads for each region. In case of CCLE, the file can be downloaded from CCLE data portal. To TCGA, it is available from Synapse websites, a research-sharing platform (https://www.synapse.org/#!Synapse:syn1695394). Hence samples or gene mutations rates can be calculated through summing up all bases where read covered as mutations per Mb. Mutation allele spectrum calculation The patterns of six trans-version distributions were searched in the sequence annotation files from CCLE and TCGA irrespectively by R programming. Then, the mutation allele mode was obtained in each of the subtypes of breast tumors and cell lines. The correlation was calculated as mutation allele spectrum in each subtype between cell lines and tumors by Pearson-correlation method. Proteins phosphorylation expression analysis and clustering All basal phosphorylation and protein abundance data were obtained by RPPA technology from reference [19] and TCGA. There are 70 phosphoproteins across 38 cell lines of breast cancer that were generated by RPPA technology and pre-processed by the Gordon Mills lab at MD Anderson. Seventy phospho-proteins in 197 patient’s tumor of breast cancer were collected from TCGA in its Level 3 dataset. The common 50 protein expressions across 38 breast cancer cell lines and 197 TCGA tumors were used as comparison analysis between cell lines and tumors. The Pearson correlation method and hierarchy clustering was used to analyze and compare the similarity and non-similarity between cell lines and tumors in breast cancer. The result about how cell lines are close to its corresponding tumors are shown in Additional file 1: Table S6 based on 50 phosphor-proteins. In mRNA and its 50-protein phosphorylation comparison for cell lines and tumors, a gene has multiple isoforms while a protein phosphorylation has multi-sites. All forms of mRNA and its phosphorylation protein are compared with Pearson correlation, 38 genes’ average correlation coefficient was calculated and compared between cell lines and tumors in Fig. 11. The cell line suitability score with breast tumors The extent to which the breast cancer cell lines match genetic characteristics shared by the TCGA tumors was assessed using a whole score by formula (2). The score can catch a cell line’s whole similarity by four molecular profiles feature to tumors in breast cancer. 2 Score=A+B+C+D Where A is the gene expression similarity between cell lines and tumors by Pearson-correlation; B is the correlation with CNV segment mean of breast tumors; C is the correlation of genes mutation variation with breast tumors; D is the protein expression-based correlation with tumors in breast cancer. The score serves to identify a better or poorer cell lines model of breast cancer in entity molecular level and rank the graduate. Software tools All data arranging was operated on Ubuntu Linux operating system by shell scripting programming. R and MATLAB was used to perform statistical analysis and plotting graphs [42]. Integrative Genomics Viewer (IGV) tools help to visualize large integrated data sets in a single frame and also supports zooming in to a particular chromosome or a certain region of the chromosome, and thus IGV (version 2.3) was used to create copy number profile plots [43]. GENE-E is a matrix visualization and analysis platform designed to support visual data exploration. Hierarchy clustering analysis used by GENE-E software on website www.broadinstitute.org/cancer/software/GENE-E/. Abbreviations CCLE, cancer cell line encyclopedia; CNV, copy number variation; ER, estrogen receptor; FGA, fraction genome altered; GEO, gene expression omnibus; HER2, human epidermal growth factor receptor 2; IGV, integrative genomics viewer; MAS5, affymetrix microarray suite 5.0; NCCN, National Comprehensive Cancer Network; PAM50, prediction analysis for microarrays 50; PR, progesterone receptor; RPPA, reverse phase protein array; SNP, single nucleotide polymorphism array Additional files Additional file 1: Table S1. The list of TCGA tumor samples used on each platform with associated subtype calls from each technology platforms, and clinical data. Table S2. The list of cell lines samples used on each platform with associated subtype calls from each technology platforms, and its annotation data. Table S3. The list of tumors samples from GEO used on gene expression comparison with associated ER, PR, HER2 status. Table S4. Mutation rate per Mb in cell lines and tumors in breast cancer. (Common genes). Table S5. Top 10 % genes of copy number variation in cell lines and tumors. Table S6. The comparison of phosphorylation protein vs gene expression in cell lines and tumors. (XLSX 2941 kb) Additional file 2: Table S7. Correlation coefficient r across 4 genomics level comparison in breast cancer. (XLSX 1527 kb) Acknowledgement The authors would like to thank Dr. Lang Li for insightful discussion and technical assistance. This work was partially supported by the National Institutes of Health Research Foundation grant numbers DK102694, GM10448301, and LM011945. Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials Functional genomics data in this manuscript can be found as the following: Microarray (Breast tumors and cell lines) from GEO, GSE36133, GSE41998; Gene expression, CNV, DNA exome mutation sequencing, RPPA protein array datasets for breast tumors are from Cancer Genome Atlas (TCGA) Data Portal (https://gdc.nci.nih.gov/); Gene expression, CNV, DNA exome mutation of breast cancer cell lines are from Cancer Cell Line Encyclopedia (CCLE) (http://www.broadinstitute.org/ccle); RPPA protein array dataset of breast cancer cell line obtains from Ref. [19]. Datasets supporting the results of this article are included in the additional files. Authors’ contributions GLJ and LJC design and drafted the manuscript; Data is collected by SJZ and AVP. Gene expression and protein expression are compared by AY; Mutation comparison and analysis by ML; Copy number variation and driver genes are selected in breast cancer by SMI; Genes are mapped and annotated in different dataset by YLL. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Center MM Jemal A Lortet-Tieulent J Ward E Ferlay J Brawley O Bray F International variation in prostate cancer incidence and mortality rates Eur Urol 2012 61 6 1079 1092 10.1016/j.eururo.2012.02.054 22424666 2. Onega T Beaber EF Sprague BL Barlow WE Haas JS Tosteson AN D Schnall M Armstrong K Schapira MM Geller B Breast cancer screening in an era of personalized regimens: a conceptual model and National Cancer Institute initiative for risk-based and preference-based approaches at a population level Cancer 2014 120 19 2955 2964 10.1002/cncr.28771 24830599 3. Prat A Pineda E Adamo B Galvan P Fernandez A Gaba L Diez M Viladot M Arance A Munoz M Clinical implications of the intrinsic molecular subtypes of breast cancer Breast 2015 24 2 S26 35 10.1016/j.breast.2015.07.008 26253814 4. Sinn P Aulmann S Wirtz R Schott S Marme F Varga Z Lebeau A Kreipe H Schneeweiss A Multigene assays for classification, prognosis, and prediction in breast cancer: a critical review on the background and clinical utility Geburtshilfe Frauenheilkd 2013 73 9 932 940 10.1055/s-0033-1350831 24771945 5. Lehmann BD Bauer JA Chen X Sanders ME Chakravarthy AB Shyr Y Pietenpol JA Identification of human triple-negative breast cancer subtypes and preclinical models for selection of targeted therapies J Clin Invest 2011 121 7 2750 2767 10.1172/JCI45014 21633166 6. Rupninder Sandhu JSP Jones WD Livasy CA Coleman WB Microarray-based gene expression profiling for molecular classification of breast cancer and identification of new targets for therapy LabMedicine 2010 41 9 8 7. Barretina J Caponigro G Stransky N Venkatesan K Margolin AA Kim S Wilson CJ Lehar J Kryukov GV Sonkin D The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity Nature 2012 483 7391 603 607 10.1038/nature11003 22460905 8. Domcke S Sinha R Levine DA Sander C Schultz N Evaluating cell lines as tumour models by comparison of genomic profiles Nat Commun 2013 4 2126 10.1038/ncomms3126 23839242 9. Holliday DL Speirs V Choosing the right cell line for breast cancer research Breast Cancer Res 2011 13 4 215 10.1186/bcr2889 21884641 10. Neve RM Chin K Fridlyand J Yeh J Baehner FL Fevr T Clark L Bayani N Coppe JP Tong F A collection of breast cancer cell lines for the study of functionally distinct cancer subtypes Cancer Cell 2006 10 6 515 527 10.1016/j.ccr.2006.10.008 17157791 11. Wilding JL Bodmer WF Cancer cell lines for drug discovery and development Cancer Res 2014 74 9 2377 2384 10.1158/0008-5472.CAN-13-2971 24717177 12. Ross DT Perou CM A comparison of gene expression signatures from breast tumors and breast tissue derived cell lines Dis Markers 2001 17 2 99 109 10.1155/2001/850531 11673656 13. Lacroix M Leclercq G Relevance of breast cancer cell lines as models for breast tumours: an update Breast Cancer Res Treat 2004 83 3 249 289 10.1023/B:BREA.0000014042.54925.cc 14758095 14. Kao J Salari K Bocanegra M Choi YL Girard L Gandhi J Kwei KA Hernandez-Boussard T Wang P Gazdar AF Molecular profiling of breast cancer cell lines defines relevant tumor models and provides a resource for cancer gene discovery PLoS One 2009 4 7 e6146 10.1371/journal.pone.0006146 19582160 15. Tomczak K Czerwinska P Wiznerowicz M The Cancer Genome Atlas (TCGA): an immeasurable source of knowledge Contemp Oncol 2015 19 1A A68 77 16. Curtis C Shah SP Chin SF Turashvili G Rueda OM Dunning MJ Speed D Lynch AG Samarajiwa S Yuan Y The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups Nature 2012 486 7403 346 352 22522925 17. Nik-Zainal S Alexandrov LB Wedge DC Van Loo P Greenman CD Raine K Jones D Hinton J Marshall J Stebbings LA Mutational processes molding the genomes of 21 breast cancers Cell 2012 149 5 979 993 10.1016/j.cell.2012.04.024 22608084 18. Kandoth C McLellan MD Vandin F Ye K Niu B Lu C Xie M Zhang Q McMichael JF Wyczalkowski MA Mutational landscape and significance across 12 major cancer types Nature 2013 502 7471 333 339 10.1038/nature12634 24132290 19. Daemen A Griffith OL Heiser LM Wang NJ Enache OM Sanborn Z Pepin F Durinck S Korkola JE Griffith M Modeling precision treatment of breast cancer Genome Biol 2013 14 10 R110 10.1186/gb-2013-14-10-r110 24176112 20. Sweeney C Bernard PS Factor RE Kwan ML Habel LA Quesenberry CP Jr Shakespear K Weltzien EK Stijleman IJ Davis CA Intrinsic subtypes from PAM50 gene expression assay in a population-based breast cancer cohort: differences by age, race, and tumor characteristics Cancer Epidemiol Biomarkers Prev 2014 23 5 714 724 10.1158/1055-9965.EPI-13-1023 24521995 21. Wang K Li M Hakonarson H ANNOVAR: functional annotation of genetic variants from high-throughput sequencing data Nucleic Acids Res 2010 38 16 e164 10.1093/nar/gkq603 20601685 22. Stephens PJ Tarpey PS Davies H Van Loo P Greenman C Wedge DC Nik-Zainal S Martin S Varela I Bignell GR The landscape of cancer genes and mutational processes in breast cancer Nature 2012 486 7403 400 404 22722201 23. Forbes SA Tang G Bindal N Bamford S Dawson E Cole C Kok CY Jia M Ewing R Menzies A COSMIC (the Catalogue of Somatic Mutations in Cancer): a resource to investigate acquired mutations in human cancer Nucleic Acids Res 2010 38 Database issue D652 657 10.1093/nar/gkp995 19906727 24. Easton DF Pooley KA Dunning AM Pharoah PD Thompson D Ballinger DG Struewing JP Morrison J Field H Luben R Genome-wide association study identifies novel breast cancer susceptibility loci Nature 2007 447 7148 1087 1093 10.1038/nature05887 17529967 25. Fanale D Amodeo V Corsini LR Rizzo S Bazan V Russo A Breast cancer genome-wide association studies: there is strength in numbers Oncogene 2012 31 17 2121 2128 10.1038/onc.2011.408 21996731 26. Cancer Genome Atlas Network Comprehensive molecular portraits of human breast tumours Nature 2012 490 7418 61 70 10.1038/nature11412 23000897 27. Shah SP Morin RD Khattra J Prentice L Pugh T Burleigh A Delaney A Gelmon K Guliany R Senz J Mutational evolution in a lobular breast tumour profiled at single nucleotide resolution Nature 2009 461 7265 809 813 10.1038/nature08489 19812674 28. Shah SP Roth A Goya R Oloumi A Ha G Zhao Y Turashvili G Ding J Tse K Haffari G The clonal and mutational evolution spectrum of primary triple-negative breast cancers Nature 2012 486 7403 395 399 22495314 29. Gujral TS Karp RL Finski A Chan M Schwartz PE MacBeath G Sorger P Profiling phospho-signaling networks in breast cancer using reverse-phase protein arrays Oncogene 2013 32 29 3470 3476 10.1038/onc.2012.378 22945653 30. Hennessy BT Lu Y Gonzalez-Angulo AM Carey MS Myhre S Ju Z Davies MA Liu W Coombes K Meric-Bernstam F A technical assessment of the utility of reverse phase protein arrays for the study of the functional proteome in non-microdissected human breast cancers Clinical proteomics 2010 6 4 129 151 10.1007/s12014-010-9055-y 21691416 31. Niepel M Hafner M Pace EA Chung M Chai DH Zhou L Muhlich JL Schoeberl B Sorger PK Analysis of growth factor signaling in genetically diverse breast cancer lines BMC Biol 2014 12 20 10.1186/1741-7007-12-20 24655548 32. Vogel C Abreu Rde S Ko D Le SY Shapiro BA Burns SC Sandhu D Boutz DR Marcotte EM Penalva LO Sequence signatures and mRNA concentration can explain two-thirds of protein abundance variation in a human cell line Mol Syst Biol 2010 6 400 10.1038/msb.2010.59 20739923 33. Greenbaum D Colangelo C Williams K Gerstein M Comparing protein abundance and mRNA expression levels on a genomic scale Genome Biol 2003 4 9 117 10.1186/gb-2003-4-9-117 12952525 34. Gazdar AF Kurvari V Virmani A Gollahon L Sakaguchi M Westerfield M Kodagoda D Stasny V Cunningham HT Wistuba II Characterization of paired tumor and non-tumor cell lines established from patients with breast cancer Int J Cancer Suppl 1998 78 6 766 774 10.1002/(SICI)1097-0215(19981209)78:6<766::AID-IJC15>3.0.CO;2-L 35. Heiser LM Sadanandam A Kuo WL Benz SC Goldstein TC Ng S Gibb WJ Wang NJ Ziyad S Tong F Subtype and pathway specific responses to anticancer compounds in breast cancer Proc Natl Acad Sci U S A 2012 109 8 2724 2729 10.1073/pnas.1018854108 22003129 36. Kennedy JJ Abbatiello SE Kim K Yan P Whiteaker JR Lin C Kim JS Zhang Y Wang X Ivey RG Demonstrating the feasibility of large-scale development of standardized assays to quantify human proteins Nat Methods 2014 11 2 149 155 10.1038/nmeth.2763 24317253 37. Horak CE Pusztai L Xing G Trifan OC Saura C Tseng LM Chan S Welcher R Liu D Biomarker analysis of neoadjuvant doxorubicin/cyclophosphamide followed by ixabepilone or Paclitaxel in early-stage breast cancer Clin Cancer Res 2013 19 6 1587 1595 10.1158/1078-0432.CCR-12-1359 23340299 38. Reis PP Waldron L Goswami RS Xu W Xuan Y Perez-Ordonez B Gullane P Irish J Jurisica I Kamel-Reid S mRNA transcript quantification in archival samples using multiplexed, color-coded probes BMC Biotechnol 2011 11 1 46 10.1186/1472-6750-11-46 21549012 39. Nielsen T Wallden B Schaper C Ferree S Liu S Gao D Barry G Dowidar N Maysuria M Storhoff J Analytical validation of the PAM50-based prosigna breast cancer prognostic gene signature assay and nCounter analysis system using formalin-fixed paraffin-embedded breast tumor specimens BMC Cancer 2014 14 177 10.1186/1471-2407-14-177 24625003 40. Tibshirani R Hastie T Narasimhan B Chu G Diagnosis of multiple cancer types by shrunken centroids of gene expression Proc Natl Acad Sci U S A 2002 99 10 6567 6572 10.1073/pnas.082099299 12011421 41. Gentleman RC Carey VJ Bates DM Bolstad B Dettling M Dudoit S Ellis B Gautier L Ge Y Gentry J Bioconductor: open software development for computational biology and bioinformatics Genome Biol 2004 5 10 R80 10.1186/gb-2004-5-10-r80 15461798 42. Team RDC R: A Language and Environment for Statistical Computing 2011 Vienna The R Foundation for Statistical Computing 43. Robinson JT Thorvaldsdottir H Winckler W Guttman M Lander ES Getz G Mesirov JP Integrative genomics viewer Nat Biotechnol 2011 29 1 24 26 10.1038/nbt.1754 21221095

PMC005xxxxxx/PMC5001207.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556419290310.1186/s12864-016-2903-zResearchCrossLink: a novel method for cross-condition classification of cancer subtypes Ma Chifeng mcfcarter@yahoo.com 1Sastry Konduru S. skonduru@sidra.org 28Flore Mario mariofloresmacias@yahoo.com 1Gehani Salah sgehani@hamad.qa 3Al-Bozom Issam ialbozom@hamad.qa 3Feng Yusheng yusheng.feng@utsa.edu 4Serpedin Erchin serpedin@ece.tamu.edu 5Chouchane Lotfi loc2008@qatar-med.cornell.edu 2Chen Yidong cheny8@uthscsa.edu 67Huang Yufei yufei.huang@utsa.edu 171 Department of Electrical and Computer Engineering, University of Texas at San Antonio, San Antonio, TX USA 2 Weill Cornell Medicine-Qatar, Doha, Qatar 3 Hamad Medical Corporation, Doha, Qatar 4 Department of Mechanical Engineering, University of Texas at San Antonio, San Antonio, TX USA 5 Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX USA 6 Department of Epidemiology and Biostatistics, University of Texas Health Science Center at San Antonio, San Antonio, TX USA 7 Greehey Children Cancer Research Institute, University of Texas Health Science Center at San Antonio, San Antonio, TX USA 8 Division of Translational Medicine, Sidra Medical and Research Center, Doha, Qatar 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.549© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background We considered the prediction of cancer classes (e.g. subtypes) using patient gene expression profiles that contain both systematic and condition-specific biases when compared with the training reference dataset. The conventional normalization-based approaches cannot guarantee that the gene signatures in the reference and prediction datasets always have the same distribution for all different conditions as the class-specific gene signatures change with the condition. Therefore, the trained classifier would work well under one condition but not under another. Methods To address the problem of current normalization approaches, we propose a novel algorithm called CrossLink (CL). CL recognizes that there is no universal, condition-independent normalization mapping of signatures. In contrast, it exploits the fact that the signature is unique to its associated class under any condition and thus employs an unsupervised clustering algorithm to discover this unique signature. Results We assessed the performance of CL for cross-condition predictions of PAM50 subtypes of breast cancer by using a simulated dataset modeled after TCGA BRCA tumor samples with a cross-validation scheme, and datasets with known and unknown PAM50 classification. CL achieved prediction accuracy >73 %, highest among other methods we evaluated. We also applied the algorithm to a set of breast cancer tumors derived from Arabic population to assign a PAM50 classification to each tumor based on their gene expression profiles. Conclusions A novel algorithm CrossLink for cross-condition prediction of cancer classes was proposed. In all test datasets, CL showed robust and consistent improvement in prediction performance over other state-of-the-art normalization and classification algorithms. The International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background The rapid development of high-throughput technologies including microarray and high-throughput sequencing have significantly advanced our understanding of disease including cancer [1]. Torrent of gene expression profiles from cancer cell lines and patient samples have been and are being made available by efforts ranging from large group projects such as The Cancer Genome Atlas to individual labs [2–4]. Significant efforts have been devoted to developing new genomic approaches using gene expression and other genomic data for cancer diagnosis and prognosis [5]. As exciting new results generated from these research efforts continue to challenge our knowledge of cancer, these efforts are also poised to revolutionize the practice of cancer therapy. A large number of gene expression based biomarkers such as PAM50 have been reported to improve cancer classification and prediction of therapy response [6–10]. As exciting as these new discoveries are, their translation from laboratories to real clinical practice remains a challenge. Overcoming systematic and condition-specific biases presented in expression data as a result of different technological platforms, varying experimental/measurement conditions, and heterogeneities in the patient age, gender and race continues to be an issue yet to be completely addressed. Although improved standards in uniform experimental and clinical protocols have and will help reduce the systematic biases, eliminating biases specific to experimental/clinical conditions, patient individuals, technology/platforms would be more effective dealt with by using computational algorithms. The well-known Microarray Quality Control project (MAQC) spearheaded the algorithm development in this front and demonstrated that through careful algorithm-based normalization, consistently differentially expressed genes can be reproduced in data produced from different platforms [11]. Since then, many algorithms have been reported to address different aspects of cross-platform data normalization [12–17]. However, removing biases from different platforms might require using different normalization algorithms. Furthermore, the problem of mitigating condition-specific bias due to differences in experimental/clinical conditions and patient characteristics has not been given sufficient attention. Therefore, a normalization algorithm may work well under one condition but not under another [12]. In this paper, we consider the problem of predicting cancer classes (e.g. subtypes) based on patient gene expression profiles. Particularly, a reference expression dataset is assumed available, where the true cancer class labels for each sample are known. However, compared with the reference dataset, the prediction dataset is generated using a different platform, from patient samples of, for instance, different races, and collected under a different condition. That is, we assume that the prediction dataset contains both systematic and condition-specific biases. Currently, the mainstream practice to this prediction starts by first normalizing the reference and prediction dataset so that both can follow the same desired characteristics (e.g. distribution). Then, a classifier is trained using the normalized reference dataset, which would produce a set of signature genes, accompanied also by their associated class-specific expression signatures [14]. This gene-signature based classifier is finally applied for cancer class prediction in the prediction dataset. The premise for the trained classifier to work well is that the distributions of the label-specific gene signatures in the reference and prediction datasets should remain similar after normalization. However, when both systematic and condition-specific biases are present in data, it cannot be guaranteed that a normalization algorithm can map the gene signatures in the reference and prediction datasets to have the same distribution for all different conditions. As a result, the trained classifier would fail under a different condition (Fig. 1), where one will have to train a new classifier after applying a different normalization algorithm.Fig. 1 General idea of CL. Due to condition-specific biases, the existing normalization algorithm might fail to normalize the distributions of class-specific gene signatures in the reference and prediction datasets. Therefore, the classifier trained using the reference dataset would not work well for the prediction dataset (top right figure). Unlike normalization based approach, CL exploits the fact that the signature is unique to its associated class under any condition and thus employs an unsupervised clustering algorithm to discover this unique signature, hence the class label (bottom right figure) To address the problems of current normalization based approaches, we propose a novel algorithm called CrossLink (CL). The CL algorithm represents a complete departure from the current normalization-classification paradigm. CL only assumes that each cancer class is associated with a set of signature genes, which are independent of the conditions. However, CL recognizes that although for a specific condition, the signature genes should define a unique, cancer class-specific gene expression signature but this signature changes under a new condition. Moreover, the change in the signature is condition-specific and there is no universal, condition-independent normalization mapping of signatures. As a result, unlike existing normalization-based algorithms, CL does not attempt to explore a mapping of the signatures across different conditions; in contrast, it exploits the fact that the signature is unique to its associated class under any condition and thus employs an unsupervised clustering algorithm to discover these unique signatures (Fig. 1). The rest of the paper is organized as follows: In Methods, the workflow of CL is discussed in details. In Results, we demonstrate the improved, robust performance of CL using both simulated and real data. The concluding remarks are drawn in Conclusion. Methods Problem definition and CL algorithm details Suppose that we are given a reference dataset that measures global gene expression of a set of known cancer classes (e.g., PAM50 subtypes). The problem that CL addresses is to predict the cancer classes for a set of new expression data samples collected under a different condition. The workflow of the CL algorithm can be divided into two steps: signature gene set identification and class prediction. For the first step, the goal is to identify the signature gene sets for each cancer classes from the reference dataset. For the second step, the signature gene sets are used to predict the class labels for the prediction dataset. The details of these two steps will be discussed next. Notice that before implementing CL, gene entries of data samples from reference and prediction datasets need to be mapped into a set of common reference entries. By default, CL uses the common Gene Symbol as the reference entries. A data entry is removed from all samples if no entry in any samples can be mapped. Signature gene set identification As commonly defined, the signature gene set of a cancer class include genes that show uniquely differential expression in that class. Analysis of Variance (ANOVA) solves such problem. Suppose that there are N classes and the reference dataset contains M genes. A one-way ANOVA model is then proposed for each gene expression 1 Gi=gi+αk+ϵ where k = 1 … N, Gi is the gene expressions of the ith gene in all samples, gi is the ambient expression of this gene while αk represents the unique effect of the kth class on the gene expression, and ε ~ N(0, σ) denotes the zero-mean noise. The ANOVA analysis identifies these signature genes for each class by testing the hypothesis 2 H0:allαk=0 against the alternative hypothesis 3 H1:atleastoneαk≠0. A multiple comparison procedure such as Benjamini-Hochberg is applied to calculate the False Discovery Rates (FDRs) for each gene and the genes that are tested significant (FDR>0.05) for a class are determined as the candidate gene of the signature gene set of that class. An additional filtering step is followed to reduce the possible false positive signature genes. The filtering imposes three expression conditions on every candidate signature genes, first, the candidate signature gene should have the largest absolute average expression in the class it test significant for; second, a lower limit is introduced on the difference of average expression between the class it tested significantly for and the rest; and third, a lower limit is introduced on the absolute average expression of the class it is testing. The leave-one-out cross-validation was applied to determine the limits that yield the best classification outcomes. Only those candidate genes that satisfy all three conditions will be retained to form the final signature gene sets. In the end, N signature gene sets will be obtained. Class prediction Once the signature gene sets are determined for each class, the next step is to predict the class labels for a new set of data samples. As opposed to using a supervised approach that performs the prediction with a classifier trained on the (normalized) training data set, CL employs a novel unsupervised solution. Since we assume that each signature gene set possesses a unique expression signature for the corresponding class, it would be natural to expect that the class-specific gene set can separate the dataset into two groups: one that belongs to the target class that the gene set is associated with and the other one that contains samples from other classes. Therefore, CL employs the K-means clustering [18] to group the dataset into two clusters and this is performed for each of the N signature gene sets independently. For each of the clustering results, the cluster with a higher absolute average expression value is determined target class, whereas the other cluster is determined as the non-target class. Now that each sample can be assigned with a target class label for any of the N classes, a sample can be associated with multiple class labels. To resolve this ambiguity, a multiple call adjustment procedure is proposed. Specifically, for each class j that is assigned to a sample i, a confidence score Si,j is calculated as: 4 Si,j=p1∗p2 Where 5 p1=−log(P(Gj|μj,nt,σj,nt)) 6 p2=−log(P(μj,t−μj,nt|t0,sp,n1+n2−2)) where j = 1, …, k. Since for each class specific signature gene set, the clustering algorithm separates the dataset into two clusters: class target clustering and class non-target cluster. We further assume that the class target cluster can be modeled by a normal distribution N1(μj,t, σj,t) and the non-target cluster by another normal distribution N2(μj,nt, σj,nt). The first criteria p1 calculate the probability of samples in the class j’s gene set under the hypothesis that the non-target cluster distribution is true. This probability essentially measures the distance from the sample to the center of the class non-target cluster center. The second criteria p2 calculate the probability of μj,t − μj,nt under the hypothesis that the difference of two means follows a student t distribution t(t0,sp,n1+n2−2) where means is 0, variance sp is the pooled variance of two normal distribution N1 and N2, and degree of freedom is n1 + n2 − 2, where n1 and n2 are the sample size of class target cluster and class non-target cluster respectfully. This probability is essentially a two-sample t test, measures the distance between the class target cluster and class non-target cluster. It is obvious that both two criteria are maximized to yield a higher confidence level. A confidence score is determined by multiplying these two criteria together. The class with the highest confidence will be assigned to the sample 7 Labeli=argmaxj(Si,j) A metric for evaluating PAM50 subtype prediction using ER and PR status We investigated CL performance for cross experiment predictions of PAM50 subtypes (See Results for details). However, the true PAM50 subtypes are rarely available for most of the prediction datasets and when this is the case, direct evaluation of the CL performance is infeasible. In contrast, the pathological biomarker assessments of the estrogen receptor (ER) progesterone receptor (PR) are much more accessible for most of the patient samples. Particularly, in a recent study, over 800 breast cancer patients were genetically profiled and their PAM50 subtypes tested by a novel RT-qPCR approach that is independent of microarray platform and their ER and PR status were recorded [19]. This study inspired us to propose an indirect evaluation of the PAM50 classification result by seeking a link between the ER-PR markers status and PAM50 subtypes. Table 1 tallies the distribution of patients from this study over a classification based on both ER-PR status and PAM50 subtypes. Based on Table 1, the empirical conditional probability of each PAM50 classification given an ER-PR status, or P(PAM50|ER, PR) can be calculated, which can be used as the confidence level of predicting a PAM50 subtype given its ER-PR status. For example, if a patient was ER+ and PR+, then from Table 1, we can infer that our confidence of PAM50 prediction as the subtype LumA is 45.64 %. Notice that another important assessment HER2-status is also available and could be included into our analysis, but it is not as commonly documented as ER and PR. Because of this reason Her2 status is not included in our assessment. However, including Her2 could further improve the performance and is very straight forward as we explained. Over all, in the absence of true PAM50 labels, we propose the Indirect Summed Evaluation Probability (ISEP) to evaluate the PAM50 prediction results and ISEP is calculated asTable 1 Distribution of patients on PAM50 subtypes and ER-PR status LumA LumB Her2 Basal Normal ER+,PR+ 246 188 78 4 23 ER+, PR- 12 51 33 3 6 ER-, PR+ 15 5 3 4 1 ER-, PR- 4 17 60 59 2 8 ISEP=∑n=1N(PAM50n|ERn,PRn) where N represents the size of the prediction dataset. Since different dataset certainly have different PAM50 class label rates, this difference in the class label rates could yield an accidentally equal ISEP. Also, because the conditional probability of each PAM50 class does not equal to each other, although the ISEPs of two experiments may differ, they could infer the exact same classification accuracy. Because of these reasons, we want to point out that the ISEPs for two datasets should not be compared. Code implementation and development environment All algorithms are designed and implemented under Matlab R2013a. Function ‘anova1’ is used in the signature selection; function ‘kmeans’ is used in the classification procedure. The designed algorithm is also implemented with R (version 3.1.1). The R code and an example demonstrating the whole pipeline are provided to show how to extract signatures from a reference dataset and how they can be used to classify independent cross-condition samples. The package can be downloaded from http://compgenomics.utsa.edu/CrossLink/CL_R.zip. Data collection for Arabic breast cancer patients The study was approved by the Institutional Review Board of Weill Cornell Medicine-Qatar and the Hamad Medical Corporation’s Ethics Committee, Doha. All subjects signed informed consent documents for participation in this study. The diagnosis of cancer was confirmed by histopathologic analyses. Expression of ER, PR and Her2 was revealed by immunohistochemistry. Human breast cancer tumor samples and non-malignant healthy breast tissues were collected, immediately placed in RNAlater solution and frozen at -80 °C until further use. RNeasy Minikit (Qiagen) was used to extract and purify RNA from these breast tissue samples. The GeneChip Human Genome U133A 2.0 Array (Affymetrix) was used to explore the differentially expressed genes according to manufacturer’s instructions. Results This section is separated into three parts: (1) the ability of CL for PAM50 classification is first demonstrated in several scenarios; (2) the application of CL on Cancer2000 classification is then demonstrated; (3) a Qatar breast cancer patients’ Microarray data analysis is conducted. Cross-experiment prediction of PAM50 breast cancer intrinsic subtype PAM50 breast cancer intrinsic subtype is a gene expression based classification system that includes five breast cancer subtypes: Luminal A (LumA), Luminal B (LumB), Her2 enriched (Her2), Basal and Normal-breast like (Normal) [20]. It has been well studied and has the ability to predict patient’s survival [19, 21, 22]. The PAM50 system is also accompanied by a 50-gene based classifier for subtype prediction based on an expression data. However, the usage of this classifier requires the prediction datasets to be also generated from the same platform as that of PAM50 (Agilent Human 1A Oligo Microarrays). Otherwise the prediction accuracy would suffer significantly due to platform bias [23]. This limitation underscores the need for a system that can faithfully map the PAM50 classification to samples generated from a different platform. PAM50 prediction of a simulated dataset We first evaluated CL on a simulated dataset, where the true class labels for the test samples are known. Breast Cancer Patient Microarray dataset (BRCA) from The Cancer Genome Atlas (TCGA) [24] was used in this experiment. This dataset includes over 500 microarray samples as well as detailed clinical information of breast cancer patients. BRCA dataset also includes PAM50 subtypes for each sample. This dataset will be used as the reference dataset for all PAM50 prediction cases. To simulate a cross-experiment prediction, a five-fold cross-validation scheme was implemented, where in each cross-validation, the four folds of dataset was considered as the reference set and the other one fold was used as the prediction set. To simulate the effect of the cross-experiment bias in the prediction set, the experimental bias was added to each gene expression value Gij according to the following model: 9 Gij=gij+αi+ϵij where Gij is the gene expression of the ith gene in the jth sample of the prediction dataset, αi ~ N(0, σ2) is the experimental bias for gene i and is constant across all the samples, and εij ~ N(0, σ12) models the sample-specific noise. Notice that the experimental biases are different for different genes. These gene-specific biases simulates the varying influence of a different experimental condition on the expression of different genes. In this experiment, we investigated the robustness of CL prediction against experimental bias, where we let σ2 equal to 0.5 and σ12 ranged from 0 to 7. The prediction performance of CL and seven state-of-the-art cross- platform normalization algorithms are shown in Fig. 2. These seven algorithms include Cross-Platform Normalization (XPN) [12], Distance Weighted Discrimination (DWD) [13], Empirical Bayes (EB) [14, 15], Median Rank Scores (MRS) [14], Quantile Discretization(QD) (Warnat, et al., 2005), Distribution Transformation(DisTran) [16], and Gene Quantiles (GQ) [17]. For each algorithm, a Support Vector Machine (SVM) based one-vs-the-rest multi-class classification algorithm was applied to the normalized data for class label prediction. In order to keep the genes used in our CL to be the same as those in SVM to obtain a fair comparison, SVM was applied on the pooled gene signature set obtained in the CL procedure. Overall, CL produced the best prediction performance at all bias levels. Interestingly, even at no bias, CL outperformed all seven other normalization algorithms, where CL obtained a classification accuracy of 0.75, which improved 0.13 percentage points over the best performing normalization algorithm (DisTran at 0.6393). The reason of this could because that the normalization algorithms actually introduced more artificial bias into the system because it assumes there was bias between training and testing datasets. Moreover, the performance of CL remained robust against the increase of the experiment biases. In contrast, four of the seven normalization algorithms suffered different degree of performance degradation with the increase of the experimental bias. Taken together, these results suggest that CL not only can obtain improved performance when no experimental bias present, but is also immune from the influence of constant, gene-specific experimental bias.Fig. 2 Comparison of CL and seven cross platform normalization + SVM algorithms for PAM50 classification accuracy. Horizontal line represents the level of experimental bias level and vertical line represents the classification accuracy PAM50 prediction for the UNC breast cancer dataset We carried out next an evaluation of CL performance on a real dataset: the UNC breast cancer patient dataset. In this case, the PAM50 subtypes are available and the prediction performance can be directly evaluated. The data samples were collected from Gene Expression Omnibus (GEO) under the data entry GSE2740 [25]. Out of all samples from the entire dataset, 349 samples from the platform GPL1390 were extracted. We used the TCGA-BRCA dataset as the reference dataset. The signature gene sets for each PAM50 subtypes (Table 2) were obtained in the signature gene sets identification step of CL.Table 2 The size of CL selected gene set for PAM50 classification Subtype Selection gene size LumA 60 LumB 60 Her2 63 Basal 299 Normal 52 In this process, the impact of different threshold (see Methods for details) was also investigated (Table 3). We can see that there is no significant trend in T1 and T2 vs. the classification performance. Because of this, the best way to locate a combination that yields the best classification performance would still be through a gradient search for a given range. In this case, two threshold were both given a range of (0.1, 1) and the combination (0.1, 0.8) was chosen for the best leave one out classification accuracy and the corresponding gene signature was obtained.Table 3 Impact of different threshold on selected size, value and corresponding classification accuracy T1 T2 combination Selected gene size Smallest absolute expression Classification accuracy 0.1, 0.1 790 0.21 79.66 % 0.3, 0.1 637 0.28 74.02 % 0.5, 0.1 441 0.37 72.99 % 0.7, 0.1 292 0.48 72.99 % 0.9, 0.1 189 0.66 74.53 % 1.1, 0.1 123 0.73 63.42 % 0.3, 0.3 634 0.30 74.02 % 0.3, 0.5 600 0.50 75.56 % 0.3, 0.7 532 0.70 73.85 % 0.3, 0.9 442 0.80 70.09 % 0.1, 0.8 (selected) 534 0.80 80.00 % This signature gene set yielded a leave one out classification rate of 80 % for the BRCA dataset. In addition, this gene set was pooled together and compared with the well-known PAM50 signature gene set (Fig. 3). Specifically, 9 genes are shared between PAM50 and CL, while the rest of the two gene set are completely different. This result suggests that while PAM50 signature gene set shows well established ability for subtype prediction in the expression pattern based algorithms [26], for some specific subtypes, the discriminative power of these genes are not as strong as CL selected gene set. The gene sets were then used in the subtype prediction step. Notice that TCGA-BRCA was also generated from the platform GPL1390, so there is no cross-platform biases. The prediction results are shown in Table 4, where CL achieves 73 % classification accuracy, which is a 16-percentage-point improvement over the best normalization algorithm (XPN: 55 %).Fig. 3 Comparison of CL selected PAM50 signature and PAM50 signature Table 4 Classification accuracy of PAM50 classification of GSE2740 Algorithm Accuracy CL 73 % EB 55 % GQ 55 % DWD 56 % XPN 57 % DisTran 53 % MRS 57 % QD 56 % PAM50 prediction for a dataset with no true PAM50 labels We then proceeded to test CL on additional datasets. This time, the true PAM50 labels were not available and we applied the proposed ISEP instead to direct assess the prediction performance. Before we proceeded to prediction, we first evaluated the relationship between the ISEP accuracy and the accuracy based on true PAM50 labels. The better the ISEP represents the true performance, the more correlated the ISEP and the true accuracy should be. ISEPs corresponding to different PAM50 classification accuracy based on the reference dataset (TCGA-BRCA) were calculated. The result shows that ISEP strongly correlated with PAM50 classification accuracy with a correlation coefficient of 0.96 (Fig. 4). The ISEPs in the previous simulation case were also calculated (Fig. 5). The result shows almost the same trend as the accuracy plot in Fig. 2. The average correlation coefficient between classification accuracy and its corresponding ISEP is 0.83. Overall, the result indicates that without the true PAM50 labels, ISEP could be used to evaluate the performance of PAM50 classification.Fig. 4 Classification accuracy vs ISEP for simulation case. Horizontal axis represents the classification accuracy and vertical axis represents the corresponding ISEP Fig. 5 Plot of ISEP with experimental bias for CL and seven cross platform normalization algorithm + SVM in the Simulation Case. Horizontal axis represents the experiment Bias level and vertical axis represents the ISEP values Next, CL was applied to a dataset that includes 28 breast cancer patients, whose stroma and epithelium cells were profiled with Affymetrix U133A 2.0 GeneChips (GSE10797). Only 20 samples with both ER and PR information were selected in order to calculate the ISEP accuracy. TCGA-BRCA was still used as the reference dataset and this time there was also a difference in platforms in addition to the apparent experimental differences. As a comparison, the original PAM50 classifier (R code) [27] was also applied in addition to the seven normalization algorithms. ISEP accuracies of each prediction outcomes were calculated and the results are summarized in Table 5. CL greatly outperforms all algorithms except QD, which has a slightly higher ISEP than CL (QD: 5.71 vs CL: 5.67). Particularly, the original PAM50 classifier greatly suffered from the platform bias and only achieved an ISEP of 3.3, which is the worst performance among all. Taken together, the results from this test and that on UNC breast cancer dataset confirm the improved performance of CL for cross-experiment predictions.Table 5 ISEP of PAM50 prediction for CL and seven cross platform normalization algorithms + SVM for GSE10797 Algorithm ISEP CL 5.67 EB 3.61 GQ 3.86 DWD 3.66 XPN 4.09 DisTran 5.27 MRS 5.12 QD 5.71 PAM50 3.3 Cross-experiment prediction of cancer 2000 subtypes Recently, over 2000 breast cancer patients (cancer2000) were profiled and a classification including 10 novel breast cancer subtypes were reported based on the integrative study of microarray gene expression, copy number variation as well as gene mutation information [28]. These novel subtypes were shown to be associated with distinct patient survival. Since Cancer2000 subtypes were recently introduced, the perdition of Cancer2000 subtypes for other patient data has not yet been extensive studied. Given this interest, we investigated how CL performed in predicting Cancer2000 subtypes. Evaluation by simulation Cancer 2000 contains two parts, where first part is a discovery dataset that includes 997 breast cancer patients samples and the second part includes 5 additional validation sets including another over 900 breast cancer patient samples. For each patient, the expression levels of 48,803 genes were measured [28]. Here we used the discovery dataset as our reference dataset for all cancer 2000 subtype classification. The same procedure as in PAM50 was conducted and 10 signature gene sets were selected by CL for all 10 classes (Table 6). As for cancer2000 prediction, we first evaluated the CL performance on the cancer2000 dataset through 5-fold cross-validation and simulation, where the same model as in (1) was applied to model the experimental bias. Notice that the prediction problem is a 10-class classification and it is extremely challenging even without any experiment bias. Once again, CL significantly outperformed all normalization algorithms at all bias levels, registering a more than 100 % improvement in prediction accuracy (~0.6 for CL vs <0.3 for others; Fig. 6). The fact that none of the normalization algorithms achieved classification accuracy higher than 30 % at any bias levels speaks for the difficulty of this classification problem and also underscores the significance of the improvement that CL achieved.Table 6 CL selected signature gene set size for cancer 2000 Subtype Selection gene size Class 1 367 Class 2 3111 Class 3 98 Class 4 207 Class 5 981 Class 6 501 Class 7 265 Class 8 247 Class 9 773 Class 10 286 Fig. 6 Comparison of Cancer 2000 Classification between CL and seven cross platform normalization algorithm + SVM in the simulation case. Horizontal axis represents the experimental bias level and vertical axis represents the classification accuracy Prediction of cancer2000 subtypes for TCGA-BRCA dataset We then used CL to predict the Cancer2000 subtypes for TCGA-BRCA dataset. A set of 10 signature gene sets was first obtained on the reference Cancer2000 dataset (Table 6) and the prediction results were shown in Fig. 7. Although there was no true Cancer2000 classification for TCGA-BRCA samples, it was shown in [28] that the 10 subtypes have unique characteristics in terms of their protein marker status, PAM50 classification, mutation and copy number variation and these characteristics provide ample evidence to assess the performance. Here we selected 4 classes with characteristics available in BRCA dataset (Table 7). Using these characteristics, we evaluated the classification performance by assessing the enrichment of the characteristics in the corresponding class. The rest 6 classes were excluded because the corresponding characteristics were not available in the BRCA dataset. The analysis results of CL predictions and the seven normalization algorithms are presented in Table 8. It is clear that the Cancer2000 characteristics are highly enriched in CL predictions. For instance, 36 of 41 patients that were predicted as Class 2 by CL are ER positive. This is highly consistent with the fact that Class 2 is mainly characterized as ER positive (Table 7). Moreover, while Class 3 is mostly Luminal A samples, 24 of 26 Class 3 samples predicted by CL are Luminal A samples. Also, Class 5 includes mostly ER negative and HER2 enriched samples and among 28 CL identified Class 5 samples, 20 samples are ER negative and 21 samples are HER2 enriched. Similarly, Class 6 samples are enriched by ER positive and Luminal samples; 26 CL identified samples are all ER positive and 24 are Luminal samples. In contrast, the predictions by all the seven normalization algorithms showed poor enrichment of desired characteristics. Specifically, EB, XPN, DisTran, MRS and QD failed to predict any samples in four out of these six selected classes. GQ and DWD did predict samples in four classes; however, the enriched characteristics of the predicted samples did not agree with the original characteristics. Particularly, GQ predicted 69 samples as Class 2 but only 37 of them are ER +. It also predicted 126 Class 3 samples but only 75 of them are Luminal A samples. Over all, CL’s predictions are much more enriched with the known characteristics and it predicted more classes.Fig. 7 Cancer2000 classification for TCGA-BRCA dataset. Horizontal axis represents the number of samples classified for each cancer 2000 cluster. Different colors label the PAM50 class label Table 7 Selected cancer 2000 classes and their characteristics Cancer 2000 cluster (Selected) Class 2 Class 3 Class 5 Class 6 Characteristics ER + Luminal A ER-,Her2 enriched Luminal Samples, ER+ Table 8 Comparison of cancer2000 prediction results between CL and 7 alternative cross platform normalization algorithm Cancer2000 class Class 2 Class 3 Class 5 ER- Class 5 Her 2 Class 6 Luminal Class 6 ER+ CL 36/41 24/26 20/28 21/28 24/26 26/26 EB 8/10 0/0 0/0 0/0 0/0 0/0 GQ 37/69 75/126 4/4 0/4 20/23 21/23 DWD 98/111 52/105 14/15 0/15 19/24 19/24 XPN 8/10 0/0 0/0 0/0 0/0 0/0 DisTran 0/0 256/256 0/0 0/0 0/1 0/1 MRS 0/0 29/68 0/0 0/0 0/0 0/0 QD 2/3 1/11 0/0 0/0 0/0 0/0 Arabic breast cancer patient’s microarray data analysis First we aimed to find genes differential expressed in Qatar breast cancer patient compare to the control sample. With two sample t test and adjusted P value set to 0.05, 116 genes showed significantly differential expression between Qatar breast cancer patients and Qatar normal breast tissue samples. We also aimed to find the genes uniquely expressed only in Aerobic species by comparing QNRF dataset with another set of breast cancer population. For comparison, dataset GSE22035 was downloaded from GEO. This dataset contains 43 Caucasian species samples. It has the same microarray platform as the QNRF dataset. Both datasets went through the same pre-process procedure and additional round of normalization was done on two datasets together. Note that this analysis was not performed on all the genes but only on the differential expressed genes detected previously. All seven cross platform normalization algorithms and quantile normalization were performed in order to detect common differential expressed gene unique to QNRF dataset. However, among all the cross-platform normalization algorithms, no common gene is reported. With Quantile normalization, 9 genes were reported but for DisTran and MRS, another set of 6 genes were reported. Although we cannot provide a consistent list of genes that differential expressed across all normalization algorithm, this 15 gene together could be our primary target of interest in future study for breast cancer in Qatar population. The PAM50 classification and Cancer 2000 classification were also reported by CL procedure (Table 9). For PAM50, the PAM50 R code classification result was also reported. PAM50 R classifies most of the QNRF samples into Lum B class, while some of them had obviously problems. For example, sample B2, B22 and B25 were both ER – and PR –, which were most likely to be Basal or Her2 subtype but PAM50 R classifies them into Lum B. On the other hand, sample B20 who is ER + and PR + was classified as Basal but is more likely to be non-Basal sample. For CL, the classified result of the above samples was much more reasonable: B2, B22 and B25 were all classified as Basal and sample B20 was classified as HER2. One interesting point is that among the 20 patients, most of the patients were identified as either Basal subtype or Her2 subtype, while only one Qatar sample was identified as Lum B. This result suggests that over all, breast cancer in Qatar population behaves more like Basal and Her2 subtypes. However, additional tests using samples from larger cohorts need to be performed to confirm this finding.Table 9 Breast cancer subtype classification of QNRF QNRF sample ER PR PAM50 R call PAM50 CL cCall Cancer 2000 CL call B10 + + LumB Basal cancer2000 icluster 1 B13 + + LumA Basal cancer2000 icluster 1 B14 NA NA LumB Basal cancer2000 icluster 3 B17 + + Normal HER2 cancer2000 icluster 3 B18 + + + Normal HER2 cancer2000 icluster 3 B19 NA NA LumB HER2 cancer2000 icluster 3 B20 + + Basal HER2 cancer2000 icluster 3 B21 + + Lum B Lum B cancer2000 icluster 3 B22 - - LumB Basal cancer2000 icluster 3 B23 + + LumB Basal cancer2000 icluster 3 B24 + + LumB Basal cancer2000 icluster 3 B25 - - LumB Basal cancer2000 icluster 3 B26 + + Basal Basal cancer2000 icluster 3 B27 + + Basal Basal cancer2000 icluster 3 B2 - - Lum B Basal cancer2000 icluster 1 B3 NA NA Lum B Basal cancer2000 icluster 1 B4 + + Lum B HER2 cancer2000 icluster 5 B5 + - Lum B Basal cancer2000 icluster 1 B6 - - Basal Basal cancer2000 icluster 1 B7 NA NA Lum A Basal cancer2000 icluster 1 Discussion and Conclusions In this paper, we proposed a novel algorithm CrossLink for cross-condition prediction of cancer classes. Unlike other normalization-based method, CL employs an unsupervised algorithm, which aims at identifying unique class-specific signatures patterns. CL was applied for cross-condition prediction of the PAM50 and Cancer2000 subtypes. In all tested datasets, CL showed robust and consistent improvement in prediction performance over other state-of-the-art normalization algorithms. Despite its advantages, CL has limitations. First, CL is better fitted for datasets of large sample size, because CL needs to perform an unsupervised learning. It cannot be applied to individual samples separately as what a classifier would do. By the same reasoning, CL would fail when there are samples from only a single class. Our future work includes to three directions. First, the result of the CL indicates that instead of choosing a common signature set for all subtypes classification, subtype specific signatures can lead to better robustness and accuracy for subtypes classification. Further investigation is needed to discover the biological insight of those signatures. By doing so, the subtype related function could be also discovered. Second, CL shows great potential for subtype classification in cross-condition breast cancer subtype classification. This ability could be further extended into other cancer genomic classification problems when condition specific bias presented. Third, the unique design of CL allows it bypassing the condition specific bias to achieve a robust classification accuracy. This advantage can be further extended to handle bias between different technical platforms, for example, between microarray and RNA-seq data. Acknowledgements This work is supported by a National Science Foundation grant (CCF-1246073) to YH and YC, a Qatar National Research Fund grant (09-874-3-235) to YC and YH, a grant from San Antonio Life Science Institute to YF, and a National Institute of Health grant (NIH-NCATS UL1TR000149) to YC. Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials R code of CL, the manual and the example files are freely available for download at http://compgenomics.utsa.edu/CrossLink/CL_R.zip. Authors’ contributions YC, LC, ES, and YH conceived the idea of the paper. KSS, SG, IA, and LC collected samples from Qatar population and performed microarray. CM, MF, and YF developed the algorithms and performed the tests. CM, YF, YC, and YH wrote the paper. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Collection of human tissues in this study was approved by the Institutional Review Board of Weill Cornell Medical College-Qatar and the Hamad Medical Corporation’s Ethics Committee, Doha. All subjects signed informed consent documents for participation in this study. ==== Refs References 1. Tseng GC Ghosh D Feingold E Comprehensive literature review and statistical considerations for microarray meta-analysis Nucleic Acids Res 2012 40 9 3785 99 10.1093/nar/gkr1265 22262733 2. Shao L Determination of minimum training sample size for microarray-based cancer outcome prediction-an empirical assessment PLoS One 2013 8 7 10.1371/journal.pone.0068579 3. Takahashi Y Microarray analysis reveals that high mobility group A1 is involved in colorectal cancer metastasis Oncol Rep 2013 30 3 1488 96 23835740 4. Liu Q Gene selection and classification for cancer microarray data based on machine learning and similarity measures BMC Genomics 2011 12 Suppl 5 S1 10.1186/1471-2164-12-S5-S1 22369383 5. Callari M Comparison of microarray platforms for measuring differential microRNA expression in paired normal/cancer colon tissues PLoS One 2012 7 9 10.1371/journal.pone.0045105 23028787 6. Gougelet A Estrogen receptor alpha and beta subtype expression and transactivation capacity are differentially affected by receptor-, hsp90- and immunophilin-ligands in human breast cancer cells J Steroid Biochem Mol Biol 2005 94 1-3 71 81 10.1016/j.jsbmb.2005.01.018 15862952 7. Nielsen TO A comparison of PAM50 intrinsic subtyping with immunohistochemistry and clinical prognostic factors in tamoxifen-treated estrogen receptor-positive breast cancer Clin Cancer Res 2010 16 21 5222 32 10.1158/1078-0432.CCR-10-1282 20837693 8. Bentink S Angiogenic mRNA and microRNA gene expression signature predicts a novel subtype of serous ovarian cancer PLoS One 2012 7 2 10.1371/journal.pone.0030269 22348002 9. Paroni G Synergistic antitumor activity of lapatinib and retinoids on a novel subtype of breast cancer with coamplification of ERBB2 and RARA Oncogene 2012 31 29 3431 43 10.1038/onc.2011.506 22056878 10. Chin SF High-resolution aCGH and expression profiling identifies a novel genomic subtype of ER negative breast cancer Genome Biol 2007 8 10 R215 10.1186/gb-2007-8-10-r215 17925008 11. MAQC Consortium Shi L The MicroArray Quality Control (MAQC) project shows inter- and intraplatform reproducibility of gene expression measurements Nat Biotechnol 2006 24 9 1151 61 10.1038/nbt1239 16964229 12. Shabalin AA Merging two gene-expression studies via cross-platform normalization Bioinformatics 2008 24 9 1154 60 10.1093/bioinformatics/btn083 18325927 13. Benito M Adjustment of systematic microarray data biases Bioinformatics 2004 20 1 105 14 10.1093/bioinformatics/btg385 14693816 14. Warnat P Eils R Brors B Cross-platform analysis of cancer microarray data improves gene expression based classification of phenotypes BMC Bioinformatics 2005 6 265 10.1186/1471-2105-6-265 16271137 15. Walker WL Empirical Bayes accomodation of batch-effects in microarray data using identical replicate reference samples: application to RNA expression profiling of blood from Duchenne muscular dystrophy patients BMC Genomics 2008 9 494 10.1186/1471-2164-9-494 18937867 16. Jiang H Joint analysis of two microarray gene-expression data sets to select lung adenocarcinoma marker genes BMC Bioinformatics 2004 5 81 10.1186/1471-2105-5-81 15217521 17. Xia XQ WebArrayDB: cross-platform microarray data analysis and public data repository Bioinformatics 2009 25 18 2425 9 10.1093/bioinformatics/btp430 19602526 18. Dembele D Kastner P Fuzzy C-means method for clustering microarray data Bioinformatics 2003 19 8 973 80 10.1093/bioinformatics/btg119 12761060 19. Bastien RR PAM50 breast cancer subtyping by RT-qPCR and concordance with standard clinical molecular markers BMC Med Genomics 2012 5 44 10.1186/1755-8794-5-44 23035882 20. Park SY Heterogeneity for stem cell-related markers according to tumor subtype and histologic stage in breast cancer Clin Cancer Res 2010 16 3 876 87 10.1158/1078-0432.CCR-09-1532 20103682 21. Martin M PAM50 proliferation score as a predictor of weekly paclitaxel benefit in breast cancer Breast Cancer Res Treat 2013 138 2 457 66 10.1007/s10549-013-2416-2 23423445 22. Prat A PAM50 assay and the three-gene model for identifying the major and clinically relevant molecular subtypes of breast cancer Breast Cancer Res Treat 2012 135 1 301 6 10.1007/s10549-012-2143-0 22752290 23. Prat A Parker J Perou CMUNCUMD Supplement: How to Perform Intrinsic Subtype Classification 9-6-*10 2010 24. Deus HF Exposing the cancer genome atlas as a SPARQL endpoint J Biomed Inform 2010 43 6 998 1008 10.1016/j.jbi.2010.09.004 20851208 25. Oh DS Estrogen-regulated genes predict survival in hormone receptor-positive breast cancers J Clin Oncol 2006 24 11 1656 64 10.1200/JCO.2005.03.2755 16505416 26. Jorgensen CL PAM50 breast cancer intrinsic subtypes and effect of gemcitabine in advanced breast cancer patients Acta Oncol 2014 53 6 776 87 10.3109/0284186X.2013.865076 24359601 27. Parker JS Supervised risk predictor of breast cancer based on intrinsic subtypes J Clin Oncol 2009 27 8 1160 7 10.1200/JCO.2008.18.1370 19204204 28. Curtis C The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups Nature 2012 486 7403 346 52 22522925

PMC005xxxxxx/PMC5001208.txt

==== Front BMC BioinformaticsBMC BioinformaticsBMC Bioinformatics1471-2105BioMed Central London 115410.1186/s12859-016-1154-yMeeting AbstractsProceedings of the 15th Annual UT-KBRIN Bioinformatics Summit 2016 Cadiz, KY, USA. 8-10 April 2016Rouchka Eric C. eric.rouchka@louisville.edu 12Chariker Julia H. julia.chariker@louisville.edu 23Harrison Benjamin J. b.harrison@louisville.edu 24Park Juw Won juw.park@louisville.edu 12Cao Xueyuan 5Pounds Stanley stanley.pounds@stjude.org 5Raimondi Susana 6Downing James 6Ribeiro Raul 7Rubnitz Jeffery 7Lamba Jatinder 8Daigle Bernie J. Jrbjdaigle@memphis.edu 910Burgess Deborah 11Gehrlich Stephanie 12Carmen John C. carmenj1@nku.edu 12Johnson Nicholas traydreein@gmail.com 13Emani Chandrakanth 13Gehrlich Stephanie 14Burgess Deborah 15Carmen John C. carmenj1@nku.edu 14De Silva Kalpani 16Heaton Michael P. 17Kalbfleisch Theodore S. ted.kalbfleisch@louisville.edu 18Viangteeravat Teeradache 1920Mudunuri Rahul rmudunur@uthsc.edu 19Ajayi Oluwaseun 19Şen Fatih 19Huang Eunice Y. 1920Mohebbi Mohammad 21Florian Luaire 21Jackson Douglas J. 21Naber John F. john.naber@louisville.edu 21Sabbir AKM 22Ellingson Sally R. sally@kcr.uky.edu 23Lu Yuping 24Phillips Charles A 24Langston Michael A. langston@tennessee.edu 24Sevakula Rahul K. 25Thirukovalluru Raghuveer 25Verma Nishchal K. 25Cui Yan ycui2@uthsc.edu 26Sayed Mohammed 27Park Juw Won juw.park@louisville.edu 2728Wang Jing 2930Liu Qi qi.liu@vanderbilt.edu 2931Shyr Yu yu.shyr@vanderbilt.edu 293032Zhang Xiaofei 33Ellingson Sally R. sally@kcr.uky.edu 34Prodduturi Naresh 35Oliver Gavin R. oliver.gavin@mayo.edu 35Grill Diane 35Na Jie 35Eckel-Passow Jeanette 35Klee Eric W. 35Goodin Michael M. 36Farman Mark 36Inocencio Harrison 37Jang Chanyong 36Jaromczyk Jerzy W. 37Moore Neil 36Sovacool Kelly 38Dent Leon 39Izban Mike 40Mandape Sammed 41Sakhare Shruti 41Pratap Siddharth 41Marshall Dana dmarshall@mmc.edu 40DePriest M Scotty michael.depriest@louisville.edu 4243MacLeod James N. 43Kalbfleisch Theodore S. 42Emani Chandrakanth chandrakanth.emani@wku.edu 44Adam Hanady 44Blandford Ethan 44Campbell Joel 44Castlen Joshua 44Dixon Brittany 44Gilbert Ginger 44Hall Aaron 44Kreisle Philip 44Lasher Jessica 44Oakes Bethany 44Speer Allison 44Valentine Maximilian 44Nagisetty Naga Satya V. Rao 45Jose Rony 45Viangteeravat Teeradache 45Rooney Robert rrooney1@uthsc.edu 45Hains David 451 Department of Computer Engineering and Computer Science, University of Louisville, Duthie Center for Engineering, Louisville, KY 40292 USA 2 Kentucky Biomedical Research Infrastructure (KBRIN) Bioinformatics Core, 522 East Gray Street, Louisville, KY 40292 USA 3 Department of Psychological and Brain Sciences, University of Louisville, Louisville, KY 40292 USA 4 Department of Anatomical Sciences and Neurobiology, University of Louisville, Louisville, KY 40292 USA 5 Department of Biostatistics, St. Jude Children’s Research Hospital, Memphis, TN 38105 USA 6 Department of Pathology, St. Jude Children’s Research Hospital, Memphis, TN 38105 USA 7 Department of Oncology, St. Jude Children’s Research Hospital, Memphis, TN 38105 USA 8 Department of Pharmacotherapy and Translational Research, College of Pharmacy, University of Florida, Gainesville, FL 32610 USA 9 Department of Biological Sciences, University of Memphis, Memphis, TN 38152 USA 10 Department of Computer Science, University of Memphis, Memphis, TN 38152 USA 11 Department of Computer Science, Northern Kentucky University, Highland Heights, KY 41099 USA 12 Department of Biological Sciences, Northern Kentucky University, Highland Heights, KY 41099 USA 13 Department of Biology, Western Kentucky University, Bowling Green, KY 42101 USA 14 Department of Biological Sciences, Northern Kentucky University, Highland Heights, KY 41099 USA 15 Department of Computer Science, Northern Kentucky University, Highland Heights, KY 41099 USA 16 Interdisciplinary Studies Program: Specialization in Bioinformatics, University of Louisville, Louisville, KY 40292 USA 17 United States Department of Agriculture, Agricultural Research Service, United States Meat Animal Research Center, Clay Center, NE 68933 USA 18 Department of Biochemistry and Molecular Genetics, School of Medicine, University of Louisville, Louisville, KY 40292 USA 19 Biomedical Informatics Core, Children’s Foundation Research Institute, Memphis, TN 38103 USA 20 Department of Pediatrics, University of Tennessee Health Science Center, Memphis, TN 38163 USA 21 Department of Electrical and Computer Engineering, University of Louisville, Louisville, KY 40292 USA 22 Department of Computer Science, University of Kentucky, Lexington, KY 40506 USA 23 Division of Biomedical Informatics, College of Medicine, University of Kentucky, Lexington, KY 40536-0093 USA 24 Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville, TN 37996-2250 USA 25 Department of Electrical Engineering, Indian Institute of Technology Kanpur, Kanpur, 208016 India 26 Department of Microbiology, Immunology and Biochemistry, University of Tennessee Health Science Center, Memphis, TN 38163 USA 27 Department of Computer Engineering and Computer Science, University of Louisville, Louisville, KY 40202 USA 28 Kentucky Biomedical Research Infrastructure Network (KBRIN) Bioinformatics Core, University of Louisville, Louisville, KY 40292 USA 29 Center for Quantitative Sciences, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 30 Department of Cancer Biology, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 31 Department of Biomedical Informatics, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 32 Department of Biostatistics, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 33 Department of Computer Science, University of Kentucky, Lexington, KY 40506 USA 34 Division of Biomedical Informatics, College of Medicine, University of Kentucky, Lexington, KY 40536-0093 USA 35 Department of Biomedical Informatics and Statistics, Health Sciences Research, Mayo Clinic, Rochester, MN 55905 USA 36 Department of Plant Pathology, University of Kentucky, Lexington, KY 40546 USA 37 Department of Computer Science, University of Kentucky, Lexington, KY 40506 USA 38 Department of Biology, University of Kentucky, Lexington, KY 40506 USA 39 Department of Surgery, Meharry Medical College, Nashville, TN 37208 USA 40 Department of Pathology, Anatomy and Cell Biology, Meharry Medical College, Nashville, TN 37208 USA 41 Bioinformatics and Proteomics Core, Microbiology and Immunology, Meharry Medical College, Nashville, TN 37208 USA 42 Department of Biochemistry and Molecular Genetics, School of Medicine, University of Louisville, Louisville, KY 40292 USA 43 Maxwell H. Gluck Equine Research Center, Department of Veterinary Science, University of Kentucky, Lexington, KY 40546 USA 44 Department of Biology, Western Kentucky University-Owensboro, Owensboro, KY 42303 USA 45 Department of Pediatrics, The University of Tennessee Health Science Center, Memphis, TN 38103 USA 19 8 2016 19 8 2016 2016 17 Suppl 10 Publication charges for this supplement were provided by NIH grant P20GM103436. The Supplement Editors declare that they have no competing interests.297© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Table of contents I1 Proceedings of the Fifteenth Annual UT- KBRIN Bioinformatics Summit 2016 Eric C. Rouchka, Julia H. Chariker, Benjamin J. Harrison, Juw Won Park P1 CC-PROMISE: Projection onto the Most Interesting Statistical Evidence (PROMISE) with Canonical Correlation to integrate gene expression and methylation data with multiple pharmacologic and clinical endpoints Xueyuan Cao, Stanley Pounds, Susana Raimondi, James Downing, Raul Ribeiro, Jeffery Rubnitz, Jatinder Lamba P2 Integration of microRNA-mRNA interaction networks with gene expression data to increase experimental power Bernie J Daigle, Jr. P3 Designing and writing software for in silico subtractive hybridization of large eukaryotic genomes Deborah Burgess, Stephanie Gehrlich, John C Carmen P4 Tracking the molecular evolution of Pax gene Nicholas Johnson; Chandrakanth Emani P5 Identifying genetic differences in thermally dimorphic and state specific fungi using in silico genomic comparison Stephanie Gehrlich, Deborah Burgess, John C Carmen P6 Identification of conserved genomic regions and variation therein amongst Cetartiodactyla species using next generation sequencing Kalpani De Silva, Michael P Heaton, Theodore S Kalbfleisch P7 Mining physiological data to identify patients with similar medical events and phenotypes Teeradache Viangteeravat, Rahul Mudunuri, Oluwaseun Ajayi, Fatih Şen, Eunice Y Huang P8 Smart brief for home health monitoring Mohammad Mohebbi, Luaire Florian, Douglas J Jackson, John F Naber P9 Side-effect term matching for computational adverse drug reaction predictions AKM Sabbir, Sally R Ellingson P10 Enrichment vs robustness: A comparison of transcriptomic data clustering metrics Yuping Lu, Charles A Phillips, Michael A Langston P11 Deep neural networks for transcriptome-based cancer classification Rahul K Sevakula, Raghuveer Thirukovalluru, Nishchal K. Verma, Yan Cui P12 Motif discovery using K-means clustering Mohammed Sayed, Juw Won Park P13 Large scale discovery of active enhancers from nascent RNA sequencing Jing Wang, Qi Liu, Yu Shyr P14 Computationally characterizing genomic pipelines and benchmarking results using GATK best practices on the high performance computing cluster at the University of Kentucky Xiaofei Zhang, Sally R Ellingson P15 Development of approaches enabling the identification of abnormal gene expression from RNA-Seq in personalized oncology Naresh Prodduturi, Gavin R Oliver, Diane Grill, Jie Na, Jeanette Eckel-Passow, Eric W Klee P16 Processing RNA-Seq data of plants infected with coffee ringspot virus Michael M Goodin, Mark Farman, Harrison Inocencio, Chanyong Jang, Jerzy W Jaromczyk, Neil Moore, Kelly Sovacool P17 Comparative transcriptomics of three Acinetobacter baumanii clinical isolates with different antibiotic resistance patterns Leon Dent, Mike Izban, Sammed Mandape, Shruti Sakhare, Siddharth Pratap, Dana Marshall P18 Metagenomic assessment of possible microbial contamination in the equine reference genome assembly M Scotty DePriest, James N MacLeod, Theodore S Kalbfleisch P19 Molecular evolution of cancer driver genes Chandrakanth Emani, Hanady Adam, Ethan Blandford, Joel Campbell, Joshua Castlen, Brittany Dixon, Ginger Gilbert, Aaron Hall, Philip Kreisle, Jessica Lasher, Bethany Oakes, Allison Speer, Maximilian Valentine P20 Biorepository Laboratory Information Management System Naga Satya V Rao Nagisetty, Rony Jose, Teeradache Viangteeravat, Robert Rooney, David Hains 15th Annual UT-KBRIN Bioinformatics Summit 2016 Cadiz, KY, USA 8-10 April 2016 http://www.bioinfosummit.org/issue-copyright-statement© The Author(s) 2016 ==== Body I1 Proceedings of the Fifteenth Annual UT- KBRIN Bioinformatics Summit 2016 Eric C. Rouchka1,2, Julia H. Chariker2,3 (julia.chariker@louisville.edu), Benjamin J. Harrison2,4 (b.harrison@louisville.edu), Juw Won Park1,2 (juw.park@louisville.edu) 1Department of Computer Engineering and Computer Science, University of Louisville, Duthie Center for Engineering, Louisville, KY 40292, USA; 2Kentucky Biomedical Research Infrastructure (KBRIN) Bioinformatics Core, 522 East Gray Street, Louisville, KY 40292, USA; 3Department of Psychological and Brain Sciences, University of Louisville, Louisville, KY 40292, USA; 4Department of Anatomical Sciences and Neurobiology, University of Louisville, Louisville, KY 40292, USA Correspondence: Eric C. Rouchka (eric.rouchka@louisville.edu) – Kentucky Biomedical Research Infrastructure (KBRIN) Bioinformatics Core, 522 East Gray Street, Louisville, KY 40292, USA The University of Tennessee (UT) and the Kentucky Biomedical Research Infrastructure Network (KBRIN) have collaborated over the past fifteen years to share research and educational expertise in bioinformatics. One result is an annual regional summit for researchers, educators and students. The Fifteenth Annual UT-KBRIN Bioinformatics Summit was held at Lake Barkley State Park from April 8-10, 2016. A total of 184 participants pre-registered, with 85 from Tennessee, 84 from Kentucky, and the remainder from various states and international locales. Among the registrants were 64 faculty, 47 staff, 46 students, and 20 postdocs. The conference program consisted of two workshops on the UCSC Genome Browser and two days of plenary presentations and short talks. In addition, a poster session with 57 posters was held on Saturday evening. Friday Workshops Robert Kuhn (University of California – Santa Cruz) opened the Summit with the workshop “UCSC Genome Browser Part I: Visualization Tool for Genomes.” This workshop focused on a lecture and live demonstration of the capabilities of the UCSC Genome Browser (http://genome.ucsc.edu/) [1-3]. Dr. Kuhn talked about many of the capabilities of the browser, including the use of pre-defined annotation tracks, the creation and use of user-defined annotation tracks and track hubs [4], connecting to the annotation tables via database connections [5] and the ability to implement protected sessions through the Genome Browser in a Box [6]. After a short break, Dr. Kuhn continued with “UCSC Genome Browser Part II: Visualization tool for genomes hands-on experience.” In this workshop, Dr. Kuhn focused on use of the UCSC Genome Browser through a set of user-guided exercises intended to give participants hands-on experience with both basic and advanced features of the genome browser. Session I GQ Zhang (University of Kentucky) opened the scientific session on Saturday morning with a presentation on “The role of ontologies in clinical and translational informatics.” His talk was broken down into three areas based on biomedical ontologies, ontologies in use, and ontology quality assurance. In the first portion, Dr. Zhang discussed biomedical ontologies in general, particularly in terms of the resources available to the University of Kentucky Center for Clinical and Translational Science, including the Appalachian Translational Research Network which consists of clinical and translational partners across Kentucky, Tennessee, Ohio, and West Virginia. Dr. Zhang discussed some of the challenges in scaling up translational informatics using big data from a patient perspective to create a learning health care system. In the second portion, Dr. Zhang discussed two resources, including the National Sleep Research Resource (sleepdata.org) [7] and the Center for Sudep Research [8-10]. In the final portion, Dr. Zhang discussed ontology quality assurance with a specific example using Gene Ontology [11] fragments and SNOMED [12, 13] data and approaches his group has taken towards developing algorithms for ontology quality assurance [14-16]. Igor Jouline (Oak Ridge National Laboratory and The University of Tennessee – Knoxville) followed with the plenary talk “Using evolutionary history for predicting functional changes in proteins.” This presentation focused on the use of evolutionary and conserved core elements within systems, such as signal transduction and chemotaxis systems within bacteria, in order to predict likely functional changes [17-21]. Dr. Jouline gave additional examples, including the structural diversity of chemoreceptor signaling domains [22-24]. Dr. Jouline brought home the point of how critical it is to consider the phylogenetic history from a sequence point of view in order to help classify disease mutations [25]. He introduced a computational approach his group has developed which looks at evolutionary changes in genes and their paralogs, and showed that changes need to be considered across all copies, in order to fully understand disease implications [26]. Session II Nancy Cox (Vanderbilt University) led the second session on Saturday morning with a presentation titled “Building a catalog of gene to medical phenome: New ways of understanding the biological mechanisms of disease.” In this presentation, Dr. Cox presented PrediXcan, an approach to medical informatics data integration [27]. Highlighted within this talk were preliminary results of applying PredicXcan to Vanderbilt University’s BioVU [28] which contains several over 215,000 subjects with DNA. Within this dataset are approximately 20,000 dense GWAS genotypes and 42,000 exome chips. Dr. Cox discussed how nearly all genes have a high correlation in at least one tissue type, with 4,000-9,000 correlating within any given tissue. Among the research results presented were several novel gene-phenotype relationships built upon disease models of genetically regulated expression (GReX) [27] and genotype tissue expression (GTEx) [29]. Dr. Cox discussed how disease from a gene expression point of view can be explained by major axes of disease risk, in which the healthiest individuals maintain a balance in the center of all of the axes. Session III Ting Wang (Washington University in St. Louis) opened up the final plenary session with a presentation “Epigenetics roadmap.” Dr. Wang’s talk was broken down into several sections. The first part of his talk focused on discussion of the Roadmap Epigenomics Project [30, 31] which collected a variety of epigenetic markers, including DNA methylation, open chromatin, and histone modification on over 100 tissue and cell types. The second portion of his talk focused on methods of accessing the Roadmap Epigenomics data, including tracks within the UCSC Genome Browser [4] and the WashU Epigenome Browser [32, 33] developed within his group. In the third section, Dr. Wang discussed project extensions, including the recent 4D Nucleosome [34] which focuses on integration of genomic and imaging data and TaRGET project which will focus on epigenomic changes relative to toxicants. The fourth and final portion of the talk was dedicated to discoveries aided by the Roadmap Epigenomics, including epigenetic annotation of genetic variants associated with disease [35] and genetic regulation due to transposable elements [36, 37]. The final plenary speaker on Sunday was Csaba Kovesdy from the University of Tennessee Health Science Center. He presented “Modeling clinical trials using observational methods: How Big Data can help us.” During the course of this presentation, Dr. Kovesdy discussed the use of Big Data in health care from the Veterans Administration (VA) in terms of making discovery in chronic kidney disease (CKD). Dr. Kovesdy proposed how Big Data might be used to supplement the use of clinical trials, particularly in cases when a study of interest is a subset of a larger study. In one particular case, Dr. Kovesdy discussed using data from large studies on hypertension to make discoveries in CKD [38-40]. In addition, he discussed strengths and weaknesses of dealing with data from the VA patient population, and potential future opportunities with the Million Veteran Program [41]. Poster Session A poster session and reception was held on Saturday evening with a total of 57 posters presented across 15 categories. The largest represented categories included high throughput sequencing, bioinformatics of health and disease, systems biology and networks, and comparative genomics. 24 of the poster abstracts are highlighted within this supplement. Eight of the poster abstracts were selected for short 10 minute presentations at the summit, including “Identifying clusters in protein structure: Comparisons between polymorphic, pathogenic, and somatic variation” (R. Michael Sivley, Vanderbilt University); “Porphyromonas gingivalis and the epithelial-to-mesenchymal transition (EMT): Signaling networks linking infection to cancer promotion” (Melissa Metzler, University of Louisville); “CC-PROMISE: Projection onto the most interesting statistical evidence (PROMISE) with canonical correlation to integrate gene expression and methylation data with multiple pharmacologic and clinical endpoints” (Xueyuan Cao, St. Jude Children’s Research Hospital); “Biochemically Aware Substructure Search (BASS) – An algorithm for finding biochemically relevant chemical subgraphs” (Joshua Mitchell, University of Kentucky); “Integration of microRNA-mRNA interaction networks with gene expression data to increase experimental power” (Bernie Daigle, University of Memphis); “Side-effect term matching for computational adverse drug predictions” (AKM Sabbir, University of Kentucky); “Lentiviral CRISPR/cas9 vector mediated miRNA editing disrupts miRNA function” (Junming Yue, University of Tennessee Health Science Center); and “CSI-UTR: An algorithm for characterizing 3’ untranslated region (3’ UTR) diversity in RNA-Seq data by employing RNA cleavage site intervals (CSIs)” (Ben Harrison, University of Louisville). Future Plans The 2017 Summit is scheduled for March 20-22 at Montgomery Bell State Park in Tennessee. Sessions will focus more on bioinformatics research in Kentucky and Tennessee to help forge more collaborative efforts. Plans are to include more Tennessee and Kentucky speakers, in terms of intermediate length (20-30 minute) presentations and short (10 minute) presentations that complement plenary session topics as well as integrating 1 minute flash talk opportunities for those submitting poster abstracts. Acknowledgements We would like to thank the Conference Program Committee members Hao Chen (University of Tennessee Health Science Center), Nigel Cooper (University of Louisville), Dan Goldowitz (University of British Columbia), Mike Langston (University of Tennessee-Knoxville), Terry Mark-Major (University of Tennessee Health Science Center), Hunter Moseley (University of Kentucky), Juw Won Park (University of Louisville), Claire Rinehart (Western Kentucky University), Arnold Stromberg (University of Kentucky), Rob Williams (University of Tennessee Health Science Center), and Zhongming Zhao (University of Texas Health Science Center - Houston) for organizing an outstanding scientific program. In addition, we wish to thank Susan Boucher, Alicia Brookins, Terry Mark-Major, Michelle Padgett, and Whitney Rogers for their efforts in handling conference organization details. Funding for the UT- KBRIN Summit is provided in part by the University of Memphis Office of the Provost, Memphis Research Consortium, Kentucky Biomedical Research Infrastructure Network (KBRIN), University of Tennessee Center for Integrative and Translational Genomics, University of Tennessee Molecular Resource Center, and NIH grant P20GM103436. References 1. Kent WJ, Sugnet CW, Furey TS, Roskin KM, Pringle TH, Zahler AM, Haussler D: The human genome browser at UCSC. Genome Res 2002, 12(6):996-1006. 2. Kuhn RM, Haussler D, Kent WJ: The UCSC genome browser and associated tools. Brief Bioinform 2013, 14(2):144-161. 3. Speir ML, Zweig AS, Rosenbloom KR, Raney BJ, Paten B, Nejad P, Lee BT, Learned K, Karolchik D, Hinrichs AS et al: The UCSC Genome Browser database: 2016 update. Nucleic Acids Res 2016, 44(D1):D717-725. 4. Raney BJ, Dreszer TR, Barber GP, Clawson H, Fujita PA, Wang T, Nguyen N, Paten B, Zweig AS, Karolchik D et al: Track data hubs enable visualization of user-defined genome-wide annotations on the UCSC Genome Browser. Bioinformatics 2014, 30(7):1003-1005. 5. Karolchik D, Hinrichs AS, Furey TS, Roskin KM, Sugnet CW, Haussler D, Kent WJ: The UCSC Table Browser data retrieval tool. Nucleic Acids Res 2004, 32(Database issue):D493-496. 6. Haeussler M, Raney BJ, Hinrichs AS, Clawson H, Zweig AS, Karolchik D, Casper J, Speir ML, Haussler D, Kent WJ: Navigating protected genomics data with UCSC Genome Browser in a Box. Bioinformatics 2015, 31(5):764-766. 7. Redline S, Sanders MH, Lind BK, Quan SF, Iber C, Gottlieb DJ, Bonekat WH, Rapoport DM, Smith PL, Kiley JP: Methods for obtaining and analyzing unattended polysomnography data for a multicenter study. Sleep Heart Health Research Group. Sleep 1998, 21(7):759-767. 8. Sahoo SS, Lhatoo SD, Gupta DK, Cui L, Zhao M, Jayapandian C, Bozorgi A, Zhang GQ: Epilepsy and seizure ontology: towards an epilepsy informatics infrastructure for clinical research and patient care. J Am Med Inform Assoc 2014, 21(1):82-89. 9. Sahoo SS, Zhang GQ, Bamps Y, Fraser R, Stoll S, Lhatoo SD, Tatsuoka C, Sams J, Welter E, Sajatovic M: Managing information well: Toward an ontology-driven informatics platform for data sharing and secondary use in epilepsy self-management research centers. Health Informatics J 2015. 10. Zhang GQ, Cui L, Lhatoo S, Schuele SU, Sahoo SS: MEDCIS: Multi-Modality Epilepsy Data Capture and Integration System. AMIA Annu Symp Proc 2014, 2014:1248-1257. 11. Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT et al: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet 2000, 25(1):25-29. 12. Cote RA, Robboy S: Progress in medical information management. Systematized nomenclature of medicine (SNOMED). JAMA 1980, 243(8):756-762. 13. Sun M, Zhu W, Tao S, Cui L, Zhang GQ: COBE: A Conjunctive Ontology Browser and Explorer for Visualizing SNOMED CT Fragments. AMIA Annu Symp Proc 2015, 2015:2092-2100. 14. Jayapandian C, Wei A, Ramesh P, Zonjy B, Lhatoo SD, Loparo K, Zhang GQ, Sahoo SS: A scalable neuroinformatics data flow for electrophysiological signals using MapReduce. Front Neuroinform 2015, 9:4. 15. Zhang GQ, Bodenreider O: Using SPARQL to Test for Lattices: application to quality assurance in biomedical ontologies. Semant Web ISWC 2010, 6497:273-288. 16. Zhang GQ, Zhu W, Sun M, Tao S, Bodenreider O, Cui L: MaPLE: A MapReduce Pipeline for Lattice-based Evaluation and Its Application to SNOMED CT. Proc IEEE Int Conf Big Data 2014, 2014:754-759. 17. Cashman DJ, Ortega DR, Zhulin IB, Baudry J: Homology modeling of the CheW coupling protein of the chemotaxis signaling complex. PLoS One 2013, 8(8):e70705. 18. Ortega DR, Zhulin IB: Evolutionary Genomics Suggests That CheV Is an Additional Adaptor for Accommodating Specific Chemoreceptors within the Chemotaxis Signaling Complex. PLoS Comput Biol 2016, 12(2):e1004723. 19. Wuichet K, Alexander RP, Zhulin IB: Comparative genomic and protein sequence analyses of a complex system controlling bacterial chemotaxis. Methods Enzymol 2007, 422:1-31. 20. Wuichet K, Zhulin IB: Origins and diversification of a complex signal transduction system in prokaryotes. Sci Signal 2010, 3(128):ra50. 21. Xie Z, Ulrich LE, Zhulin IB, Alexandre G: PAS domain containing chemoreceptor couples dynamic changes in metabolism with chemotaxis. Proc Natl Acad Sci U S A 2010, 107(5):2235-2240. 22. Alexander RP, Zhulin IB: Evolutionary genomics reveals conserved structural determinants of signaling and adaptation in microbial chemoreceptors. Proc Natl Acad Sci U S A 2007, 104(8):2885-2890. 23. Ortega DR, Yang C, Ames P, Baudry J, Parkinson JS, Zhulin IB: A phenylalanine rotameric switch for signal-state control in bacterial chemoreceptors. Nat Commun 2013, 4:2881. 24. Upadhyay AA, Fleetwood AD, Adebali O, Finn RD, Zhulin IB: Cache Domains That are Homologous to, but Different from PAS Domains Comprise the Largest Superfamily of Extracellular Sensors in Prokaryotes. PLoS Comput Biol 2016, 12(4):e1004862. 25. Adebali O, Reznik AO, Ory DS, Zhulin IB: Establishing the precise evolutionary history of a gene improves prediction of disease-causing missense mutations. Genet Med 2016. 26. Adebali O, Ortega DR, Zhulin IB: CDvist: a webserver for identification and visualization of conserved domains in protein sequences. Bioinformatics 2015, 31(9):1475-1477. 27. Gamazon ER, Wheeler HE, Shah KP, Mozaffari SV, Aquino-Michaels K, Carroll RJ, Eyler AE, Denny JC, Consortium GT, Nicolae DL et al: A gene-based association method for mapping traits using reference transcriptome data. Nat Genet 2015, 47(9):1091-1098. 28. Pulley J, Clayton E, Bernard GR, Roden DM, Masys DR: Principles of human subjects protections applied in an opt-out, de-identified biobank. Clin Transl Sci 2010, 3(1):42-48. 29. Wang J, Gamazon ER, Pierce BL, Stranger BE, Im HK, Gibbons RD, Cox NJ, Nicolae DL, Chen LS: Imputing Gene Expression in Uncollected Tissues Within and Beyond GTEx. Am J Hum Genet 2016, 98(4):697-708. 30. Bernstein BE, Stamatoyannopoulos JA, Costello JF, Ren B, Milosavljevic A, Meissner A, Kellis M, Marra MA, Beaudet AL, Ecker JR et al: The NIH Roadmap Epigenomics Mapping Consortium. Nat Biotechnol 2010, 28(10):1045-1048. 31. Roadmap Epigenomics C, Kundaje A, Meuleman W, Ernst J, Bilenky M, Yen A, Heravi-Moussavi A, Kheradpour P, Zhang Z, Wang J et al: Integrative analysis of 111 reference human epigenomes. Nature 2015, 518(7539):317-330. 32. Zhou X, Li D, Lowdon RF, Costello JF, Wang T: methylC Track: visual integration of single-base resolution DNA methylation data on the WashU EpiGenome Browser. Bioinformatics 2014, 30(15):2206-2207. 33. Zhou X, Wang T: Using the Wash U Epigenome Browser to examine genome-wide sequencing data. Curr Protoc Bioinformatics 2012, Chapter 10:Unit10 10. 34. Tashiro S, Lanctot C: The International Nucleome Consortium. Nucleus 2015, 6(2):89-92. 35. Zhou X, Li D, Zhang B, Lowdon RF, Rockweiler NB, Sears RL, Madden PA, Smirnov I, Costello JF, Wang T: Epigenomic annotation of genetic variants using the Roadmap Epigenome Browser. Nat Biotechnol 2015, 33(4):345-346. 36. Sundaram V, Cheng Y, Ma Z, Li D, Xing X, Edge P, Snyder MP, Wang T: Widespread contribution of transposable elements to the innovation of gene regulatory networks. Genome Res 2014, 24(12):1963-1976. 37. Xie M, Hong C, Zhang B, Lowdon RF, Xing X, Li D, Zhou X, Lee HJ, Maire CL, Ligon KL et al: DNA hypomethylation within specific transposable element families associates with tissue-specific enhancer landscape. Nat Genet 2013, 45(7):836-841. 38. Gosmanov AR, Lu JL, Sumida K, Potukuchi PK, Rhee CM, Kalantar-Zadeh K, Molnar MZ, Kovesdy CP: Synergistic association of combined glycemic and blood pressure level with risk of complications in US veterans with diabetes. J Hypertens 2016, 34(5):907-913. 39. Kovesdy CP, Bleyer AJ, Molnar MZ, Ma JZ, Sim JJ, Cushman WC, Quarles LD, Kalantar-Zadeh K: Blood pressure and mortality in U.S. veterans with chronic kidney disease: a cohort study. Ann Intern Med 2013, 159(4):233-242. 40. Li L, Streja E, Rhee CM, Mehrotra R, Soohoo M, Brunelli SM, Kovesdy CP, Kalantar-Zadeh K: Hypomagnesemia and Mortality in Incident Hemodialysis Patients. Am J Kidney Dis 2015, 66(6):1047-1055. 41. Gaziano JM, Concato J, Brophy M, Fiore L, Pyarajan S, Breeling J, Whitbourne S, Deen J, Shannon C, Humphries D et al: Million Veteran Program: A mega-biobank to study genetic influences on health and disease. J Clin Epidemiol 2016, 70:214-223. P1 CC-PROMISE: Projection onto the Most Interesting Statistical Evidence (PROMISE) with Canonical Correlation to integrate gene expression and methylation data with multiple pharmacologic and clinical endpoints Xueyuan Cao1, Stanley Pounds1, Susana Raimondi2, James Downing2, Raul Ribeiro3, Jeffery Rubnitz3, Jatinder Lamba4 1Department of Biostatistics, St. Jude Children's Research Hospital, Memphis, TN 38105, USA; 2Department of Pathology, St. Jude Children's Research Hospital, Memphis, TN 38105, USA; 3Department of Oncology, St. Jude Children's Research Hospital, Memphis, TN 38105, USA; 4Department of Pharmacotherapy and Translational Research, College of Pharmacy, University of Florida, Gainesville, FL 32610, USA Correspondence: Stanley Pounds (stanley.pounds@stjude.org) – Department of Biostatistics, St. Jude Children's Research Hospital, Memphis, TN 38105, USA Background Projection onto the most interesting statistical evidence (PROMISE) is a general procedure to identify genomic variables that exhibit a specific biologically interesting pattern of association with multiple endpoint variables. It has been successfully applied to multiple studies with gene expression profiling or SNP data to identify genes that are associated with intrinsically related clinical endpoint variables in cancer. In the context of genetic studies in clinical trials, the clinical endpoint variables are related to one another and different forms of genetic data (such as methylation and expression) are also related to one another. Materials and methods Here, CC-PROMISE, PROMISE with canonical correlation, is proposed to extend the PROMISE procedure to integrate DNA methylation, gene expression and multiple pharmacologic and/or clinical variables into one test at the gene level. First, canonical correlation analysis is performed on the multiple probe-sets measuring DNA methylation and gene expression in one gene. Second, the methylation and expression scores of the gene are calculated as the first canonical correlates of the signals of individual methylation probes and expression probes, respectively. Third, perform PROMISE analysis with multiple endpoints on the methylation scores and expression scores separately. Next, the CC-PROMISE statistic is defined as the average of the PROMISE statistics of the methylation and expression scores. Finally, the significance of CC-PROMISE statistics is determined by permuting the endpoint data. Results We applied the CC-PROMISE procedure to the Affymetrix U133A gene expression array and Illumina 450K methylation array data of the multi-center AML02 clinical trial (NCT00136084). The methylation and expression of 202 genes showed a meaningful pattern of association with in vitro drug sensitivity, minimal residual disease, and event-free survival (p ≤ 0.001, q ≤ 0.05). Several of the identified genes are of known relevance to disease biology, showing that CC-PROMISE can make meaningful discoveries by effectively integrating expression, methylation, and clinical data. P2 Integration of microRNA-mRNA interaction networks with gene expression data to increase experimental power Bernie J Daigle, Jr.1,2 (bjdaigle@memphis.edu) 1Department of Biological Sciences, University of Memphis, Memphis, TN 38152, USA; 2Department of Computer Science, University of Memphis, Memphis, TN 38152, USA Background The detection of differentially expressed (DE) genes between two or more biological conditions is an essential step in the search for candidate disease genes, drug targets, and discriminative biomarkers. Although widely used for this task, DNA microarrays are notorious for generating noisy data. One strategy for mitigating the effects of noise is to assay many experimental replicates. However, as this approach can be costly and sometimes impossible with limited resources, analytical methods are needed which improve DE gene identification at no additional cost. Materials and methods An important source of information for differential expression analysis comes from experiments performed on microRNAs (miRNAs). While the transcriptional roles of miRNAs are well documented, principled methods for incorporating miRNA datasets with traditional gene expression assays are lacking. To this end, I developed Noisy-Or Optimization for DifferentiaL Expression analysis (NOODLE), a novel Bayesian network-based approach for integrating miRNA-mRNA interaction networks with microarray data to improve DE gene identification. Given a dataset of interest, NOODLE provides an efficient mechanism for increasing belief that an mRNA is DE if one or more interacting miRNAs are themselves DE (and vice versa). Results I first apply NOODLE to synthetic datasets, achieving more accurate DE gene identification than the popular limma method over a wide range of network topologies and configurations. Using two publicly available human cancer datasets, I next demonstrate how use of NOODLE increases experimental power by as much as a factor of four. Finally, I apply NOODLE to a recent dataset interrogating expression differences between human induced pluripotent and embryonic stem cells. My results uncover important biological differences between these stem cell types that would be missed using existing methods. Acknowledgements Supported in part by the Institute for Collaborative Biotechnologies through grant W911NF-10-2-0111 from the U.S. Army Research Office. P3 Designing and writing software for in silico subtractive hybridization of large eukaryotic genomes Deborah Burgess1, Stephanie Gehrlich2, John C Carmen2 1Department of Computer Science, Northern Kentucky University, Highland Heights, KY 41099, USA; 2Department of Biological Sciences, Northern Kentucky University, Highland Heights, KY 41099, USA Correspondence: John C Carmen (carmenj1@nku.edu) – Department of Biological Sciences, Northern Kentucky University, Highland Heights, KY 41099, USA Background In silico subtractive genome hybridization is a method used to identify unique genetic sequences in a genome of interest by deleting regions of similarity with a control genome. The program Genomic Organismal Subtractive Hybridization (GOSH) was designed and created to complete the subtractive step of in silico subtractive hybridization of large eukaryotic genomes using BLASTn output. A genome file, in the .fasta format, is effectively a word document containing a long string of characters. BLASTn takes two genome files and identifies areas of alignment characterized by significant sequence overlap. BLASTn can only compare two genomes. In order to identify genes found specifically in a group of organisms (i.e. the thermally dimorphic fungi), we need to compare multiple genomes or perform sequential subtractive steps. Materials and methods GOSH was designed to subtract areas of alignment from genomes using BLASTn output files. The program accesses the alignment provided by BLASTn, identifies the nucleotides making up the shared sequences in the genome, and replaces them with Ns in the genome file. Through iterative subtractions, a database containing genomic sequence shared by the thermal dimorphs Blastomyces dermatitidis, Paracoccidioides brasieliensis, and Histoplasma capsulatum is created which can then be aligned with the genomes of fungi which do not exhibit thermal dimorphism (e.g. Saccharomyces cerevisiae). Results GOSH successfully used the alignment data provided by BLASTn to modify a genome sequence file by replacing nucleotides in one genome with Ns when they aligned with another fungal genome. Conclusions Initially attempts to perform in silico subtractive genome hybridization using GOSH to perform iterative subtractions generated a database of sequence. However, further investigation found that the file contained multiple replicates of genomic sequence. Currently, efforts are underway to determine whether this is due to GOSH, to an inherent flaw in the design of the subtractive steps, or to the numerous repeats found in the genomes of B. dermatitidis and H. capsulatum. P4 Tracking the molecular evolution of Pax gene Nicholas Johnson, Chandrakanth Emani Department of Biology, Western Kentucky University, Bowling Green, KY 42101, USA Correspondence: Nicholas Johnson (traydreein@gmail.com) – Department of Biology, Western Kentucky University, Bowling Green, KY 42101, USA Background The purpose of this study is to trace the molecular evolution of the Paired Box (PAX) gene that was frequently expressed in diverse cancers and was crucial for growth and survival of cancer cells. The long-term goal of the study is to identify conserved domains of the PAX protein in a model organism that can be suitable drug targets for plant based anti-cancerous pharmaceutical compounds. The paired-box (PAX) genes encode a family of nine well-characterized paired-box transcription factors, with important roles in development and disease. PAX genes are primarily expressed during embryo development, but very frequent expression was observed in diverse tumor cell lines such as lymphoma, breast, ovarian, lung, and colon cancer. A phylogenetic analysis of the PAX protein will identify crucial molecular elements that are implicated in cancer cell survival. Materials and methods In the present study the Human PAX1 protein sequence was used as a reference to retrieve homologous sequences using PSI-BLAST. A neighbor joining phylogenetic tree was developed using MEGA6. Results The conserved domains of PAX gene were identified as HTH – Helix-Turn-Helix that are shown to mediate responses to stress including exposure to heavy metals, drugs, or oxygen radicals across life forms. These could be novel targets for treatment as expression of PAX2 domain has been linked with cell survival cell migration and invasion. The generated neighbor-joining phylogenetic tree showed that the ancestral form of PAX traces back to the American alligator with another close relative in a model organism, zebrafish. PAX homologs were present extensively in birds with the most evolved form in the Golden-collared Manakin. Conclusions The bioinformatics analysis proved crucial in identifying a model organism for researching a plant based cancer treatment in the form of zebrafish, which is already being used as a model organism to study cancer. Present research in our lab already established the anti-cancerous action of plant based extracts on diverse cancers in terms of reduced cell proliferation. The study thus suggests an effective tool to research plant based treatments of diverse cancers by identifying the crucial molecular domains as targets. P5 Identifying genetic differences in thermally dimorphic and state specific fungi using in silico genomic comparison Stephanie Gehrlich1, Deborah Burgess2, John C Carmen1 1Department of Biological Sciences, Northern Kentucky University, Highland Heights, KY 41099, USA; 2Department of Computer Science, Northern Kentucky University, Highland Heights, KY 41099, USA Correspondence: John C Carmen (carmenj1@nku.edu) – Department of Biological Sciences, Northern Kentucky University, Highland Heights, KY 41099, USA Background Thermally dimorphic fungi such as Blastomyces dermatitidis and Histoplasma capsulatum have the rare, amongst fungi, ability to cause invasive infections in otherwise healthy individuals. In order to better understand what makes these fungi different from state specific fungi, we have used two bioinformatics methods to compare dimorphic and non-dimorphic fungal genomes. Materials and methods First, we compiled a list of “missing” enzymes using BLAST to search for homologs of Saccharomyces cerevisiae proteins in B. dermatitidis and H. capsulatum. Students searched for the components of 24 S. cerevisiae pathways in H. capsulatum and 38 S. cerevisiae pathways in B. dermatitidis. Second, biological sciences and computer science students collaborated to create GOSH, software designed to be used to complete in silico subtractive genome hybridization of large fungal genomes when combined with BLASTn. A file containing thermal dimorph-shared genome sequences was created by performing multiple alignments and subtractions between the thermally dimorphic fungi of interest. This list of sequences was use to interrogate the genome the yeast S. cerevisiae and the genome of the mold A. fumigatus. Results We found 29 S. cerevisiae proteins from various different pathways lacking a clear homolog in H. capsulatum. Similarly, we found 39 missing components when we searched for S. cerevisiae. Preliminary results of in silico subtractive genome hybridization using GOSH combined with BLASTn are promising. The iterative analysis pipeline identified multiple genome sequences found in dimorphic fungi but not one of the state-specific fungi. Conclusions We started two separate avenues of in silico genome comparison to identify genomic differences between thermally dimorphic fungi and state specific fungi. Both methods yielded results confirming that the dimorphic fungal genomes differ from state-specific fungi. Efforts are currently underway to expand on and characterize these differences. P6 Identification of conserved genomic regions and variation therein amongst Cetartiodactyla species using next generation sequencing Kalpani De Silva1, Michael P Heaton2, Theodore S Kalbfleisch3 1Interdisciplinary Studies Program: Specialization in Bioinformatics, University of Louisville, Louisville, KY 40292, USA; 2United States Department of Agriculture, Agricultural Research Service, United States Meat Animal Research Center, Clay Center, NE 68933, USA; 3Department of Biochemistry and Molecular Genetics, School of Medicine, University of Louisville, Louisville, KY 40292, USA Correspondence: Theodore S Kalbfleisch (ted.kalbfleisch@louisville.edu) – Department of Biochemistry and Molecular Genetics, School of Medicine, University of Louisville, Louisville, KY 40292, USA Background Next Generation Sequencing has created an opportunity to genetically characterize an individual both inexpensively and comprehensively. In earlier work produced in our collaboration [1], it was demonstrated that, for animals without a reference genome, their Next Generation Sequence data can be mapped to the reference genome of another animal from which it has recently evolutionarily diverged producing a wealth of data on regions which have been evolutionarily conserved, and variation therein which has been tolerated. Materials and methods Since then, 16 different animals spanning 10 different Cetartiodactyla species have been sequenced and mapped to the Bos taurus genome at an equivalent coverage of 10X. Here, we describe this resource which is publicly available and may be found at 10x WGS of Cetartiodactyla Species Mapped to Cattle [http://www.ars.usda.gov/Research/docs.htm?docid=25590]. Results Analysis of these mappings identifies genomic regions in the respective species that are highly conserved relative to cattle. Within these conserved regions, species specific alleles selected for by evolution can be identified as well as sites that vary within the respective species. Here we present a summary of non-bovine alleles that can be measured across these species relative to the Bovine reference genome, and identify those which appear to be common to the species, and those which are likely variant within the species. References 1. Kalbfleisch T, Heaton M: Mapping whole genome shotgun sequence and variant calling in mammalian species without their reference genomes.F1000Res. 2014; 2:244. P7 Mining physiological data to identify patients with similar medical events and phenotypes Teeradache Viangteeravat1, 2, Rahul Mudunuri1, Oluwaseun Ajayi1, Fatih Şen1, Eunice Y Huang1,2 1Biomedical Informatics Core, Children’s Foundation Research Institute, Memphis, TN 38103, USA; 2Department of Pediatrics, University of Tennessee Health Science Center, Memphis, TN 38163, USA Correspondence: Rahul Mudunuri (rmudunur@uthsc.edu) – Biomedical Informatics Core, Children’s Foundation Research Institute, Memphis, TN 38103, USA Background The volume of data that is being generated in the hospital is very large and this large volume is due to the continuous collection of sequential time series data. Clinicians are expected to examine large volumes of data along with readily available electronic medical records of patients and identify correlations between dozens to hundreds of variables based on their own clinical experience to detect significant medical events. Here, we demonstrate a data mining approach applied on physiological data to identify symbolic patterns to derive patient similar matrices that will allow clinicians to identify patients with similar events and phenotypes for the purpose of predicting patient outcomes. Materials and methods We employed symbolic aggregate approximation (SAX) and piecewise aggregate approximation (PAA) techniques to convert oxygen saturation (SpO2) to symbolic patterns (Fig. 1a). We then applied data mining techniques called Low Rank Matrix Decomposition (LRMD) on these symbolic patterns to produce a concept vector space in which query vector and symbolic term-to-patient were projected. Resulting patients with similar events for each query are determined and compared with a control reference of 563 de-identified patients with asthma or related conditions using Precision and Recall measurements at various time intervals prior to outcomes (i.e., death, intubation, or transfer (DIT) to Intensive Care Unit). Results We found that this approach identified specific common patterns of SpO2 level among cases. For example, patterns ‘cab’, ‘bac’ and ‘acb’ exist among the patients 4, 11, 13, 15, 9, 2, 5, 7, 12, 14, 10, 3 and 6 with DIT outcomes (Fig. 1b). This results indicate their possible clinical use in the future to predict impending respiratory failure. Acknowledgements The authors would like to thank the Children’s Foundation Research Institute (CFRI), Le Bonheur Children’s Hospital, who supported this work.Fig. 1 (abstract P7) a Representative SAX results for SpO2 levels; b Clustering and Heatmap analysis of LRMD P8 Smart brief for home health monitoring Mohammad Mohebbi, Luaire Florian, Douglas J Jackson, John F Naber Department of Electrical and Computer Engineering, University of Louisville, Louisville, KY 40292, USA Correspondence: John F Naber (john.naber@louisville.edu) – Department of Electrical and Computer Engineering, University of Louisville, Louisville, KY 40292, USA Background A system for monitoring the moisture level and temperature of a disposable brief will be presented. The system includes a Bluetooth wireless transmitter that works in tandem with a smart phone/tablet application. Materials and methods An inexpensive disposable moisture sensor combined with a re-usable temperature sensor are used as the wireless sensors. The Bluetooth Low-Energy (BTLE) transceiver is powered by a 2032 coin cell battery having a lifetime greater than 3 years. The BTLE device collects data from the sensors and transmits the data to a smart phone/tablet. The data can be monitored locally on a smart phone/tablet or stored on a server to be analyzed remotely. Results A working prototype developed in the Wireless and IC design Laboratory at the University of Louisville (Fig. 2) has the following features and demonstrated characteristics: Features:BTLE transmitter module sits externally on a standard disposable brief and is connected to the brief using snaps. These snaps provide electrical connection to the embedded moisture sensor. The BTLE equipped smart phone or tablet easily interfaces with the internet. The module has a low-battery alert consisting of a beep, LED or text message. The smart phone / tablet sends data to a server that is time stamped for statistical analysis. Performance and Characteristics:The module includes a built-in temperature sensor with an accuracy of +/- 0.5 C. The module contains a non-replaceable coin cells that last up to 3 years with a 10 second data sampling period. The module has a range of at least 50’ indoors. The module can monitor approximate distance from the smart phone / tablet. The module costs approximately $15 in low volume.Fig. 2 (abstract P8) (Left): Top view of the hermetically sealed module with 2 brief snaps on the bottom. (Right): internal view showing the custom circuit board, BTLE module and coin cell holder. P9 Side-effect term matching for computational adverse drug reaction predictions AKM Sabbir1, Sally R Ellingson2 1Department of Computer Science, University of Kentucky, Lexington, KY 40506, USA; 2Division of Biomedical Informatics, College of Medicine, University of Kentucky, Lexington, KY 40536-0093, USA Correspondence: Sally R Ellingson (sally@kcr.uky.edu) – Division of Biomedical Informatics, College of Medicine, University of Kentucky, Lexington, KY 40536-0093, USA Background The establishment of polypharmacological networks, all the interactions between a collection of drugs and proteins, will allow for the exploration of drug re-purposing, side effect prediction, and the development of more efficacious drugs, targeting multiple proteins in a disease pathway. This project will help pave the way for the computational prediction of adverse drug reactions using a polypharmacological network built using molecular docking scores, an efficient prediction of how and how well drugs bind to a protein. The research presented here is an effort to map side-effect terms associated with a toxicity screen [1], known proteins in which drugs should not interact, to known side-effects associated with FDA-approved drugs [2] in order to test the accuracy of predictions. Materials and methods Multiple methods were tested for term matching.Edit Distance determines the minimum amount of editing required to transform a source string into a destination string. Editing operations include insertion, deletion and substitution. Each operation involves a cost and a minimum cost path is found. Knuth-Morris-Pratt is a pattern-matching algorithm that finds the number of times a given text pattern occurs within a text document, which has been modified to include a cost function. Sense Disambiguation using meta map takes advantage of UMLS (Unified medical language system) metathesaurus, which is a large multi-purpose thesaurus containing medical and health related concepts, and builds concept networks, relating different concepts and their synonyms. Language Model takes a partially formed sentence and tries to find the most appropriate word or words to form the full and proper sentence. Pubmed journal abstracts and titles were used to train the language model. Results Sense disambiguation appears to be the most accurate methods but only resolved around 75% of the terms. Multi-approach methods are being investigated to achieve the maximum number of terms matched with a high-degree of accuracy. Conclusion This significant research will help alleviate the current economic burden of developing new pharmaceuticals by innovatively utilizing massive computational power. This research will lead to the establishment of structures to use in virtual drug screenings that will predict side-effects quickly and efficiently, resulting in safer clinical trials, as fewer drugs with negative off-target effects will advance to this stage, and more affordable therapies, as drugs destined for failure will be predicted earlier in the drug discovery process. This project will address important public health concerns by providing safer and more affordable drugs. Acknowledgements This work was supported by the National Institutes of Health (NIH) National Center for Advancing Translational Science grant KL2TR000116. References 1. Bowes J, Brown AJ, Hamon J, Jarolimek W, Sridhar A, Waldron G, Whitebread S: Reducing safety-related drug attrition: the use of in vitro pharmacological profiling.Nature rev drug discov. 2012; 11(12), 909-922. 2. Kuhn M, Letunic I, Jensen LJ, Bork P: The SIDER database of drugs and side effects. Nucleic acids res. 2016; 44(D1):D1075-D1079. P10 Enrichment vs robustness: A comparison of transcriptomic data clustering metrics Yuping Lu, Charles A Phillips, Michael A Langston Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville, TN 37996-2250, USA Correspondence: Michael A Langston (langston@tennessee.edu) – Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville, TN 37996-2250, USA Background Transcriptomic graph density and community structure remain hallmarks of putative biological fidelity. Yet these very graphs frequently have numerous maximum cliques, forcing top-down, density-based algorithms to choose a starting clique in some fashion, either randomly or by some often-arbitrary tie-breaking scheme. Materials and methods In order to help evaluate the potential effectiveness of selection strategies, we investigate the impacts of clique choice on cluster ontology enrichment and robustness. We employ yeast gene co-expression data obtained from the Gene Expression Omnibus, and create graphs in the usual fashion, by calculating all pairwise correlations and placing edges between pairs correlated at or above a selected threshold. We then run the noise-resilient paraclique algorithm to generate gene clusters. For enrichment, we use GO p-values obtained from DAVID, and compare clusters obtained by repeatedly using the maximum clique with the highest average edge weight (correlation) to clusters obtained using the lowest average edge weight. While the use of higher weights has much intuitive appeal, we find that with proper threshold selection this choice seems to have at most a negligible effect on paraclique enrichment. For robustness, we introduce a new metric defined as t/(dr), expressed as a percentage, where t denotes the total number of (not necessarily distinct) gene pairs appearing together across all clusters, d represents the number of distinct pairs appearing together in at least one cluster, and r is the number of runs. Robustness thus falls between 0% and 100%. We find that paraclique scores are generally 80% or higher, demonstrating that it produces highly repeatable cluster profiles regardless of the particular starting clique chosen. P11 Deep neural networks for transcriptome-based cancer classification Rahul K Sevakula1, Raghuveer Thirukovalluru1, Nishchal K. Verma1, Yan Cui2 1Department of Electrical Engineering, Indian Institute of Technology Kanpur, Kanpur, India 208016; 2Department of Microbiology, Immunology and Biochemistry, University of Tennessee Health Science Center, Memphis, TN 38163 USA Correspondence: Yan Cui (ycui2@uthsc.edu) – Department of Microbiology, Immunology and Biochemistry, University of Tennessee Health Science Center, Memphis, TN 38163 USA Background Deep neural networks are shown to learn complex relationships in data and provide greater generalization performance in diagnostic informatics [1, 2]. Using deep learning techniques for large-scale omics data can be computationally expensive, as it involves learning of millions of network weights. This abstract focuses on designing efficient deep learning models that are sufficient to perform transcriptome-based cancer classification with minimal computational elements. This includes the use of a two layered stacked de-noising sparse auto-encoders (SDSAE) [3] to generate a feature representation that is 200 fold smaller than the input representation, and then use a novel fine-tuning method for improved classification performance. Materials and methods One of the main challenges in solving complex supervised machine learning problems is to come up with good features. SDSAE in context of Deep Learning is widely used in a variety of tasks for generation of useful feature representations. SDSAE removes redundancies while learning compact feature representations and helps the network learn important statistical regularities. The reduction in number of computational elements was achieved by drastically reducing the feature representation from input layer to 1st hidden layer, and a moderate reduction from 1st to 2nd hidden layer. The fine-tuning method on the other hand forces data i.e. feature values in the new representation to converge towards the median of the respective class samples. Results The entire procedure was validated upon the two class Prostate Tumor data [4] containing 102 samples and 10509 features. The 2 layered neural network had 10509 nodes in input layer, 100 nodes in 1st hidden layer and 49 nodes in 2nd hidden layer. Network weights connecting the layers were initialized with denoising sparse auto-encoders and fine-tuned. Support Vector Machine (C-SVM) was used for classifying the data samples in the final feature representation. Our results show that while the network had fewer computational elements, the deep learning model with fine-tuning method gave better performance than support vector machines and random forests. Table 1 reports the four-fold cross validation AUC values (Area Under ROC Curve) of several methods. References 1. Qiao C, Lin DD, Cao SL, Wang YP: The effective diagnosis of schizophrenia by using multi-layer RBMs deep networks,IEEE bioinformatics and biomedicine 2015 (BIBM 2015). 2015; 603-606. 2. Lena PD, Nagata K, Baldi PF: Deep spatio-temporal architectures and learning for protein structure prediction, Advances in neural information processing systems (NIPS’12) 2012. 2012; 512-520. 3. Vincent P, Larochelle H, Lajoie L, Bengio Y, Manzagol PA: Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion, J mach learn res. 2010; 11:3371–3408. 4. Singh D, Febbo PG, Ross K, Jackson DG, Manola J, Ladd L, Tamayo P et al. Gene expression correlates of clinical prostate cancer behavior, Cancer cell. 2002; 1(2):203-209.Table 1 (abstract P11) Results on prostate tumor dataset over four-fold cross validation. Method AUC Random Forest 0.9340 Linear Kernel SVM 0.9565 RBF Kernel SVM 0.8950 SDSAE + Fine Tuning + Linear kernel SVM 0.9569 SDSAE + Fine Tuning + RBF kernel SVM 0.9747 P12 Motif discovery using K-means clustering Mohammed Sayed1, Juw Won Park1,2 1Department of Computer Engineering and Computer Science, University of Louisville, Louisville, KY 40202, USA; 2Kentucky Biomedical Research Infrastructure Network (KBRIN) Bioinformatics Core, University of Louisville, Louisville, KY 40292, USA Correspondence: Juw Won Park (juw.park@louisville.edu) – Kentucky Biomedical Research Infrastructure Network (KBRIN) Bioinformatics Core, University of Louisville, Louisville, KY 40292, USA Background DNA motifs are short patterns in DNA sequence and are usually associated with a biological function. The existing techniques for identifying these motifs are either computationally prohibitive or stuck at a local minimum. In this study, we propose a hybrid technique which combines both profile and word-based approaches. The proposed technique has comparable performance with classical tools such as MEME [1] and Weeder [2]. Materials and methods Two types of motif features were used to discriminate motifs from background subsequences. First, relative complexity (ratio of N-mer complexity to average of background N-mers complexities) was used to capture motif structure. Second, contextual features like a position-specific scoring matrix (PSSM)-based score was used to isolate a motif from its background sequence. Unlike the expectation maximization (EM) algorithm, PSSM is computed from only one sequence. In turn, the score for each motif is constant and can be computed independent of the clustering technique. To simplify the clustering step, features were multiplied to form one composite feature. This will reduce feature space and consequently increase the speed. K-means clustering technique was used to cluster all possible N-mers into two cluster (background and candidate). To capture over-represented candidate motifs, N-mers in candidate cluster are counted and ones with high count were determined. Finally, candidates with more than one occurrence per sequence were removed. To evaluate the proposed technique, benchmark proposed by Sandve [3] was used. This benchmark includes 50 datasets from TRANSFAC database [4]. Different performance measures were calculated in nucleotide-level of both known and predicted sites [5]. Based on true positives (TP), true negatives (TN), false positives (FP) and false negatives (FN), the correlation coefficient was calculated: CorrelationCoefficient(CC)=(TP*TN−FP*FN)(TP+FN)(TP+FP)(TN+FN)(TN+FP) Results Results generated from the benchmark are presented in Table 2. These results represent the average of each performance measure over the 50 datasets. Weeder showed the highest sensitivity followed by our technique and MEME. Weeder has high sensitivity because it allows few mismatched in pattern search. On the other hand, our technique searches for the exact match. It was also interesting to see that Weeder has the highest sensitivity with the lowest specificity. In contrast, MEME has the highest specificity with the lowest sensitivity. In conclusion, our proposed technique was comparable with the existing techniques and its sensitivity can be improved with allowing mismatches in pattern search. Acknowledgements This work was supported by National Institutes of Health (NIH) grants P20GM103436 (Nigel Cooper, PI) References 1. Bailey TL, Elkan C: Fitting a mixture model by expectation maximization to discover motifs in biopolymers. Proc int conf intell syst mol biol. 1994; 2:28-36. 2. Pavesi G, Mereghetti P, Mauri G, Pesole G: Weeder Web: discovery of transcription factor binding sites in a set of sequences from co-regulated genes. Nucleic acids res. 2004; 32(suppl 2):W199-W203. 3. Sandve GK, Abul O, Walseng V, Drabløs F: Improved benchmarks for computational motif discovery. BMC bioinformatics. 2007; 8(1):193. 4. Wingender E, Dietze P, Karas H, Knüppel R: TRANSFAC: a database on transcription factors and their DNA binding sites. Nucleic acids res. 1996; 24(1):238-241. 5. Tompa M, Li N, Bailey TL, Church GM, De Moor B, Eskin E, Favorov AV, Frith MC, Fu Y, Kent WJ: Assessing computational tools for the discovery of transcription factor binding sites. Nat biotechnol. 2005; 23(1):137-144.Table 2 (abstract P12) Results of the three methods: MEME, Weeder and K-means approach Sensitivity (SN) Specificity (SP) Correlation Coefficient (CC) MEME 0.103 0.982 0.083 Weeder 0.202 0.960 0.096 K-means approach 0.176 0.971 0.090 P13 Large scale discovery of active enhancers from nascent RNA sequencing Jing Wang1,2, Qi Liu1,3, Yu Shyr1,2,4 (yu.shyr@vanderbilt.edu) 1Center for Quantitative Sciences, Vanderbilt University School of Medicine, Nashville, TN 37232, USA; 2Department of Cancer Biology, Vanderbilt University School of Medicine, Nashville, TN 37232, USA; 3Department of Biomedical Informatics, Vanderbilt University School of Medicine, Nashville, TN 37232, USA; 4Department of Biostatistics, Vanderbilt University School of Medicine, Nashville, TN 37232, USA Correspondence: Qi Liu (qi.liu@vanderbilt.edu) – Department of Biomedical Informatics, Vanderbilt University School of Medicine, Nashville, TN 37232, USA Background Global nuclear run-on sequencing (GRO-seq) and precision nuclear run-on sequencing (PRO-seq) are techniques for mapping and quantifying transcriptionally engaged polymerase density genome-wide. They have been widely used for measuring RNA polymerase pausing and elongation, and condition-dependent transcription response. In addition, they provide a sensitive way to identify and quantify enhancer-derived RNAs (eRNAs), which is a robust indicator of enhancer activity. Results We developed a method to identify active enhancers from GRO/PRO-seq data. Applied the method to human 15 cell lines, over ten thousands of active enhancers were uncovered including 80% novel enhancers. Aligning histone modification data to the enhancer centers, we found that novel enhancers were flanked by expected histone modification markers of H3K4me1 and H3K27ac. Moreover, the signal flanking novel enhancers was as strong as that around known enhancers, indicating the reliability of novel active enhancers identified from GRO/PRO-seq data. In 12 of the 15 cell lines, the transcription abundance of enhancer-linked genes was found significantly higher than the expression of non-linked genes. In addition, the tissue specific genes were observed to locate remarkably closer to tissue specific enhancers than to universal enhancers, suggesting the regulation role of tissue specific enhancers on tissue specific expression. Conclusions The method provides efficient analysis of GRO/PRO-seq data for active enhancer identification. The large-scale discovery of active enhancers across multiple human cell lines provides valuable source for enhancer study. P14 Computationally characterizing genomic pipelines and benchmarking results using GATK best practices on the high performance computing cluster at the University of Kentucky Xiaofei Zhang1, Sally R Ellingson2 1Department of Computer Science, University of Kentucky, Lexington, KY 40506, USA; 2Division of Biomedical Informatics, College of Medicine, University of Kentucky, Lexington, KY 40536-0093, USA Correspondence: Sally R Ellingson (sally@kcr.uky.edu) – Division of Biomedical Informatics, College of Medicine, University of Kentucky, Lexington, KY 40536-0093, USA Background As genetic sequence data is now being used to make health care decisions, analysis tools needed for personalized medicine must be well tested and verified while establishing and maintaining competency in the state-of-the-art in both the technology and analysis. This study demonstrates the usefulness of high-confident call sets (validated genomic variations) in testing and optimizing bioinformatics pipelines. Materials and methods The Genome Analysis Tool Kit (GATK) [1, 2] best practices pipeline for genomic variation detection was used on two Illumina Hi-Seq genomic datasets obtained from a sample originating from NA12878, a participant in the HapMap [3, 4] project. One of the test datasets consists of four pairs of paired end data from different runs with an average depth of coverage of 14. The other consists of one pair of paired end data with an average depth of coverage of 58. Two high-confident call sets are used to detect the accuracy of the pipeline. The National Institute for Standards and Technology call set developed by the Genome in a Bottle Consortium incorporates several sequencing technologies and analysis methods [5] and the Illumina Platinum Genome call set requires concordance across multiple analysis methods and incorporates an inheritance structure. Results In this study, several types of alternatives in the entire workflow were evaluated. 1. Experimental conditions: one sequencing run with a higher depth of coverage has about 1% lower true positive rate and .1% higher positive predictive power than four runs with lower coverage each. 2. Computational architecture: threading to efficiently use a 16 CPU node gives a speed-up of almost 4.5 times that of using only one CPU; however, utilizing a 32 CPU node only gives a speed-up of 1.1 over that of a 16 CPU node. 3. Analysis tools: UnifiedGenotyper is about 7 times faster than HaplotypeCaller which only has about a 1-2% increase in true positive rates. 4. Comparison tools: GATK VariantEval, Useq vcfcomparator, and RTG vcfeval all produce similar comparison results. Conclusion A workflow that easily and reproducibly tests the accuracy and efficiency of a given method on a given computational platform is critical in order to confidently and cost-effectively utilize genomic sequencing in a clinical setting. Acknowledgements This work was supported by the National Institutes of Health (NIH) National Center for Advancing Translational Science grant KL2TR000116. References 1. GATK Best Practiceshttps://www.broadinstitute.org/gatk/guide/best-practices (accessed December 17, 2015). 2. McKenna A, Hanna M, Banks E, Sivachenko A, Cibulskis K, Kernytsky A, Garimella K, Altshuler D, Gabriel S, Daly M, DePristo MA: The Genome Analysis Toolkit: a MapReduce framework for analyzing next-generation DNA sequencing data. Genome res. 2010; 20(9), 1297-303. 3. Gibbs RA, Belmont JW, Hardenbol P, Willis TD, Yu F, Yang H, Ch'ang L-Y, Huang W, Liu B, Shen Y: The international HapMap project. Nature. 2003; 426(6968), 789-796. 4. The International Hapmap Consortium: Integrating common and rare genetic variation in diverse human populations. Nature. 2010; 467(7311), 52-58. 5. Zook JM, Chapman B, Wang J, Mittelman D, Hofmann O, Hide W, Salit M: Integrating human sequence data sets provides a resource of benchmark SNP and indel genotype calls.Nature biotech. 2014; 32:246-251. P15 Development of approaches enabling the identification of abnormal gene expression from RNA-Seq in personalized oncology Naresh Prodduturi1†, Gavin R Oliver1†, Diane Grill1, Jie Na1, Jeanette Eckel-Passow1, Eric W Klee1 1Department of Biomedical Informatics and Statistics, Health Sciences Research, Mayo Clinic, Rochester, MN 55905, USA Correspondence: Gavin R Oliver (oliver.gavin@mayo.edu) – Department of Biomedical Informatics and Statistics, Health Sciences Research, Mayo Clinic, Rochester, MN 55905, USA †Contributed equally Background While high-throughput DNA-based clinical assays are becoming increasingly common, RNA-based approaches primarily remain a research activity. Nonetheless, the potential for clinical benefit offered by RNA-Seq, particularly in the field of personalized oncology, is substantial. RNA-Seq expression profiling of tumor samples may reinforce or deconvolute the findings of DNA-based testing. This is done by revealing the existence and magnitude of deviations from normal levels of gene expression in the tumor’s complex molecular landscape. Such evidence could ultimately be used to highlight and select targeted treatment options to improve patient care. However, multiple challenges exist to the routine implementation of these methods. One notable obstacle, particularly related to metastatic cancer, is the lack of normal tissue from the same site with which to compare RNA expression levels. Other challenges include the impact of reference sample size, reference tissue source, and the appropriate normalization and differential expression (DE) method for the N = 1 tumor sample. These challenges must be solved in order to categorize gene expression as normal or abnormal and therefore to unlock the potential of RNA-Seq profiling in the personalized oncology setting. We describe an approach taken within Mayo Clinic’s Center for Individualized Medicine (CIM) to compile normal reference expression ranges and thus evaluate DE in a N = 1 tumor sample. Utilizing data generated in house and by the The Cancer Genome Atlas (TCGA), we formulated a workflow and analytical methods that should inform and enable other researchers working in the personalized oncology field. Results We developed a bootstrap based confidence interval method to identify DE genes in a N = 1 patient tumor using different references. RLE (Relative Log Expression), 75th Quantile and RPKM (reads per kilo base per million mapped reads) normalization methods has a minimal effect upon DE calling. An increase in the number of normal reference samples stabilized the number of DE calls. We observed higher gene level variance in the TCGA reference samples in comparison with in-house reference samples. DE calls using a reference from the same site were compared against mixed reference tissue samples from bladder, breast, uterine and lung tissues. While not ideal, if reference samples from the same tissue site are not available, mixed tissue reference samples can be used to recover some of the DE genes. Acknowledgements This work was supported by the Center for Individualized Medicine, Mayo Clinic. References The results shown here are in whole or part based upon data generated by the TCGA Research Network: http://cancergenome.nih.gov/. P16 Processing RNA-Seq data of plants infected with coffee ringspot virus Michael M Goodin1, Mark Farman1, Harrison Inocencio2, Chanyong Jang1, Jerzy W Jaromczyk2, Neil Moore1, Kelly Sovacool3 1Department of Plant Pathology, University of Kentucky, Lexington, KY 40546, USA; 2Department of Computer Science, University of Kentucky, Lexington, KY 40506, USA; 3Department of Biology, University of Kentucky, Lexington, KY 40506, USA Correspondence: Jerzy W Jaromczyk (jurek@cs.uky.edu) – Department of Computer Science, University of Kentucky, Lexington, KY 40506, USA Background Coffee is a widely traded agricultural commodity across the globe. The emerging coffee ringspot virus (CoRSV) reduces the quality of beans harvested and amount produced by infected plants [1]. Besides coffee, CoRSV also infects Chenopodium quinoa when incubated at 28° C (4° C above typical conditions for this plant) – an expanded host range which may indicate increasing risk to crops as global temperatures continue to rise [1]. Examining the differences in expression levels between infected and uninfected C. quinoa may shed light into the effect of CoRSV on gene expression and the effect of temperature on host susceptibility. Material and methods As an initial stage toward this goal, we developed methods for processing RNA-Seq data of virus-infected plants for which there is no reference genome. Our methods began with paired-end Illumina RNA-Seq data from three samples of Chenopodium quinoa: one sample infected with CoRSV and incubated at 28° C (28V), another sample uninfected and incubated at 28° C (28H), and the third uninfected and incubated at 24° C (24H). First, we analyzed the quality of the RNA-Seq data with fastqc [2], and trimmed low quality reads with Trimmomatic-0.30 [3]. We then aligned the trimmed reads to the viral RNA genome (GenBank accession numbers KF812525.1 and KF812526.1) [1] using Bowtie 2 [4]. We used HTSeq-count [5] to determine the number of reads that mapped to each viral gene. Reads from the uninfected samples which mapped to viral genes were examined for the possibility of artifacts arising from sample bleeding. To assist with this examination, we wrote a Python program to visualize the layout of reads as they were arranged on the Illumina flow cell. Finally, we are in the process of building a de novo transcriptome assembly of the non-viral reads using trinityrnaseq-2.1.1 [6]. Results The results of processing the data with the Python program indicate that the three data sets were well dispersed across the flow cell. Pixels in the image are colored based on the percent composition of each data set in a particular area of the flow cell (Fig. 3). Gray areas indicate an even proportion of reads from each data set. Additionally, for each read that mapped to the viral genome from the healthy plants (28H and 24H), the identity of the nearest neighbor on the flow cell was determined. Only 3.01% of 28H reads and 2.75% of 24H reads had viral reads from the virus-infected plant as their nearest neighbors (Table 3). Thus, sample bleeding during sequencing cannot explain the vast majority of viral reads obtained from the healthy plants. Alternative hypotheses must be investigated to account for these results. Top image: Tile 1301 at scale 100. Bottom image: red-outlined section from Tile 1301 at scale 1. Red pixels represent 28V reads, green pixels 28H reads, blue pixels 24H reads, and black pixels represent viral reads from any of the three samples. Acknowledgements This research is supported by a CAFE Seed Proposal awarded to M. Goodin. References 1. Ramalho TO, Figueira AR, Sotero AJ, et al: Characterization of Coffee ringspot virus-Lavras: a model for an emerging threat to coffee production and quality. Virology. 2014; 464-465:385-96. 2. Patel RK, Jain M: NGS QC Toolkit: A toolkit for quality control of next generation sequencing data.PLoS one. 2016; 7(2): e30619. 3. Bolger, AM, Lohse M, Usadel, B: Trimmomatic: A flexible trimmer for Illumina sequence data.Bioinformatics. 2014; 30(15)2114-2120. 4. Langmead B, Salzberg S: Fast gapped-read alignment with Bowtie 2.Nat methods. 2012; 9:357-359. 5. Anders S, Pyl PT, Huber W: HTSeq--a Python framework to work with high-throughput sequencing data.Bioinformatics 2015, 31(2):166-9. 6. Grabherr MG, Haas BJ, Yassour M, et al: Full-length transcriptome assembly from RNA-Seq data without a reference genome.Nat biotechnol. 2011; 29(7):644-52.Fig. 3 (abstract P16) Scaled output of Tile 1301. Table 3 (abstract P16) Analysis of sample bleeding plausibility. Total viral reads Number of viral reads plausibly from bleeding Fraction of viral reads plausibly from bleeding 28H reads 2,692 81 3.01 % 24H reads 2,037 56 2.75 % P17 Comparative transcriptomics of three Acinetobacter baumanii clinical isolates with different antibiotic resistance patterns Leon Dent1, Mike Izban2, Sammed Mandape3, Shruti Sakhare3, Siddharth Pratap3, Dana Marshall2 1Department of Surgery, Meharry Medical College, Nashville, TN 37208, USA; 2Department of Pathology, Anatomy and Cell Biology, Meharry Medical College, Nashville, TN 37208, USA; 3Bioinformatics and Proteomics Core, Microbiology and Immunology, Meharry Medical College, Nashville, TN 37208, USA Correspondence: Dana Marshall (dmarshall@mmc.edu) – Department of Pathology, Anatomy and Cell Biology, Meharry Medical College, Nashville, TN 37208, USA Background Acinetobacter baumanii is an important nosocomial pathogen in the US and worldwide. It is of great concern because it rapidly acquires antibiotic resistance (multi-drug resistant A. baumanii or MDRAB) and is resistant to all antibiotics, except the polymixins, in many medical centers. The time to effective antibiotic therapy correlates with patient survival, thus the initial antibiotic selection is critical. Transcriptomic biomarkers of resistance could lead to more rapid diagnosis and treatment for MDRAB infections. Materials and methods Three MDRAB clinical isolates were selected for RNAseq analysis based on varying drug resistance phenotypes (Table 4). They were cultured, in duplicate, in LB broth, and RNA was isolated using the Qiagen RNeasy kit. Illumina Genome Analyzer 2 (GAIIx) sequencing was performed using 49 bp single end reads at COFACTOR Genomics (St. Louis, MO). Assembly to reference genome (A. baumannii ATCC 17978) was done using CLC Genomics Workbench version 8.5.1. Transcripts mapping to genes with positive RPKM values were classified as present and a VENN overlap diagram was constructed using VENNY 2.0. Annotations were downloaded from NCBI. Results There were 27 – 35 million transcript read mappings to the reference genome for each of the two replicate runs per isolate. Numbers of unique and overlapping transcripts are presented in Fig. 4. Annotation of unique transcripts identified genes associated with aminoglycoside and other types of antibiotic and stress resistance, as well as virulence. As an example, only isolate 13 expresses the acetyltransferase that confers resistance to aminoglycosides, and isolate 13 is resistant to all tested aminoglycosides. Conclusions Transcriptomic analysis of three MDRAB isolates identified transcripts for genes associated with specific antibiotic resistance mechanisms. Acetyltransferase is known to confer aminoglycoside resistance in A. baumanii, thus its expression could be used as a biomarker in the rapid identification of aminoglycoside resistance in A. baumanii. These transcripts could be used as biomarkers for resistance and rapidly identified by PCR, decreasing the time required to begin effective antibiotic therapy. Acknowledgements This research was supported by The Meharry Translational Research Center (MeTRC) Grant Number U54RR026140-01 and U54 MD007593 and The Research Centers in Minority Institutes (RCMI) Grant Number G12RR003032-24S1 and G12 MD007586.Table 4 (abstract P17) Antibiotic resistance profiles of A. baumannii clinical isolates obtained from Nashville General Hospital at Meharry. (R = Resistant, S = Susceptible, I = intermediate). Antibiotics are: AK (Amikacin), GM (Gentamycin), TO (Tobramycin), A/S (Ampicillin/Sulbactam), ETP (Ertapenem), MER (Meropenem), CAX (Ceftriaxone), CAZ (Ceftazidime), CFT (Cefitaxime), CPE (Cefepime), CP (Ciprofloxacin), LVX (Levofloxacin), T/S (Trimethoprim/Sulfamethoxazole), PI (Piperacillin), TIM (Ticaricillin/K Clavulanate), TE (Tetracycline). Fig. 4 (abstract P17) RNA-Seq Transcript overlap of three MDRAB c linical isolates shows 91% common and 0.7 – 1.2% unique transcripts among the isolates. P18 Metagenomic assessment of possible microbial contamination in the equine reference genome assembly M Scotty DePriest1,2, James N MacLeod2, Theodore S Kalbfleisch1 1Department of Biochemistry and Molecular Genetics, School of Medicine, University of Louisville, Louisville, KY 40292, USA; 2Maxwell H. Gluck Equine Research Center, Department of Veterinary Science, University of Kentucky, Lexington, KY 40546, USA Correspondence: M Scotty DePriest (michael.depriest@louisville.edu) – Maxwell H. Gluck Equine Research Center, Department of Veterinary Science, University of Kentucky, Lexington, KY 40546, USA Background In a genome sequencing project, contaminating DNA from non-target organisms can result in errors in downstream analyses. These non-target organisms may include pathogens or parasites present in the original sample, as well as contaminants introduced in the sequencing process. As a genome assembly algorithm cannot distinguish between target and contaminant DNA, sequence reads from contaminants can and will be assembled into contigs. To prepare the new equine reference genome (EquCab3) for publication, contaminant contigs must be identified and screened out. Identification of sequences as contaminants requires a metagenomic approach. Many software packages can be used to identify sequences taxonomically in metagenomics studies, but they often require very long run times and can result in many false positives. Kraken [1] addresses both of these problems by using exact matches to k-mers, rather than similarity, to identify sequences. Although Kraken was not designed specifically to identify contaminants in a eukaryotic genome assembly, it has been shown to be effective for screening the bovine genome [2]. In the current study we similarly use and test Kraken to screen EquCab3. Materials and methods To build the Kraken database, we downloaded all bacterial and viral genomes from NCBI RefSeq (11,061 genomes total) and used a k-mer length of 31. We used Kraken to search EquCab3 contigs (n = 106,319). For each flagged contig, we calculated the number of k-mer hits per kilobase of contig length. Then, we downloaded all nonredundant bacterial proteins from RefSeq to build a BLAST database (46,173,990 proteins total). We queried this database for Kraken-flagged contigs with >1 k-mer hit per kilobase using BLASTX. Contig hits with >90% sequence identity with a bacterial protein sequence along >50% of its length were considered significant. Finally, we mapped 30X Illumina HiSeq 2 × 100bp genomic DNA reads from three other horses (Thoroughbreds TB03 and TB10, Standardbred ST22) to EquCab3 using BWA MEM [3], and we calculated mapping coverage for each Kraken-flagged contig. Reads with mapping quality <10 were excluded from the calculation. Results Out of the 106,319 total EquCab3 contigs, 7,565 were identified as possible contaminants based on k-mer matches. Of the Kraken-flagged contigs, 2,257 matched more than 1 hit per kilobase, and 697 of these contained significant BLAST hits to bacterial proteins. When the short reads from three horses were mapped to these EquCab3 bacterial contigs, 217 contigs had very low (<5%) mapping coverage, indicating that the sequence was not found in the other horses. In addition, 31 bacterial contigs were mapped with >80% coverage from one or more of the test horses. Conclusion The 217 contigs in EquCab3 containing significant BLAST hits to bacterial proteins, as well as very few mapped short reads from equine genomic DNA, are likely to be contaminants. These contigs are all under 3 kb in length. Thirty-one contigs with significant BLAST hits to bacterial proteins are also well-mapped with equine short reads, suggesting that some contamination may be present in the three equine samples. Alternatively, some of these “shared contaminants” may be misidentified equine DNA, which warrants further investigation. Taxonomically, most of the contaminant sequences can be identified as commensals, including Escherichia coli, Campylobacter jejuni, and Enterobacter spp., which are known to inhabit the mammalian digestive tract. These bacteria could be introduced during sampling or any other step involving a human or horse. Because so few Kraken-flagged contigs were confirmed as bacteria with BLAST, we conclude that Kraken alone was not sufficient to accurately identify contigs as bacteria. The screening results of metagenomics tools such as Kraken need to be further corroborated using other independent analytical methods. References 1. Wood DE, Salzberg SL: Kraken: ultrafast metagenomic sequence classification using exact alignments.Genome biol. 2014;15:R46. 2. Merchant S, Wood DE, Salzberg SL: Unexpected cross-species contamination in genome sequencing projects.PeerJ. 2014; 2: e675. 3. Li H, Durban R: Fast and accurate short read alignment with Burrows-Wheeler Transform.Bioinformatics. 20029; 25:1754–60. P19 Molecular evolution of cancer driver genes Chandrakanth Emani; Hanady Adam; Ethan Blandford; Joel Campbell; Joshua Castlen; Brittany Dixon; Ginger Gilbert; Aaron Hall; Philip Kreisle; Jessica Lasher; Bethany Oakes; Allison Speer; Maximilian Valentine Department of Biology, Western Kentucky University-Owensboro, Owensboro, KY 42303, USA Correspondence: Chandrakanth Emani (chandrakanth.emani@wku.edu) – Department of Biology, Western Kentucky University-Owensboro, Owensboro, KY 42303, USA Background The present study traces the molecular evolution of specific cancer driver genes selected from a list of 125 genes identified by the cancer genome landscapes study [1]. The purpose of the study is to identify ancestral forms of the cancer driver genes and identify the specific conserved domains during the molecular evolution in terms of gene duplications and mutational changes. Materials and methods The randomly chosen genes chosen in the specific study were ABL1, BRACA1, CASP8, DAXX, EZH2, FOXL2, GATA1, HRAS, IDH1, JAK1, MAP2K1, NOTCH1 and TP53. Protein sequences retrieved from the NCBI database by the PSI-BLAST program were subjected to multiple alignment and neighbor joining phylogenetic trees were constructed using the MEGA6 program [2]. Results Comprehensive bioinformatics analysis of the resulting multiple alignments and the generated phylogenetic trees gathered valuable insights in identifying the specific molecular elements that form the basis of the specific cancer types, the related molecular processes affected across diverse life forms during the molecular evolution of the genes and suggest specific molecular targets for cancer treatment. References 1. Vogelstein B, Papadopoulos N, Velculescu VE, Zhou S, Diaz LA, Kinzler KW: Cancer genome landscapes. Science. 2013; 339:1546–1558. 2. Tamura K, Stecher G, Peterson D, Filipski A, Kumar S: MEGA6: Molecular evolutionary genetics analysis version 6.0.Mol biol evol. 2013; 30:2725-2729. P20 Biorepository Laboratory Information Management System Naga Satya V Rao Nagisetty, Rony Jose, Teeradache Viangteeravat, Robert Rooney, David Hains Department of Pediatrics, The University of Tennessee Health Science Center, Memphis, TN 38103, USA Correspondence: Robert Rooney (rrooney1@uthsc.edu) – Department of Pediatrics, The University of Tennessee Health Science Center, Memphis, TN 38103, USA Background Healthcare organizations are increasingly moving towards personalized medicine and integrating genomic information into day to day clinical decision making [1]. Biorepositories help facilitate this movement by providing the means to store, link, and analyze biological samples, clinical information, and large-scale data sets, essentially creating a platform through which clinicians and researchers in biomedical and pharmaceutical industries can identify genetic variants and mutations that cause or are associated with disease symptomology, susceptibility and/or prognosis, variation in drug and therapeutic responses, and new disease subtypes. Such information is necessary for the development and evaluation of new targeted drugs and treatment modalities, new biomarkers or more accurate diagnostic tests, proper clinical trial design, and informed life decisions by patients and their families. Effective biorepository development and operation, particularly in a hospital setting, is a complex task requiring a cohesive effort from multiple groups and technologies within the organization. At the heart of this process is a laboratory information management system (LIMS) that supports a workflow dependent upon real-time patient consenting and integration of patient data from the Electronic Medical Record (EMR), accurate and efficient sample collection, processing, storage, and distribution, and reliable integration of analysis data. The LIMS must be customizable to fit a laboratories’ equipment and procedures while still effectively protecting personal health information. Materials and methods In order to facilitate functionality tailored to biorepository needs we have built an agile and streamlined LIMS infrastructure that is cost effective, provides improved flexibility for high-throughput laboratory workflow, and has a modular design to facilitate modification and installation of new equipment and data systems (Fig. 5). The BLIMS has two major components: EMR interfaces and a template driven LIMS application. EMR interfaces were developed using open source Mirth Connect interface software using HL7 [2]. The LIMS system was developed with an interface built using PHP, JQuery & Bootstrap for flexibility and responsiveness, MVC (Model View Controller) paradigm is used to abstract various functional components of the system for extendibility, security and understandability. MySQL with PDO (PHP Data Objects) is used for data storage and manipulation. Server side functionality provides features like customizable sample templates, batch mode sample collection, and support for data import from custom laboratory equipment, multiform import in XML, CSV or TXT formats. QA/QC, tracking equipment, sample locations, statuses, and strong security with grid based access control and completed audit log management. References 1. Ginsburg GS, Kuderer NM: Comparative effectiveness research, genomics-enabled personalized medicine, and rapid learning health care: A common bond.J clinical oncol. 2012; 30(34):4233-4242. 2. Viangteeravat T, Anyanwu MN, Nagisetty VR, Kuscu E, Sakauye ME, Wu D: Clinical data integration of distributed data sources using health level seven (HL7) v3-RIM mapping.J clin bioinform.a 2011; 1(1):32.Fig. 5 (abstract P20) Overall infrastructure for BIG initiative.

PMC005xxxxxx/PMC5001209.txt

==== Front BMC Public HealthBMC Public HealthBMC Public Health1471-2458BioMed Central London 27556802337810.1186/s12889-016-3378-1Meeting ReportInfluenza epidemiology and influenza vaccine effectiveness during the 2014–2015 season: annual report from the Global Influenza Hospital Surveillance Network Puig-Barberà Joan +34 961925948puig_joa@gva.es 1Burtseva Elena 2Yu Hongjie 3Cowling Benjamin J. 4Badur Selim 5Kyncl Jan 6Sominina Anna 7on behalf of the GIHSNAfanasieva Olga Afanasieva Veronica Ciblak Meral Akcay Aktas F. Badur Selim Belenguer-Varea Ángel Borekci S. Boza Fernando Bursteva Elena Buzitskaya Zhanna Caetano Braulia Cai Jian Çakir B. Carballido-Fernández Mario Carbonell-Franco Empar Carratalá-Munuera Concha Çelebi S. Chai C. de Paula Ivamara Changas de Lima Porto Chen Enfu Cowling Ben Cui Yunjie Deniz D. B. Dondurei Elena Dong H. Dong X. Durusu Mine El Guerche-Séblain Clotilde Enda-Moura Fernanda Eren-Şensoy A. Fadeev Artem Feng Luzhao Feng Shuo César Raimundo Fisch Patricia Garina Ekaterina Gencer S. Gil-Guillén Vicente Go Alexa Gonchar Vitaly Golovacheva Ekaterina Grudinin Mikhail Hacımustafaoğlu M. Hancerli S. Havlickova Martina Herrmannova Kristyna Huang L. Jiang Hui Ip Dennis Jirincova Helena Jurzykowska Lucie Kisteneva Lidiya Kolobukhina Ludmila Komissarov Andrey Kralova Radka Krasnoslobotsev Kirill Kyncl Jan Labrador Xavier Li Chao Li Xiangxin Limón-Ramírez Ramón Liu Jianhua Carmen Mari Mahé Cédric Mandakova Zdenka Merkulova L. Mese Sevim Iglesias Ainara Mira Mukasheva Evgenia Sancho Angels Natividad Nováková Lucia Obraztsova Elena Osidak Ludmila del Carmen Maria Özer S. Ozisik L. Picot Valentina Pisarev Maria Pradel Florence Prochazkova Jitka Puig-Barberá Joan Qin Ying Raboni Sonia Roháčová Hana Rozhkova Elena Li Sa Sadikhova M.I. Schwarz-Chavarri Germán Siqueira Marilda Smorodintseva Elizaveta Sominina Anna Stolyarov Kirill Sukhovetskaya Vera Sun G. Tang Y. Tortajada-Girbés Miguel Trushakova Svetlana Tuells José Tumina Tatiana Vartanyan R. Voloshuk Lubov Wang Quanyi Wen D. Wu Peng Xiao Wen Yang Peng Yanina Marina Yi Bo Yu Hongjie Yurtcu Kubra Zarishnyuk Pavel Zhang S. Zhang Yi Zhang Tiebiano Zheng Jiandong Peng Zhibin 1 Foundation for the Promotion of Health and Biomedical Research in the Valencia Region FISABIO – Public Health, Avda Catalunya 21, 46020 Valencia, Spain 2 D.I. Ivanovsky Institute of Virology FGBC “N.F. Gamaleya FRCEM” Ministry of Health of Russian Federation, Moscow, Russian Federation Russia 3 Division of Infectious Disease, Key Laboratory of Surveillance and Early-Warning on Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing, China 4 School of Public Health, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong, Special Administrative Region China 5 National Influenza Reference Laboratory, Istanbul Faculty of Medicine, Istanbul University, Istanbul, Turkey 6 National Institute of Public Health (NIPH), Prague, Czech Republic 7 Research Institute of Influenza, Saint Petersburg, Russian Federation Russia 22 8 2016 22 8 2016 2016 16 Suppl 1 Publication of this supplement was supported by Sanofi Pasteur.757© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.The Global Influenza Hospital Surveillance Network (GIHSN) has established a prospective, active surveillance, hospital-based epidemiological study to collect epidemiological and virological data for the Northern and Southern Hemispheres over several consecutive seasons. It focuses exclusively on severe cases of influenza requiring hospitalization. A standard protocol is shared between sites allowing comparison and pooling of results. During the 2014–2015 influenza season, the GIHSN included seven coordinating sites from six countries (St. Petersburg and Moscow, Russian Federation; Prague, Czech Republic; Istanbul, Turkey; Beijing, China; Valencia, Spain; and Rio de Janeiro, Brazil). Here, we present the detailed epidemiological and influenza vaccine effectiveness findings for the Northern Hemisphere 2014–2015 influenza season. Electronic supplementary material The online version of this article (doi:10.1186/s12889-016-3378-1) contains supplementary material, which is available to authorized users. Keywords InfluenzaVirusSurveillanceVaccineHospitalizationEpidemiological studyGIHSN Annual Meeting 2015 Annecy, France 19-20 October 2015 http://www.gihsn.org/?page=homeissue-copyright-statement© The Author(s) 2016 ==== Body Introduction Every year, between 5 % and 10 % of adults and 20 – 30 % of children have symptomatic influenza illness [1, 2], and 3 to 5 million individuals suffer from severe influenza, leading to 250,000 to 500,000 deaths [2–4]. Influenza illness can result in hospitalization and death, mainly among high-risk groups but also in a substantial proportion of previously healthy individuals [5]. In recent years, especially after the 2009 pandemic season, influenza surveillance has been expanded, as recommended by the World Health Organization (WHO), to include additional epidemiological data [6]. The Global Influenza Hospital Surveillance Network (GIHSN) is an international public-private collaboration initiated in 2012 by Sanofi Pasteur and the Fundación para el Fomento de la Investigación Sanitaria y Biomédica de la Comunitat Valenciana (FISABIO), a regional public health institution in Valencia, Spain. The aim of the GIHSN is to improve understanding of influenza epidemiology to better inform public health policy decisions. It is the first global network focusing exclusively on severe cases of influenza requiring hospitalization. The GIHSN runs a prospective, active surveillance, hospital-based epidemiological study to collect epidemiological and virological data for the Northern and Southern Hemispheres over several consecutive seasons. A standardised protocol and standard operating procedures are shared between sites allowing comparison and pooling of results [7]. The GIHSN is coordinated by FISABIO and is made up of several country sites affiliated with national health authorities. Each site coordinates several hospitals in its region. The network currently includes 27 hospitals coordinated by 7 sites in 6 countries (St. Petersburg and Moscow, Russian Federation; Prague, Czech Republic; Istanbul, Turkey; Beijing, China; Valencia, Spain; and Rio de Janeiro, Brazil). The surveillance data collected by the GIHSN are used to describe the circulating strains related to severe disease, estimate the burden of severe influenza disease, and evaluate the benefit of influenza vaccination to prevent severe disease. Results have been published from the network’s first two seasons, 2012–2013 [5, 8] and 2013–2014 [9]. In this report, we describe the influenza epidemiology and vaccine effectiveness results from the GIHSN during the 2014–2015 influenza season. Complete data from the Southern Hemisphere was not available at the time of the meeting or during analysis and writing, so only data provided by sites in the Northern hemisphere during the 2014–2015 season are presented. Methods Summary of overall methodology As described in detail elsewhere [7], patients admitted in the participating hospitals are included, after written consent, if they are residents in the predefined hospital’s catchment area, present with an acute illness possibly related to influenza, are not institutionalised, and the onset of symptoms was within 7 days of admission. Swabs are collected from patients meeting the inclusion criteria and tested by reverse transcription-polymerase chain reaction (RT-PCR) for influenza (Fig. 1). Influenza-positive samples are sub-typed by RT-PCR to identify A(H1N1)pdm09, A(H3N2), B/Yamagata-lineage, and B/Victoria-lineage strains. Vaccine effectiveness is assessed using a test-negative design in which vaccine coverage is compared between admissions with and without laboratory-confirmed influenza.Fig. 1 Overview of the methodology used by the GIHSN Epidemiological analysis Epidemiological and virological data were collected from 7 coordinating sites and a total of 27 hospitals in 6 countries (Additional file 1). Briefly, eligible admissions included non-institutionalised residents in the predefined catchment areas of the participating hospitals, hospitalised in the last 48 h, and with presenting illness potentially associated with influenza (Additional file 2 and Additional file 3). The study activities were performed over influenza circulation periods defined using pre-specified criteria (Additional file 3). Nasopharyngeal swabs (all subjects), pharyngeal swabs (subjects ≥14 years) or nasal swabs (subjects <14 years) were tested by semi-quantitative RT-PCR for influenza A (subtypes H3 and H1pdm09) and B (Yamagata and Victoria lineages). The distribution of hospital admission according to RT-PCR result was described by site and risk group. Secondary outcomes included hospital admissions by subtype for influenza A(H1N1)pdm09, A(H3N2), and B-lineage, by site and risk group. The significance of differences among groups or categories was estimated by the likelihood ratio test, t-test, or nonparametric tests as required. A P-value <0.05 was considered to indicate statistical significance. To describe the major determinants for admission with influenza (vs. influenza-negative admission), a stepwise logistic regression model was fitted by including all risk factors at P < 0.2. Adjusted odds ratios (aORs) for RT-PCR-positive vs. RT-PCR-negative admissions in the presence of major risk factors of interest were estimated by multivariate logistic regression using minimal sufficient adjustment sets of covariates identified as confounders by causal diagrams. To account for the possible effect of study site, data were fitted to a random effects logistic regression model including site as a cluster variable. Likelihood ratio tests were used to check for the potential effect of clustering by site [10]. The adjusted effect of site in the probability of influenza with admission was estimated. Heterogeneity in the effects of risk factors by influenza strain and site were quantified using the I2 test. Heterogeneity was defined as an I2 > 50 % [11, 12]. Further details are provided elsewhere [5, 7, 8]. Influenza vaccine effectiveness analysis Influenza vaccine effectiveness (IVE) was estimated as (1 ˗ OR) × 100, where the OR compared the vaccine coverage rate between influenza-positive and influenza-negative patients. Patients were considered vaccinated if they had received the current season’s influenza vaccine at least 14 days before symptom onset. The types of vaccines used at each site are summarised in Additional file 4. IVE overall (irrespective of vaccine type) was determined in patients who had been swabbed within 7 days of the onset of ILI symptoms. Records for which outcome, exposure, or confounding variables were missing were excluded from the multivariate IVE analyses. The adjusted IVE was estimated by logistic regression using a random effects model with study site as a shared parameter for the pooled analysis and including week of symptom onset as a continuous variable, and age group, sex, hospitalisation in the previous 12 months, presence of chronic conditions, and smoking habits as potential confounding factors. A P-value <0.05 was considered to indicate statistical significance. Heterogeneity in IVE estimates was assessed using the I2. Potential sources of heterogeneity, including coordinating site, age, and influenza subgroup were examined in ad-hoc analyses. Heterogeneity was defined as low if I2 statistic <25 %, moderate if 25 – 49 %, and high if ≥50 %. Further details of the methodology are described elsewhere [8]. Results Epidemiology of influenza in the GIHSN during the 2014–2015 influenza season Patients included in the epidemiology analysis Twenty thousand five hundred fifty-one eligible admissions were identified between November 16, 2014 and May 23, 2015, of which 9614 met the selection criteria and were included (Table 1). Based on RT-PCR, 2177 (23 %) were positive for influenza. Major reasons for exclusion included no ILI symptoms before admission (15 %), previous admission fewer than 30 days from the current episode (13 %), admission more than 7 days after the onset of symptoms (6 %), recruitment outside periods of continuous admissions with influenza (6 %).Table 1 Selection of patients and results of RT-PCR St. Petersburg Moscow Czech Republic Turkey Beijing Valencia Total Category n % n % n % n % n % n % n % Screened admissions 3164 1934 123 1409 1425 12,496 20,551 Exclusion criteria  Non resident 21 0.7 95 4.9 12 9.8 73 5.2 5 0.4 50 0.4 256 1.2  Institutionalised 14 0.4 14 0.7 2 1.6 17 1.2 2 0.1 800 6.4 849 4.1  Previous discharge <30 days 31 1.0 51 2.6 8 6.5 216 15.3 13 0.9 2283 18.3 2602 12.7  Unable to communicate 20 0.6 47 2.4 2 1.6 125 8.9 0 0.0 782 6.3 976 4.7  Not giving consent 100 3.2 32 1.7 14 11.4 47 3.3 15 1.1 504 4.0 712 3.5  No ILI symptoms ≥5 years of age 19 0.6 25 1.3 1 0.8 131 9.3 18 1.3 2903 23.2 3097 15.1  Admission within 7 days of symptoms onset 181 5.7 150 7.8 4 3.3 110 7.8 44 3.1 745 6.0 1234 6.0  Previous influenza infection 1 0.0 0 0.0 0 0.0 7 0.5 0 0.0 1 0.0 9 0.0  Onset of symptoms to swab >9 days 0 0.0 1 0.1 0 0.0 2 0.1 0 0.0 1 0.0 4 0.0  Sample inadequate 0 0.0 0 0.0 0 0.0 0 0.0 0 0.0 5 0.0 5 0.0  Sample lost 0 0.0 0 0.0 0 0.0 0 0.0 1 0.1 1 0.0 2 0.0  Recruited outside periods with continuous influenza positive admissions 31 1.0 115 5.9 1 0.8 65 4.6 178 12.5 764 6.1 1154 5.6 Included with valid laboratory results 2715 85.8 1400 72.4 79 64.2 614 43.6 1149 80.6 3657 29.3 9614 46.8 RT-PCR result  Influenza negative 2113 77.8 966 69.0 20 25.3 543 88.4 875 76.2 2920 79.8 7437 77.4  Influenza positive 602 22.2 434 31.0 59 74.7 71 11.6 274 23.8 737 20.2 2177 22.6  Subtype and lineagea   A(H1N1)pdm09 47 7.8 30 6.9 7 11.9 26 36.6 1 0.4 10 1.4 121 5.6   A(H3N2) 267 44.3 163 37.6 33 55.9 6 8.5 163 59.5 611 82.9 1243 57.1   A not subtyped 48 8.0 9 2.1 2 3.3 0 0.0 0 0.0 47 6.4 106 4.9   B/Yamagata lineage 258 42.9 175 40.3 16 27.1 0 0.0 109 39.8 65 8.8 623 28.6   B/Victoria lineage 0 0 10 2.3 0 0 0 0.0 1 0.4 0 0.0 11 0.5   B not subtypedb 0 0 52 12.0 2 3.4 39 54.9 0 0.0 4 0.5 97 4.5 Abbreviations: ILI, influenza-like disease; RT-PCR, reverse transcriptase-polymerase chain reaction aBecause there were 24 mixed infections, each involving two different influenza viruses, the sum by strain may be greater than the number of patients included with lab results. Percentages are reported by total of influenza-positive cases bFor Turkey and Valencia, all B not subtyped were assumed to be B/Yamagata lineage based on virus circulation at these sites. This assumption was not applied for Moscow because of a mixed pattern of influenza B circulation Influenza viruses identified in admissions In the 2177 included influenza-positive patients, A(H3N2) (n = 1243; 57 %) was the most commonly identified type of influenza, followed by B/Yamagata-lineage (n = 623; 29 %), A(H1N1)pdm09 (n = 121;6 %), A not subtyped (106; 5 %), B lineage not determined (n = 97; 5 %), and B/Victoria-lineage (n = 11; 0.5 %) (Table 1 and Fig. 2a and b). Mixed influenza infections were found in 24 cases. Influenza B lineage not determined were considered B/Yamagata-lineage for 39 cases in Turkey and four in Valencia. Due to the mixed circulation of B/Yamagata and B/Victoria lineages in Moscow, this assumption was not applied at that site to cases where B-lineage was not determined.Fig. 2 Admissions with influenza by epidemiological week and virus type, subtype, or lineage overall and by site. B strains included B not subtyped and mixed influenza infections including influenza B but excluded B/Victoria lineage The 2014–2015 influenza season at the GIHSN sites Influenza was detected over a span of 27 weeks, from week 47 of 2014 to week 20 of 2015, with the peak at week 7 of 2015 (Fig. 2). The earliest start of the influenza season was reported in Beijing, where influenza-positive admissions occurred over a span of 23 weeks in two waves, the first due to A(H3N2) and the second due to B/Yamagata-lineage (Fig. 2). The latest influenza-positive admission was in St. Petersburg, where continuous weekly admissions with influenza were observed over a span of 19 weeks. A(H3N2) was the most frequently detected influenza virus in St. Petersburg (44 % of positives), Czech Republic (56 %), Beijing (60 %), and Valencia (83 %) (Table 1). B/Yamagata-lineage was the second-most frequently detected influenza virus in St. Petersburg (43 %), Czech Republic (27 %), and Beijing (39 %). With the exception of Beijing and Turkey, A(H3N2) and B/Yamagata-lineage co-circulated at all sites (Fig. 2). In Turkey, A(H3N2) accounted for only 8.5 % of positives, and instead, B influenza viruses predominated (55 %), followed by A(H1N1)pdm09 (37 %), with co-circulation of these two viruses (Table 1 and Fig. 2). Main characteristics of included patients Overall, all age groups were represented. Approximately one-third of included admissions were patients less than 5 years of age, one-third were 5 to 64 years of age, and one-third were 65 years of age or older (Table 2). More than half of the included patients were male (n = 5417; 56 %). Most (n = 5867; 61 %) did not have an underlying chronic condition, and most (n = 6756; 70 %) had not been hospitalised in the 12 months before the current episode. Among the 39 % (n = 3747) of admissions in patients with underlying chronic conditions, the most frequent were cardiovascular disease (n = 1998; 21 %), chronic respiratory conditions (including chronic obstructive pulmonary disease [COPD; n = 1459] and asthma [n = 446]; 20 %), diabetes (n = 1048; 11 %), and renal disease (n = 606; 6 %). Few patients had active neoplasms (3 %), neuromuscular diseases (3 %), autoimmune diseases (2 %), liver disease (2 %), or immunodeficiency (1 %).Table 2 Characteristics of included patients overall and by site St. Petersburg Moscow Czech Republic Turkey Beijing Valencia Total N = 2715 N = 1400 N = 79 N = 614 N = 1149 N = 3657 N = 9614 Characteristic n % n % n % n % n % n % n % Age in years, median (range) 3 (0–94) 19 (0–90) 51 (19–91) 12 (0–98) 8 (0–96) 73 (0–106) 21 (0–106) Age group  0–1 y 714 26.3 137 9.8 0 0.0 112 18.2 76 6.6 476 13.0 1515 15.8  2–4 y 1034 38.1 371 26.5 0 0.0 133 21.7 403 35.1 265 7.2 2206 22.9  5–17 y 357 13.1 171 12.2 0 0.0 80 13.0 147 12.8 72 2.0 827 8.6  18–49 y 426 15.7 632 45.1 38 48.1 38 6.2 106 9.2 221 6.0 1461 15.2  50–64 y 110 4.1 59 4.2 14 17.7 75 12.2 131 11.4 359 9.8 748 7.8  65–74 y 39 1.4 13 0.9 12 15.2 69 11.2 93 8.1 593 16.2 819 8.5  75–84 y 31 1.1 11 0.8 9 11.4 80 13.0 127 11.1 969 26.5 1227 12.8  ≥85 y 4 0.1 6 0.4 6 7.6 27 4.4 66 5.7 702 19.2 811 8.4 Sex  Male 1489 54.8 827 59.1 40 50.6 356 58.0 696 60.6 2009 54.9 5417 56.3  Female 1226 45.2 573 40.9 39 49.4 258 42.0 453 39.4 1648 45.1 4197 43.7 Chronic conditions  0 2380 87.7 1246 89.0 29 36.7 182 29.6 820 71.4 1210 33.1 5867 61.0  1 244 9.0 116 8.3 31 39.2 196 31.9 233 20.3 1026 28.1 1846 19.2  >1 91 3.4 38 2.7 19 24.1 236 38.4 96 8.4 1421 38.9 1901 19.8 Previously hospitalised (last 12 months)  No 1781 65.6 1123 80.2 56 70.9 341 55.5 964 85.1 2491 68.1 6756 70.4  Yes 934 34.4 277 19.8 23 29.1 273 44.5 169 14.9 1166 31.9 2842 29.6 Underlying chronic conditions  Cardiovascular disease 140 5.2 68 4.9 29 36.7 215 35.0 218 19.0 1328 36.3 1998 20.8  Chronic obstructive pulmonary disease 51 1.9 19 1.4 6 7.6 153 24.9 137 11.9 1093 29.9 1459 15.2  Asthma 60 2.2 19 1.4 3 3.8 74 12.1 8 0.7 282 7.7 446 4.6  Immunodeficiency/organ transplant 30 1.1 0 0.0 3 3.8 48 7.8 0 0.0 25 0.7 106 1.1  Diabetes 32 1.2 14 1.0 12 15.2 96 15.6 34 3.0 860 23.5 1048 10.9  Renal impairment 18 0.7 26 1.9 4 5.1 61 9.9 12 1.0 485 13.3 606 6.3  Neuromuscular disease 68 2.5 15 1.1 1 1.3 79 12.9 12 1.0 92 2.5 267 2.8  Neoplasm 7 0.3 9 0.6 9 11.4 79 12.9 7 0.6 190 5.2 301 3.1  Cirrhosis/liver disease 34 1.3 21 1.5 5 6.3 19 3.1 5 0.4 118 3.2 202 2.1 Autoimmune disease 13 0.5 14 1.0 4 5.1 22 3.6 0 0.0 122 3.3 175 1.8 Pregnant (women 15–45 y) 0 0.0 291 95.7 1 7.1 1 4.8 0 0.0 5 5.7 298 45.4 Obese (all ages) 263 9.7 162 11.6 12 15.2 109 17.8 155 13.5 957 26.2 1658 17.2 Outpatient consultations last 3 months  0 1215 44.8 492 35.1 23 29.1 113 18.4 4 0.3 649 17.7 2496 26.0  1 895 33.0 314 22.4 19 24.1 100 16.3 697 60.9 678 18.5 2703 28.1  >1 605 22.3 594 42.4 37 46.8 401 65.3 443 38.7 2330 63.7 4410 45.9 Smoking habits (patients ≥18 y)  Never smoker 325 53.3 345 47.9 40 50.6 135 46.7 269 51.4 1363 47.9 2477 48.9  Past smoker 76 12.5 136 18.9 19 24.1 117 40.5 162 31 1034 36.4 1544 30.5  Current smoker 209 34.3 240 33.3 20 25.3 37 12.8 92 17.6 447 15.7 1045 20.6 Functional status impairment (Barthel score; patients ≥65 y)  Total (0–15) 1 1.4 0 0.0 3 11.1 6 3.4 23 10.0 166 7.3 199 7.1  Severe (20–35) 1 1.4 0 0.0 1 3.7 3 1.7 19 8.3 71 3.1 95 3.4  Moderate (40–55) 1 1.4 1 3.3 1 3.7 6 3.4 37 16.1 140 6.2 186 6.6  Mild (60–90) 25 33.8 6 20.0 9 33.3 76 43.2 136 59.1 414 18.3 666 23.8  Minimal (95–100) 46 62.2 23 76.7 13 48.1 85 48.3 15 6.5 1473 65.1 1655 59.1 Sampling time  0–2 days 1351 49.8 655 46.8 20 25.3 125 20.4 324 28.2 896 24.5 3371 35.1  3–4 days 915 33.7 523 37.4 26 32.9 213 34.7 382 33.2 1572 43.0 3631 37.8  5–7 days 449 16.5 221 15.8 24 30.4 239 38.9 358 31.2 1058 28.9 2349 24.4  8–9 days 0 0.0 1 0.1 9 11.4 37 6.0 85 7.4 131 3.6 263 2.7 Influenza vaccination ≥14 days from symptom onset 59 2.2 39 2.8 1 1.3 28 4.6 127 11.1 1759 48.1 2013 20.9 Just under half (n = 298; 45 %) of the admitted women 15–45 years of age were pregnant. Obese patients represented 17 % (n = 1658) of admissions. Among admissions in adult patients (≥18 years; n = 5066), 1045 (21 %) were current smokers, 1544 (30 %) were past smokers, and 2477 (49 %) had never smoked. Among elderly patients (≥65 years; n = 2857), 17 % (n = 480) had severe functional impairment as defined by a Barthel index <60. Finally, 2013 (21 %) admissions were in patients that had received the current season’s influenza vaccine at least 14 days before the onset of symptoms. Overall, swabs were obtained within 4 days after the onset of symptoms onset in 7002 (73 %) of included admissions. Site-related characteristics of included patients Patients included in St. Petersburg were younger than patients included at other sites (Table 2). The difference in age of included patients was especially marked when comparing St. Petersburg with the Czech Republic and Valencia. Patients were most frequently young adults in Moscow and the Czech Republic. Ages were homogeneously distributed in Turkey and Beijing (P = 0.9480). By contrast, in Valencia most (62 %) admissions were in elderly patients (≥65 years). Patients without comorbidities represented 88 % of admissions in St. Petersburg, 89 % in Moscow, 71 % in Beijing, 37 % in Czech Republic, 30 % in Turkey, and 33 % in Valencia. Of the different chronic conditions, cardiovascular disease, respiratory disease, and diabetes were the most common, and their relative importance at each site corresponded to the proportion of patients with one or more underlying chronic condition. In Moscow, among admissions in patients with known risk factors for influenza, pregnant women represented the majority of admissions (n = 291; 96 %). Obese patients represented 10 – 15 % of admissions in St. Petersburg, Moscow, the Czech Republic, and Beijing, whereas 18 % in Turkey and 26 % in Valencia were obese. The proportion of who never smoked ranged from 47 – 53 % in adult (≥18 years) admissions and was similar across sites (p = 0.1520). The overall proportion of current smokers, however, differed, with the highest rate (34 %) in St. Petersburg and Moscow (33 %), followed by Czech Republic (25 %), Beijing (18 %), Valencia (16 %), and Turkey (13 %) (p < 0.0001). For elderly adults, functional impairment status was mild or minimal in 83–97 % of included admissions at all sites except Beijing, where 30 % of admissions in elderly patients had moderate to severe functional impairment. Rates of influenza vaccination were below 5 % for all sites except Beijing (11 %) and Valencia (48 %). Heterogeneity between sites The proportion of samples with positive results differed between sites, from as low as 12 % for Turkey to as high as 75 % for the Czech Republic (Table 1; p < 0.0001 by test of homogeneity for equal odds). This difference persisted after excluding pregnant women and excluding the two sites with extreme results: proportions with positive results were 22 % for St. Petersburg, 31 % for Moscow, 24 % for Beijing, and 20 % for Valencia (p < 0.0014 by test of homogeneity for equal odds). After excluding pregnant women, however, proportions were homogenous in St. Petersburg, Moscow, and Beijing (p < 0.1464 by test of homogeneity for equal odds). After adjusting for sex, age, comorbidity, previous admissions, time to swab, influenza vaccination, and calendar time, the heterogeneity of aORs for a positive result were similar to the unadjusted results (Additional file 5 and Additional file 6; I2 = 96.4 %; p <0.0001). Risk of admission with influenza according to age and sex and variability by influenza virus Influenza positivity was related to age. Overall, influenza-positive admissions tended to be older than influenza-negative admissions (Table 3). Admissions positive for A(H1N1)pdm09 were younger than those negative for influenza, those positive for A(H3N2), and those positive for B/Yamagata-lineage. Also, admissions positive for A(H3N2) were older than influenza-negative admissions, those positives for A(H1N1)pdm09, and those positive for B/Yamagata-lineage (Table 3 and Fig. 3).Table 3 Characteristics of included patients according to PCR result Influenza-negative Influenza-positive A(H1N1)pdm09 A(H3N2) B/Yamagata lineage N = 7437 N = 2177 N = 115 N = 1231 N = 646 n % n % P vs. negative n % P vs. negative n % P vs. negative n % P vs. negative Age in years, median (range) 18.4 (0–106) 32.8 (0–100) 0.0001 5.6 (0–85) 0.0861 54.5 (0–100) 0.0001 26.2 (0–96) 0.0013 Age group <0.0001 <0.0001 <0.0001 <0.0001  0–1 y 1371 18.4 144 6.6 14 12.2 74 6.0 37 5.7  2–4 y 1777 23.9 429 19.7 36 31.3 212 17.2 133 20.6  5–17 y 547 7.4 280 12.9 14 12.2 121 9.8 118 18.3  18–49 y 1038 14 423 19.4 28 24.3 183 14.9 168 26.0  50–64 y 557 7.5 191 8.8 9 7.8 92 7.5 73 11.3  65–74 y 608 8.2 211 9.7 5 4.3 153 12.4 45 7.0  75–84 y 933 12.5 294 13.5 8 7.0 235 19.1 38 5.9  ≥85 606 8.1 205 9.4 1 0.9 161 13.1 34 5.3 Sex <0.0001 0.0390 0.0003 0.0040  Male 4276 57.5 1141 52.4 55 47.8 651 52.9 333 51.5  Female 3161 42.5 1036 47.6 60 52.2 580 47.1 313 48.5 Chronic conditions 0.0940 0.1600 <0.0001 0.004  0 4572 61.5 1295 59.5 78 67.8 643 52.2 434 67.2  ≥1 2865 38.5 882 40.5 37 32.2 588 47.8 212 32.8 Underlying chronic conditions  Cardiovascular disease 1529 20.6 469 21.5 0.3210 18 15.7 0.1820 319 25.9 <0.0001 109 16.9 0.02200  Chronic obstructive pulmonary disease 1153 15.5 306 14.1 0.0570 7 6.1 0.0001 222 18.0 0.0270 60 9.3 <0.0001  Asthma 346 4.7 100 4.6 0.9080 5 4.3 0.8760 65 5.3 0.3440 24 3.7 0.2600  Immunodeficiency/organ transplant 92 1.2 14 0.6 0.0130 0 0.0 0.0920 8 0.6 0.0550 6 0.9 0.4750  Diabetes 814 10.9 234 10.7 0.7960 8 7.0 0.1480 180 14.6 <0.0001 34 5.3 <0.0001  Renal impairment 463 6.2 143 6.6 0.5640 2 1.7 0.0200 108 8.8 0.0010 27 4.2 0.0280  Neuromuscular disease 215 2.9 52 2.4 0.2020 8 7.0 0.0290 19 1.5 0.0040 22 3.4 0.4670  Neoplasm 238 3.2 63 2.9 0.4660 4 3.5 0.8680 38 3.1 0.8330 17 2.6 0.4160  Cirrhosis/liver disease 168 2.3 34 1.6 0.0390 0 0.0 0.0220 19 1.5 0.0950 11 1.7 0.3390  Autoimmune disease 127 1.7 48 2.2 0.1360 1 0.9 0.4470 36 2.9 0.0060 8 1.2 0.3510 Pregnant (women 15–45 y) 138 33.7 160 64.8 0.0000 11 68.8 0.0050 64 55.7 <0.0001 68 70.1 <0.0001 Obese (all ages) 1300 17.5 358 16.4 0.2590 15 13.0 0.1970 223 18.1 0.5890 92 14.2 0.0320 Smoking habits (patients ≥18 y)  Never smoked 1760 47.0 717 54.2 23 45.1 451 54.7 191 53.4  Past smoker 1164 31.1 380 28.7 14 27.5 252 30.6 96 26.8  Current smoker 818 21.9 227 17.1 14 27.5 121 14.7 71 19.8 Functional status impairment (Barthel score; patients ≥65 y)  Total (0–15) 152 7.2 47 6.7 0 0.0 41 7.6 5 4.3  Severe (20–35) 76 3.6 19 2.7 0 0.0 13 2.4 4 3.5  Moderate (40–55) 140 6.7 46 6.6 0 0.0 32 5.9 11 9.6  Mild (60–90) 518 24.7 148 21.1 1 7.1 111 20.5 32 27.8  Minimal (95–100) 1214 57.8 441 62.9 13 92.9 345 63.7 63 54.8 Influenza vaccination ≥14 d since onset of symptoms 1566 21.1 447 20.5 0.5960 6 5.2 <0.0001 356 28.9 <0.0001 57 8.8 <0.0001 Not subtyped A and B were 24 patients with mixed influenza infections were not included in the analysis by strain Fig. 3 Proportion of admissions by strain and age group After adjusting for sex, occupational class, comorbidity, influenza vaccination, time to swab, and the clustering effect of site, heterogeneity due to strain was significant for admissions in subjects ≥5 years of age due to a decrease in aOR with age for admission with A(H1N1)pdm09 (Table 4 and Additional file 7). After excluding admissions with A(H1N1)pdm09, the aOR for admission with influenza was homogeneous for elderly patients but heterogeneous for patients 5–64 years of age (I2 = 75–77 %) due to a higher aOR for admissions with B/Yamagata-lineage than for A(H3N2) (Additional file 7).Table 4 Subject characteristics and risk of admission with influenza All admissions Influenza-positive Crude OR Heterogeneity by strain (I2)a aORb N = 9164 N = 2177 Characteristic n n % Value 95 % CI Value 95 % CI Age  0–1 y 1515 144 9.5 1.00 - 32.5 % 1.00 -  2–4 y 2206 429 19.4 2.30 1.88-2.81 0.0 % 2.14 1.74-2.64  5–17 y 827 280 33.9 4.87 3.90-6.10 73.3 % 4.34 3.42-5.51  18–49 y 1461 423 29.0 3.88 3.16-4.77 59.2 % 3.11 2.49-3.90  50–64 y 748 191 25.5 3.26 2.57-4.14 72.5 % 4.08 3.11-5.36  65–74 y 819 211 25.8 3.30 2.62-4.17 67.8 % 4.99 3.76-6.64  75–84 y 1227 294 24.0 3.00 2.42-3.72 62.4 % 4.51 3.43-5.92  ≥85 811 205 25.3 3.22 2.55-4.07 71.2 % 4.79 3.59-6.40 Sexc  Male 5417 1141 21.1 1.00 - 0.0 % 1.00 -  Female 4197 1036 24.7 1.24 1.13-1.37 0.0 % 1.21 1.09-1.34  Female non-pregnant 3899 876 22.5 1.11 1.00-1.22 0.0 % 1.10 0.99-1.23 Other risk factors (excludes pregnant women)  Comorbidityd 3709 856 23.1 1.15 1.04-1.27 62.8 % 1.48 1.30-1.69  Cardiovascular disease 1996 468 23.4 1.17 1.04-1.33 22.2 % 1.47 1.25-1.72  Chronic obstructive pulmonary disease 1458 306 21.0 1.02 0.88-1.17 57.6 % 1.39 1.15-1.68  Asthma 440 96 21.8 1.07 0.84-1.35 0.0 % 1.37 1.04-1.80  Immunosuppression 106 14 13.2 0.58 0.33-1.03 0.0 % 0.76 0.40-1.46  Diabetes 1048 234 22.3 1.10 0.94-1.29 0.0 % 1.36 1.10-1.70  Renal disease 588 129 21.9 1.08 0.88-1.32 48.2 % 1.23 0.95-1.59  Neuromuscular 167 52 31.1 0.93 0.68-1.26 38.9 % 1.13 0.80-1.58  Neoplasm (active) 301 63 20.9 1.01 0.76-1.35 55.6 % 1.29 0.92-1.81  Liver disease 200 33 16.5 0.76 0.52-1.11 50.8 % 0.79 0.52-1.21  Autoimmune disease 161 39 24.2 1.22 0.85-1.77 0.0 % 1.44 0.95-2.18  Obesee 1620 337 20.8 1.0 0.9-1.2 0.0 % 0.87 0.73-1.03 Pregnancyf 298 160 53.7 3.45 2.23-5.34 0.0 % 2.08 1.43-3.03  Associated comorbidity 38 26 68.4 7.07 3.09-16.18 0.0 % 4.29 2.65-6.94  No comorbidity 260 126 48.5 3.05 2.08-4.47 0.0 % 1.80 1.22-2.66 aStrains considered: A(H3N2), A(H1N1)pdm09 and B/Yamagata bMinimal sufficient adjustment sets for estimating the exposure or risk factor effect on the risk of admission with influenza vs. all included admissions without underlying conditions or pregnant cFemale or female non-pregnant vs. male. aORs were adjusted for age, occupational social class group, underlying comorbidity, obesity, influenza vaccination, time to swab, calendar time, and site as a clustering factor dOne or more underlying conditions or individual comorbidities vs. no comorbidity. aORs were adjusted for sex, occupational social class group, obesity, influenza vaccination, time to swab, calendar time, and site as a clustering factor eaOR adjusted for sex, age, occupational social class group, influenza vaccination, time to swab, calendar time, and site as a clustering factor fWomen 15-45 years of age included in Moscow, St. Petersburg, Czech Republic, Turkey and Valencia. aOR adjusted for smoking habits, time to swab, calendar time, comorbidity, and site as a clustering factor. For results stratified by comorbidity, aORs were adjusted by the same covariates and were estimated taking into account the interaction between pregnancy and comorbidity Female patients had a higher risk than male patients of being influenza-positive (aOR, 1.21 [95 % CI, 1.09-1.34]), irrespective of strain (I2 = 0 %). However, after excluding pregnant women, the risk was more similar for males and females (aOR, 1.10 [95 % CI, 0.99–1.23]) (Table 4). Risk of admission with influenza according presence of comorbidity Similar proportions of influenza-positive admissions (882/2177; 41 %) and influenza-negative admissions (2865/7437; 39 %) had one or more chronic underlying condition (p = 0.0940) (Table 3). After excluding pregnant women, 42 % of influenza-positive admissions had comorbidity compared to 39 % of influenza-negative admissions (p = 0.006) (data not shown). The aOR for admission with influenza was 1.5 (95 % CI, 1.3–1.7) for patients with comorbidities, although the values were heterogeneous by strain (I2 = 63 %) (Table 4) due to a higher aOR for admission with A(H3N2) or B/Yamagata-lineage in patients with comorbidities compared to patients with no underlying conditions (Additional file 8). Irrespective of the involved strain (I2 = 22 %), the risk of admission with influenza was significantly increased in patients with cardiovascular disease (aOR = 1.5 [95 % CI, 1.3–1.7), asthma (1.4 [95 % CI, 1.0–1.8]), or diabetes (1.4 [95 % CI, 1.1–1.7]) (Table 4, Fig. 4, and Additional file 9). The aOR was heterogeneous for the risk of admission with influenza in patients with COPD (aOR 1.4 [95 % CI, 1.2–1.7]; I2 = 58 %) due to lower aOR for admission with A(H1N1)pdm09 (Additional file 9). Point values for aORs were above 1.0 for admission with influenza for patients with renal, neuromuscular, or autoimmune disease, but 95 % CIs overlapped 1.0. In patients with active neoplasms, the overall aOR for influenza-positive admission was heterogeneous and not significant (I2 = 56 %; aOR = 1.3 [95 % CI, 0.9–1.8]), although for B/Yamagata-lineage, the risk was significantly elevated (aOR = 2.2 [95 % CI, 1.1–4.1]) (Additional file 9).Fig. 4 Adjusted odds ratio (aOR) and number of admissions with influenza according to comorbidity. CVD, cardiovascular disease. COPD: chronic obsructive pulmonary disease Risk of admission with influenza according pregnancy A total of 298 included admissions were pregnant women 15–45 years of age, 291 of whom were included in Moscow, one in the Czech Republic, one in Turkey, and five in Valencia (Table 2). No pregnant women were included in Beijing. Non-pregnant women in this age group accounted for another 359 included admissions, of which 197 were in St. Petersburg, 13 in Moscow, 13 in the Czech Republic, 20 in Turkey, 33 in Beijing, and 83 in Valencia (data not shown). The probability of laboratory-confirmed influenza was higher in included pregnant women than included same age non-pregnant women (54 % vs. 24 %; p < 0.0001; data not shown). After taking into account clustering by site (and not considering data from Beijing), the crude OR of admission with influenza was 3.5 (95 % CI, 2.2–5.3) (Table 4). This crude estimated OR was higher in pregnant women with associated comorbidity (OR 7.1 [95 % CI, 3.1–16.2]), with moderate evidence of an interaction between comorbidity and pregnancy before adjustment (p = 0.0659) and a significant interaction after adjustment (p < 0.0001). Taking into account the modifying effect of associated comorbidity, the aOR for admission with influenza in pregnant women was 4.3 (95 % CI, 2.7–6.9) in presence of associated comorbidity and 2.1 (95 % CI, 1.4–3.0) for pregnant women with no comorbidity. In both cases, the values were homogenous (I2 = 0 %) for A(H3N2), A(H1N1)pdm09, and B/Yamagata-lineage infections. The probability of admission with influenza was higher in all three trimesters for pregnant women without associated comorbidities than for non-pregnant women in the same age group without comorbidity. In pregnant women with comorbidities, the risk of admission with influenza was highest in the first trimester (Fig. 5 and Additional file 10).Fig. 5 Predicted probability of admission with influenza in non-pregnant 15 – 45 years old women and by pregnancy trimester in same age pregnant women Risk of admission with influenza and complications by strain Intensive care unit (ICU) admissions, extracorporeal membrane oxygenation, and mechanical ventilation were more frequent for influenza-negative than for influenza-positive admissions (p ≤ 0.002), whereas rates of in-hospital death were similar (p = 0.3460) (Table 5). By strain, the point estimate of rates of ICU admission and extracorporeal membrane oxygenation were higher in admissions with A(H1N1)pdm09, although differences were not significant. In contrast, rates of in-hospital death were significantly higher in admissions with A(H3N2) (p = 0.0080). Less than 4 % of admissions in these categories experienced a severe outcome. Finally, length of stay did not differ between influenza-positives and influenza-negative admissions for influenza overall or between strains (Table 5).Table 5 Influenza severity and complications by RT-PCR result Category Influenza-negative Influenza-positive P-value influenza-negative vs. positive A(H1N1)pdm09 A(H3N2) B/Yamagata lineage P-value for distribution by strain N = 7437 N = 2177 N = 115 N = 1231 N = 646 n % n % n % n % n % Severity indicator  Intensive care unit admission 184 2.5 31 1.4 0.0020 4 3.5 15 1.2 9 1.4 0.2400  Mechanical ventilation 123 1.7 20 0.9 <0.0001 1 0.9 14 1.1 4 0.6 0.5230  Extracorporeal membrane oxygenation 184 2.8 25 1.3 0.0020 3 2.6 9 0.8 9 1.7 0.1600  Death during hospitalisation 131 1.8 32 1.5 0.3460 1 0.9 26 2.1 3 0.5 0.0080  Length of stay (days), median (interquartile range) 6 (4-9) 6 (4-8) 0.0612 6 (3-8) 6 (3-8) 6 (4-8) 0.2835 Pulmonary complications <0.0001 <0.0001  None 1939 26.1 1212 55.7 53 46.1 697 56.6 353 54.6  Pneumonia 1545 20.8 364 16.7 18 15.7 214 17.4 109 16.9  COPD exacerbation 265 3.6 87 4.0 3 2.6 66 5.4 15 2.3  Respiratory failure 55 0.7 32 1.5 0 0.0 23 1.9 5 0.8  Asthma exacerbation 28 0.4 12 0.6 0 0.0 11 0.9 1 0.2  pulmonary collapse 5 0.1 1 0.0 0 0.0 1 0.1 0 0.0  Acute respiratory distress syndrome 7 0.1 2 0.1 0 0.0 2 0.2 0 0.0  Bronchiolitis 416 5.6 201 9.2 15 13.0 91 7.4 75 11.6  Upper respiratory infection 3172 42.7 266 12.2 26 22.6 126 10.2 88 13.6 Metabolic failure 0.4690 0.3530  Acute renal failure 87 1.2 32 1.5 1 0.9 24 1.9 5 0.8  Diabetic coma 4 0.1 2 0.1 0 0.0 1 0.1 0 0.0  Fluid/electrolyte/acid-base/balance disorders 80 1.1 29 1.3 1 0.9 19 1.5 7 1.1 Cardiovascular events 0.3390 <0.0001  None 6335 85.2 1883 86.5 107 93.0 991 80.5 612 94.7  Acute myocardial infarction 8 0.1 5 0.2 0 0.0 5 0.4 0 0.0  Acute heart failure 1 0.0 1 0.0 0 0.0 1 0.1 0 0.0  Cardiac arrest 4 0.1 3 0.1 0 0.0 3 0.2 0 0.0  Malignant hypertension 37 0.5 10 0.5 0 0.0 9 0.7 1 0.2  Any cardiovascular condition 1050 14.1 275 12.6 8 7.0 222 18.0 33 5.0 Systemic inflammatory response syndrome, shock, or disseminated intravascular coagulation 76 1.0 12 0.6 0.0320 2 1.7 9 0.7 1 0.2 0.0810 Neurologic events  No 7423 99.8 2173 99.8 0.3140 114 99.1 1228 99.8 646 100.0 0.1249  Altered mental status 10 0.1 4 0.2 1 0.9 3 0.2 0 0.0  Convulsions 4 0.1 0 0.0 0 0.0 0 0.0 0 0.0 Major discharge diagnoses <0.0001 <0.0001  Influenza 124 1.7 1266 58.2 76 66.1 603 49.0 456 70.6  Pneumonia 1807 24.3 223 10.2 10 8.7 145 11.8 59 9.1  Other respiratory disease 3653 49.1 290 13.3 20 17.4 188 15.3 57 8.8  Cardiovascular 603 8.1 117 5.4 0 0.0 105 8.5 7 1.1  Other 1250 16.8 281 12.9 9 7.8 190 15.4 67 10.4 Exacerbation of chronic obstructive pulmonary disease, respiratory failure, exacerbation of asthma, and bronchiolitis were more frequently reported for influenza-positive admissions than for influenza-negative admissions (Table 5). These were associated with A(H3N2), except in the case of bronchiolitis, where the proportions for admission with all three strains (A(H3N2), A(H1N1)pdm09, and B/Yamagata-lineage) were higher than the proportion for influenza-negative admissions. Cardiovascular events were more frequently reported for admissions with influenza A(H3N2) than for admissions with influenza A(H1N1)pdm09 or B (OR 1.3 [95 % CI, 1.1–1.6]; p = 0.0004; data not shown), whereas, shock was more frequent in admissions with influenza A(H1N1)pdm09 (adjusted p < 0.0001; Table 5). Figure 6 shows the estimated marginal probabilities by strain and age for severe outcomes after adjusting by sex, comorbidity, calendar time, age, and clustering by site. We found several non-significant associations: A(H1N1)pdm09 was associated with intensive care unit admission and shock; A(H3N2) was associated with an increased probability of COPD exacerbation, respiratory failure, cardiovascular complications, and death; B/Yamagata-lineage was related to respiratory failure; and all three strains were related to death at both extremes of age (Fig. 6). We found similar non-significant associations for complications when influenza-negative admissions were included (Additional file 11).Fig. 6 Predicted probability of severe outcome by strain (not subtyped, mixed influenza with influenza infections and B/Victoria lineage excluded) Influenza vaccine effectiveness in the GIHSN during the 2014–2015 influenza season Patients included in the influenza vaccine effectiveness analysis After applying exclusions related to vaccine contraindication (egg allergy and <6 months of age), 8455 specimens collected from November, 2014 through May, 2015 were included in the IVE analyses. Of all collected specimens, 2027 (24 %) were positive for influenza, of which 1165 (57 %) were positive for A(H3N2), 104 (5 %) for A(H1N1)pdm09, and 625 (31 %) for B Yamagata-lineage (Table 6). Overall, 22 % (n = 446) of influenza-positive admissions and 24 % (n = 1556) of influenza-negative admissions were vaccinated (p = 0.042) (Table 7). The proportion of patients vaccinated with the seasonal influenza vaccine in 2014–2015 ≥ 14 days before symptom onset was 3 % in St. Petersburg (n = 43) and Moscow (n = 30), 5 % in Turkey (n = 22), 11 % (n = 94) in Beijing, and 54 % (n = 1367) in Valencia (data not shown).Table 6 IVE for all cases and for targeted groups only by age and strain Influenza-positive Influenza-negative Crude IVE Adjusted IVE Population Strain Age Total Vaccinated Total Vaccinated Percent (95 % CI) P interaction Percent (95 % CI) P interaction Overall Any Any 2027 446 6428 1556 -1 (-17, 12) 22 (8, 33) <65 y 1334 78 4299 289 -4 (-36, 20) 0.090 -5 (-38, 20) 0.054 ≥65 y 693 368 2129 1267 21 (5, 34) 24 (9, 37) A(H3N2) Any 1165 356 6428 1556 -6 (-24, 10) 20 (4, 33) <65 y 630 50 4299 289 -15 (-59, 17) 0.036 -16 (-64, 17) 0.031 ≥65 y 535 306 2129 1267 24 (7, 37) 25 (8, 39) A(H1N1) Any 104 7 6428 1556 25 (-85, 69) 27 (-82, 71) <65 y 91 3 4299 289 16 (-173, 74) 0.996 21 (-161, 76) 0.993 ≥65 y 13 4 2129 1267 47 (-128, 88) 59 (-83, 91) B/Yamagata Any 625 57 6428 1556 16 (-17, 39) 31 (2, 52) <65 y 509 20 4299 289 7 (-51, 42) 0.266 29 (-17, 58) 0.273 ≥65 y 116 37 2129 1267 38 (-2, 62 33 (-12, 61) Targeted groups only Any Any 1670 425 5077 1462 13 (-2, 26) 23 (8, 35) <65 y 977 57 2948 195 -21 (-70. 14) 0.037 -12 (-58, 20) 0.019 ≥65 y 693 368 2129 1267 26 (16, 42) 28 (14, 41) H3N2 Any 994 344 5077 1462 13 (-4, 28) 22 (5, 36) <65 y 459 38 2948 195 -12 (-65, 24) 0.051 -20 (-80. 21) 0.030 ≥65 y 535 306 2129 1267 27 (11, 41) 28 (11, 42) H1N1 Any 84 6 5077 1462 44 (-58, 80) 46 (-52, 81) <65 y 71 2 2948 195 33 (-198, 85) 0.793 39 (-167, 86) 0.770 ≥65 y 13 4 2129 1267 47 (-128, 88) 50 (-111, 89) B/Yamagata Any 486 49 5077 1462 21 (-18, 46) 30 (-5, 53) <65 y 370 12 2948 195 -8 (-105, 44) 0.139 8 (-79, 53) 0.250 ≥65 y 116 37 2129 1267 38 (-2, 62) 33 (-12, 60) Abbreviation: IVE influenza vaccine effectiveness Table 7 Characteristics of patients included in the primary analysis by vaccination status Risk variables Category Unvaccinated Vaccinated P value n % N % Number of patients, n (%) Controls 4872 75.5 1556 77.7 0.042 Cases 1581 24.5 446 22.3 Age (y) Median (range) 14.6 (0.8-84.0) 78.7 (9.0-91.9) <0.001 Age group, n (%) 6–11 mo 496 7.7 3 0.1 1–4 y 2120 32.9 49 2.4 5–17 y 712 11.0 102 5.1 18–49 y 1347 20.9 75 3.7 50–64 y 591 9.2 138 6.9 65–74 y 388 6.0 416 20.8 75–84 y 512 7.9 704 35.2 ≥85 y 287 4.4 515 25.7 Female, n (%) - 2825 43.8 843 42.1 0.188 Comorbidities, n (%) None 4505 70 366 18.3 <0.001 1 1077 16.7 647 32.3 >1 871 13.5 989 49.4 Pregnant, n (%) - 294 4.6 3 0.1 <0.001 Obesity, n (%) - 982 15.2 564 28.2 <0.001 Morbid obesity, n (%) - 86 1.3 54 2.7 <0.001 Previous hospitalisation within 12 months, n (%) - 1887 29.2 761 38 <0.001 GP visit within 3 months, n (%) None 1815 28.1 343 17.1 <0.001 1 2027 31.4 389 19.4 >1 2606 40.4 1272 63.5 Smoking, n (%) Current 1527 23.7 225 11.2 <0.001 Past 1069 16.6 742 37.1 Never 3856 59.8 1035 51.7 Functional impairment in ≥65 y, n (%) None or minimal 619 54.2 1021 62.4 <0.001 Mild 324 28.4 330 20.2 Moderate 95 8.3 88 5.4 Severe 32 2.8 62 3.8 Total 72 6.3 123 7.5 Sampling interval (days) Median (range) 3 (1-7) 4(1-7) <0.001 Sampling interval, n (%) ≤4 d 3703 57.4 990 49.5 <0.001 5-7 d 2587 40.1 936 46.8 8-9 d 163 2.5 76 3.8 Site, n (%) St. Petersburg 2138 33.1 59 2.9 <0.001 Moscow 1306 20.2 39 1.9 Turkey 503 7.8 26 1.3 Beijing 996 15.4 127 6.3 Valencia 1510 23.4 1751 87.5 Vaccinated, n (%) In 2012–2013 473 7.5 1471 73.3 <0.001 In 2013–2014 513 8.1 1722 87.1 <0.001 Overall, 1709 of 2002 (85 %) influenza vaccinations among study patients were both self-reported and confirmed from registries. Self-report captured 156 of 2002 vaccinations (8 % overall, 67 % in Moscow, 8 % in Turkey, 1 % in Beijing, and 7 % in Valencia; data not shown). Another 137 patients (7 % overall, 12 % in St. Petersburg, 42 % in Turkey, and 7 % in Valencia; data not shown) with vaccination records failed to self-report vaccination. The proportion of participants with comorbidity was significantly higher in vaccinated than in non-vaccinated admissions (82 % vs. 30 %, p < 0.001) (Table 7). Vaccination was also more common among elderly (median age = 79 years for vaccinated patients vs. 15 years for non- vaccinated patients, p < 0.001), obese patients (28 % obese for vaccinated patients vs. 15 % for non-vaccinated patients, p < 0.001), elderly patients with impairment or minimal functional impairment (28 % impaired for vaccinated patients vs. 15 % for non-vaccinated patients, p < 0.001), patients with outpatient visits (83 % for vaccinated patients vs. 72 % for non-vaccinated patients, p < 0.001), and patients admitted to a hospital in the previous 12 months (38 % for vaccinated patients vs. 29 % for non-vaccinated patients, p < 0.001) (Table 7). Three (0.1 %) pregnant women had received the current season’s vaccine. Most patients vaccinated in 2014–2015 reported prior vaccination: 87 % of vaccinated patients had received the 2013–2014 vaccine (p < 0.001) and 73 % had received the 2012–2013 vaccine (p < 0.001) (Table 7). Both the 2011-2012 and 2012–2013 vaccines were received by 90 % (26/29) of cases and 77 % (89/116) of controls (p = 0.12). Influenza vaccine effectiveness Against all-age influenza-related hospitalisation, the overall crude IVE was −1 % (95 % confidence interval [CI], −17–12), and the adjusted IVE was 22 % (95 % CI, 8–33) (Table 6). Age at admission, presence of comorbidities, and degree of functional impairment were the covariates with the largest confounding effect on crude IVE (data not shown), raising the crude IVE with adjustment. The adjusted IVE for patients of all ages was higher against influenza B (31 % [95 % CI, 2–52]) than for influenza A(H3N2) (20 % [95 % CI, 4–33]) and influenza A(H1N1)pdm09 (27 % [95 % CI, −82–71]) (Table 6), although confidence intervals overlapped (I2 for adjusted IVE across strains = 0 %, p = 0.762; data not shown). Age-specific estimates suggested that vaccination against any influenza was less effective in patients <65 years of age (IVE [95 % CI] = −5 % [−38–20]) than in patients ≥65 years of age (IVE = 24 % [95 % CI, 9–37]) (P value for effect modification of age = 0.054). This pattern of lower IVE in the younger patients was consistent across strains, but only age-specific estimates for A(H3N2) were significantly different (Table 6). Estimates were similar when the analyses were restricted to patients belonging to the target group for vaccination (crude IVE against overall influenza for all ages = 13 % [95 % CI, −2–26], adjusted IVE = 23 % [95 % CI, 8–35]) (Table 6). IVE estimates were consistently higher for recipients of the 2012–2013 influenza vaccine, the 2013–2014 influenza vaccine, or both vaccines than for recipients of only the current season’s vaccine, although confidence intervals overlapped (Additional file 12). Statistical heterogeneity across sites in the estimates of IVE against influenza-related hospitalisation was relatively low, with site-specific adjusted point estimates ranging from -27 – 35 % [I2 = 0 %; P = 0.835) (Additional file 13). Sensitivity analyses were performed to assess the effects of excluding pregnant women, participants vaccinated within 14 days before symptom onset, and without medical vaccination records. In all cases, IVE estimates remained similar to those of primary analysis (Additional file 14). Further sensitivity analyses using various statistical methods to account for potential data clustering by site showed consistent results, with no evidence of heterogeneity (I2 = 0 %) in estimates of IVE across methods (Additional file 15). Discussion According to data collected by active surveillance within the GIHSN sites, the 2014–2015 influenza season was characterised by a predominance of A(H3N2) and B/Yamagata-lineage, and to a lesser extent, A(H1N1)pdm09, while B/Victoria-lineage was relatively rare. Reports of severe influenza, defined as hospitalisation with laboratory (i.e., PCR)-confirmed influenza, spanned 6 months and affected all ages, although influenza-related admissions were most common in older individuals. Among patients with laboratory-confirmed influenza, those with A(H1N1)pdm09 were younger than those with A(H3N2) or B/Yamagata-lineage, whereas those with B/Yamagata-lineage were most frequently young and middle-aged adults. This pattern of influenza circulation is consistent with that reported by the WHO [13]. Likewise, the age distribution of the A(H1N1)pdm09, A(H3N2) and B/Yamagata-lineage strains agrees with others’ reports [14, 15]. According to our data, comorbidity increased the risk of admission with influenza, irrespective of the strain involved. This was also the case for pregnant women. Furthermore, the combination of pregnancy and comorbidity increased the risk of admission several-fold, suggesting an interaction. Remarkably, however, nearly 60 % of eligible admissions with influenza were patients without known risk factors. The probability of ICU admission and shock were higher in patients infected with A(H1N1)pdm09 than with other strains. Also, A(H3N2) infection was associated with respiratory failure and cardiac complications, whereas B/Yamagata-lineage was associated with an increased probability of respiratory failure. Influenza infection overall was associated with in-hospital death at both age extremes. These findings agree with other reports [15–17], although there may be differences in the absolute percentage of admissions with influenza in patients with comorbidity, patterns of severity, lengths of hospital stay, rates of ICU admission, use of supportive measures, or estimates of in-hospital death rates [15, 18, 19]. Although vaccination coverage was low at the participating sites (2.8–48 %; average 20.9 %), we found that vaccination conferred a low to moderate protective effect (adjusted IVE = 22 %). This protective effect was greater for adults ≥65 years of age than for adults <65 years of age and was greater for B/Yamagata-lineage than for A(H3N2). The low influenza vaccine effectiveness for the 2014–2015 season is similar to others’ reports and appears to be due mostly to a mismatch between the main A(H3N2) circulating strain and the vaccine strain [20–23]. Across all strains, the IVE was lower in young patients, although only age-specific estimates for A(H3N2) were significantly different due to few cases of B/Yamagata-lineage and A(H1N1)pdm09 and a higher IVE in patients vaccinated during the 2012–2013, 2013–2014, or both seasons than in those vaccinated during the 2014–2015 season, a finding also reported by others [24]. This lower IVE in young patients, however, contrasts with previous reports where the opposite was found [25]. Thus, there appears to be variability in the interference or protection conferred by vaccination in previous seasons. This could be explained by the differences between the various strains circulating in different seasons and their distance from the vaccine strains, combined with inhibition of the immunological response when the vaccine strains are similar to those in previous seasons’ vaccines [26]. Limitations and considerations Our results are to be interpreted with caution due to the heterogeneity and bias of multi-centric observational studies. We assumed heterogeneity in the circulating strains, socio-demographic diverse populations observed, their health care seeking behaviour, the characteristics of the different health care systems involved, the types of participating hospitals, and by calendar time along the season. We took account of this heterogeneity by thoroughly describing the season, the sites, and included admissions, as well as by quantifying the heterogeneity of our estimates. In this way, we are able to visualise the relative impact of the different influenza strains on diverse risk factors, including age, comorbidity, pregnancy, and obesity [12]. Furthermore, we restricted our analysis to periods with influenza circulation [27], took into account risk by calendar date [28], as well as the clustering effect of site [10] by adjusting and modelling and, finally, compared PCR-detected influenza-positive admissions with influenza-negative admissions. We consider this a reasonable approach for describing the effect of influenza in individuals according to their risk profile [29]. In addition, to reduce bias and to allow us to describe the severe consequences of community-acquired influenza, we accepted only data from patients admitted within 7 days of onset of ILI symptoms and for whom swabbing was performed within 48 h of admission. Even with a large dataset as the one accrued annually by the GIHSN sites, small numbers are a limitation. Splitting the data by strain and risk group can decrease group sizes, so that sufficient power is available only for detecting large differences (i.e., OR ≥2). This limitation can be only dealt with by increasing the number of participating sites and by pooling data across influenza seasons. In fact, the GIHSN continues to grow, and data pooling across seasons is underway. Most hospital studies rely on the criteria of the physician providing care for influenza confirmation and employ historical database searching [15, 17, 18, 30–32]. This combined with different case definitions and laboratory methods can complicate comparisons between sites and seasons and between different studies. Our approach of using active surveillance, a shared core protocol, and PCR confirmation of influenza avoids these limitations. This approach has very recently begun to be employed by others and for other respiratory viruses [33]. Conclusions This report describes the results from the GIHSN during the 2014–2015 influenza season that were presented at the 2015 GIHSN Annual Meeting. During the 2014–2015 influenza season, the network included 27 hospitals in six countries (Russian Federation, Czech Republic, Turkey, China, Spain, and Brazil). This offered us the opportunity to describe the characteristics of severe disease related to influenza by time, person, and strain and to describe IVE across a wide geographical area in the Northern Hemisphere. We found that influenza is associated with severe outcomes during an extended period in the Northern Hemisphere and that comorbidity and pregnancy were significant risk factors for severe influenza illness. The distribution and impact of the three influenza virus types (A(H1N1)pdm09, A(H3N2), and B) were similar to others’ reports. An important finding was that approximately 60 % of influenza-related hospital admissions were in healthy subjects with no known comorbidity. Our results support the current WHO recommendations on the use of influenza vaccine [4], although for the 2014–2015, IVE was low due to a significant mismatch between the circulating and vaccine viruses. We also found that IVE was affected by age and the circulating strain. These findings highlight the need to develop vaccines that are more effective and cover a broader spectrum of influenza viruses. Abbreviations AOR, adjusted odds ratio; CI, confidence interval; GIHSN, Global Influenza Hospital Surveillance Network; IVE, influenza vaccine effectiveness; OR, odds ratio; RT-PCR, reverse transcription-polymerase chain reaction Additional files Additional file 1: Table S1. Characteristics of participating hospitals during the 2014–2015 season. (PDF 101 kb) Additional file 2: Table S2. Diagnoses and presenting complaints used to identify admissions possibly related with an influenza infection. (PDF 86 kb) Additional file 3: Table S3. Protocol application across sites. (PDF 85 kb) Additional file 4: Table S4. Types of vaccines available at each site. (PDF 69 kb) Additional file 5: Table S5. Site heterogeneity in the risk of a positive influenza result in included admissions. (PDF 11 kb) Additional file 6: Figure S1. Heterogeneity between sites in the OR of admission with a positive influenza result. (PDF 33 kb) Additional file 7: Figure S2. aOR and number of admissions with influenza by age group and virus strain. (PDF 86 kb) Additional file 8: Figure S3. aOR and number of admissions with influenza by virus strain in patients with one or more comorbidity compared to patients without comorbidity. (PDF 77 kb) Additional file 9: Figure S4. aOR and number of admissions with influenza by chronic underlying comorbidity and virus strain. (PDF 52 kb) Additional file 10: Table S6. Predicted probability of admission with influenza for women 15 to 45 years of age. (PDF 8 kb) Additional file 11: Figure S5. Probability of severe outcome by RT-PCR result. (PDF 83 kb) Additional file 12: Figure S6. aOR by vaccination the current year (2014–2015) and the two previous years (2013–2014 and 2012–2013). (PDF 31 kb) Additional file 13: Figure S7. Site-specific IVE against all influenza types for all ages. (PDF 43 kb) Additional file 14: Table S7. Sensitivity analysis. (PDF 106 kb) Additional file 15: Figure S8. Statistical methods to account for data clustering by site. (PDF 31 kb) The authors thank Dr. Phillip Leventhal (4Clinics, Paris, France) for medical writing, which was funded by Sanofi Pasteur. Members of the GIHSN as of October 19, 2015 Olga Afanasieva (Research Institute of Influenza, St. Petersburg, Russian Federation), Veronica Afanasieva (Research Institute of Influenza, St. Petersburg, Russian Federation), Meral Akcay Ciblak (National Influenza Reference Laboratory Capa-Istanbul, Istanbul, Turkey), F Aktas (Faculty of Medicine Gazi University, Ankara, Turkey), Selim Badur (National Influenza Reference Laboratory Capa-Istanbul, Istanbul, Turkey), Ángel Belenguer-Varea (Hospital de La Ribera, Alzira, Spain), S Borekci (Cerrahpaşa Faculty of Medicine, Istanbul University, Istanbul, Turkey), Fernando Boza (Hospital Quinta D’Or, Rio de Janeiro, Brazil), Elena Bursteva (D.I. Ivanovsky Institute of Virology FGBC “N.F. Gamaleya FRCEM” Ministry of Health, Moscow, Russian Federation), Zhanna Buzitskaya (Research Institute of Influenza, St. Petersburg, Russian Federation), Braulia Caetano (Oswaldo Cruz Foundation (FIOCRUZ), Rio de Janeiro, Brazil), Jian Cai (Zhejiang Provincial Center for Disease Control and Prevention, Hangzhou, China), B Çakir (Department of Public Health, Faculty of Medicine, Sihhiye, Ankara, Turkey), Mario Carballido-Fernández (Hospital General, Castellón, Spain), Empar Carbonell-Franco (Hospital Arnau de Vilanova, Valencia, Spain), Concha Carratalá-Munuera (Universidad Miguel Hernández, San Juan de Alicante, Spain), S Çelebi (Uludağ University Faculty of Medicine, Bursa, Turkey), C Chai (Zhejiang Provincial Center for Disease Control and Prevention, Hangzhou, China), Ivamara Changas de Lima Porto de Paula (Universidade Federal do Ceará, Fortaleza, Brazil), Enfu Chen (Zhejiang Provincial Center for Disease Control and Prevention, Hangzhou, China), Ben Cowling (School of Public Health, Li Ka Shing Faculty of Medicine, University of Hong Kong, Hong Kong Special Administrative Region, China), Yunjie Cui (Changping District Hospital, Beijing, China), DB Deniz (Dr. Siyami Ersek Gögüs Kalp ve Damar Cerrahisi Egitim ve Arastirma Hastanesi, Istanbul, Turkey), Elena Dondurei (Research Institute of Influenza, St. Petersburg, Russian Federation), H Dong (Ningbo Center for Diseases Prevention and Control, Ningbo, China), X Dong (The First Peoples’ Hospital of Huzhou, Huzhou, China), Mine Durusu (Hacettepe University, Ankara, Turkey), Clotilde El Guerche-Séblain (SANOFI Pasteur, Lyon, France), Fernanda Enda-Moura (Universidade Federal do Ceará, Fortaleza, Brazil), A Eren-Şensoy (Dr. Siyami Ersek Göğüs Kalp ve Damar Cerrahisi Eğitim ve Araştırma Hastanesi, Istanbul, Turkey), Artem Fadeev (Research Institute of Influenza, St. Petersburg, Russian Federation), Luzhao Feng (Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing, China), Shuo Feng (School of Public Health, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong Special Administrative Region), Raimundo César (Ferreira de Silva Filho, Universidade Federal do Ceará, Fortaleza, Brazil), Patricia Fisch (Hospital Nossa Senhora da Conceição, Brazil), Ekaterina Garina (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), S Gencer (Dr. Lütfi Kırdar Kartal Training and Research Hospital, Istanbul, Turkey), Vicente Gil-Guillén (Hospital de Elda, Elda, Spain), Alexa Go (Research Institute of Influenza, St. Petersburg, Russian Federation), Vitaly Gonchar (Research Institute of Influenza, St. Petersburg, Russian Federation), Ekaterina Golovacheva (Research Institute of Influenza, St. Petersburg, Russian Federation), Mikhail Grudinin (Research Institute of Influenza, St. Petersburg, Russian Federation), M Hacımustafaoğlu (Uludağ University, Bursa, Turkey), S Hancerli (Istanbul Faculty of Medicine, Istanbul University, Istanbul, Turkey), Martina Havlickova (National Institute of Public Health, Prague, Czech Republic), Kristyna Herrmannova (Hospital Na Bulovce, Prague, Czech Republic), L Huang (Ningbo Women and Children Hospital, Ningbo, China), Hui Jiang (Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing, China), Dennis Ip (School of Public Health, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong Special Administrative Region), Helena Jirincova (National Institute of Public Health, Prague, Czech Republic), Lucie Jurzykowska (National Institute of Public Health, Prague, Czech Republic), Lidiya Kisteneva (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), Ludmila Kolobukhina (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), Andrey Komissarov (Research Institute of Influenza, St. Petersburg, Russian Federation), Radka Kralova (National Institute of Public Health, Prague, Czech Republic), Kirill Krasnoslobotsev (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), Jan Kyncl (National Institute of Public Health, Prague, Czech Republic), Xavier Labrador (FISABIO-Salud Publica, Valencia, Spain and Consorcio de Investigación Biomédica de Epidemiología y Salud Pública, Spain, Instituto Carlos II, Madrid, Spain), Chao Li (Huairou District Center for Diseases Prevention and Control, Beijing, China), Xiangxin Li (Chanping District Hospital, Beijing, China), Ramón Limón-Ramírez (Hospital de la Plana, Vila-real, Spain), Jianhua Liu (The First Hospital in Huairou District, Beijing, China), Mari Carmen (Llopis Garcia, FISABIO-Salud Pública, Valencia, Spain), Cédric Mahé (SANOFI Pasteur, Lyon, France), Zdenka Mandakova (National Institute of Public Health, Prague, Czech Republic), L Merkulova (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), Sevim Mese (Istanbul University, Istanbul, Turkey), Ainara Mira Iglesias (FISABIO-Salud Pública, Valencia, Spain), Evgenia Mukasheva (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), Angels Natividad Sancho (FISABIO-Salud Pública, Valencia, Spain), Lucia Nováková (Hospital Na Bulovce, Prague, Czech Republic), Elena Obraztsova (Research Institute of Influenza, St. Petersburg, Russian Federation), Ludmila Osidak (Research Institute of Influenza, St. Petersburg, Russian Federation), Maria del Carmen (Otero-Reigada Hospital Universitario y Politécnico La Fe, Valencia, Spain), S Özer (Dr. Lütfi Kırdar Kartal Training and Research Hospital, Istanbul, Turkey), L Ozisik (Hacettepe University, Ankara, Turkey), Valentina Picot (Fondation Mérieux, Lyon, France), Maria Pisarev (Research Institute of Influenza, St. Petersburg, Russian Federation), Florence Pradel (Fondation Mérieux, Lyon, France), Jitka Prochazkova (National Institute of Public Health, Prague, Czech Republic), Joan Puig-Barberá (FISABIO-Salud Pública, Valencia, Spain), Ying Qin (Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing, China), Sonia Raboni (Hospital de Clínicas/Universidade Federal do Paraná, Curritiba, Brazil), Hana Roháčová (Hospital Na Bulovce, Prague, Czech Republic), Elena Rozhkova (Research Institute of Influenza, St. Petersburg, Russian Federation), Sa Li (Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing, China), MI Sadikhova (Research Institute of Influenza, St. Petersburg, Russian Federation), Germán Schwarz-Chavarri (Hospital General de Alicante, Alicante, Spain), Marilda Siqueira (FIOCRUZ, Brazil), Elizaveta Smorodintseva (Research Institute of Influenza, St. Petersburg, Russian Federation), Anna Sominina (Research Institute of Influenza, St. Petersburg, Russian Federation), Kirill Stolyarov (Research Institute of Influenza, St. Petersburg, Russian Federation), Vera Sukhovetskaya (Research Institute of Influenza, St. Petersburg, Russian Federation), G Sun (Beijing Huairou Hospital, Beijing, China), Y Tang (Changping District Center for Diseases Prevention and Control, Beijing, China), Miguel Tortajada-Girbés (Hospital Doctor Peset, Valencia, Spain), Svetlana Trushakova (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), José Tuells (Hospital Universitario del Vinalopó, Elche, Spain), Tatiana Tumina (Research Institute of Influenza, St. Petersburg, Russian Federation), R Vartanyan (D.I. Ivanovsky Institute of Virology, Moscow, Russian Federation), Lubov Voloshuk (Research Institute of Influenza, St. Petersburg, Russian Federation), Quanyi Wang (Beijing Center for Disease Prevention and Control, Beijing, China), D Wen (Huzhou Center for Diseases Prevention and Control, Huzhou, China), Peng Wu (School of Public Health, Li Ka Shing Faculty of Medicine, University of Hong Kong, Hong Kong Special Administrative Region), Wen Xiao (The First Hospital in Huairou District, Beijing, China), Peng Yang (Beijing Center for Disease Prevention and Control, Beijing, China), Marina Yanina (Research Institute of Influenza, St. Petersburg, Russian Federation), Bo Yi (Ningbo Center for Diseases Prevention and Control, Ningbo, China), Hongjie Yu (Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing, China), Kubra Yurtcu (National Influenza Reference Laboratory Capa-Istanbul, Istanbul, Turkey), Pavel Zarishnyuk (Research Institute of Influenza, St. Petersburg, Russian Federation), S Zhang (The Third Hospital of Yinzhou District, Ningbo, China), Yi Zhang (Beijing Center for Disease Prevention and Control, Beijing, China), Tiebiano Zhang (The First Hospital in Huairou District, Beijing, China), Jiandong Zheng (Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing), Zhibin Peng (Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Disease, Chinese Center for Disease Control and Prevention, Beijing, China). Funding The study and its publication was funded by Sanofi Pasteur and the participating institutions. Availability of data and materials The datasets supporting the conclusions of this article are included within the article and its appendices. Authors’ contributions All authors participated in the collection and analysis of data, preparation of the manuscript, and approval of the final version. Competing interests The authors’ institutions received funding from Sanofi Pasteur for the conduct of this trial. The authors declare no other competing interests related to this article. Ethics approval and consent to participate The protocol used by the GIHSN was approved by each site’s Ethics Research Committee. All patients provided written informed consent. ==== Refs References 1. Lee N Chan PK Lui GC Wong BC Sin WW Choi KW Complications and outcomes of pandemic 2009 Influenza A (H1N1) virus infection in hospitalized adults: how do they differ from those in seasonal influenza? J Infect Dis. 2011 203 1739 47 10.1093/infdis/jir187 21606532 2. Jernigan DB Cox NJ Webster RG Monto AS Braciale TJ Lamb RA Human influenza: One health, one world Textbook of influenza 2013 Chichester, West Sussex Wiley Blackwell 3 19 3. Hayward AC Fragaszy EB Bermingham A Wang L Copas A Edmunds WJ Comparative community burden and severity of seasonal and pandemic influenza: results of the Flu Watch cohort study Lancet Respir Med. 2014 2 445 54 10.1016/S2213-2600(14)70034-7 24717637 4. World Health Organization Vaccines against influenza WHO position paper - November 2012 Wkly Epidemiol Rec 2012 87 461 76 23210147 5. Puig-Barbera J Tormos A Sominina A Burtseva E Launay O Ciblak MA First-year results of the Global Influenza Hospital Surveillance Network: 2012–2013 Northern hemisphere influenza season BMC Public Health. 2014 14 564 10.1186/1471-2458-14-564 24903737 6. Ortiz JR Sotomayor V Uez OC Oliva O Bettels D McCarron M Strategy to enhance influenza surveillance worldwide Emerg Infect Dis. 2009 15 1271 8 10.3201/eid1508.081422 19751590 7. Puig-Barbera J, Tormos A, Trushakova S, Sominina A, Pisareva M, Ciblak MA, et al. The Global Influenza Hospital Surveillance Network (GIHSN): A new platform to describe the epidemiology of severe influenza. Influenza Other Respir Viruses. 2015;9:277-86. 8. Puig-Barbera J Natividad-Sancho A Launay O Burtseva E Ciblak MA Tormos A 2012–2013 Seasonal influenza vaccine effectiveness against influenza hospitalizations: results from the global influenza hospital surveillance network PLoS One. 2014 9 e100497 10.1371/journal.pone.0100497 24945510 9. Puig-Barbera J Natividad-Sancho A Trushakova S Sominina A Pisareva M Ciblak MA Epidemiology of Hospital Admissions with Influenza during the 2013/2014 Northern Hemisphere Influenza Season: Results from the Global Influenza Hospital Surveillance Network PLoS One. 2016 11 e0154970 10.1371/journal.pone.0154970 27196667 10. Kirkwood BR Sterne JAC Analysis of clustered data Essential medical statistics. edn 2003 Malden Blackwell Science 355 70 11. Higgins JP Thompson SG Quantifying heterogeneity in a meta-analysis Stat Med. 2002 21 1539 58 10.1002/sim.1186 12111919 12. Higgins JP Thompson SG Deeks JJ Altman DG Measuring inconsistency in meta-analyses BMJ. 2003 327 557 60 10.1136/bmj.327.7414.557 12958120 13. World Health Organization Review of the 2014–2015 influenza season in the northern hemisphere Wkly Epidemiol Rec 2015 90 281 96 26050269 14. Barr IG, Vijaykrishna D, Sullivan SG. Differential age susceptibility to influenza B/Victoria lineage viruses in the 2015 Australian influenza season. Euro Surveill. 2016;21. pii: 19692. 15. Reed C Chaves SS Perez A D'Mello T Daily Kirley P Aragon D Complications among adults hospitalized with influenza: a comparison of seasonal influenza and the 2009 H1N1 pandemic Clin Infect Dis. 2014 59 166 74 10.1093/cid/ciu285 24785230 16. Van Kerkhove MD Vandemaele KA Shinde V Jaramillo-Gutierrez G Koukounari A Donnelly CA Risk factors for severe outcomes following 2009 influenza A (H1N1) infection: a global pooled analysis PLoS Med. 2011 8 e1001053 10.1371/journal.pmed.1001053 21750667 17. Chaves SS Aragon D Bennett N Cooper T D'Mello T Farley M Patients hospitalized with laboratory-confirmed influenza during the 2010–2011 influenza season: exploring disease severity by virus type and subtype J Infect Dis. 2013 208 1305 14 10.1093/infdis/jit316 23863950 18. Dao CN Kamimoto L Nowell M Reingold A Gershman K Meek J Adult hospitalizations for laboratory-positive influenza during the 2005–2006 through 2007–2008 seasons in the United States J Infect Dis. 2010 202 881 8 10.1086/655904 20677944 19. Wong JY Kelly H Cheung CM Shiu EY Wu P Ni MY Hospitalization Fatality Risk of Influenza A(H1N1)pdm09: A Systematic Review and Meta-Analysis Am J Epidemiol. 2015 182 294 301 10.1093/aje/kwv054 26188191 20. Skowronski DM, Chambers C, Sabaiduc S, De Serres G, Dickinson JA, Winter AL, et al. Interim estimates of 2014/15 vaccine effectiveness against influenza A(H3N2) from Canada's Sentinel Physician Surveillance Network, January 2015. Euro Surveill. 2015;20. pii: 21022. 21. Flannery B Clippard J Zimmerman RK Nowalk MP Jackson ML Jackson LA Early estimates of seasonal influenza vaccine effectiveness - United States, January 2015 MMWR Morb Mortal Wkly Rep. 2015 64 10 5 25590680 22. Puig-Barbera J, Mira-Iglesias A, Tortajada-Girbes M, Lopez-Labrador FX, Belenguer-Varea A, Carballido-Fernandez M, et al. Effectiveness of influenza vaccination programme in preventing hospital admissions, Valencia, 2014/15 early results. Euro Surveill. 2015;20. pii: 21044. 23. World Health Organization Recommended composition of influenza virus vaccines for use in the 2015–2016 northern hemisphere influenza season Wkly Epidemiol Rec 2015 90 97 108 25771542 24. Castilla J Navascues A Fernandez-Alonso M Reina G Pozo F Casado I Effectiveness of subunit influenza vaccination in the 2014–2015 season and residual effect of split vaccination in previous seasons Vaccine. 2016 34 1350 7 10.1016/j.vaccine.2016.01.054 26854911 25. Ohmit SE Petrie JG Malosh RE Fry AM Thompson MG Monto AS Influenza vaccine effectiveness in households with children during the 2012–2013 season: assessments of prior vaccination and serologic susceptibility J Infect Dis. 2015 211 1519 28 10.1093/infdis/jiu650 25416812 26. Smith DJ Forrest S Ackley DH Perelson AS Variable efficacy of repeated annual influenza vaccination Proc Natl Acad Sci U S A. 1999 96 14001 6 10.1073/pnas.96.24.14001 10570188 27. Sullivan SG Tay EL Kelly H Variable definitions of the influenza season and their impact on vaccine effectiveness estimates Vaccine. 2013 31 4280 3 10.1016/j.vaccine.2013.06.103 23850417 28. Foppa IM Haber M Ferdinands JM Shay DK The case test-negative design for studies of the effectiveness of influenza vaccine Vaccine. 2013 31 3104 9 10.1016/j.vaccine.2013.04.026 23624093 29. Rothman KJ Greenland S Lash TL Validity in epidemiological studies Modern Epidemiology 2008 Philadelphia Wolters Kluwer Health/Lippincott Williams & Wilkins 128 47 30. Ortiz JR Neuzil KM Rue TC Zhou H Shay DK Cheng PY Population-based incidence estimates of influenza-associated respiratory failure hospitalizations, 2003 to 2009 Am J Respir Crit Care Med. 2013 188 710 5 10.1164/rccm.201212-2341OC 23855650 31. Chan PK Chan MC Cheung JL Lee N Leung TF Yeung AC Influenza B lineage circulation and hospitalization rates in a subtropical city, Hong Kong, 2000–2010 Clin Infect Dis. 2013 56 677 84 10.1093/cid/cis885 23074315 32. Ortiz JR Neuzil KM Shay DK Rue TC Neradilek MB Zhou H The burden of influenza-associated critical illness hospitalizations Crit Care Med. 2014 42 2325 32 10.1097/CCM.0000000000000545 25148596 33. Cohen C Walaza S Moyes J Groome M Tempia S Pretorius M Epidemiology of severe acute respiratory illness (SARI) among adults and children aged >/=5 years in a high HIV-prevalence setting, 2009–2012 PLoS One. 2015 10 e0117716 10.1371/journal.pone.0117716 25706880

PMC005xxxxxx/PMC5001210.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556295289310.1186/s12864-016-2893-xIntroductionIntelligent biology and medicine in 2015: advancing interdisciplinary education, collaboration, and data science Huang Kun Kun.Huang@osumc.edu 1Liu Yunlong yunliu@iupui.edu 23Huang Yufei YHuang@utsa.edu 4Li Lang lali@iupui.edu 23Cooper Lee lee.cooper@emory.edu 56Ruan Jianhua Jianhua.Ruan@utsa.edu 7Zhao Zhongming zhongming.zhao@uth.tmc.edu 81 Department of Biomedical Informatics, The Ohio State University, Columbus, OH 43210 USA 2 Department of Medical and Molecular Genetics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 3 Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 4 Department of Electrical and Computer Engineering, The University of Texas at San Antonio, San Antonio, TX 78249 USA 5 Department of Biomedical Informatics, Emory University, Atlanta, GA 30322 USA 6 Department of Biomedical Engineering, Emory University / Georgia Institute of Technology, Atlanta, GA 30322 USA 7 Department of Computer Science, The University of Texas at San Antonio, San Antonio, TX 78249 USA 8 Center for Precision Health, School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, TX 77030 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.524© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.We summarize the 2015 International Conference on Intelligent Biology and Medicine (ICIBM 2015) and the editorial report of the supplement to BMC Genomics. The supplement includes 20 research articles selected from the manuscripts submitted to ICIBM 2015. The conference was held on November 13–15, 2015 at Indianapolis, Indiana, USA. It included eight scientific sessions, three tutorials, four keynote presentations, three highlight talks, and a poster session that covered current research in bioinformatics, systems biology, computational biology, biotechnologies, and computational medicine. The International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Introduction The 2015 International Conference on Intelligent Biology and Medicine (ICIBM 2015) brought together more than one hundred twenty researchers and trainees from multiple countries with highly interdisciplinary background such as biology, medicine, computer science, biomedical engineering, statistics, mathematics, and chemistry, among others, participated this exciting three-day event. The event provided a forum for attendees to exchange new ideas, demonstrate new tools, showcase recent innovative work, and foster interdisciplinary research collaborations. It was also an important educational and training venue for students and junior investigators in bioinformatics, systems biology, intelligent computing, and computational medicine. The scientific program included eight scientific sessions covering seven important areas, four tutorials, four keynote presentations, nine highlight talks, and a poster session. Among the scientific sessions, since there are many interesting and important work related to the area of next generation sequencing data analysis, it was split into two sessions held on different days. The detailed information of all presentations and posters can be found on the conference website [1] and in the conference program book. Here we briefly review the keynote speakers’ presentation, followed by the tutorials, and the papers selected by BMC Genomics in the regular scientific sessions. Keynote lectures Four keynote speakers who are world-renowned leaders in bioinformatics, medical informatics, integrative genomics, systems biology and computational medicine delivered lectures on their cutting-edge research, provided insightful views for their research fields, and shared their perspectives on the future of these fields in the era of big data. These speakers were Dr. Jiajie Zhang from The University of Texas Health Science Center at Houston, Dr. Christopher Sanders from Harvard Medical School, Dr. Keith Dunker from Indiana University, and Dr. Sylvia Plevritis from Stanford University. Below, we briefly introduce the four keynote presentations. “Beyond Big Data: New Trends in Biomedical Informatics Research and Education” by Dr. Jiajie Zhang. Dr. Zhang is the Dean, the Glassell Family Foundation Distinguished Chair in Informatics Excellence, and the Dr. Doris L. Ross Professor at The University of Texas School of Biomedical Informatics. He is an elected fellow of the American College of Medical Informatics. Dr. Zhang in his presentation provided insights on the rapidly emerging new trends in biomedical informatics beyond the current Big Data Revolution. He pointed out that Big Data itself cannot solve any problems, instead what the community needs are precise, small scale, transparent, live knowledge and intelligence for the right problem of a right individual at the right time and the right location, with the power of prediction and intervention. In addition, Dr. Zhang discussed these new trends in biomedical informatics research and education as the promise of data in medicine and biology will not be realized without a new generation of students, researchers, and developers who are trained with the state-of-the-art tools and equipped with the most up-to-date knowledge in biomedical informatics. “Systems Biology in Action: Prediction of Large Protein 3D Structures and Design of Cancer Combination Therapy” by Dr. Christopher Sander. Dr. Sander was Head of the Computational Biology Center at Memorial Sloan Kettering Cancer Center and Tri-Institutional Professor at Rockefeller and Cornell Universities. Dr. Sander is a pioneer and world leader in computational biology. In this presentation, he presented his recent high impact work on systems pharmacology, computational genomics, and protein structure and function. First, he introduced perturbation biology, a method for computing responses to combinatorial therapy aiming to block the emergence of resistance to targeted cancer therapies. In addition, he presented cBioPortal for cancer genomics in collaboration with Niki Schultz and Ethan Cerami being information resource architects under his leadership. Finally, he showed recent work on solving the 3D protein folding prediction problem using statistical physics and information from next-generation sequencing in collaboration with Debora Marks at Harvard Medical School and the Zecchina group at the Politecnico di Torino. “Intrinsically Disordered Protein and the Origins of Complex Multicellular Organisms” by Dr. Keith Dunker. Dr. Dunker is a renowned protein scientist and a professor of Biochemistry and Molecular Biology and was the Director of the Center for Computational Biology and Bioinformatics at Indiana University. In this presentation, Dr. Dunker focused on understanding two important biological questions for evolution of multicellular organisms: First, which molecular functions underpinned the evolution of multicellular organisms? Secondly, which of these molecular functions depend on the intrinsically disordered proteins (IDP)? Answers to the first question involve the advent of molecules for cellular adhesion, cell-cell communication, developmental programs, spatial and temporal regulation of the developmental programs as well as cell-specific biochemistry. For the second question, Dr. Dunker and his colleagues used key-words in Swiss Protein ranked for associations with predictions of protein structure or disorder and found that “differentiation” was the biological process most strongly associated with IDPs. In addition, all of the aforementioned five underpinning molecular functions for multicellularity were found to strongly depend on IDP-based mechanisms. These findings lead to new direction in characterizing the evolution of complex multicellular organisms and using gene regulatory network models currently used to explain cellular differentiation. “Optimizing Combination Cancer Therapy Based on Single Cell Analysis” Presented by Dr. Sylvia Plevritis. Dr. Plevritis is a Professor of Radiology in the School of Medicine and (by courtesy) of Management Science and Engineering in the School of Engineering at Stanford University. She is co-Section Chief of Integrative Biomedical Imaging Informatics (IBIIS) at Stanford University, Director of the Stanford Center for Cancer Systems Biology (CCSB) and Director of the Cancer Systems Biology Scholars (CSBS) postdoctoral training program. Tumors often evolve under exposure to different treatment regimes, leading to intratumor heterogeneity and different clinical outcomes. Personalized treatment strategies are thus required to improve the effectiveness of cancer drug combination therapy. In this study, Dr. Plevritis and her colleagues aim to identify optimal cancer drug combinations based on characterization of an individual patient’s intratumoral heterogeneity in response to a screening panel of single drugs at the single cell level. Specifically, Dr. Plevritis presented mass cytometry, the state-of-the-art single cell technology, to elucidate the intratumoral response to drug exposure. This approach is based on a set of novel algorithms to analyze the high- dimensional data generated by the mass cytometry. The ultimate goal of this study is to optimize cancer drug combinations for each patient by capturing an individual patient’s intratumoral heterogeneity through integrative single-cell based analyses based on proteomic, genomic and imaging data. Tutorials ICIBM 2015 included three tutorial sessions covering frontier and emerging research topics such as next-generation sequencing data analysis, proteomics, and bioimage informatics. These tutorials were well attended and appreciated by the conference participants. They provided a wealth of information on these cutting-edge areas and techniques in bioinformatics and computational biology. “Advanced Data Mining and Quality Control of High Throughput Sequencing Data” (provided by Dr. Yan Guo from Vanderbilt University): Illumina high throughput sequencing (HTS) technology is one of the most prevalent high throughput sequencing technologies driving genomic studies. However HTS data creates numerous bioinformatics challenges due to its complexity, yet simultaneously offer exciting new opportunities for mining new biomedical informatics from the vast amount of data. In this tutorial, Dr. Guo discussed two major aspects of HTS data analysis: quality control (QC) and data mining. QC is critical for downstream analysis but a common myth is that quality control is only needed at the raw data. In fact, QC for HTS data needs to be carried out in at least three stages: raw data, alignment, and variant calling. In this tutorial, Dr. Guo introduced the concept of multi-perspective QC and discussed the detailed QC strategies for both RNA and DNA sequencing data. HTS data with proper QC also provide unique opportunities for data mining, and the specific type of data mining depends on the specific sequencing methods. Dr. Guo specifically introduced potential minable components from both exome and RNA-seq data and discussed the appropriate approach for each of the components. “Novel genotype-phenotype associations in human cancers enabled by advanced molecular platforms and computational analysis of whole slide images” This tutorial was organized by Dr. Lee Cooper from Emory University and Dr. Kun Huang from the Ohio State University. Advances in computing, imaging, and genomics have created new opportunities for exploring relationships between histology, molecular events, and clinical outcomes using quantitative methods. Slide scanning devices are now capable of rapidly producing massive digital image archives that capture histological details in high resolution. Commensurate advances in computing and image analysis algorithms enable mining of archives to extract descriptions of histology, ranging from basic human annotations to automatic and precisely quantitative morphometric characterization of hundreds of millions of cells. These imaging capabilities represent a new dimension in tissue-based studies, and when combined with genomic and clinical endpoints, can be used to explore biologic characteristics of the tumor microenvironment and to discover new morphologic biomarkers of genetic alterations and patient outcomes. In this tutorial, Drs. Cooper and Huang described recent developments in quantitative pathology imaging and illustrate how image features can be integrated with clinical and genomic data to investigate fundamental problems in cancer. Using motivating examples from the study of gliomas (GBMs), Dr. Cooper also demonstrated how public data from The Cancer Genome Atlas (TCGA) can serve as an open platform to conduct in silico tissue-based studies that integrate existing data resources. “Computational Challenges in Top-Down Proteomics” This tutorial is offered by Dr. Xiaowen Chen from School of Informatics and Computing at IUPUI. While the genome provides the blueprint of gene products, proteins are the bricks and mortar of biology. Mass spectrometry (MS) is the core technology for the studies of proteins and post-translational modifications. Over the past decade, proteomics has been dominated by bottom-up MS that digests proteins into fragments and analyzes the resulting short peptides. Since information about intact proteins is lost during digestion, recent studies advocated top-down MS that analyzes intact proteins and gives rise to many computational challenges. While top-down MS researchers have made great progress, the algorithms for interpreting top-down MS data are still in their infancy. We describe computational challenges and combinatorial algorithms for the analysis of top-down MS data and show how they enable new biological applications. Scientific sessions and BMC Genomics / Systems Biology supplement issues ICIBM 2015 had eight regular scientific sessions covering recent research in the areas of bioinformatics, systems biology, intelligent computing, and computational medicine. The detailed information of the sessions, including session chairs, authors, presenters, and the title and abstract of each talk were made available on the conference website [1] and in the conference program brochure. The presentations were selected through a rigorous review process from more than 60 submissions based on their scientific merit and technical quality by a program committee of more than 90 experts in the field (see the Conference Organization section below). These sessions were:Session I: NGS Data Analysis, I Session II: Systems Biology Session III: NGS Data Analysis, II Session IV: Integrative Genomics Session V: Genomics and Genetics Session VI: Epigenomics, Proteomics and Metabolomics Session VII: Biomarker Discovery and Precision Medicine Session VIII: Pharmacogenomics and Systems Medicine Here we present the editorial report for the supplement to BMC Genomics that includes 20 research papers. Each selected manuscript was reviewed for two rounds. The first round of reviewed was carried out by at least three reviewers and was substantially revised according to reviewers’ critiques. The revision was further reviewed by at least two reviewers before being accepted into the supplement issue. These papers cover a wide spectrum of topics in bioinformatics and computational biology. Below we group the papers in this special issue of BMC Genomics into five categories and summarize them. Next and the third generation sequencing data analysis methods While next generation sequencing (NGS) techniques have been widely adopted by biomedical community during the past decade, new experimental techniques and even the third-generation sequencing technology keep proposing new challenges for data analysis. During the conference, several groups proposed new methods or methods comparison for processing and analyzing NGS and third generation sequencing data. In [2], Feng et al. proposed a Bayesian inference-based method that takes advantage of the signal distributions of electrical voltages measured for all the homopolymers for third generation sequencing technology such as the Nanopore sequencer. By cross-referencing the length of homopolymers in the reference genome and the voltage signal distributions, the proposed integrated model significantly improves the alignment accuracy around the homopolymer regions. Cherukuri and Janga [3] then benchmarked available assembler algorithms such as de Bruijn graphs, Overlap Layout Consensus (OLC) and Greedy extension approaches to find an appropriate framework that can efficiently assemble Nanopore sequenced reads. Their analysis unveiled that OLC-based algorithms could generate a high quality assembly with ten times higher N50 & mean contig values as well as one-fifth the number of total number of contigs compared to other tools. The findings should help in stimulating the development of novel assemblers for handling Nanopore sequence data. In [4], the study aimed to evaluate the appropriateness of different statistical distributions on modeling sequence-context-dependent DNA sequencing error rates generated by different NGS technologies. Using a generalized linear model framework, Hao et al. found that zero-inflated negative binomial distribution fits the sequencing errors the best and also performed the best in identifying low-frequency single nucleotide variants (SNVs), especially within the 0.5 % to 1 % ranges in two commonly used sequencing platforms with completely different biochemistries - Ion Proton and Illumina MiSeq. This work provides guidance in predicting sequencing errors and facilitates low-frequency SNV detection as well as their downstream applications. The development of single-cell RNA sequencing enables tracking cell heterogeneity and determination of cell subpopulations. Chen et al. [5] developed a gene expression variation model (GEVM), utilizing the relation between coefficient of variation (CV) and average expression level to address the over-dispersion of single-cell data, and its corresponding statistical significance to quantify the variably expressed genes (VEGs). Obtaining VEGs allowed researchers to detect possible subpopulations, providing further evidences of cell heterogeneity. In Cui et al. [6], the authors developed a novel algorithm for uncovering the potential types of m6A methylation by clustering the degree of m6A methylation peaks in MeRIP-Seq data. This algorithm utilizes a hierarchical graphical model to model the reads account variance and the underlying clusters of the methylation peaks. It was applied to two different MeRIP-seq datasets and revealed a novel pattern that methylation peaks with less peak enrichment tend to clustered in the 5’ end of both in both mRNAs and lncRNAs, whereas those with higher peak enrichment are more likely to be distributed in CDS and towards the 3’end of mRNAs and lncRNAs. These results suggest that m6A’s functions could be location specific. Applications of sequence and NGS data analysis methods In addition to developing novel methods for analyzing data generated from new NGS techniques or the third-generation technology, application of NGS technologies to biological and medical problems also calls for extensive bioinformatics research. Bai et al. [7] has investigated the variation of gene expression in blood transcriptome profile of Chinese Holstein cows associated to the milk yield traits. Totally, 100 differentially expressed genes (DEGs) between 13 high yielders and 10 low yielders were obtained, which were shown to be significantly enriched in immune response processes. Furthermore, alternative splicing analysis demonstrated that the alternative 3’ splicing site was the major splicing pattern in high yielders, however, in low yielders was exon skipping. This study allowed us to explore associations between immune traits and production traits related to milk production. In [8], Zhou et al. identified 197 exons whose BMSC splicing patterns were altered by LPS via comparing RNA-seq data from LPS-treated samples versus the control. Functional analysis of these alternatively spliced genes demonstrated significant enrichment of phosphoproteins, zinc finger proteins, and proteins undergoing acetylation. Additional bioinformatics analysis strongly suggest that LPS-induced alternatively spliced exons could have major effects on protein functions by disrupting key protein functional domains, protein-protein interactions, and post-translational modifications. The study provides greater understanding of the intracellular mechanisms that underlie the therapeutic potential of BMSCs. The evolution of exceptionally powerful transporter systems in Streptomyces has enabled their adaptation to the complex soil environment. A better understanding of transport systems will allow enhanced optimization of production processes for both pharmaceutical and industrial applications of Streptomyces. In [9], Zhou et al. presented a catalog of transport systems in eleven Streptomyces species and found that each of the species possesses a rich repertoire of transport proteins, which can be divided into a wide range of transporter families. To characterize the biological and medical significance of Bacillus sp. NRRL B-14911, in particular, cardiac autoimmunity, Massilamany et al. [10] sought to analyze the complete genome sequence of this bacterium. The genome was found to encode several virulence factors like adhesins, invasins, colonization factors, siderophores and transporters. The availability of complete genome sequence of this bacterium may facilitate genetic manipulations to assess gene functions associated with bacterial survival and virulence, and also to establish a disease model to study the immune pathogenesis of bacterial myocarditis. In [11], Bai et al. presented an improved algorithm “Read-Split-Run” (RSR) for detecting genome-wide Ire1α-targeted genes with non-canonical spliced regions at a faster speed. They compared the RSR algorithm to the “Read-Split-Walk” (RSW) algorithm when applied to mouse embryonic fibroblast cells (MEF) and the human Encyclopedia of DNA Elements (ENCODE) RNA-seq data. The new RSR algorithm outperformed others in the defined context and showed a higher efficiency in identifying novel splice junctions genome-wide. Integrative genomics and precision medicine One important issue in precision medicine is how to effectively integrate multiple modalities of biomedical data, especially different sets of high throughput molecular data to better stratify patients into groups with distinctive clinical outcomes such as different prognosis and response to treatment. During this conference, a series of papers proposed different integrative genomic methods to achieve this goal. While whole exome-sequencing is widely used to screen for somatic mutations in cancer patients, the mutations often do not occur on the same genes among patients. In [12], Zhang et al. developed a novel approach of integrating patient somatic mutation, transcriptome and clinical data to mine underlying functional gene groups that can be used to stratify cancer patients into groups with different clinical outcomes. Specifically, distance correlation metric was used to mine the correlations between expression profiles of mutated genes from different patients. With this method, a stable subgroup of breast cancer patients that are highly enriched with ER-negative and triple-negative subtypes were identified, and the somatic mutation genes they harbor were capable of acting as potential biomarkers to predict patient survival in several different breast cancer datasets, especially in ER-negative cohorts which has lacked of reliable biomarkers. The method provides a novel and promising approach of integrating genotype and gene expression data in patient stratification in complex diseases. Proper cell models for breast cancer primary tumors have long been the focal point in the cancer’s research. In [13], a comprehensive comparison in copy number variation (CNV), mutation, mRNA expression and protein expression between 68 breast cancer cell lines and 1375 primary breast tumors is conducted and presented. The important drug targets, ESR1, PGR, HER2, EGFR and AR have a high similarity in mRNA and protein in both tumors and cell lines. A total score developed from the four correlations among four molecular profiles suggests that cell lines, BT483, T47D and MDAMB453 have the highest similarity with tumors. In [14], Wang et al. proposed an integrative genomics approach to explore the functional consequences of a key driver gene, PBRM1 through its truncated mutations in clear cell renal cell carcinoma (ccRCC) by incorporating somatic mutations, mRNA expression, DNA methylation, and microRNA expression profiles from The Cancer Genome Atlas (TCGA). Their results suggested that methylation and microRNA alterations were likely the downstream events associated with the PBRM1 truncation mutations. This study provided some important insights into the understanding of tumorigenesis driven by PBRM1 truncated mutations in ccRCC. Cancer biomarker discovery and pan cancer study Discovery of biomarkers and signatures is an important issue in translational research for cancers. During this conference, several studies focused on the methods for identifying cancer specific or common signatures and markers predicting clinical outcomes for cancer patients. For instance, to classify cancer classes (e.g. subtypes) using patient gene expression profiles when both systematic and condition-specific biases presented, Ma et al. [15] developed a novel algorithm called CrossLink (CL). CL exploits the fact that the signature is unique to its associated class under any condition and thus employs an unsupervised clustering algorithm to discover this unique signature. The results showed that CL can achieve robust and improved performance than state-of-the-art normalization algorithms. In [16], the authors performed a pan-cancer analysis of copy number of variants (CNVs) and gene expression in one of the most important gene categories, tumor suppressor genes (TSGs), in order to provide a systematic view of CNV and gene expression concordant changes in TSGs across all the major cancers. They found that 81 TSGs with concordant copy number loss events and decreased gene expression in the tumor samples and provided a draft landscape of CNV in pan-cancer. In [17], Zhang and Chen presented a peptidomics method for identifying cancer-related and isoform-specific peptide for clinical proteomics application from LC-MS/MS. They showed that the method for identifying cancer-specific protein isoform biomarkers from clinical proteomics application is an effective one for increasing the number of identified alternative splicing isoform markers in clinical proteomics. Translational bioinformatics and pharmacogenomics Translational bioinformatics methods including network analysis are widely applied to human disease studies and pharmacogenomics applications. In [18], Chen et al. developed a drug repositioning approach combining human disease genomics and mouse phenotype data towards predicting targeted therapies for glioblastoma (GBM). For existing GBM drugs, this approach achieved a significantly higher median rank than a recent approach (9.2 % vs. 45.6 %). In addition, many top predictions have been demonstrated effective in inhibiting the growth of human GBM cells. In [19], Li et al. extracted the functional modules and identified 19 key rifampin-response genes that are associated with seven function pathways that include drug response and metabolism, and cancer pathways. In addition, six genes functioning as gene hubs in the gene networks that are regulated by rifampin was identified. The results suggest that rifampin contributes to changes in the expression of genes by regulating key molecules in the protein interaction networks. In [20], Xu and Wang presented an integrated approach for drug repurposing for rheumatoid arthritis (RA). They developed a network-based ranking algorithm to find diseases that shared high degrees of genetic commonality with RA and then implemented a drug prioritization algorithm to reposition drugs from RA-related diseases to treat RA. This approach performed significantly better in novel predictions than the existing approach when evaluated using 165 not-yet-FDA-approved RA drugs. While gene co-expression network analysis is widely adopted, there is a lack of a rigorous way to evaluate the concordance of the expression profiles for the genes in co-expressed modules. Han et al. [21] presented a linear algebraic based Centralized Concordance Index (CCI) for evaluating the concordance of co-expressed gene modules from gene co-expression network analysis. The CCI can be used to evaluate the performance for co-expression network analysis algorithms as well as for detecting condition specific co-expression modules. Conference organization 2015 International conference on intelligent biology and medicine (ICIBM 2015) (November 13–15, 2015, Indianapolis, Indiana, USA) Our sincerest thanks to the members of our Steering, Program, Publication, Workshop/Tutorial, Award, Publicity, and Local Organization committees, as well as our numerous reviewers and volunteers, for their countless hours and energy spent on making ICIBM 2015 a success! The conference would not make so many accomplishments if the support and efforts from those people were not provided. Sponsors Indiana University, Center for Computational Biology and Bioinformatics at Indiana University School of Medicine, Vanderbilt University, Bioinformatics Resource Center at Vanderbilt-Ingram Cancer Center, The University of Texas at San Antonio, Shanghai Center for Bioinformation Technology, China. General Chairs Zhongming Zhao (Vanderbilt University and now The University of Texas Health Science Center at Houston) and Yunlong Liu (Indiana University). Steering Committee Lang Li (Indiana University), Yufei Huang (The University of Texas at San Antonio), Mathew Palakal (Indiana University), Tony Hu (Drexel University), Han Liang (MD Anderson Cancer Center), Subha Madhavan (Georgetown University). Program Committee Chair: Kun Huang (The Ohio State University), Co-Chair: Lee Cooper (Emory University), Members: Kristen Anton (Dartmouth College), Yong-sheng Bai (Indiana State University), William S. Bush (Vanderbilt University), Jake Chen (Indiana University -Purdue University Indianapolis), Xue-Wen Chen (Wayne State University), Yidong Chen (University of Texas Health Science Center at San Antonio), Jianlin Cheng (University of Missouri Columbia), Juan Cui (University of Georgia), Qinghua Cui (Peking University, China), Youping Deng (Rush University Medical Center), Joshua Denny (Vanderbilt University), Jeremy Edwards (University of New Mexico Health Sciences Center), Weixing Feng (Harbin Engineering University), Jennifer M. Fettweis (Virginia Commonwealth University), Marcelo Fiszman (National Library of Medicine, National Institutes of Health), Jan Freudenberg (Feinstein Medical Research Institute), Ge Gao (Peking University, China), Chittibabu (Babu) Guda (University of Nebraska Medical Center), Yan Guo (Vanderbilt University), Hao Han (Singapore Bioinformatics Insistute), Zhi Han (The Ohio State University),, Tzu-Hung Hsiao (University of Texas Health Science Center at San Antonio), Weichun Huang, (National Institute of Environmental Health Sciences), Yang Huang (Kaiser Permanente), Yufei Huang (The University of Texas at San Antonio), Jenn-Kang Hwang (National Chiao Tung University, Taiwan), Peilin Jia (The University of Texas Health Science Center at Houston), Yufang Jin (The University of Texas at San Antonio), Victor Jin (The University of Texas Health Science Center at San Antonio), Sun Kim (Seoul National University, Korea), Dmitry Korkin (University of Missouri - Columbia), K.B. Kulasekera (University of Louisville), Jun Kong (Emory University), Fuhai Li (Houston Methodist hospital Research Institute), Lang Li (Indiana Univeristy), Leping Li (National Institute of Environmental Health Sciences), Li Liao (University of Delaware), Honghuang Lin (Boston University), Chunyu Liu (University of Chicago), Hongfang Liu (Georgetown University), Qi Liu (Vanderbilt University), Tianming Liu (University of Georgia), Yin Liu (The University of Texas Health Science Center at Houston), Yunlong Liu (Indiana University), Zhandong Liu (Baylor College of Medicine), Xinghua Lu (University of Pittsburgh), Zhiyong Lu (NCBI, National Library of Medicine), Patricio A. Manque (Universidad Mayor, Chile), Tabrez Mohammad (The University of Texas Health Science Center at San Antonio), Hatice Gulcin Ozer (The Ohio State University), Ranadip Pal (Texas Tech University), Yonghong Peng (University of Bradford, United Kindom), Horacio Perez-Sanchez (Catholic University of Murcia, Spain), Jiang Qian (Johns Hopkins University), Thomas Rindflesch (National Institutes of Health), Jianhua Ruan (The University of Texas at San Antonio), Bairong Shen (Soochow University, China), Alexander Statnikov (New York University Langone Medical Center), Jingchun Sun (The University of Texas Health Science Center at Houston), Wing-Kin Sung (National University of Singapore, Singapore), Manabu Torii (Georgetown University Medical Center), Vladimir Uversky (University of South Florida), Jun Wan (Johns Hopkins University), Edwin Wang (National Council of Canada/McGill University), Jing Wang (Vanderbilt University), Junbai Wang (Oslo University Hospital, Norway), Qingguo Wang (Sloan-Kettering Memorial Institute), Yufeng Wang (The University of Texas at San Antonio), Xiaoyan Wang (University of Connecticut), Chaochun Wei (Shanghai Jiao Tong University), Xiwei Wu (City of Hope National Medical Center), Yonghui Wu (The University of Texas Health Science Center at Houston), Junfeng Xia (Anhui University, China), Lu Xie (Shanghai Center for Bioinformation Technology, China), Hua Xu (The University of Texas Health Science Center at Houston), Jianhua Xuan (Virginia Tech), Bin Xue (University of South Florida), Zhenqing Ye (The University of Texas Health Science Center at San Antonio), Sungroh Yoon (Seoul National University, Korea), Bing Zhang (Vanderbilt University), Jie Zhang (The Ohio State University), Michelle Zhang (The University of Texas at San Antonio), Yanqing Zhang (Georgia State University), Min Zhao (University of the Sunshine Coast, Australia), Yanjun Zhao (Troy University), Zhongming Zhao (Vanderbilt University), Huiru (Jane) Zheng (University of Ulster, United Kingdom), Wenjin Jim Zheng (The University of Texas Health Science Center at Houston), and Dongxiao Zhu (Wayne State University). Publication Committee Chair: Jianhua Ruan (The University of Texas at San Antonio), Co-Chair: Sarath Janga (Indiana University), Milan Radovich (Indiana University). Workshop/Tutorial Committee Chair: Ting Wang (Washington University), Co-Chair: Xiaowen Liu (Indiana University). Award Committee Chair: Hua Xu (The University of Texas Health Science Center at Houston), Co-Chair: Yufei Huang (The University of Texas at San Antonio). Publicity Committee Chair: Dongxiao Zhu (Wayne State University). Local Organization Committee Chair: Yunlong Liu (Indiana University). Acknowledgement We thank numerous reviewers for judging the scientific merits of the manuscripts submitted to ICIBM 2015 and the related special issues. We would like to acknowledge Drs. Yan Guo, Xiaowen Liu, and others for organizing the tutorials in ICIBM 2015. We thank Indiana University and many volunteers for the local support. We thank the National Science Foundation (NSF grant IIS-1451135) for financial support of previous ICIBM conferences. Declarations This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Authors' contributions KH, YL, JR and ZZ managed and participated in the peer-review of ICIBM'15 manuscripts, excluding those on which they were authors, and handled the editorial process of this supplement. YH, LL and LC supported the post-acceptance manuscript processing. KH, YL, JR and ZZ wrote the article, which was read and approved by all authors. All authors have made substantial contribution to the collection of the manuscripts and abstracts and, thus, the success of the ICIBM 2015 conference. Competing interests The authors declare that they have no competing interests. ==== Refs References 1. 2015 International Conference on Intelligent Biology and Medicine. [http://watson.compbio.iupui.edu/yunliu/icibm/index.html] 2. Feng W, Xue D, Song F, Zhao S, Li Z, Chen D, Hao Y, Liu Y. Improving alignment accuracy on homopolymer regions for semiconductor-based sequencing technologies. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 3. Cherukuri Y,Janga S. Benchmarking of de novo assembly algorithms for nanopore data reveals optimal performance of OLC approaches. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 4. Hao Y, Zhang P, Xuei X, Nakshatri H, Edenberg H, Li L, Liu Y. Statistical modeling for sensitive detection of low-frequency single nucleotide variants. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 5. Chen H, Jin Y, Huang Y,Chen Y. Detection of high variability in gene expression from single-cell RNA-seq profiling. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 6. Cui X, Meng J, Rao M, Chen Y, Huang Y. An HMM-based hierarchical model for detecting methylation sites in MeRIP-Seq data. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 7. Bai X, Liu B, Ji X, Zhang W, Bai Y. Whole blood transcriptional profiling comparison between different milk yield of Chinese Holstein cows using RNA-seq data. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 8. Zhou A, Li M, He B, Feng W, Huang F, Xu B, Dunker K, Balch C, Li B, Liu Y, Wang Y. Lipopolysaccharide treatment induces genome-wide pre-mRNA splicing pattern changes in mouse bone marrow stromal stem cells. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 9. Zhou Z, Sun N, Wu S, Li Y, Wang Y. Genomic data mining reveals a rich repertoire of transport proteins in Streptomyces. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 10. Massilamany C, Mohammed A, Loy J, Purvis T, Krishnan B, Basavalingappa R, Kelley C, Chittibabu Guda, Barletta R, Moriyama E, Smith T, Reddy J. Whole genomic sequence analysis of Bacillus infantis: defining the genetic blueprint of strain NRRL B-14911, an emerging cardiopathogenic microbe. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 11. Bai Y, Kinne J, Donham B, Hassler J, Kaufman R. Read-Split-Run: An improved bioinformatics pipeline for identification of genome-wide non-canonical spliced regions using RNA-Seq data. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 12. Zhang J, Abrams Z, Parvin J, Huang K. Integrative analysis of somatic mutations and transcriptomic data to functionally stratify breast cancer patients. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 13. Cheng L, Jiang G, Zhang S, Yazdanparast A, Li M, Liu Y, Inavolu S. Comprehensive comparison of molecular portraits between cell lines and tumor in breast cancer. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 14. Wang Y, Guo X, Bray M, Zhao Z. An integrative genomics approach for identifying novel functional consequences of PBRM1 truncated mutations in clear cell renal cell carcinoma (ccRCC). BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 15. Ma C, Sastry K, Flores M, Gehani S, Al-Bozom I, Feng Y, Serpedin E, Chouchane L, Chen Y, Huang Y. CrossLink. A novel method for cross-condition classification of cancer subtypes. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 16. Zhao M, Zhao Z. Concordance of copy number loss and down-regulation of tumor suppressor genes: a pan-cancer study. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 17. Zhang F, Chen J. A method for identifying discriminative isoform-specific peptides for clinical proteomics application. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 18. Chen Y, Gao Z, Wang B, Xu R. Towards precision medicine-based therapies for glioblastoma: interrogating human disease genomics and mouse phenotypes. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 19. Li J, Wang Y, Wang L, Dai X, Cong W, Feng W, Xu C, Deng Y, Skaar T, Liang H, Liu Y. Identification of rifampin-regulated functional modules and related microRNAs in human hepatocytes based on the protein interaction network. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 20. Xu R, Wang Q. A genomics-based systems approach towards drug repositioning for rheumatoid arthritis. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known 21. Han Z, Zhang J, Sun G, Liu G, Huang K. A matrix rank based concordance index for evaluating and detecting conditional specific co-expressed gene modules. BMC Genomics 2016, 17 (Suppl 7): 10.1186/s12864-016-XXXX-X. Complete DOI once known

PMC005xxxxxx/PMC5001211.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556636289510.1186/s12864-016-2895-8ResearchBenchmarking of de novo assembly algorithms for Nanopore data reveals optimal performance of OLC approaches Cherukuri Yesesri ycheruku@iupui.edu 1Janga Sarath Chandra scjanga@iupui.edu 1231 Department of Bio Health Informatics, School of Informatics and Computing, Indiana University Purdue University, 719 Indiana Ave Ste 319, Walker Plaza Building, Indianapolis, IA 46202 USA 2 Centre for Computational Biology and Bioinformatics, Indiana University School of Medicine, 5021 Health Information and Translational Sciences (HITS), 410 West 10th Street, Indianapolis, IA 46202 USA 3 Department of Medical and Molecular Genetics, Indiana University School of Medicine, Medical Research and Library Building, 975 West Walnut Street, Indianapolis, IA 46202 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.507© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Improved DNA sequencing methods have transformed the field of genomics over the last decade. This has become possible due to the development of inexpensive short read sequencing technologies which have now resulted in three generations of sequencing platforms. More recently, a new fourth generation of Nanopore based single molecule sequencing technology, was developed based on MinION® sequencer which is portable, inexpensive and fast. It is capable of generating reads of length greater than 100 kb. Though it has many specific advantages, the two major limitations of the MinION reads are high error rates and the need for the development of downstream pipelines. The algorithms for error correction have already emerged, while development of pipelines is still at nascent stage. Results In this study, we benchmarked available assembler algorithms to find an appropriate framework that can efficiently assemble Nanopore sequenced reads. To address this, we employed genome-scale Nanopore sequenced datasets available for E. coli and yeast genomes respectively. In order to comprehensively evaluate multiple algorithmic frameworks, we included assemblers based on de Bruijn graphs (Velvet and ABySS), Overlap Layout Consensus (OLC) (Celera) and Greedy extension (SSAKE) approaches. We analyzed the quality, accuracy of the assemblies as well as the computational performance of each of the assemblers included in our benchmark. Our analysis unveiled that OLC-based algorithm, Celera, could generate a high quality assembly with ten times higher N50 & mean contig values as well as one-fifth the number of total number of contigs compared to other tools. Celera was also found to exhibit an average genome coverage of 12 % in E. coli dataset and 70 % in Yeast dataset as well as relatively lesser run times. In contrast, de Bruijn graph based assemblers Velvet and ABySS generated the assemblies of moderate quality, in less time when there is no limitation on the memory allocation, while greedy extension based algorithm SSAKE generated an assembly of very poor quality but with genome coverage of 90 % on yeast dataset. Conclusion OLC can be considered as a favorable algorithmic framework for the development of assembler tools for Nanopore-based data, followed by de Bruijn based algorithms as they consume relatively less or similar run times as OLC-based algorithms for generating assembly, irrespective of the memory allocated for the task. However, few improvements must be made to the existing de Bruijn implementations in order to generate an assembly with reasonable quality. Our findings should help in stimulating the development of novel assemblers for handling Nanopore sequence data. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2895-8) contains supplementary material, which is available to authorized users. Keywords ContigsDe novo assemblyDe BruijnGreedy Extension graphMinION®NanoporeN50Oxford NanoporeThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background In recent years, next generation sequencing technologies have been evolving rapidly with the potential to accelerate the research in sequencing biology [1–3]. However, today’s next generation sequencing technologies such as Illumina, 454 Roche, Ion Torrent, SMRT (single –molecule real time sequencing) from Pacific biosciences, have various significant limitations [4] especially amplification biases, short read lengths and genome assembly complexities. For example, Illumina – one of the most commonly used technologies for sequencing in recent years, produces read length of 75–100 base pairs (bp) [5] and hence is believed to suffer from short read lengths resulting in poor assembly of complex regions like long repeats and duplications [4, 6, 7]. However, the application of the SMRT platform to small microbial as well as complex eukaryotic genomes have improved the quality of genome assembly but the commercial availability and price of sequencing are the major limitations of this approach [8–10]. Similar improvements were also accomplished by the Illumina Truseq synthetic long-read sequencing strategy [11, 12], but the long range polymerase chain reaction step included in the library preparation will be a limitation in time-constrained projects, thus making it inaccessible to the whole research community. To overcome such limitations efforts are now been made to develop an inexpensive single-molecule Nanopore-based fourth generation DNA sequencing technology [13–17]. In fact, the concept of single molecule sequencing using biological Nanopores was first proposed by Deamer & Akeson [6, 14, 18] in the year 1996, since then intense efforts have been made to overcome the formidable technical challenges and finally in the year 2014, Oxford Nanopore Technologies Ltd released the first commercial Nanopore sequencer [19, 20] to early access customers. The MinION® device, which is no larger than a typical smartphone consists of pores embedded into a membrane which is placed over an electric grid, as the DNA bases i.e. A(adenine), T(thymine), G(guanine) & C(cytosine) pass through the pores they generate a particular intensity of ionic current, which are further base called using metricorr software [21, 22]. The reads generated by this sequencer, can be classified into three types: 2D reads, template reads and complement reads [23]. In our study, we analyzed all three types of reads but mainly focussed on 2D reads since they are optimal reads that consist of consensus information of both the strands [22]. However, similar results were observed upon analyzing all three types of reads, illustrating the reproducibility of our results irrespective of the type of reads analyzed. Despite the high error content of the MinION reads [20, 24], Aston et al. [25] have demonstrated the utility of these reads in microbial sequencing, which incited the need for the development of new tools either to correct the erroneous reads or for the downstream analysis. The error correcting algorithms have already emerged [24, 26] while, development of downstream pipelines is at nascent stage. A major computational step in any of the DNA sequencing pipelines is assembly and can be defined as a hierarchical data structure that maps the sequence data for the reconstruction of the target genome. This process involves initially grouping the reads into contigs and then contigs into scaffolds thereby generating the assembly. Currently, the most common algorithmic frameworks on which assembly algorithms are developed include the Overlap Layout Consensus (OLC) [27], de Bruijn Graph (DBG) [28] which uses some form of k-mer graph method and greedy extension graphs which use either OLC or DBG [29]. There are about 24 academically available de novo assemblers [29] which have been developed by implementing one of these three assembler algorithms. Most of the assembler algorithms, generally take a file of sequence reads and a quality-score file as input, but for Nanopore data, the quality scores are not available so we failed to test assemblers which insist on the requirement of the quality score file as a compulsory input. An example of one such assembler is PCAP, which although is specifically developed for long read data does not accept reads without quality score information [30]. On the other hand, most of the assemblers such as Newbler failed to assemble Nanopore reads due to the length of the reads. Due to these constraints we finally employed in our study one or two assemblers for each type of assembly algorithm and analyzed the quality, accuracy and efficiency of each assembler on whole genome Nanopore sequencing data for E. coli and yeast. Our study unveiled OLC as the optimal algorithm, in multiple contexts benchmarked in this study, providing a direction for further development of assembly tools for Nanopore data. Methods Data retrieval Through an early access program of Nanopore sequencer (MAP), Quick et al. [23] sequenced the genome of the model organism, Escherichia coli K12 substr. Initially, we have used this dataset to benchmark various assembler algorithms but due to high error rate all the existing assemblers failed to assemble such long erroneous reads (5000 – 50,000 bp) (~35 % error) [24]. Later, in January 2015, Schatz and co-workers, developed a novel hybrid algorithm called Ncorr to correct these erroneous reads. By implementing this Ncorr algorithm [24] they have error corrected the reads of the E. coli dataset sequenced by Quick et al. [23] and the reads of the yeast dataset sequenced by Goodwin et al. [24]. These error corrected datasets have been retrieved from the Schatzlab website [24] in FASTA format for all subsequent analysis. Composition of the nanopore sequencing datasets used in this study The E. coli dataset consisted of 1,8842 2D reads (when two strands are read accurately and consensus is built from them) as well as 25,432 template and 11,130 complement reads (which cannot be converted to 2D reads) while, the yeast dataset consisted of 28,258 2D reads, 56,046 and 20,506 template and complement reads respectively. Assembler algorithms employed in this study Velvet Velvet (developed in C) is a secure and reliable de Bruijn graph-based assembler. It extensively uses graph simplification strategy to scale down non-intersecting paths into single nodes. This simplification compresses the graph without much loss of information. To reduce the time-complexity of the algorithm, Velvet implements bubble search (to narrow down the candidate bubbles) and read threading (removal of paths that represent fewer reads than the threshold) [29, 31]. ABySS ABySS (developed in C++) is a de Bruijn graph based assembler mainly developed to address the memory issues while assembling mammalian–size genome. ABySS implements a partition approach at the level of individual graph nodes (for efficiency each graph node is processed separately as each node is individually assigned to a CPU). To overcome the memory requirements, the assignment of graph node to CPU is attained by converting K-mer to an integer using strand-neutral formula i.e. k-mer and its reverse complement map to same integer. ABySS also implements graph simplification like Velvet and then performs bubble smoothening by bounded search where priority is given to the path supported by more reads [32]. Celera Celera is an Overlap Layout Consensus (OLC)-based assembler, which was developed at the time of Sanger sequencing by Celera Genomics. In recent years, the algorithm has been modified to handle long Pac Bio reads whose nature is similar to nanopore reads. The revised pipeline is named as CABOG (Celera assembler with best overlap graph). CABOG constructs an overlay graph from the reads and reports the best overlaps, which are then used to build unitigs. These unitigs are joined to build contigs and finally these contigs are connected to form scaffolds [27]. SSAKE SSAKE is a greedy graph-based assembler. It does not use the graph explicitly. Instead, it iteratively searches for reads with overlap to build the contigs. Initially, it will look for reads with end –to –end confirmation by favoring error-free reads and then performs the extension [29]. Binning of reads In order to test the performance of various metrics in relation to the size of the datasets, we have divided the total reads in a dataset into four bins i.e. 25 %, 50 %, 75 % and 100 % of the reads. To avoid the prejudice in selecting the reads, the binning of the data was performed by randomly generating the bins of the reads ten times using python script, and finally the average result of all the ten trials after processing the each trial is reported in the figures. Implementation of the benchmarking pipeline According to our survey there are at least 24 de novo assemblers which can be accessed with a free academic license [33], that have been developed by implementing one among the following three algorithms namely de Bruijn graphs, Overlap Layout Consensus (OLC) and greedy extension. Out of these only few assemblers i.e. Velvet [24] (de Bruijn graph), Abyss (de Bruijn graph) [25], Celera (OLC method) [34] and SSAKE (greedy extension based method) [29] could be successfully run for assembling the nanopore reads. All the other assemblers failed to assemble most likely due to the length of the reads and/or due to their expectation for quality scores or other input parameters not available for nanopore reads. It is important to note that most of the assemblers were developed in view of short read sequencing data whose read lengths range from 500-3000 bp [5] whereas the read length of Nanopore reads range from 5000–50,000 bp. All the assemblers in this study were run on UNIX command line with default parameters (to evaluate true potential of each tool), and the obtained results were analyzed for reliability, quality and accuracy of assemblies. To evaluate and compare the efficiency of various assembly algorithms the following metrics were employed [33]: Calculation of assembly metrics The contig files which were generated as a result of successful assembly by each assembler were used for statistical analysis of an assembly. The assembly metrics i.e. N50 value,which represents 50 % content of the assembly and all the contig metrics including the mean, total length of all generated contigs as well as the number of contigs obtained, were calculated using a perl script. Calculation of performance metrics Running times and memory consumed by each assembler was captured during the assembly process using UNIX utilities. While limiting the memory usage of each assembler was accomplished using the Ulimit program. Calculation of accuracy metrics To assess the accuracy and quality of the generated assemblies, genome coverage i.e. defined as the percentage of the genome covered when the generated contigs are mapped onto reference genome and the percentage of alignment i.e. the number of contigs mapped to the genome out of the total generated contigs, were computed using shell scripts while mapping to the reference genome was performed using a fast gapped aligner tool Bowtie [24]. Results and discussion Pipeline implemented for the analysis Our analysis pipeline shown in Fig. 1, illustrates the step wise protocol followed for benchmarking the various assembler algorithms for Nanopore sequencing data (see Materials and Methods). Initially, we have retrieved the non-error corrected and Ncorr-error corrected [24] datasets to perform preliminary analysis with all the available assemblers, which helped us to identify few assemblers that can potentially assemble Nanopore sequenced reads. After this initial analysis to identify potential assemblers, we analyzed the efficiency of these assemblers as well as the accuracy and quality of the generated assemblies using various metrics on the error-corrected reads. Our benchmarking analysis enabled us to unveil the ideal algorithmic frameworks for addressing the various needs in the assembly of Nanopore sequencing data.Fig. 1 Illustrates the pipeline implemented in this study for benchmarking various assembler algorithms on Nanopore sequenced datasets Comparison of the assembly metrics generated by various assemblers reveals Celera as an optimal assembler The main features that can best explain the quality of an assembly from sequencing reads include the N50 value, number of contigs, mean length of contigs and the total sum of the lengths of all the contigs identified in an assembly. Hence, we have calculated all of these metrics for the Nanopore sequencing reads for the E. coli and Yeast genomes to understand the relative performance of the assemblers (see Materials and Methods). In the following sections, we summarize these comparisons. 1) N50 value: Upon analyzing the 2D reads from E.coli dataset (see Fig. 2a) we observed consistent increase in the N50 values of the assemblies generated by various assemblers with an increase in the dataset size measured as the percentage of the total number of reads employed in the analysis. In particular, assembly generated by Celera had the highest N50 value ranging from 20,000 bp to 80,000 bp (as we move from 25 % to 100 % of the total reads) while, the assemblies generated by Velvet, ABySS and SSAKE consisted of an average N50 value of 10,000 bp (approximately), which is eight times lower than N50 value of Celera generated assembly for E. coli (Fig. 2a). The differences were found to be even more striking when the results were compared between the assemblers for the yeast genome (Fig. 2b). Since, N50 value represents the 50 % content of the assembly, higher the N50 value better would be the quality of an assembly. Hence, from the above observations based on N50 values, it can be concluded that Celera assembler is likely to generate better assembly compared to the other assemblers studied here. It is possible to speculate that since OLC-based algorithms like the Celera assembler have traditionally been used for longer read technologies like sanger sequencing and probably due to recent modifications made to this specific implementation to make it compatible with even longer reads like PacBio reads of length 3000-15000 bp, they are likely to outperform in terms of assembly quality, most short-read assembler implementations for nanopore sequencing data. This is especially likely to be true if the number of allowed mis-matches for building the contigs can be increased - due to high error rates in the ends of the nanopore reads. While, SSAKE generated assembly is of poor quality with an N50 value of approximately 100 for the yeast genome, which is 100 times less than N50 value of Celera generated assembly (Fig. 2b). A similar trend was observed when template (see Additional file 1) and complement reads (see Additional file 2) of the E. coli dataset were analyzed, confirming the reproducibility of the results. Notably, when 2D reads of Yeast dataset were analyzed, we observed that the N50 value of an assembly generated by Celera for 50 % of the reads is much higher than for the whole dataset. Even though the binning of the reads was performed by randomly generating the bins of reads ten times, it is possible to associate this variation due to selection bias or genome-specific variations as this trend was seen only for yeast and not E. coli. However, it is still evident that the N50 value of Celera generated assembly is much higher than the N50 values of the assemblies generated by any other assembler (see Fig. 2b). We found very similar trends for 1D reads namely template and complement reads, further confirming the reproducibility of the results (see Additional files 1 and 2).Fig. 2 Each pair of plots give an overview of the comparisons of the quality of the assemblies across assemblers for E. coli and yeast datasets. a&b: Histograms with error bars plotted between % of 2D reads and N50_value of an assembly show the variation in N50 value of an assembly among different assembler algorithms and how it varies with respect to the data size. c&d: Histograms with error bars plotted between % of 2D reads and number of contigs generated from an assembly, shows how the number of contigs generated vary with respect to the mean contig length for each respective assembler algorithm across various bins of respective datasets. e&f: Histograms showing the percentage of 2D reads employed on X-axis versus the average length of the contigs obtained using each algorithm. g&h: Histograms showing the sum of the lengths of all the contigs generated by an assembler as a function of the percentage of the total reads employed in the assembly. In each set of plots, left panel corresponds to E. coli dataset while the plots in the right panel correspond to the Yeast dataset. In all the plots labeled numeric values on histograms indicate corresponding values of the metric in respective color representing each tool 2) Number of contigs: We observed, that the number of contigs and mean contig length of an assembly are inversely proportional (Fig. 2c–f). Ideally, a good assembler should generate less number of contigs with a high mean and N50 values. We found this generally held true for assemblies generated by Celera compared to the other assemblers studied here. For instance, Velvet followed by ABySS were found to consistently show high number of contigs compared to other assemblers at different percentage of reads employed in the assembly. This was in contrast to the assemblies generated from Celera and SSAKE, which were found to show low number of contigs indicating the possibility of low but more comprehensive assemblies from the latter two (Fig. 2e and f). These results suggest that Velvet and ABySS are likely to produce very fragmented assemblies. We found similar trends for all the three types of reads in both the datasets (see Fig. 2(c, d), Additional files 1 and 2). Since Celera assembler initially constructs overlay graph among reads and reports the best overlaps, which are further used to build untigs, which are joined to generate contigs, it is possible that data from error-corrected long read sequencing technologies like nanopore are likely better assembled using OLC-based methods as the read error-correction methods further improve. Indeed, less number of contigs with longer lengths identified in our analysis by Celera’s assembler further supports this trend. 3) Mean length of the contigs: It is similar to N50 value but the weightage is not given to contigs with longer length while calculating the mean. We found that it followed similar trend in both the E. coli and yeast datasets i.e. Celera assembler generating contigs with high mean values followed by Velvet, ABySS and SSAKE respectively (see Fig. 2e and f)). The analysis of 1D reads revealed the same overall trend but the increasing trend of mean values were not found to be proportional with data size unlike that seen for 2D read data (see Additional files 1 and 2). 4) Total sum of lengths of all contigs: While this metric does not play a specific role in assessing the quality of an assembly mainly when the genomes have several duplicated regions, nevertheless it can provide information which can be useful for downstream analysis and prioritization in the assembly framework. So we compared the total length of the contigs obtained, at varying percentages of sequence data employed, using various assemblers (Fig. 2g and h). Not surprisingly, this analysis revealed that the assemblers which showed high number of contigs also exhibited a high total contig length suggesting that these assemblers are likely to produce too many fragmented and/or repetitive contigs thereby causing erroneous assemblies. Upon analyzing the assembly metrics of the generated assemblies we observe that, irrespective of the data size and its complexity across genomes, OLC based Celera assembler generates better quality assembly than other assemblers. Evaluation of the memory and run time requirements of various assemblers reveals Celera to be the fastest when sufficient memory is provided Major parameters that can be measured to assess the performance of any computational tool or algorithm are memory (virtual and RAM) and time consumed by the tool to complete the assigned task. In this study, we observed that irrespective of size of the dataset, the RAM and virtual memory required for each tool to perform the task is ~26.5 KB and ~1.2 KB respectively. While the time required by each tool to complete the task significantly varies with the size of the dataset and complexity of the genome. For 2D reads of the E. coli dataset, the wall time as well the CPU time consumed by Velvet is the lowest with ~15–30 sec of wall time and 15–30 sec of CPU time followed by Celera with ~90 sec each of wall time and CPU time, ABySS with ~50–100 sec of wall time and ~60–100 sec of CPU time and SSAKE with ~1000–1500 sec of wall time and ~1500 sec of CPU time (see Fig. 3a and c, Additional file 3). Values for run times are log transformed in the plots to facilitate easy comparison across tools and datasets. Across the assemblers, the time taken to run by each tool increased with the increase in the data size. For yeast dataset, the trend was found to be same but the time consumed by each tool was approximately 3 times higher than the time consumed to assemble the E. coli genome, likely due to the differences in the complexity of the genomes and size of the datasets (see Fig. 3b and d, Additional file 4). In addition, we analyzed the performance of the assemblers, by restricting the memory allotment using Ulimit utility on UNIX environment, to study how the run times vary across them when memory allotted is altered between different runs. We observed that the time taken by each tool remains same when more amount of memory is provided except for Celera, for which we found that the run times significantly decreased when more memory is provided and this resulted in a trend with Celera consuming the lowest time followed by Velvet, ABySS and SSAKE (see Fig. 3e–h, Additional files 3 and 4). The analysis of 1D reads further confirmed the reproducibility of these results (see Additional files 5 and 6).Fig. 3 Each pair of plots give an overview of the computational requirements of each assembler for assembling E. coli and Yeast datasets. a&b: Histogram with error bars plotted between % of 2D reads and log values of wall time which represents the actual time consumed by each assembler to execute the task with respect to gradual increase in data size. c&d: Histograms with error bars plotted between % of 2D reads and log values of CPU time which represents amount of time the CPU is actually executing instructions for each assembler with variation in data size. e&f: Histograms with error bars plotted between varying amount of allotted memory on X-axis and log values of the wall time, showing the influence of memory allocation on wall time consumption by various assembler algorithms. g&h: Histograms with error bars plotted between varying amount of memory and log values of the CPU time, illustrating the influence of memory allocation on the CPU time consumed by various assembler algorithms. In each set of these plots, left panel corresponds to E. coli dataset while the plots in the right panel correspond to the Yeast dataset Overall, our performance metric analysis revealed that the time taken by the de Bruijn graph and OLC-based algorithms to generate assembly is low, while the time consumed by greedy-extension algorithms to generate the assembly are likely to be relatively higher for nanopore data. This might be due to the extensive search made by the greedy-extension algorithms to find the end-to-end overlap of the reads while assembling. It is possible that indexing in greedy extension methods might reduce the run times to some extent. On other hand, de Bruijin graph based assemblers take less time as they implement bubble search which narrow down the candidate bubbles and help in speeding up the assembly process. While, Celera implements OLC algorithm which looks for overlap among the reads to join them together. Since, nanopore reads are longer but fewer, it is not only easy to find overlaps but are also likely to exhibit longer overlaps among the reads, which facilitates more accurate construction of the Contigs. Thus, it is possible that OLC-based approaches like Celera will take lesser run time to generate more accurate assemblies with nanopore data. However, in order to improve performance of these methods, it is important to note that error rates in nanopore reads need to be decreased while allowing increased mismatches in the assembly process. Evaluation of the quality of the generated assemblies reveals OLC-based algorithms to be ideal for nanopore data Two specific metrics which can help in assessing the accuracy of an assembly are genome coverage and alignment percentage (see Materials and Methods). Surprisingly, the genome coverage of all the generated assemblies was very low, but comparatively the assembly generated by Celera for the E. coli 2D read data exhibited better genome coverage (12–13 % versus 2 % for all other assemblies) (see Fig. 4a). For the yeast dataset, the percentage of genome coverage for the assemblies generated by ABySS, Celera and Velvet were found to be 80 %, 70 % and 50 % respectively. In contrast, it was found to be only 2 % for the assembly generated by SSAKE (see Fig. 4b). When the percentage of alignment was compared between the assemblers, the contigs generated by Celera and ABySS for the E. coli 2D read data showed 100 % alignment to the reference genome while the alignment percentage of the contigs generated by Velvet and SSAKE was found to be 80 % and 0 % respectively (see Fig. 4c). For Yeast 2D read data the alignment percentage ranged between 60 %–90 %, with contigs generated by ABySS having highest alignment percentage when aligned to the reference genome followed by Celera, Velvet and SSAKE (see Fig 4d). Further evaluation of 1D reads for coverage and alignment showed a similar trend, confirming the reproducibility of these results (see Additional files 7 and 8).Fig. 4 Each pair of plots show the accuracy of the assembly generated by various assembler algorithms for E.coli (Panels A and C) and Yeast (Panels B and D) datasets. a&b: Line graphs plotted between % of 2D reads and the % of genome covered, showing the extent of genome assembled by each assembler algorithm. c&d: Line graphs between the % of 2D reads and % of alignment showing the confidence level of the contigs being assembled by various assembler algorithms Conclusion In this study, we implemented a computational pipeline for the benchmarking of assembler algorithms which revealed several observations which can aid in the development and improvement of frameworks for assembling genomes using nanopore data. In particular, we found that OLC-based assembler Celera generates an assembly with ten times higher N50 value & mean value and five times lower number of contigs. Our analysis also confirmed that OLC-based approaches can result in high genome coverages with 12 % in E. coli and 70 % in Yeast along with moderate alignment percentages of approximately 85 % when compared to other assemblies, indicating a relatively high quality of the assembly compared to other tools studied here. Moreover, Celera was found to exhibit lesser run times when increased memory was provided to perform the task. Thus, Overlap Layout Consensus (OLC) based algorithms would be ideal frameworks for building de novo assemblers for nanopore reads followed by de Bruijn graph based algorithms since assemblies generated by ABySS were found to show high accuracy, moderate quality and reasonable run times and memory requirements. Our results also suggest that improvements in greedy-extension algorithms can be implemented by indexing in order to decrease the run times. Although this step might reduce the run times for greedy extension methods, accuracy and quality of an assembly generated will be potential issues to be addressed for these methods. There are several challenges that currently exist in dealing with the Nanopore sequencing data. These include high error rate of the long reads and lack of automated computational pipelines for error correction, assembly/alignment as well as downstream analysis of the reads. Developing efficient algorithms which can automate the process of error correction and assembly of the reads would pose some potential opportunities in this domain. For instance, an automated pipeline can be developed by implementing HGAP (Hierarchical Genome Assembly Process) algorithm for error correction, which is already proven to be an optimal algorithm for the error correction in the context to PacBio reads. However, the implementation of HGAP algorithm restricts the application of the tool to specific genomes i.e., only those for which short read data is already available in the public domain. Hence, there is a need to develop methods which can correct the reads from single molecule sequencing methods without using short read or reference genome sequences and using such implementation in the assembly and alignment process for downstream analysis. Indeed, we anticipate rapid development of automated computational pipelines to address various aspects of nanopore sequencing data analysis as new datasets spanning multiple species become available to the scientific community in the coming years. Hence, some of the opportunities for computational biologists include:Enhancing the error correcting algorithms which don’t require short read sequencing data or reference genomes. Development of OLC based assembler algorithms which can consider error-rates in the assembly process, since our results confirm the performance of these methods to be significantly better than other algorithms. Developing automated pipelines for pre- processing of the long reads and downstream analysis. Abbreviations Bp, base pair; kbp - kilobasepair; E. coli, Escherichia coli; MAP, MinION® early Access Program; OLC, overlap layout consensus; Yeast, Saccharomyces cerevisiae Additional files Additional file 1: Each pair of plots give an overview of the comparisons of the quality of the assemblies across assemblers for nanopore sequenced template reads from E. coli and yeast datasets. A&B: Histograms with error bars plotted between % of template reads and N50_value of an assembly show the variation in N50 value of an assembly among different assembler algorithms and how it varies with respect to the data size. C&D: Histograms with error bars plotted between % of template reads and number of contigs generated from an assembly showing how the number of contigs generated vary for each respective assembler algorithm across various bins of respective datasets. E&F: Histograms showing the percentage of template reads employed on X-axis versus the average length of the contigs represented as mean of the contigs, obtained using each algorithm. Mean of the Contigs is the average value of the total sum of lengths of all the contigs. G&H: Histograms showing the sum of the lengths of all the contigs generated by an assembler as a function of the percentage of the total reads employed in the assembly. In each set of plots, left panel corresponds to E. coli dataset while the plots in the right panel correspond to the Yeast dataset. In all the plots labeled numeric values on histograms indicate corresponding values of the metric in respective color representing each tool. (PDF 670 kb) Additional file 2: Each pair of plots give an overview of the comparisons of the quality of the assemblies across assemblers for nanopore sequenced complement reads from E. coli and yeast datasets. A&B: Histograms with error bars plotted between % of complement reads and N50_value of an assembly show the variation in N50 value of an assembly among different assembler algorithms and how it varies with respect to the data size. C&D: Histograms with error bars plotted between % of complement reads and number of contigs generated from an assembly showing how the number of contigs generated vary for each respective assembler algorithm across various bins of respective datasets. E&F: Histograms showing the percentage of complement reads employed on X-axis versus the average length of the contigs represented as mean of the contigs, obtained using each algorithm. Mean of the Contigs is the average value of the total sum of lengths of all the contigs. G&H: Histograms showing the sum of the lengths of all the contigs generated by an assembler as a function of the percentage of the total reads employed in the assembly. In each set of plots, left panel corresponds to E. coli dataset while the plots in the right panel correspond to the Yeast dataset. In all the plots labeled numeric values on histograms indicate corresponding values of the metric in respective color representing each tool. (PDF 700 kb) Additional file 3: Overview of the running times for various assemblers. (PDF 236 kb) Additional file 4: Overview of the influence of memory allocation on the running times for various assemblers. (PDF 155 kb) Additional file 5: Each pair of plots give an overview of the computational requirements of each assembler for assembling nanopore sequenced template reads from E. coli and Yeast datasets. A&B: Histogram with error bars plotted between % of template reads and log values of wall time which represents the actual time consumed by each assembler to execute the task with respect to gradual increase in data size. C&D: Histograms with error bars plotted between % of template reads and log values of CPU time which represents amount of time the CPU is actually executing instructions for each assembler with variation in data size. E&F: Histograms with error bars plotted between varying amount of allotted memory on X-axis and log values of the wall time, showing the influence of memory allocation on wall time consumption by various assembler algorithms. G&H: Histograms with error bars plotted between varying amount of memory and log values of the CPU time, illustrating the influence of memory allocation on the CPU time consumed by various assembler algorithms. In each set of these plots, left panel corresponds to E. coli dataset while the plots in the right panel correspond to the Yeast dataset. (PDF 813 kb) Additional file 6: Each pair of plots give an overview of the computational requirements of each assembler for assembling nanopore sequenced complement reads from E. coli and Yeast datasets. A&B: Histogram with error bars plotted between % of complement reads and log values of wall time which represents the actual time consumed by each assembler to execute the task with respect to gradual increase in data size. C&D: Histograms with error bars plotted between % of complement reads and log values of CPU time which represents amount of time the CPU is actually executing instructions for each assembler with variation in data size. E&F: Histograms with error bars plotted between varying amount of allotted memory on X-axis and log values of the wall time, showing the influence of memory allocation on wall time consumption by various assembler algorithms. G&H: Histograms with error bars plotted between varying amount of memory and log values of the CPU time, illustrating the influence of memory allocation on the CPU time consumed by various assembler algorithms. In each set of these plots, left panel corresponds to E. coli dataset while the plots in the right panel correspond to the Yeast dataset. (PDF 838 kb) Additional file 7: Each pair of plots show the accuracy of the assembly generated by various assembler algorithms for nanopore sequenced template reads from E.coli (Panels A and C) and Yeast (Panels B and D) datasets. A&B: Line graphs plotted between % of template reads and the % of genome covered, showing the extent of genome assembled by each assembler algorithm. C&D: Line graphs between the % of template reads and % of alignment showing the confidence level of the contigs being assembled by various assembler algorithms. (PDF 584 kb) Additional file 8 Each pair of plots show the accuracy of the assembly generated by various assembler algorithms for nanopore sequenced complement reads from E.coli (Panels A and C) and Yeast (Panels B and D) datasets. A&B: Line graphs plotted between % of complement reads and the % of genome covered, showing the extent of genome assembled by each assembler algorithm. C&D: Line graphs between the % of complement reads and % of alignment showing the confidence level of the contigs being assembled by various assembler algorithms. (PDF 798 kb) Acknowledgements This work is supported by the School of Informatics and Computing at IUPUI in the form of startup funds. The authors are grateful to Mr. Vishal Kumar Sarsani for proposing the initial idea for the study and for helpful discussions and to Mr. Rajneesh Srivastava for his constructive critiques and comments on a previous version of the manuscript. The authors acknowledge all the members of the Janga Lab for their feedback during the course of the project. Declarations Publication charges for this article have been funded by support from open access funds from IUPUI Library. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials All datasets on which the conclusions of the manuscript rely, are available in the publicly accessible repositories. Also, additional files, which may be needed to reproduce the results presented in the manuscript, are made available as supplementary material [35]. Authors’ contributions SCJ and YC conceived study. SCJ supervised the study. YC performed the analysis. SCJ and YC wrote the manuscript. Both the authors read and approved the manuscript. Competing interests The authors declared that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Rhee M Burns MA Nanopore sequencing technology: research trends and applications Trends Biotechnol 2006 24 12 580 6 10.1016/j.tibtech.2006.10.005 17055093 2. Ku CS Roukos DH From next-generation sequencing to nanopore sequencing technology: paving the way to personalized genomic medicine Expert Rev Med Devices 2013 10 1 1 6 10.1586/erd.12.63 23278216 3. Metzker ML Sequencing technologies - the next generation Nat Rev Genet 2010 11 1 31 46 10.1038/nrg2626 19997069 4. Alkan C Sajjadian S Eichler EE Limitations of next-generation genome sequence assembly Nat Methods 2011 8 1 61 5 10.1038/nmeth.1527 21102452 5. Quail MA Smith M Coupland P Otto TD Harris SR Connor TR Bertoni A Swerdlow HP Gu Y A tale of three next generation sequencing platforms: comparison of Ion Torrent, Pacific Biosciences and Illumina MiSeq sequencers BMC Genomics 2012 13 341 10.1186/1471-2164-13-341 22827831 6. Bayley H Nanotechnology: Holes with an edge Nature 2010 467 7312 164 5 10.1038/467164a 20829786 7. Onishi-Seebacher M Korbel JO Challenges in studying genomic structural variant formation mechanisms: the short-read dilemma and beyond Bioessays 2011 33 11 840 50 10.1002/bies.201100075 21959584 8. Koren S Schatz MC Walenz BP Martin J Howard JT Ganapathy G Wang Z Rasko DA McCombie WR Jarvis ED Hybrid error correction and de novo assembly of single-molecule sequencing reads Nat Biotechnol 2012 30 7 693 700 10.1038/nbt.2280 22750884 9. Chin CS Alexander DH Marks P Klammer AA Drake J Heiner C Clum A Copeland A Huddleston J Eichler EE Nonhybrid, finished microbial genome assemblies from long-read SMRT sequencing data Nat Med 2013 10 6 563 9 10. Chaisson MJ Huddleston J Dennis MY Sudmant PH Malig M Hormozdiari F Antonacci F Surti U Sandstrom R Boitano M Resolving the complexity of the human genome using single-molecule sequencing Nature 2015 517 7536 608 11 10.1038/nature13907 25383537 11. Kuleshov V Xie D Chen R Pushkarev D Ma Z Blauwkamp T Kertesz M Snyder M Whole-genome haplotyping using long reads and statistical methods Nat Biotechnol 2014 32 3 261 6 10.1038/nbt.2833 24561555 12. McCoy RC Taylor RW Blauwkamp TA Kelley JL Kertesz M Pushkarev D Petrov DA Fiston-Lavier AS Illumina TruSeq synthetic long-reads empower de novo assembly and resolve complex, highly-repetitive transposable elements PLoS One 2014 9 9 10.1371/journal.pone.0106689 25188499 13. Gaylin W Some problems never go away Ann N Y Acad Sci 1988 530 163 6 10.1111/j.1749-6632.1988.tb35310.x 3408072 14. Deamer DW Akeson M Nanopores and nucleic acids: prospects for ultrarapid sequencing Trends Biotechnol 2000 18 4 147 51 10.1016/S0167-7799(00)01426-8 10740260 15. Gupta PK Single-molecule DNA, sequencing technologies for future genomics research Trends Biotechnol 2008 26 11 602 11 10.1016/j.tibtech.2008.07.003 18722683 16. Niedringhaus TP Milanova D Kerby MB Snyder MP Barron AE Landscape of next-generation sequencing technologies Anal Chem 2011 83 12 4327 41 10.1021/ac2010857 21612267 17. von Bubnoff A Next-generation sequencing: the race is on Cell 2008 132 5 721 3 10.1016/j.cell.2008.02.028 18329356 18. Kasianowicz JJ Brandin E Branton D Deamer DW Characterization of individual polynucleotide molecules using a membrane channel Proc Natl Acad Sci U S A 1996 93 24 13770 3 10.1073/pnas.93.24.13770 8943010 19. Eisenstein M Oxford Nanopore announcement sets sequencing sector abuzz Nat Biotechnol 2012 30 4 295 6 10.1038/nbt0412-295 22491260 20. Mikheyev AS Tin MM A first look at the Oxford Nanopore MinION sequencer Mol Ecology Resour 2014 14 6 1097 102 10.1111/1755-0998.12324 21. Stoddart D Heron AJ Mikhailova E Maglia G Bayley H Single-nucleotide discrimination in immobilized DNA oligonucleotides with a biological nanopore Proc Natl Acad Sci U S A 2009 106 19 7702 7 10.1073/pnas.0901054106 19380741 22. Clarke J Wu HC Jayasinghe L Patel A Reid S Bayley H Continuous base identification for single-molecule nanopore DNA sequencing Nat Nanotechnol 2009 4 4 265 70 10.1038/nnano.2009.12 19350039 23. Quick J Quinlan AR Loman NJ A reference bacterial genome dataset generated on the MinION portable single-molecule nanopore sequencer GigaSci 2014 3 22 10.1186/2047-217X-3-22 24. Goodwin S, Gurtowski J, Ethe-Sayers S, Deshpande P, Schatz MC, McCombie WR. Oxford Nanopore sequencing, hybrid error correction, and de novo assembly of a eukaryotic genome. Genome Res. 2015;25(11):1750–56. 25. Ashton PM Nair S Dallman T Rubino S Rabsch W Mwaigwisya S Wain J O'Grady J MinION nanopore sequencing identifies the position and structure of a bacterial antibiotic resistance island Nat Biotechnol 2015 33 3 296 300 10.1038/nbt.3103 25485618 26. Madoui MA Engelen S Cruaud C Belser C Bertrand L Alberti A Lemainque A Wincker P Aury JM Genome assembly using Nanopore-guided long and error-free DNA reads BMC Genomics 2015 16 327 10.1186/s12864-015-1519-z 25927464 27. Myers EW Toward simplifying and accurately formulating fragment assembly J Comput Biol 1995 2 2 275 90 10.1089/cmb.1995.2.275 7497129 28. Idury RM Waterman MS A new algorithm for DNA sequence assembly J Comput Biol 1995 2 2 291 306 10.1089/cmb.1995.2.291 7497130 29. Miller JR Koren S Sutton G Assembly algorithms for next-generation sequencing data Genomics 2010 95 6 315 27 10.1016/j.ygeno.2010.03.001 20211242 30. Huang X Wang J Aluru S Yang SP Hillier L PCAP: a whole-genome assembly program Genome Res 2003 13 9 2164 70 10.1101/gr.1390403 12952883 31. Zerbino DR Birney E Velvet: algorithms for de novo short read assembly using de Bruijn graphs Genome Res 2008 18 5 821 9 10.1101/gr.074492.107 18349386 32. Simpson JT Wong K Jackman SD Schein JE Jones SJ Birol I ABySS: a parallel assembler for short read sequence data Genome Res 2009 19 6 1117 23 10.1101/gr.089532.108 19251739 33. Zhang W Chen J Yang Y Tang Y Shang J Shen B A practical comparison of de novo genome assembly software tools for next-generation sequencing technologies PLoS One 2011 6 3 e17915 10.1371/journal.pone.0017915 21423806 34. Myers EW Sutton GG Delcher AL Dew IM Fasulo DP Flanigan MJ Kravitz SA Mobarry CM Reinert KH Remington KA A whole-genome assembly of Drosophila Science 2000 287 5461 2196 204 10.1126/science.287.5461.2196 10731133 35. http://www.iupui.edu/~sysbio/nanopore

PMC005xxxxxx/PMC5001212.txt

==== Front BMC NeurosciBMC NeurosciBMC Neuroscience1471-2202BioMed Central London 2753439328310.1186/s12868-016-0283-6Meeting Abstracts25th Annual Computational Neuroscience Meeting: CNS-2016 Sharpee Tatyana O. sharpee@snl.salk.edu 1Destexhe Alain destexhe@unic.cnrs-gif.fr 23Kawato Mitsuo kawato@hip.atr.co.jp 4Sekulić Vladislav vlad.sekulic@utoronto.ca 56Skinner Frances K. 567179180182Wójcik Daniel K. d.wojcik@nencki.gov.pl 8Chintaluri Chaitanya 8Cserpán Dorottya 9Somogyvári Zoltán 9Kim Jae Kyoung jaekkim@kaist.ac.kr 10Kilpatrick Zachary P. 11Bennett Matthew R. 12Josić Kresimir 1113Elices Irene irene.elices@uam.es 14Arroyo David 14Levi Rafael 1415Rodriguez Francisco B. 14Varona Pablo pablo.varona@uam.es 14Hwang Eunjin 16414418Kim Bowon 1617415Han Hio-Been 1618Kim Tae 19McKenna James T. 20Brown Ritchie E. 20McCarley Robert W. 20Choi Jee Hyun jeechoi@kist.re.kr 1617412414416417420Rankin James james.rankin@nyu.edu 21421Popp Pamela Osborn 21Rinzel John 2122Tabas Alejandro atabas@bournemouth.ac.uk 23Rupp André 2425Balaguer-Ballester Emili 2326Maturana Matias I. 2728Grayden David B. 2829249335423424Cloherty Shaun L. 30Kameneva Tatiana tkam@unimelb.edu.au 28Ibbotson Michael R. 2731Meffin Hamish hmeffin@unimelb.edu.au 2731252253334Koren Veronika veronika.koren@bccn-berlin.de 3233Lochmann Timm 3233Dragoi Valentin 34Obermayer Klaus 3233Psarrou Maria m.psarrou@herts.ac.uk 35Schilstra Maria 35Davey Neil 35Torben-Nielsen Benjamin 35Steuber Volker 35Ju Huiwen 36Yu Jiao 37Hines Michael L. 38Chen Liang 39Yu Yuguo yuyuguo@fudan.edu.cn 36Kim Jimin 40Leahy Will 41Shlizerman Eli shlizee@uw.edu 4042Birgiolas Justas justas@asu.edu 43Gerkin Richard C. 43Crook Sharon M. 244Viriyopase Atthaphon a.viriyopase@science.ru.nl 454647Memmesheimer Raoul-Martin 454748Gielen Stan 4546Dabaghian Yuri dabaghian@rice.edu 4950DeVito Justin 49Perotti Luca 51Kim Anmo J. anmo.kim@gmail.com 52Fenk Lisa M. 52Cheng Cheng 51Maimon Gaby 52Zhao Chang zhao@pyl.unibe.ch 53Widmer Yves 54Sprecher Simon 54Senn Walter 43Halnes Geir geir.halnes@nmbu.no 55Mäki-Marttunen Tuomo tuomomm@uio.no 56361Keller Daniel 57Pettersen Klas H. 5859Andreassen Ole A. 56361Einevoll Gaute T. 5560365Yamada Yasunori ysnr@jp.ibm.com 61Steyn-Ross Moira L. msr@waikato.ac.nz 62Alistair Steyn-Ross D. 62Mejias Jorge F. jorge.f.mejias@gmail.com 21Murray John D. 63Kennedy Henry 64Wang Xiao-Jing 2165Kruscha Alexandra alexandra.kruscha@bccn-berlin.de 6667Grewe Jan 6869Benda Jan 6869Lindner Benjamin 6667169Badel Laurent laurent@brain.riken.jp 70Ohta Kazumi 70Tsuchimoto Yoshiko 70Kazama Hokto 70Kahng B. bkahng@snu.ac.kr 71Tam Nicoladie D. nicoladie.tam@unt.edu 72Pollonini Luca 73Zouridakis George 74Soh Jaehyun 75Kim DaeEun daeeun@yonsei.ac.kr 75Yoo Minsu minsu@uchicago.edu 76Palmer S. E. 77Culmone Viviana v.culmone@pgr.reading.ac.uk 78Bojak Ingo 78Ferrario Andrea andrea.ferrario@plymouth.ac.uk 79Merrison-Hort Robert 79Borisyuk Roman 79Kim Chang Sub cskim@jnu.ac.kr 80Tezuka Taro tezuka@slis.tsukuba.ac.jp 81Joo Pangyu pangyu32@postech.ac.kr 82Rho Young-Ah yarho75@gmail.com 8384Burton Shawn D. 8586Bard Ermentrout G. 8386Jeong Jaeseung jsjeong@kaist.ac.kr 84163164190313314316Urban Nathaniel N. 8586Marsalek Petr petr.marsalek@lf1.cuni.cz 8788Kim Hoon-Hee 84Moon Seok-hyun 89Lee Do-won 89Lee Sung-beom 89Lee Ji-yong 89Molkov Yaroslav I. ymolkov@gsu.edu 90Hamade Khaldoun 91Teka Wondimu 92Barnett William H. 90Kim Taegyo 91Markin Sergey 91Rybak Ilya A. 91Forro Csaba forro@biomed.ee.ethz.ch 93Dermutz Harald 93Demkó László 93Vörös János 93Babichev Andrey 4950Huang Haiping physhuang@gmail.com 94Verduzco-Flores Sergio sergio.verduzco@oist.jp 95Dos Santos Filipa f.d.s.brandao@keele.ac.uk 96Andras Peter 96Metzner Christoph c.metzner@herts.ac.uk 97364Schweikard Achim 98Zurowski Bartosz 99Roach James P. roachjp@umich.edu 100Sander Leonard M. 101102Zochowski Michal R. 101102Skilling Quinton M. 103Ognjanovski Nicolette 104Aton Sara J. 104Zochowski Michal michalz@umich.edu 103105Wang Sheng-Jun 106107Ouyang Guang 107Guang Jing 108Zhang Mingsha 108Michael Wong K. Y. 109Zhou Changsong cszhou@hkbu.edu.hk 107110111Robinson Peter A. 12112113206Sanz-Leon Paula paula.sanz-leon@sydney.edu.au 112113206Drysdale Peter M. 112113Fung Felix 112113Abeysuriya Romesh G. 112Rennie Chris J. 112113Zhao Xuelong 112113Choe Yoonsuck choe@tamu.edu 114Yang Huei-Fang 115Mi Yuanyuan 116188Lin Xiaohan 116Wu Si wusi@bnu.edu.cn 116188Liedtke Joscha joscha@nld.ds.mpg.de 117118Schottdorf Manuel manuel@nld.ds.mpg.de 117118Wolf Fred 117118Yamamura Yoriko yoriko@oist.jp 119Wickens Jeffery R. 119Rumbell Timothy 120Ramsey Julia 121Reyes Amy 121Draguljić Danel 121Hof Patrick R. 122Luebke Jennifer 123Weaver Christina M. christina.weaver@fandm.edu 111He Hu 124Yang Xu yangxu@tsinghua.edu.cn 125Ma Hailin 124Xu Zhiheng 124Wang Yuzhe 124Baek Kwangyeol kb567@cam.ac.uk 126127Morris Laurel S. 126Kundu Prantik 128Voon Valerie 126Agnes Everton J. everton.agnes@cncb.ox.ac.uk 129Vogels Tim P. 129Podlaski William F. william.podlaski@cncb.ox.ac.uk 130Giese Martin martin.giese@uni-tuebingen.de 131344Kuravi Pradeep 132Vogels Rufin 131Seeholzer Alexander alex.seeholzer@epfl.ch 133Podlaski William 134Ranjan Rajnish 135Vogels Tim 133Torres Joaquin J. 136Baroni Fabiano 137Latorre Roberto 14Gips Bart bart.gips@donders.ru.nl 45Lowet Eric 45138Roberts Mark J. 45138de Weerd Peter 138Jensen Ole 45van der Eerden Jan 45Goodarzinick Abdorreza a.goodarzinick@iasbs.ac.ir 139Niry Mohammad D. 139140Valizadeh Alireza 139141142143Pariz Aref a.pariz@iasbs.ac.ir 142Parsi Shervin S. 142Warburton Julia M. julia.warburton@bristol.ac.uk 144Marucci Lucia 145Tamagnini Francesco 146147Brown Jon 148149Tsaneva-Atanasova Krasimira 150Kleberg Florence I. kleberg@fias.uni-frankfurt.de 151Triesch Jochen 151154Moezzi Bahar bahar.moezzi@unisa.edu.au 152Iannella Nicolangelo 152153Schaworonkow Natalie 154Plogmacher Lukas 154Goldsworthy Mitchell R. 155Hordacre Brenton 155McDonnell Mark D. mark.mcdonnell@unisa.edu.au 152336Ridding Michael C. 155Zapotocky Martin zapotocky@biomed.cas.cz 156157Smit Daniel 156157158Fouquet Coralie 158Trembleau Alain 158Dasgupta Sakyasingha sdasgup@jp.ibm.com 159200401402Nishikawa Isao 161Aihara Kazuyuki 161440Toyoizumi Taro 160Robb Daniel T. robb@roanoke.edu 162Mellen Nick 163Toporikova Natalia 164Tang Rongxiang 165Tang Yi-Yuan yiyuan.tang@ttu.edu 166Liang Guangsheng 166Kiser Seth A. 167Howard James H. Jr.168Goncharenko Julia i.goncharenko@herts.ac.uk 35Voronenko Sergej O. sergej@physik.hu-berlin.de 66169Ahamed Tosif tosif.ahamed@oist.jp 170Stephens Greg 170171Yger Pierre pierre.yger@inserm.com 172Lefebvre Baptiste 172Spampinato Giulia Lia Beatrice 172Esposito Elric 172et Olivier Marre Marcel Stimberg 172Choi Hansol 173Song Min-Ho minho.song@imv.uio.no 174Chung SueYeon schung@fas.harvard.edu 175Lee Dan D. 176Sompolinsky Haim 175177Phillips Ryan S. Ryan.Phillips@nih.gov 178179Smith Jeffrey 178Chatzikalymniou Alexandra Pierri alexandra.chatzikalymniou@mail.utoronto.ca 179180Ferguson Katie 179181Alex Cayco Gajic N. natasha.gajic@ucl.ac.uk 183Clopath Claudia 184410Angus Silver R. 183Gleeson Padraig p.gleeson@ucl.ac.uk 183271Marin Boris 183Sadeh Sadra 183Quintana Adrian 183271Cantarelli Matteo 185Dura-Bernal Salvador salvadordura@gmail.com 186260267Lytton William W. 186260266267Davison Andrew 187Li Luozheng 188Zhang Wenhao 188Wang Dahui 188189Song Youngjo 190Park Sol 190191Choi Ilhwan 191Shin Hee-sup 191Choi Hannah hannahch@uw.edu 40192193Pasupathy Anitha 192193Shea-Brown Eric 40193425427Huh Dongsung huh@salk.edu 194Sejnowski Terrence J. 194195Vogt Simon M. simonsunimail@gmail.com 196Kumar Arvind 197198287291298302Schmidt Robert 196197Van Wert Stephen szv124@psu.edu 199Schiff Steven J. 199200Veale Richard richard@nips.ac.jp 201Scheutz Matthias 202Lee Sang Wan sangwan@kaist.ac.kr 84203204Gallinaro Júlia julia.gallinaro@bcf.uni-freiburg.de 205Rotter Stefan 205Rubchinsky Leonid L. lrubchin@iupui.edu 207208Cheung Chung Ching 207Ratnadurai-Giridharan Shivakeshavan 207Shomali Safura Rashid safura@ipm.ir 209Ahmadabadi Majid Nili 209210Shimazaki Hideaki 211Nader Rasuli S. 212213Zhao Xiaochen 116Rasch Malte J. malte.rasch@bnu.edu.cn 116Wilting Jens jwilting@nld.ds.mpg.de 214Priesemann Viola 118214215217218224Levina Anna anna.levina@ist.ac.at 216Rudelt Lucas l.rudelt@gmail.com 218Lizier Joseph T. 219220Spinney Richard E. 220Rubinov Mikail 221222Wibral Michael 223Bak Ji Hyun jhbak@princeton.edu 225Pillow Jonathan 226Zaho Yuan 228229Park Il Memming memming.park@stonybrook.edu 223227Kang Jiyoung 230Park Hae-Jeong parkhj@yuhs.ac 231Jang Jaeson jaesonjang@kaist.ac.kr 84Paik Se-Bum 84225232233Choi Woochul choiwc1128@kaist.ac.kr 84232233Lee Changju lcj110808@kaist.ac.kr 84Song Min night@kaist.ac.kr 84233Lee Hyeonsu hslee9305@kaist.ac.kr 84Park Youngjin yodamaster@kaist.ac.kr 84Yilmaz Ergin erginyilmaz@yahoo.com 234Baysal Veli 234Ozer Mahmut 235Saska Daniel research@saska.io 236Nowotny Thomas 236237Chan Ho Ka hc338@sussex.ac.uk 237Diamond Alan 237Herrmann Christoph S. 238Murray Micah M. 239Ionta Silvio 239Hutt Axel 240Lefebvre Jérémie jeremie.lefebvre@uhnresearch.com 241Weidel Philipp p.weidel@fz-juelich.de 242Duarte Renato 242243244Morrison Abigail 242243245246292298301Lee Jung H. jungl@alleninstitute.org 247433Iyer Ramakrishnan 247433Mihalas Stefan 247433Koch Christof 247Petrovici Mihai A. mpedro@kip.uni-heidelberg.de 248Leng Luziwei 248Breitwieser Oliver 247Stöckel David 248Bytschok Ilja 248Martel Roman 248Bill Johannes 248Schemmel Johannes 248Meier Karlheinz 248Esler Timothy B. tesler@student.unimelb.edu.au 249Burkitt Anthony N. 2128249Kerr Robert R. 250Tahayori Bahman 251Nolte Max max.nolte@epfl.ch 254Reimann Michael W. 254Muller Eilif 254Markram Henry 254Parziale Antonio anparziale@unisa.it 255256Senatore Rosa 255256257Marcelli Angelo 255256Skiker K. skiker.kaoutar85@gmail.com 258Maouene M. 259Neymotin Samuel A. samn@neurosim.downstate.edu 260261Seidenstein Alexandra 260262Lakatos Peter 263Sanger Terence D. 264265Menzies Rosemary J. 268McLauchlan Campbell 268van Albada Sacha J. 269292Kedziora David J. 268Neymotin Samuel 267Kerr Cliff C. cliff@thekerrlab.com 268Suter Benjamin A. 270Shepherd Gordon M. G. 270Ryu Juhyoung jh67753737@snu.ac.krvisionsl@snu.ac.kr 272Lee Sang-Hun 272273274275Lee Joonwon jwl89@snu.ac.krvisionsl@snu.ac.kr 273Lee Hyang Jung hyangjung.lee@snu.ac.krvisionsl@snu.ac.kr 274Lim Daeseob daeseob@snu.ac.krvisionsl@snu.ac.kr 275Wang Jisung 276Lee Heonsoo beafool@postech.ac.kr 276Jung Nam 277Anh Quang Le 277Maeng Seung Eun 277Lee Tae Ho 277Lee Jae Woo jaewlee@inha.ac.kr 277Park Chang-hyun park.changhyun@gmail.com 278279Ahn Sora 280282Moon Jangsup 278279Choi Yun Seo 279Kim Juhee 280Jun Sang Beom 280281282Lee Seungjun slee@ewha.ac.kr 280282Lee Hyang Woon 278279283Jo Sumin 282Jun Eunji 282Yu Suin 282Goetze Felix afgoetze@gmail.com 284285Lai Pik-Yin 284Kim Seonghyun 286Kwag Jeehyun jkwag@korea.ac.kr 286Jang Hyun Jae 286Filipović Marko marko.filipovic@bcf.uni-freiburg 287288Reig Ramon 289Aertsen Ad 287288Silberberg Gilad 290Bachmann Claudia c.bachmann@fz-juelich.de 292Buttler Simone 292Jacobs Heidi 293294295Dillen Kim 296Fink Gereon R. 296297Kukolja Juraj 296297Kepple Daniel akula@cshl.edu 299Giaffar Hamza 299Rinberg Dima 300Shea Steven 299Koulakov Alex 299Bahuguna Jyotika j.bahuguna@fz-juelich.de 298301302Tetzlaff Tom 301Kotaleski Jeanette Hellgren 302Kunze Tim tkunze@cbs.mpg.de 303304Peterson Andre 305Knösche Thomas 303Kim Minjung 306Kim Hojeong hojeong.kim03@gmail.com 306Park Ji Sung 307Yeon Ji Won 307Kim Sung-Phil spkim@unist.ac.kr 307308Kang Jae-Hwan 308Lee Chungho 308Spiegler Andreas Andreas.Spiegler@univ-amu.fr 309Petkoski Spase 309310Palva Matias J. 311Jirsa Viktor K. 309Saggio Maria L. marisa.saggio@gmail.com 309Siep Silvan F. 309Stacey William C. 312351352Bernar Christophe 309Choung Oh-hyeon iohyeonki@kaist.ac.kr 84Jeong Yong 84Lee Yong-il 84313Kim Su Hyun 84313Jeong Mir 84Lee Jeungmin 84314Kwon Jaehyung 84313Kralik Jerald D. jerald.kralik@raphe.kaist.ac.kr 84314Jahng Jaehwan jahngjh.627@gmail.com 84313Hwang Dong-Uk 315Kwon Jae-Hyung jh2393@kaist.ac.kr 84316Park Sang-Min 84316Kim Seongkyun 84Kim Hyoungkyu 84Kim Pyeong Soo 84Yoon Sangsup 84313Lim Sewoong 84313Park Choongseok issaf14@ecu.edu 317Miller Thomas 317Clements Katie 317Ahn Sungwoo 318Ji Eoon Hye 319Issa Fadi A. cpark@ncat.edu 317Baek JeongHun ku21fang@gmail.com 320Oba Shigeyuki 320Yoshimoto Junichiro 321322Doya Kenji 321Ishii Shin 320Mosqueiro Thiago S. 323Strube-Bloss Martin F. 324Smith Brian brian.h.smith@asu.edu 325Huerta Ramon 323Hadrava Michal hadrava@cs.cas.cz 325326327Hlinka Jaroslav 326Bos Hannah h.bos@fz-juelich.de 292Helias Moritz 292328Welzig Charles M. welzig@mcw.edu 329Harper Zachary J. 329330Kim Won Sup 331Shin In-Seob 331Baek Hyeon-Man 332Han Seung Kee skhan@chungbuk.ac.kr 331Richter René rene.richter@cs.tu-chemnitz.de 331Vitay Julien 331Beuth Frederick 331Hamker Fred H. 331332Toppin Kelly 333Guo Yixin yixin@math.drexel.edu 333Graham Bruce P. 337Kale Penelope J. Penelope.Kale@qimr.edu.au 338Gollo Leonardo L. leonardo.l.gollo@gmail.com 338409431Stern Merav merav.stern@mail.huji.ac.il 339Abbott L. F. 340Fedorov Leonid A. leonid.fedorov@uni-tuebingen.de 341342Giese Martin A. 341342Ardestani Mohammad Hovaidi Mohammad.Hovaidi-Ardestani@uni-tuebingen.de 343344Faraji Mohammad Javad mohammadjavad.faraji@epfl.ch 345Preuschoff Kerstin 346Gerstner Wulfram 345van Gendt Margriet J. m.j.van_gendt@lumc.nl 347Briaire Jeroen J. 347Kalkman Randy K. 347Frijns Johan H. M. 347348Lee Won Hee wonhee.lee@mssm.edu 349Frangou Sophia 349Fulcher Ben D. ben.fulcher@monash.edu 350Tran Patricia H. P. 350Fornito Alex 350Gliske Stephen V. 351Lim Eugene 353Holman Katherine A. 354Fink Christian G. cgfink@owu.edu 353355Kim Jinseop S. jinseop.s.kim@kbri.re.kr 356357Mu Shang 358Briggman Kevin L. 359Sebastian Seung H. 356358the EyeWirers http://eyewire.org Wegener Detlef wegener@brain.uni-bremen.de 360Bohnenkamp Lisa 353361Ernst Udo A. 360Devor Anna 363364Dale Anders M. 362363368Lines Glenn T. 367Edwards Andy 366Tveito Aslak 366Hagen Espen e.hagen@fz-juelich.de 292Senk Johanna 292Diesmann Markus 292369370376377378Schmidt Maximilian max.schmidt@fz-juelich.de 371Bakker Rembrandt 45370Shen Kelly 372Bezgin Gleb 373Hilgetag Claus-Christian 374375van Albada Sacha Jennifer 370Sun Haoqi hsun004@e.ntu.edu.sg 379380381382Sourina Olga 379381Huang Guang-Bin 379381Klanner Felix 381383Denk Cornelia 381Glomb Katharina katharina.glomb@upf.edu 384Ponce-Alvarez Adrián 383Gilson Matthieu matthieu.gilson@upf.edu 383390Ritter Petra 385386387388Deco Gustavo 383389390Witek Maria A. G. 391Clarke Eric F. 392Hansen Mads 393Wallentin Mikkel 394Kringelbach Morten L. 391394395Vuust Peter 391394Klingbeil Guido guido.klingbeil@oist.jp 396De Schutter Erik 396397398399400Chen Weiliang w.chen@oist.jp 397Zang Yunliang yunliang.zang@oist.jp 398Hong Sungho shhong@oist.jp 403Takashima Akira 399Zamora Criseida criseida.chimal@oist.jp 400Gallimore Andrew R. 400Goldschmidt Dennis dennis.goldschmidt@neuro.fchampalimaud.org 400Manoonpong Poramate 401Karoly Philippa J. pkaroly@student.unimelb.edu.au 404407Freestone Dean R. deanrf@unimelb.edu.au 404405Soundry Daniel 405Kuhlmann Levin 406Paninski Liam 405Cook Mark 404Lee Jaejin jaejin@mail.med.upenn.edu 408Fishman Yonatan I. 409Cohen Yale E. 408Roberts James A. james.roberts@qimrberghofer.edu.au 410432Cocchi Luca 410Sweeney Yann y.sweeney@imperial.ac.uk 411Lee Soohyun 412413Jung Woo-Sung 412414Kim Youngsoo 416Jung Younginha 418419Song Yoon-Kyu 419Chavane Frédéric 422Soman Karthik 428Muralidharan Vignesh 428Srinivasa Chakravarthy V. schakra@iitm.ac.in 428Shivkumar Sabyasachi 428Mandali Alekhya 428Pragathi Priyadharsini B. 428Mehta Hima 428Davey Catherine E. cedavey@unimelb.edu.au 424Brinkman Braden A. W. bradenb@uw.edu 40426Kekona Tyler 40Rieke Fred 426427Buice Michael 26De Pittà Maurizio maurizio.depitta@gmail.com 429430435Berry Hugues 430431436Brunel Nicolas 430431Breakspear Michael 432Marsat Gary gary.marsat@mail.wvu.edu 437Drew Jordan 417Chapman Phillip D. 417Daly Kevin C. 417Bradle Samual P. 417Seo Sat Byul satbyul.seo@mavs.uta.edu 438Su Jianzhong 434Kavalali Ege T. 439Blackwell Justin 434Shiau LieJune shiau@uhcl.edu 440Buhry Laure 441Basnayake Kanishka 442Lee Sue-Hyun suelee@kaist.ac.kr 84443Levy Brandon A. 444Baker Chris I. 444446Leleu Timothée timothee@sat.t.u-tokyo.ac.jp 445Philips Ryan T. 427Chhabria Karishma 4271 Computational Neurobiology Laboratory, The Salk Institute for Biological Studies, San Diego, CA USA 2 UNIC, CNRS, Gif sur Yvette, France 3 The European Institute for Theoretical Neuroscience (EITN), Paris, France 4 ATR Computational Neuroscience Laboratories, 2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0288 Japan 5 Krembil Research Institute, University Health Network, Toronto, ON M5T 2S8 Canada 6 Department of Physiology, University of Toronto, Toronto, ON M5S 1A8 Canada 7 Department of Medicine (Neurology), University of Toronto, Toronto, ON M5T 2S8 Canada 8 Department of Neurophysiology, Nencki Institute of Experimental Biology, Warsaw, Poland 9 Department of Theory, Wigner Research Centre for Physics of the Hungarian Academy of Sciences, Budapest, 1121 Hungary 10 Department of Mathematical Sciences, KAIST, Daejoen, 34141 Republic of Korea 11 Department of Mathematics, University of Houston, Houston, TX 77004 USA 12 Department of Biochemistry & Cell Biology and Institute of Biosciences and Bioengineering, Rice University, Houston, TX 77005 USA 13 Department of Biology and Biochemistry, University of Houston, Houston, TX 77004 USA 14 Grupo de Neurocomputación Biológica, Dpto. de Ingeniería Informática, Escuela Politécnica Superior, Universidad Autónoma de Madrid, Madrid, Spain 15 Department of Biological Sciences, University of Southern California, Los Angeles, CA USA 16 Center for Neuroscience, Korea Institute of Science and Technology, Hwarang-ro 14-gil 5, Seongbuk-gu, Seoul, 02792 South Korea 17 Department of Neuroscience, University of Science and Technology, 217 Gajeong-ro, Yuseong-gu, Daejon, 34113 South Korea 18 Department of Psychology, Yonsei University, 50 Yonsei-ro, Seodaemun-gu, Seoul, 03722 South Korea 19 Department of Psychiatry, Kyung Hee University Hospital at Gangdong, 892, Dongnam-ro, Gangdong-gu, Seoul, 05278 South Korea 20 Department of Psychiatry, Veterans Administration Boston Healthcare System and Harvard Medical School, Brockton, MA 02301 USA 21 Center for Neural Science, New York University, New York, NY 10003 USA 22 Courant Institute of Mathematical Sciences, New York University, New York, NY 10012 USA 23 Faculty of Science and Technology, Bournemouth University, Bournemouth, England, UK 24 Heidelberg University, Baden-Württemberg, Germany 25 Heidelberg University, Baden-Württemberg, Germany 26 Bernstein Center for Computational Neuroscience, Heidelberg-Mannheim, Baden-Württemberg, Germany 27 National Vision Research Institute, Australian College of Optometry, Carlton, 3053 Australia 28 NeuroEngineering Laboratory, Dept. Electrical & Electronic Engineering, University of Melbourne, Parkville, VIC 3010 Australia 29 Centre for Neural Engineering, University of Melbourne, Parkville, 3010 Australia 30 Department of Physiology, Monash University, Melbourne, 3800 Australia 31 ARC Centre of Excellence for Integrative Brain Function, Dept. Optometry and Vision Sciences, University of Melbourne, Parkville, 3010 Australia 32 Institute of Software Engineering and Theoretical Computer Science, Technische Universitaet Berlin, 10587 Berlin, Germany 33 Bernstein Center for Computational Neuroscience Berlin, Humboldt-Universitaet zu Berlin, 10115 Berlin, Germany 34 Department of Neurobiology and Anatomy, University of Texas-Houston Medical School, Houston, TX 77030 USA 35 Centre for Computer Science and Informatics Research, University of Hertfordshire, Hatfield, AL10 9AB UK 36 School of Life Science and the Collaborative Innovation Center for Brain Science, Fudan University, Shanghai, 200438 China 37 LinyiHospitalof TraditionalChineseMedicine, 211 Jiefang Road, Lanshan, Linyi, 276000 Shandong Province China 38 Department of Neuroscience, Yale University School of Medicine, New Haven, CT 06520 USA 39 Department of Neurosurgery, Huashan Hospital, Shanghai Medical College, Fudan University, Shanghai, China 40 Department of Applied Mathematics, University of Washington, Seattle, WA 98195 USA 41 Amazon.com lnc, Seattle, WA 98108 USA 42 Department of Electrical Engineering, University of Washington, Seattle, WA 98195 USA 43 School of Life Science, Arizona State University, Tempe, AZ 85287 USA 44 School of Mathematical and Statistical Sciences, Arizona State University, Tempe, AZ 85287 USA 45 Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen (Medical Centre), Nijmegen, The Netherlands 46 Department for Biophysics, Faculty of Science, Radboud University Nijmegen, Nijmegen, The Netherlands 47 Department for Neuroinformatics, Faculty of Science, Radboud University Nijmegen, Nijmegen, The Netherlands 48 Center for Theoretical Neuroscience, Columbia University, New York, NY USA 49 Department of Neurology Pediatrics, Baylor College of Medicine, Houston, TX 77030 USA 50 Department of Computational and Applied Mathematics, Rice University, Houston, TX 77005 USA 51 Physics Department, Texas Southern University, 3100 Cleburne St, Houston, TX 77004 USA 52 Laboratory of Integrative Brain Function, The Rockefeller University, New York, NY 10065 USA 53 Department of Physiology, University of Bern, 3012 Bern, Switzerland 54 Department of Biology, University of Fribourg, 1700 Fribourg, Switzerland 55 Department of Mathematical Sciences and Technology, Norwegian University of Life Sciences, Ås, Norway 56 NORMENT, Institute of Clinical Medicine, University of Oslo, Oslo, Norway 57 The Blue Brain Project, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland 58 Letten Centre and Glialab, Dept. of Molecular Medicine, Inst. of Basic Medical Sciences, University of Oslo, Oslo, Norway 59 Centre for Molecular Medicine Norway, University of Oslo, Oslo, Norway 60 Department of Physics, University of Oslo, Oslo, Norway 61 IBM Research - Tokyo, Tokyo, Japan 62 School of Engineering, University of Waikato, Hamilton, 3240 New Zealand 63 Department of Psychiatry, Yale School of Medicine, New Haven, CT 06511 USA 64 INSERM U846, Stem Cell and Brain Research Institute, Bron Cedex, France 65 NYU-ECNU Institute of Brain and Cognitive Science, NYU Shanghai, Shanghai, China 66 Bernstein Center for Computational Neuroscience, 10115 Berlin, Germany 67 Institute for Physics, Humboldt-Universität zu Berlin, 12489 Berlin, Germany 68 Institute for Neurobiology, Eberhardt Karls Universität Tübingen, Tübingen, Germany 69 Bernstein Center for Computational Neuroscience, Munich, Germany 70 Riken Brai Science Institute, 2-1 Hirosawa, Wako, Saitama Japan 71 Department of Physics and Astronomy, Seoul National University, Seoul, 08826 Korea 72 Department of Biological Sciences, University of North Texas, Denton, TX 76203 USA 73 College of Technology, the University of Houston, Houston, TX 77204 USA 74 Departments of Engineering Technology, Computer Science, and Electrical and Computer Engineering, University of Houston, Houston, TX 77204 USA 75 Biological Cybernetics, School of Electrical and Electronic Engineering, Yonsei University, Shinchon, Seoul, 120-749 South Korea 76 Committee on Computational Neuroscience, University of Chicago, Chicago, IL USA 77 Department of Organismal Biology and Anatomy, University of Chicago, Chicago, IL USA 78 School of Psychology, University of Reading, Reading, Berkshire RG1 6AY UK 79 School of Computing and Mathematics, Plymouth University, Plymouth, PL4 8AA UK 80 Department of Physics, Chonnam National University, Gwangju, 61186 Republic of Korea 81 Faculty of Library, Information and Media Science, University of Tsukuba, Tsukuba, 305-0821 Japan 82 Physics, POSTECH, Pohang, 37673 Republic of Korea 83 Department of Mathematics, University of Pittsburgh, Pittsburgh, PA 15260 USA 84 Department of Bio and Brain Engineering/Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141 South Korea 85 Department of Biological Sciences, Carnegie Mellon University, Pittsburgh, PA 15213 USA 86 Center for the Neural Basis of Cognition, Pittsburgh, PA 15213 USA 87 Inst. of Pathological Physiology, First Faculty of Medicine, Charles University in Prague, 128 53 Prague, Czech Republic 88 Czech Technical University in Prague, Zikova 1903/4, 166 36 Prague, Czech Republic 89 Korea Science Academy of KAIST, Busan, 10547 South Korea 90 Department of Mathematics and Statistics, Georgia State University, Atlanta, GA 30303 USA 91 Department of Neurobiology and Anatomy, Drexel University, Philadelphia, PA 19129 USA 92 Department of Mathematical Sciences, Indiana University – Purdue University, Indianapolis, IN 46202 USA 93 LBB, ETH Zürich, 8051 Zurich, Switzerland 94 RIKEN Brain Science Institute, Wako-shi, Saitama, Japan 95 Computational Neuroscience Unit, Okinawa Institute of Science and Technology, Okinawa, 1919-1 Japan 96 School of Computing and Mathematics, Keele University, Newcastle-under-Lyme, ST5 5BG UK 97 Science and Technology Research Institute, University of Hertfordshire, Hatfield, UK 98 Institute for Robotics and Cognitive Systems, University of Luebeck, Luebeck, Germany 99 Department of Psychiatry, University of Luebeck, Schleswig-Holstein, Luebeck, Germany 100 Neuroscience Graduate Program, University of Michigan, Ann Arbor, MI 48109 USA 101 Center for the Study of Complex Systems, University of Michigan, Ann Arbor, MI 48109 USA 102 Department of Physics, University of Michigan, Ann Arbor, MI 48109 USA 103 Biophysics Program, University of Michigan, Ann Arbor, MI 48109 USA 104 Department of Molecular, Cellular, and Developmental Biology, University of Michigan, Ann Arbor, MI 48109 USA 105 Department of Physics, University of Michigan, Ann Arbor, MI 48109 USA 106 Department of Physics, Shaanxi Normal University, Xi’An City, ShaanXi Province China 107 Department of Physics and Centre for Nonlinear Studies, Institute of Computational and Theoretical Studies, Hong Kong Baptist University, Kowloon Tong, Hong Kong 108 State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing, China 109 Department of Physics, Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong 110 Beijing Computational Science Research Center, Beijing, 100084 People’s Republic of China 111 Research Centre, HKBU Institute of Research and Continuing Education, Shenzhen, China 112 School of Physics, University of Sydney, Sydney, NSW 2006 Australia 113 Center for Integrative Brain Function, University of Sydney, Sydney, NSW 2006 Australia 114 Department of Computer Science & Engineering, Texas A&M University, College Station, TX 77845 USA 115 Research Center for Information Technology Innovation, Academia Sinica, Taipei, Taiwan 116 State Key Lab of Cognitive Neuroscience & Learning, IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing, 100875 China 117 Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany 118 Bernstein Center for Computational Neuroscience, Göttingen, Germany 119 Neurobiology Research Unit, Okinawa Institute of Science and Technology, Onna-son, Okinawa, 904-0412 Japan 120 Computational Biology Center, IBM Research, Thomas J. Watson Research Center, Yorktown Heights, NY 10598 USA 121 Department of Mathematics, Franklin and Marshall College, Lancaster, PA 17604 USA 122 Fishberg Department of Neuroscience and Friedman Brain Institute, Icahn School of Medicine at Mount Sinai, New York, NY 10029 USA 123 Department of Anatomy and Neurobiology, Boston University School of Medicine, Boston, MA 02118 USA 124 Institute of Microelectronics, Tsinghua University, Beijing, 100081 China 125 School of Software, Beijing Institute of Technology, Beijing, 100083 China 126 Department of Psychiatry, University of Cambridge, Cambridge, CB2 0QQ UK 127 Department of Biomedical Engineering, Ulsan National Institute of Science and Technology, Ulsan, South Korea 128 Departments of Radiology and Psychiatry, Icahn School of Medicine at Mount Sinai, New York City, NY 10029 USA 129 Centre for Neural Circuits and Behaviour, University of Oxford, Oxford, OX1 3SR UK 130 Centre for Neural Circuits and Behaviour, University of Oxford, Oxford, UK 131 Section Computational Sensomotorics, CIN & HIH, Department of Cognitive Neurology, University Clinic Tübingen, Tübingen, Germany 132 Lab. Neuro en Psychofysiologie, Dept. Neuroscience, KU Leuven, Louvain, Belgium 133 Laboratory of Computational Neuroscience, EPF Lausanne, Lausanne, Switzerland 134 Centre for Neural Circuits and Behaviour, University of Oxford, Oxford, UK 135 The Blue Brain Project, EPF Lausanne, Lausanne, Switzerland 136 Departamento de Electromagnetismo y Física de la Materia, and Institute “Carlos I” for Theoretical and Computational Physics, University of Granada, Granada, Spain 137 School of Psychological Sciences, Faculty of Biomedical and Psychological Sciences, Monash University, Parkville, Australia 138 Faculty of Psychology and Neuroscience, Maastricht University, 6200 MD Maastricht, The Netherlands 139 Department of Physics, Institute for Advanced Studies in Basic Sciences (IASBS), Zanjan, 45137-66731 Iran 140 Center for Research in Climate Change and Global Warming (CRCC), Institute for Advanced Studies in Basic Sciences (IASBS), Zanjan, 45137-66731 Iran 141 School of Cognitive Sciences, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran 142 Department of Physics, Institute for Advanced Studies in Basic Sciences, Zanjan, Iran 143 School of Cognitive Sciences, Institute for Studies in Theoretical Physics and Mathematics, Niavaran, Tehran, Iran 144 Bristol Centre for Complexity Sciences, University of Bristol, Bristol, BS8 1TR UK 145 Department of Engineering Mathematics, University of Bristol, Bristol, BS8 1UB UK 146 School of Physiology and Pharmacology, University of Bristol, Bristol, BS8 1TD UK 147 Medical School, University of Exeter, Exeter, EX4 4PE UK 148 School of Physiology and Pharmacology, University of Bristol, Bristol, BS8 1TD UK 149 Medical School, University of Exeter, Exeter, EX4 4PE UK 150 Department of Mathematics, University of Exeter, Exeter, EX4 4QF UK 151 Frankfurt Institute for Advanced Studies, 60438 Frankfurt am Main, Hessen, Germany 152 Computational and Theoretical Neuroscience Laboratory, School of Information Technology and Mathematical Sciences, University of South Australia, Adelaide, Australia 153 School of Mathematical Sciences, University of Nottingham, Nottingham, UK 154 Frankfurt Institute for Advanced Studies, Goethe-Universität, Frankfurt am Main, Germany 155 Robinson Research Institute, School of Medicine, University of Adelaide, Adelaide, Australia 156 Institute of Physiology of the Czech Academy of Sciences, Prague, Czech Republic 157 Institute of Biophysics and Informatics, First Faculty of Medicine, Charles University in Prague, Prague, Czech Republic 158 IBPS, Neuroscience Paris Seine, CNRS UMR8246, Inserm U1130, UPMC UM 119, Université Pierre et Marie Curie, Paris, France 159 IBM Research - Tokyo, Tokyo, Japan 160 RIKEN Brain Science Institute, Tokyo, Japan 161 The University of Tokyo, Tokyo, Japan 162 Department of Mathematics, Computer Science and Physics, Roanoke College, Salem, VA 24153 USA 163 Department of Pediatrics, University of Louisville, Louisville, KY 40208 USA 164 Department of Biology, Washington and Lee University, Lexington, VA 24450 USA 165 Department of Psychology, Washington University in St. Louis, St. Louis, MO 63130 USA 166 Department of Psychological Sciences, Texas Tech University, Lubbock, TX 79409 USA 167 The Department of Veteran Affairs, District of Columbia VA Medical Center, Washington, DC 20420 USA 168 Department of Psychology, The Catholic University of America, Washington, DC 20064 USA 169 Department of Physics, Humboldt University, 10099 Berlin, Germany 170 Biological Physics Theory Unit, Okinawa Institute of Science and Technology, Okinawa, 904-0495 Japan 171 Department of Physics and Astronomy, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands 172 Institut de la Vision, INSERM UMRS 968, CNRS UMR 7210, Paris, France 173 Bernstein Center Freiburg, Institute of Biology III, University of Freiburg, 79100 Freiburg, Germany 174 fourMs group, Dept. Musicology, University of Oslo, 0371 Oslo, Norway 175 Center for Brain Science, Harvard University, Cambridge, MA 02138 USA 176 Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104 USA 177 Edmond and Lily Safra Center for Brain Sciences, Hebrew University, 91904 Jerusalem, Israel 178 NINDS, NIH, Bethesda, MD 20892 USA 179 Department of Physics, University of New Hampshire, Durham, NH 03824 USA 180 Krembil Research Institute, University Health Network, Toronto, ON Canada 181 Department of Physiology, University of Toronto, Toronto, ON Canada 182 Department of Neuroscience, Yale School of Medicine, New Haven, CT 06520 USA 183 Department of Medicine (Neurology), University of Toronto, Toronto, ON Canada 184 Department of Neuroscience, Physiology and Pharmacology, University College London, London, UK 185 Department of Bioengineering, Imperial College London, London, UK 186 Metacell LLC, San Diego, CA USA 187 State University of New York Downstate Medical Center, Brooklyn, NY USA 188 Neuroinformatics group Unité de Neurosciences, Information et Complexité, CNRS, Gif sur Yvette, France 189 State Key Laboratory of Cognitive Neuroscience & Learning, IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing, 100875 China 190 School of System Science, Beijing Normal University, Beijing, 100875 China 191 Bio and Brain Engineering, KAIST, Daejeon, 34141 Republic of Korea 192 Center for Cognition and Sociality, IBS, Daejeon, 34047 Republic of Korea 193 Department of Biological Structure, University of Washington, Seattle, WA 98195 USA 194 UW Institute for Neuroengineering, University of Washington, Seattle, WA 98195 USA 195 The Salk Institute for Biological Studies, La Jolla, CA 92037 USA 196 Division of Biological Sciences, University of California at San Diego, La Jolla, CA 92095 USA 197 BrainLinks-BrainTools, Cluster of Excellence, University of Freiburg, Freiburg, Germany 198 Faculty of Biology and Bernstein Center Freiburg, University of Freiburg, Freiburg, Germany 199 Department of Computational Biology, Royal Institute of Technology Stockholm, Stockholm, Sweden 200 Center for Neural Engineering, Department of Engineering Science and Mechanics, The Pennsylvania State University, University Park, PA 16802 USA 201 Departments of Neurosurgery and Physics, The Pennsylvania State University, University Park, PA 16802 USA 202 National Institute for Physiological Sciences, Okazaki, Aichi Japan 203 Department of Computer Science, Tufts University, Medford, MA USA 204 Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea 205 KAIST Institute for Health Science and Technology, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea 206 Bernstein Center Freiburg & Faculty of Biology, University of Freiburg, Baden-Württember, Freiburg, 79194 Germany 207 School of Physics, University of Sydney, Sydney, NSW Australia 208 Department of Mathematical Sciences, Indiana University-Purdue University Indianapolis, Indianapolis, IN USA 209 Stark Neurosciences Research Institute, Indiana University School of Medicine, Indianapolis, IN USA 210 School of Cognitive Sciences, Institute for Research in Fundamental Sciences (IPM), Tehran, 19395-5746 Iran 211 School of ECE, College of Engineering, University of Tehran, Tehran, 14155-6619 Iran 212 RIKEN Brain Science Institute, Wako, Saitama 351-0198 Japan 213 Department of Physics, University of Guilan, Rasht, 41335-1914 Iran 214 School of Physics, Institute for Research in Fundamental Sciences (IPM), Tehran, 19395-5531 Iran 215 Max-Planck-Institute for Dynamics and Self-Organization, 37077 Göttingen, Germany 216 Bernstein Center for Computational Neuroscience, University of Göttingen, 37075 Göttingen, Germany 217 IST Austria, 3400 Klosterneuburg, Austria 218 BCCN & MPI for Dynamics and Self-Organization, 37077 Göttingen, Germany 219 Dept. of Non-linear Dynamics, Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany 220 School of Civil Engineering, The University of Sydney, Sydney, NSW Australia 221 Complex Systems Research Group, Faculty of Engineering & IT, The University of Sydney, Sydney, NSW 2006 Australia 222 Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, VA 20147 USA 223 Department of Psychiatry, University of Cambridge, Cambridge, UK 224 MEG Unit, Brain Imaging Center, Goethe University, 60528 Frankfurt am Main, Germany 225 Department of Nonlinear Dynamics, Max Planck Institute for Dynamics & Self-Organization, Göttingen, Germany 226 Department of Physics & Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton, NJ 08544 USA 227 Department of Psychology & Princeton Neuroscience Institute, Princeton University, Princeton, NJ 08544 USA 228 Department of Neurobiology and Behavior, Stony Brook University, Stony Brook, NY 11794 USA 229 Department of Applied Mathematics and Statistics, Stony Brook University, Stony Brook, NY 11794 USA 230 Institute for Advanced Computational Science, Stony Brook University, Stony Brook, NY 11794 USA 231 Graduate School of Life Science, University of Hyogo, 3-2-1 Koto, Kamigori, Ako, Hyogo 678-1297 Japan 232 Department of Nuclear Medicine, Radiology and Psychiatry, Yonsei University College of Medicine, Department of Cognitive Science, Yonsei University, 50 Yonsei-ro, Sinchon-dong Seodaemoon-gu, Seoul, 120-752 Republic of Korea 233 Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology, Daejeon, 34141 Republic of Korea 234 Program of Brain and Cognitive Engineering, KAIST, Daejeon, 34141 Republic of Korea 235 Department of Biomedical Engineering, Bülent Ecevit University, 67100 Zonguldak, Turkey 236 Department of Electrical and Electronics Engineering, Bülent Ecevit University, 67100 Zonguldak, Turkey 237 School of Engineering and Informatics, Sussex Neuroscience, University of Sussex, Falmer, Brighton, BN1 9QJ UK 238 School of Engineering and Informatics, University of Sussex, Falmer, Brighton BN1 9QJ UK 239 Research Center Neurosensory Science, Carl-von-Ossietzky University Oldenburg, Oldenburg, Germany 240 The Laboratory for Investigative Neurophysiology (The LINE), Department of Clinical Neurosciences and Department of Radiology, University Hospital Center and University of Lausanne, 1011 Lausanne, Switzerland 241 Deutscher Wetterdienst, 63067 Offenbach, Germany 242 Krembil Research Institute, University Health Network, Toronto, ON M5T 2S8 Canada 243 Institute of Advanced Simulation (IAS-6) & Institute of Neuroscience and Medicine (INM-6) & JARA BRAIN Institute I, Jülich Research Center, 52425 Jülich, Germany 244 Bernstein Center Freiburg, Albert-Ludwig University of Freiburg, Freiburg im Breisgau 79104, Germany 245 Institute for Adaptive and Neural Computation, School of Informatics, University of Edinburgh, Edinburgh, EH8 9AB UK 246 Institute of Cognitive Neuroscience, Faculty of Psychology, Ruhr-University Bochum, 44801 Bochum, Germany 247 Simulation Laboratory Neuroscience – Bernstein Facility for Simulation and Database Technology, Institute for Advanced Simulation, Jülich Aachen Research Alliance, Jülich Research Center, Jülich, Germany 248 Allen Institute for Brain Science, Seattle, WA 98109 USA 249 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany 250 NeuroEngineering Laboratory, Electrical & Electronic Engineering, The University of Melbourne, Parkville, VIC 3010 Australia 251 IBM Research, Melbourne, Australia 252 Monash Institute of Medical Engineering, Monash University, Melbourne, Australia 253 National Vision Research Institute, Melbourne, Australia 254 School of Mathematical Sciences, University of Nottingham, Nottingham, UK 255 Blue Brain Project, École Polytechnique fédérale de Lausanne (EPFL), Geneva, Switzerland 256 Department of Information and Electrical Engineering, University of Salerno, 84084 Fisciano, SA Italy 257 Department of Information and Electrical Engineering and Applied Mathematics, University of Salerno, Fisciano, SA 81100 Italy 258 Laboratory of Neural Computation, Istituto Italiano di Tecnologia, Rovereto, TN 38068 Italy 259 LIST Laboratory, FST, Abdelmalek Essaadi’s University, Tangier, Morocco 260 Department of Computer Science, ENSAT, Abdelmalek Essaadi’s University, Tangier, Morocco 261 Department Physiology & Pharmacology, SUNY Downstate, Brooklyn, NY 11203 USA 262 Department Neuroscience, Yale University School of Medicine, New Haven, CT USA 263 Department of Chemical & Biomedical Engineering, Tandon School of Engineering, NYU, Brooklyn, NY USA 264 Nathan Kline Institute for Psychiatric Research, Orangeburg, NY USA 265 Department Biomedical Engineering, University of Southern California, Los Angeles, CA USA 266 Div Neurology, Child Neurology and Movement Disorders, Children’s Hospital Los Angeles, Los Angeles, CA USA 267 Department Neurology, Kings County Hospital Center, Brooklyn, NY 11203 USA 268 Department of Physiology & Pharmacology, SUNY Downstate Medical Center, Brooklyn, NY 11023 USA 269 Complex Systems Group, School of Physics, University of Sydney, Sydney, NSW 2006 Australia 270 Institute of Neuroscience and Medicine (INM-6), Jülich Research Centre and JARA, Jülich, Germany 271 Department Physiology, Northwestern University, Chicago, IL 60611 USA 272 Department of Neuroscience, Physiology & Pharmacology, University College London, London, WC1E6BT UK 273 Brain and Cognitive Science, Seoul National University, Seoul, 151-742 Republic of Korea 274 Department of Brain and Cognitive Sciences, Seoul National University, Seoul, 151-742 Korea 275 Department of Brain and Cognitive Neuroscience, Seoul National University, Gwanak-gu, South Korea 276 Department of Brain and Cognitive Sciences, Seoul National University, Seoul, 08826 South Korea 277 Physics Department, Pohang University of Science and Technology, Pohang, South Korea 278 Department of Physics, Inha University, Namgu, Incheon, 22212 Korea 279 Departments of Neurology, Ewha Womans University School of Medicine, Seoul, Korea 280 Department of Medical Science, Ewha Womans University School of Medicine, Seoul, Korea 281 Department of Electronics Engineering, Ewha Womans University College of Engineering, Seoul, Korea 282 Brain & Cognitive Sciences, Ewha Womans University College of Scranton, Seoul, Korea 283 Department of Electronics Engineering, Ewha Womans University, Seoul, 120-750 Korea 284 Department of Neurology, Ewha Womans University, Seoul, 120-750 Korea 285 Department of Physics, National Central University, Chung-Li, Taiwan, ROC 286 Taiwan International Graduate Program for Molecular Science and Technology, Institute for Atomic and Molecular Sciences, Academia Sinica, Taipei, Taiwan, ROC 287 Department of Brain and Cognitive Engineering, Korea University, Seoul, Korea 288 Bernstein Center Freiburg, Freiburg, Germany 289 Faculty of Biology, University of Freiburg, 79104 Freiburg, Germany 290 Instituto de Neurociencias de Alicante, University of Alicante, Alicante, Spain 291 Department of Neuroscience, Karolinska Institute, Stockholm, 17177 Sweden 292 Dept. of Computational Science and Technology, School of Computer Science and Communication, KTH Royal Institute of Technology, 10040 Stockholm, Sweden 293 Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, 52425 Jülich, Germany 294 Faculty of Health, Medicine and Life Science, School for Mental Health and Neuroscience (MHeNS),Alzheimer Centre Limburg, Maastricht University Medical Centre, PO Box 616, 6200 MD Maastricht, The Netherlands 295 Department of Radiology &Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School, Boston, MA 02114 USA 296 Department of Cognitive Neuroscience, Faculty of Psychology and Neuroscience, Maastricht University, PO BOX 616, 6200 MD Maastricht, The Netherlands 297 Cognitive Neuroscience, Inst. of Neuroscience and Medicine (INM-3), Jülich Research Centre, Jülich, Germany 298 Department of Neurology, University Hospital of Cologne, Cologne, Germany 299 Computational Neuroscience, Bernstein Center Freiburg, 79104 Freiburg, Germany 300 Cold Spring Harbor Laboratory, Cold Spring Harbor, NY 11724 USA 301 NYU Neuroscience Institute, New York, NY 10016 USA 302 Institute of Neuroscience and Medicine (INM-6), Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany 303 Computational Brain Science, Dept. of Computational Science and Technology, School of Computer Science and Communication, KTH, Royal Institute of Technology, Stockholm, Sweden 304 Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany 305 Institute of Biomedical Engineering and Informatics, Ilmenau University of Technology, Ilmenau, Germany 306 Department of Medicine, University of Melbourne, Melbourne, Australia 307 Division of IoT and Robotics Convergence Research, DGIST, Daegu, 42988 Korea 308 Department of Human Factors Engineering, Ulsan National Institute of Science and Technology, Ulsan, 689-798 South Korea 309 Department of Human and Systems Engineering, Ulsan National Institute of Science and Technology, Ulsan, South Korea 310 INSERM UMR 1106 Institut de Neurosciences des Systèmes - Aix-Marseille Université, Marseille, France 311 Aix-Marseille Université, CNRS, ISM UMR 7287, 13288 Marseille, France 312 Neuroscience Center, University of Helsinki, 00014 Helsinki, Finland 313 Dept of Neurology, Dept of Biomedical Engineering, University of Michigan, Ann Arbor, MI 48109 USA 314 Program of Brain and Cognitive Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141 South Korea 315 Program of Brain Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141 Republic of Korea 316 Division of Computational Mathematics, National Institute for Mathematical Sciences (NIMS), Daejeon, 34047 South Korea 317 Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141 South Korea 318 Department of Mathematics, North Carolina A&T State University, Greensboro, NC 27411 USA 319 Department of Biology, East Carolina University, Greenville, NC 27858 USA 320 Department of Mathematics, East Carolina University, Greenville, NC 27858 USA 321 David Geffen School of Medicine, UCLA, Los Angeles, CA 90095 USA 322 Graduate School of Informatics, Kyoto University, Yoshidahonmachi 36-1, Sakyo, Kyoto, Japan 323 Neural Computation Unit, Okinawa Institute of Science and Technology Graduate University, 1919-1 Tancha, Onna-son, Kunigami-gun, Okinawa Japan 324 Graduate School of Information Science, Nara Institute of Science and Technology, 8916-5 Takayama, Ikoma, Nara Japan 325 University of California San Diego, La Jolla, CA USA 326 Biocenter University of Würzburg, Würzburg, Germany 327 School of Life Sciences, Arizona State University, Tempe, AZ USA 328 Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague, 166 27 Prague, Czech Republic 329 Department of Nonlinear Dynamics and Complex Systems, Institute of Computer Science, The Czech Academy of Sciences, 182 07 Prague, Czech Republic 330 National Institute of Mental Health, 250 67 Klecany, Czech Republic 331 Department of Physics, Faculty 1, RWTH Aachen University, 52074 Aachen, Germany 332 Departments of Neurology and Physiology, Medical College of Wisconsin, Milwaukee, WI 53226 USA 333 College of Engineering & Applied Science, University of Wisconsin-Milwaukee, Milwaukee, WI 53211 USA 334 Department of Physics, Chungbuk National University, Cheongju, Chungbuk 28644 Republic of Korea 335 Korea Basic Science Institute, Cheongju, Chungbuk 28119 Republic of Korea 336 Department of Computer Science, Chemnitz University of Technology, Chemnitz, Germany 337 Bernstein Center for Computational Neuroscience, Charité University Medicine, Berlin, Germany 338 Department of Mathematics, Drexel University, Philadelphia, PA 19104 USA 339 National Vision Research Institute, Australian College of Optometry, Carlton, VIC 3053 Australia 340 Centre for Neural Engineering, University of Melbourne, Parkville, VIC 3010 Australia 341 Computational and Theoretical Neuroscience Laboratory, School of Information Technology and Mathematical Sciences, University of South Australia, Mawson Lakes, SA 5095 Australia 342 Computing Science & Mathematics, School of Natural Sciences, University of Stirling, Stirling, FK9 4LA UK 343 Systems Neuroscience Group, QIMR Berghofer Medical Research Institute, Brisbane, QLD 4006 Australia 344 Faculty of Medicine, Technion, Haifa, Israel 345 Department of Neuroscience and Department of Physiology and Cellular Biophysics, Columbia University, New York, NY USA 346 Section for Computational Sensomotorics, Dept. Cognitive Neurology, CIN&HIH, Tübingen, Germany 347 GTC, International Max Planck Research School, University of Tübingen, Tübingen, Germany 348 Section Computational Sensomotorics, Department of Cognitive Neurology, CIN & HIH, 72076 Tübingen, Germany 349 IMPRS for Cognitive and Systems Neuroscience, University Clinic Tübingen, 72076 Tübingen, Germany 350 School of Life Sciences, Brain Mind Institute and School of Computer and Communication Sciences, Ecole Polytechnique Federal de Lausanne (EPFL), 1015 Lausanne, Switzerland 351 Geneva Finance Research Institute (GFRI) and Swiss Center for Affective Sciences (CISA), University of Geneva, 1211 Geneva, Switzerland 352 ENT-Department, Leiden University Medical Centre, Leiden, 2300 RC The Netherlands 353 Leiden Institute for Brain and Cognition, 2300 RC Leiden, The Netherlands 354 Department of Psychiatry, Icahn School of Medicine at Mount Sinai, New York, NY 10029 USA 355 Monash Institute of Cognitive and Clinical Neurosciences, Monash University, Clayton, VIC 3168 Australia 356 Department of Neurology, University of Michigan, Ann Arbor, MI 48104 USA 357 Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI 48104 USA 358 Department of Physics, Ohio Wesleyan University, Delaware, OH 43015 USA 359 Department of Physics, Towson University, Towson, MD 21252 USA 360 Neuroscience Program, Ohio Wesleyan University, Delaware, OH 43015 USA 361 Department of Structure and Function of Neural Networks, Korea Brain Research Institute, Daegu, 41068 Republic of Korea 362 Princeton Neuroscience Institute, Princeton University, Princeton, NJ 08544 USA 363 Circuit Dynamics and Connectivity Unit, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, MD 20824 USA 364 Computer Science Department, Princeton University, Princeton, NJ 08544 USA 365 Brain Research Institute, University of Bremen, 28334 Bremen, Germany 366 Institute for Neurophysics, University of Bremen, 28334 Bremen, Germany 367 NORMENTa, Institute of Clinical Medicine, University of Oslo, Oslo, Norway 368 Department of Neurosciences, University of California San Diego, La Jolla, CA USA 369 Department of Radiology, University of California San Diego, La Jolla, CA USA 370 Biocomputation Research Group, University of Hertfordshire, Hatfield, UK 371 Department of Physics, University of Oslo, Oslo, Norway 372 Simula Research Laboratory and Center for Cardiological Innovation, Oslo, Norway 373 Multimodal Imaging Laboratory, UC San Diego, La Jolla, CA USA 374 Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, 52074 Aachen, Germany 375 Department of Physics, Faculty 1, RWTH Aachen University, 52074 Aachen, Germany 376 Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6), and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany 377 Rotman Research Institute, Baycrest, Toronto, ON M6A 2E1 Canada 378 McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University, Montreal, QC Canada 379 Department of Computational Neuroscience, University Medical Center Eppendorf, Hamburg, Germany 380 Department of Health Sciences, Boston University, Boston, MA USA 381 Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, Aachen, Germany 382 Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany 383 Energy Research Institute @ NTU (ERI@N), Interdisciplinary Graduate School, Nanyang Technological University, Singapore, 639798 Singapore 384 Fraunhofer IDM @ NTU, Nanyang Technological University, Singapore, 639798 Singapore 385 School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, 639798 Singapore 386 Future Mobility Research Lab, A Joint Initiative of BMW Group & NTU, Nanyang Technological University, Singapore, 639798 Singapore 387 School of Computer Engineering, Nanyang Technological University, Singapore, 639798 Singapore 388 Center for Brain and Cognition, Universitat Pompeu Fabra, 08018 Barcelona, Spain 389 Minerva Research Group Brain Modes, Max Planck Institute for Human Cognitive and Brain Sciences, 04103 Leipzig, Germany 390 Dept. of Neurology, Charité - University Medicine, 10117 Berlin, Germany 391 Bernstein Focus State Dependencies of Learning & Bernstein Center for Computational Neuroscience, 10115 Berlin, Germany 392 Berlin School of Mind and Brain & Mind and Brain Institute, Humboldt University, 10117 Berlin, Germany 393 Catalan Institution for Advanced Studies (ICREA), Universitat Barcelona, 08010 Barcelona, Spain 394 Center for Brain Cognition, Universitat Pompeu Fabra, Barcelona, Spain 395 Center for Music in the Brain, Aarhus University & Royal Academy of Music, Aarhus/Aalborg, Denmark 396 Faculty of Music, University of Oxford, Oxford, UK 397 Department of Psychology and Behavioural Sciences, Aarhus University, Aarhus, Denmark 398 Center of Functionally Integrative Neuroscience, Aarhus University, Aarhus, Denmark 399 Department of Psychiatry, University of Oxford, Oxford, UK 400 Computational Neuroscience Unit, Okinawa Institute of Science and Technology, 1919-1 Tancha, Onna-son, Kunigami-gun, Okinawa 904-0495 Japan 401 Computational Neuroscience Unit, Okinawa Institute of Science and Technology, Okinawa, 904-0411 Japan 402 Computational Neuroscience Unit, Okinawa Institute of Science and Technology Graduate University, Onna-son, Okinawa Japan 403 Computational Neuroscience Unit, Okinawa Institute of Science and Technology Graduate University, Onna-son, Okinawa, 904-0495 Japan 404 Computational Neuroscience Unit, Okinawa Institute of Science and Technology Graduate University, Okinawa, 904-0895 Japan 405 Champalimaud Neuroscience Programme, Champalimaud Center for the Unknown, Lisbon, Portugal 406 Center of Biorobotics, Mærsk Mc-Kinney Møller Institute, University of Southern Denmark, Odense, Denmark 407 Riken Brain Science Institute, 2-1 Hirosawa, Wako, Saitama Japan 408 IBM, IBM Research - Tokyo, Tokyo, 103-8510 Japan 409 Department of Medicine, The University of Melbourne, Parkville, VIC 3010 Australia 410 Department of Statistics, Columbia University, New York, NY USA 411 Swinburne University of Technology, Hawthorn, VIC 3122 Australia 412 Department of Electrical and Electronic Engineering, The University of Melbourne, Parkville, VIC 3010 Australia 413 Department of Otorhinolaryngology – Head and Neck Surgery, University of Pennsylvania, Philadelphia, PA 19104 USA 414 Department of Neurology, Albert Einstein College of Medicine, Bronx, NY 10461 USA 415 Systems Neuroscience Group, QIMR Berghofer Medical Research Institute, Herston, QLD 4006 Australia 416 Department of Bioengineering, Imperial College London, London, UK 417 Department of Physics, POSTECH, Pohang, 37673 South Korea 418 Center for Neuroscience, KIST, Seoul, 02792 South Korea 419 Department of Industrial and Management Engineering, POSTECH, Pohang, 37673 South Korea 420 Center for Neuroscience, Korea Institute of Science and Technology, Seoul, South Korea 421 Department of Psychiatry, VA Boston Healthcare System & Harvard Medical School, Brockton, MA USA 422 Department of Neuroscience, University of Science and Technology, Daejon, South Korea 423 Center for Neuroscience, Korea Institute of Science and Technology, Seoul, 02792 Korea 424 Program in Nano Science and Technology, Seoul National University, Seoul, 08826 Korea 425 Department of Neuroscience, University of Science and Technology, Daejon, 34113 Korea 426 Center for Neural Science, New York University, 4 Washington Place, 10003, New York NY USA 427 Institut de Neuroscienes de la Timone (INT), CNRS & Aix-Marseille University, 27 Boulevard Jean Moulin, 13005 Marseille, France 428 Department of Biotechnology, Indian Institute of Technology Madras, Chennai, Tamil Nadu 600036 India 429 Department of Electrical and Electronic Engineering, University of Melbourne, Parkville, VIC 3010 Australia 430 Centre for Neural Engineering, University of Melbourne, Parkville, VIC 3010 Australia 431 Department of Physiology and Biophysics, University of Washington, Seattle, WA 98195 USA 432 Howard Hughes Medical Institute, University of Washington, Seattle, WA 98195 USA 433 Allen Institute for Brain Science, Seattle, WA 98109 USA 434 Department of Neurobiology, University of Chicago, Chicago, IL 60637 USA 435 Project-Team BEAGLE, INRIA Rhône-Alpes, 69603 Villeurbanne, France 436 Department of Statistics, University of Chicago, Chicago, IL 60637 USA 437 Systems Neuroscience Group, QIMR Berghofer Medical Research Institute, Brisbane, QLD 4006 Australia 438 Department of Biology, West Virginia University, Morgantown, WV 26506 USA 439 Department of Mathematics, University of Texas at Arlington, Arlington, TX 76019 USA 440 Department of Neuroscience, University of Texas Southwestern Medical Center, Dallas, TX 75390 USA 441 Department of Mathematics, University of Houston, Clear Lake, Houston, TX 77059 USA 442 Department of Computational Neurosciences, University of Lorraine, 54600 Nancy, France 443 Computational Neurosciences Laboratory, Ecole Polytechnique Federale de Lausanne, 1015 Lausanne, Switzerland 444 Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141 Republic of Korea 445 Laboratory of Brain and Cognition, National Institute of Mental Health, National Institutes of Health, Bethesda, MD 20892 USA 446 Institute of Industrial Science, the University of Tokyo, Tokyo, Japan 18 8 2016 18 8 2016 2016 17 Suppl 1 The publication charges for this supplement were funded by the Organization for Computational Neurosciences. The Supplement Editors declare that they have no competing interests.54© The Author(s) 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Table of contents A1 Functional advantages of cell-type heterogeneity in neural circuits Tatyana O. Sharpee A2 Mesoscopic modeling of propagating waves in visual cortex Alain Destexhe A3 Dynamics and biomarkers of mental disorders Mitsuo Kawato F1 Precise recruitment of spiking output at theta frequencies requires dendritic h-channels in multi-compartment models of oriens-lacunosum/moleculare hippocampal interneurons Vladislav Sekulić, Frances K. Skinner F2 Kernel methods in reconstruction of current sources from extracellular potentials for single cells and the whole brains Daniel K. Wójcik, Chaitanya Chintaluri, Dorottya Cserpán, Zoltán Somogyvári F3 The synchronized periods depend on intracellular transcriptional repression mechanisms in circadian clocks. Jae Kyoung Kim, Zachary P. Kilpatrick, Matthew R. Bennett, Kresimir Josić O1 Assessing irregularity and coordination of spiking-bursting rhythms in central pattern generators Irene Elices, David Arroyo, Rafael Levi, Francisco B. Rodriguez, Pablo Varona O2 Regulation of top-down processing by cortically-projecting parvalbumin positive neurons in basal forebrain Eunjin Hwang, Bowon Kim, Hio-Been Han, Tae Kim, James T. McKenna, Ritchie E. Brown, Robert W. McCarley, Jee Hyun Choi O3 Modeling auditory stream segregation, build-up and bistability James Rankin, Pamela Osborn Popp, John Rinzel O4 Strong competition between tonotopic neural ensembles explains pitch-related dynamics of auditory cortex evoked fields Alejandro Tabas, André Rupp, Emili Balaguer-Ballester O5 A simple model of retinal response to multi-electrode stimulation Matias I. Maturana, David B. Grayden, Shaun L. Cloherty, Tatiana Kameneva, Michael R. Ibbotson, Hamish Meffin O6 Noise correlations in V4 area correlate with behavioral performance in visual discrimination task Veronika Koren, Timm Lochmann, Valentin Dragoi, Klaus Obermayer O7 Input-location dependent gain modulation in cerebellar nucleus neurons Maria Psarrou, Maria Schilstra, Neil Davey, Benjamin Torben-Nielsen, Volker Steuber O8 Analytic solution of cable energy function for cortical axons and dendrites Huiwen Ju, Jiao Yu, Michael L. Hines, Liang Chen, Yuguo Yu O9 C. elegans interactome: interactive visualization of Caenorhabditis elegans worm neuronal network Jimin Kim, Will Leahy, Eli Shlizerman O10 Is the model any good? Objective criteria for computational neuroscience model selection Justas Birgiolas, Richard C. Gerkin, Sharon M. Crook O11 Cooperation and competition of gamma oscillation mechanisms Atthaphon Viriyopase, Raoul-Martin Memmesheimer, Stan Gielen O12 A discrete structure of the brain waves Yuri Dabaghian, Justin DeVito, Luca Perotti O13 Direction-specific silencing of the Drosophila gaze stabilization system Anmo J. Kim, Lisa M. Fenk, Cheng Lyu, Gaby Maimon O14 What does the fruit fly think about values? A model of olfactory associative learning Chang Zhao, Yves Widmer, Simon Sprecher,Walter Senn O15 Effects of ionic diffusion on power spectra of local field potentials (LFP) Geir Halnes, Tuomo Mäki-Marttunen, Daniel Keller, Klas H. Pettersen,Ole A. Andreassen, Gaute T. Einevoll O16 Large-scale cortical models towards understanding relationship between brain structure abnormalities and cognitive deficits Yasunori Yamada O17 Spatial coarse-graining the brain: origin of minicolumns Moira L. Steyn-Ross, D. Alistair Steyn-Ross O18 Modeling large-scale cortical networks with laminar structure Jorge F. Mejias, John D. Murray, Henry Kennedy, Xiao-Jing Wang O19 Information filtering by partial synchronous spikes in a neural population Alexandra Kruscha, Jan Grewe, Jan Benda, Benjamin Lindner O20 Decoding context-dependent olfactory valence in Drosophila Laurent Badel, Kazumi Ohta, Yoshiko Tsuchimoto, Hokto Kazama P1 Neural network as a scale-free network: the role of a hub B. Kahng P2 Hemodynamic responses to emotions and decisions using near-infrared spectroscopy optical imaging Nicoladie D. Tam P3 Phase space analysis of hemodynamic responses to intentional movement directions using functional near-infrared spectroscopy (fNIRS) optical imaging technique Nicoladie D.Tam, Luca Pollonini, George Zouridakis P4 Modeling jamming avoidance of weakly electric fish Jaehyun Soh, DaeEun Kim P5 Synergy and redundancy of retinal ganglion cells in prediction Minsu Yoo, S. E. Palmer P6 A neural field model with a third dimension representing cortical depth Viviana Culmone, Ingo Bojak P7 Network analysis of a probabilistic connectivity model of the Xenopus tadpole spinal cord Andrea Ferrario, Robert Merrison-Hort, Roman Borisyuk P8 The recognition dynamics in the brain Chang Sub Kim P9 Multivariate spike train analysis using a positive definite kernel Taro Tezuka P10 Synchronization of burst periods may govern slow brain dynamics during general anesthesia Pangyu Joo P11 The ionic basis of heterogeneity affects stochastic synchrony Young-Ah Rho, Shawn D. Burton, G. Bard Ermentrout, Jaeseung Jeong, Nathaniel N. Urban P12 Circular statistics of noise in spike trains with a periodic component Petr Marsalek P14 Representations of directions in EEG-BCI using Gaussian readouts Hoon-Hee Kim, Seok-hyun Moon, Do-won Lee, Sung-beom Lee, Ji-yong Lee, Jaeseung Jeong P15 Action selection and reinforcement learning in basal ganglia during reaching movements Yaroslav I. Molkov, Khaldoun Hamade, Wondimu Teka, William H. Barnett, Taegyo Kim, Sergey Markin, Ilya A. Rybak P17 Axon guidance: modeling axonal growth in T-Junction assay Csaba Forro, Harald Dermutz, László Demkó, János Vörös P19 Transient cell assembly networks encode persistent spatial memories Yuri Dabaghian, Andrey Babichev P20 Theory of population coupling and applications to describe high order correlations in large populations of interacting neurons Haiping Huang P21 Design of biologically-realistic simulations for motor control Sergio Verduzco-Flores P22 Towards understanding the functional impact of the behavioural variability of neurons Filipa Dos Santos, Peter Andras P23 Different oscillatory dynamics underlying gamma entrainment deficits in schizophrenia Christoph Metzner, Achim Schweikard, Bartosz Zurowski P24 Memory recall and spike frequency adaptation James P. Roach, Leonard M. Sander, Michal R. Zochowski P25 Stability of neural networks and memory consolidation preferentially occur near criticality Quinton M. Skilling, Nicolette Ognjanovski, Sara J. Aton, Michal Zochowski P26 Stochastic Oscillation in Self-Organized Critical States of Small Systems: Sensitive Resting State in Neural Systems Sheng-Jun Wang, Guang Ouyang, Jing Guang, Mingsha Zhang, K. Y. Michael Wong, Changsong Zhou P27 Neurofield: a C++ library for fast simulation of 2D neural field models Peter A. Robinson, Paula Sanz-Leon, Peter M. Drysdale, Felix Fung, Romesh G. Abeysuriya, Chris J. Rennie, Xuelong Zhao P28 Action-based grounding: Beyond encoding/decoding in neural code Yoonsuck Choe, Huei-Fang Yang P29 Neural computation in a dynamical system with multiple time scales Yuanyuan Mi, Xiaohan Lin, Si Wu P30 Maximum entropy models for 3D layouts of orientation selectivity Joscha Liedtke, Manuel Schottdorf, Fred Wolf P31 A behavioral assay for probing computations underlying curiosity in rodents Yoriko Yamamura, Jeffery R. Wickens P32 Using statistical sampling to balance error function contributions to optimization of conductance-based models Timothy Rumbell, Julia Ramsey, Amy Reyes, Danel Draguljić, Patrick R. Hof, Jennifer Luebke, Christina M. Weaver P33 Exploration and implementation of a self-growing and self-organizing neuron network building algorithm Hu He, Xu Yang, Hailin Ma, Zhiheng Xu, Yuzhe Wang P34 Disrupted resting state brain network in obese subjects: a data-driven graph theory analysis Kwangyeol Baek, Laurel S. Morris, Prantik Kundu, Valerie Voon P35 Dynamics of cooperative excitatory and inhibitory plasticity Everton J. Agnes, Tim P. Vogels P36 Frequency-dependent oscillatory signal gating in feed-forward networks of integrate-and-fire neurons William F. Podlaski, Tim P. Vogels P37 Phenomenological neural model for adaptation of neurons in area IT Martin Giese, Pradeep Kuravi, Rufin Vogels P38 ICGenealogy: towards a common topology of neuronal ion channel function and genealogy in model and experiment Alexander Seeholzer, William Podlaski, Rajnish Ranjan, Tim Vogels P39 Temporal input discrimination from the interaction between dynamic synapses and neural subthreshold oscillations Joaquin J. Torres, Fabiano Baroni, Roberto Latorre, Pablo Varona P40 Different roles for transient and sustained activity during active visual processing Bart Gips, Eric Lowet, Mark J. Roberts, Peter de Weerd, Ole Jensen, Jan van der Eerden P41 Scale-free functional networks of 2D Ising model are highly robust against structural defects: neuroscience implications Abdorreza Goodarzinick, Mohammad D. Niry, Alireza Valizadeh P42 High frequency neuron can facilitate propagation of signal in neural networks Aref Pariz, Shervin S. Parsi, Alireza Valizadeh P43 Investigating the effect of Alzheimer’s disease related amyloidopathy on gamma oscillations in the CA1 region of the hippocampus Julia M. Warburton, Lucia Marucci, Francesco Tamagnini, Jon Brown, Krasimira Tsaneva-Atanasova P44 Long-tailed distributions of inhibitory and excitatory weights in a balanced network with eSTDP and iSTDP Florence I. Kleberg, Jochen Triesch P45 Simulation of EMG recording from hand muscle due to TMS of motor cortex Bahar Moezzi, Nicolangelo Iannella, Natalie Schaworonkow, Lukas Plogmacher, Mitchell R. Goldsworthy, Brenton Hordacre, Mark D. McDonnell, Michael C. Ridding, Jochen Triesch P46 Structure and dynamics of axon network formed in primary cell culture Martin Zapotocky, Daniel Smit, Coralie Fouquet, Alain Trembleau P47 Efficient signal processing and sampling in random networks that generate variability Sakyasingha Dasgupta, Isao Nishikawa, Kazuyuki Aihara, Taro Toyoizumi P48 Modeling the effect of riluzole on bursting in respiratory neural networks Daniel T. Robb, Nick Mellen, Natalia Toporikova P49 Mapping relaxation training using effective connectivity analysis Rongxiang Tang, Yi-Yuan Tang P50 Modeling neuron oscillation of implicit sequence learning Guangsheng Liang, Seth A. Kiser, James H. Howard, Jr., Yi-Yuan Tang P51 The role of cerebellar short-term synaptic plasticity in the pathology and medication of downbeat nystagmus Julia Goncharenko, Neil Davey, Maria Schilstra, Volker Steuber P52 Nonlinear response of noisy neurons Sergej O. Voronenko, Benjamin Lindner P53 Behavioral embedding suggests multiple chaotic dimensions underlie C. elegans locomotion Tosif Ahamed, Greg Stephens P54 Fast and scalable spike sorting for large and dense multi-electrodes recordings Pierre Yger, Baptiste Lefebvre, Giulia Lia Beatrice Spampinato, Elric Esposito, Marcel Stimberg et Olivier Marre P55 Sufficient sampling rates for fast hand motion tracking Hansol Choi, Min-Ho Song P56 Linear readout of object manifolds SueYeon Chung, Dan D. Lee, Haim Sompolinsky P57 Differentiating models of intrinsic bursting and rhythm generation of the respiratory pre-Bötzinger complex using phase response curves Ryan S. Phillips, Jeffrey Smith P58 The effect of inhibitory cell network interactions during theta rhythms on extracellular field potentials in CA1 hippocampus Alexandra Pierri Chatzikalymniou, Katie Ferguson, Frances K. Skinner P59 Expansion recoding through sparse sampling in the cerebellar input layer speeds learning N. Alex Cayco Gajic, Claudia Clopath, R. Angus Silver P60 A set of curated cortical models at multiple scales on Open Source Brain Padraig Gleeson, Boris Marin, Sadra Sadeh, Adrian Quintana, Matteo Cantarelli, Salvador Dura-Bernal, William W. Lytton, Andrew Davison, R. Angus Silver P61 A synaptic story of dynamical information encoding in neural adaptation Luozheng Li, Wenhao Zhang, Yuanyuan Mi, Dahui Wang, Si Wu P62 Physical modeling of rule-observant rodent behavior Youngjo Song, Sol Park, Ilhwan Choi, Jaeseung Jeong, Hee-sup Shin P64 Predictive coding in area V4 and prefrontal cortex explains dynamic discrimination of partially occluded shapes Hannah Choi, Anitha Pasupathy, Eric Shea-Brown P65 Stability of FORCE learning on spiking and rate-based networks Dongsung Huh, Terrence J. Sejnowski P66 Stabilising STDP in striatal neurons for reliable fast state recognition in noisy environments Simon M. Vogt, Arvind Kumar, Robert Schmidt P67 Electrodiffusion in one- and two-compartment neuron models for characterizing cellular effects of electrical stimulation Stephen Van Wert, Steven J. Schiff P68 STDP improves speech recognition capabilities in spiking recurrent circuits parameterized via differential evolution Markov Chain Monte Carlo Richard Veale, Matthias Scheutz P69 Bidirectional transformation between dominant cortical neural activities and phase difference distributions Sang Wan Lee P70 Maturation of sensory networks through homeostatic structural plasticity Júlia Gallinaro, Stefan Rotter P71 Corticothalamic dynamics: structure, number of solutions and stability of steady-state solutions in the space of synaptic couplings Paula Sanz-Leon, Peter A. Robinson P72 Optogenetic versus electrical stimulation of the parkinsonian basal ganglia. Computational study Leonid L. Rubchinsky, Chung Ching Cheung, Shivakeshavan Ratnadurai-Giridharan P73 Exact spike-timing distribution reveals higher-order interactions of neurons Safura Rashid Shomali, Majid Nili Ahmadabadi, Hideaki Shimazaki, S. Nader Rasuli P74 Neural mechanism of visual perceptual learning using a multi-layered neural network Xiaochen Zhao, Malte J. Rasch P75 Inferring collective spiking dynamics from mostly unobserved systems Jens Wilting, Viola Priesemann P76 How to infer distributions in the brain from subsampled observations Anna Levina, Viola Priesemann P77 Influences of embedding and estimation strategies on the inferred memory of single spiking neurons Lucas Rudelt, Joseph T. Lizier, Viola Priesemann P78 A nearest-neighbours based estimator for transfer entropy between spike trains Joseph T. Lizier, Richard E. Spinney, Mikail Rubinov, Michael Wibral, Viola Priesemann P79 Active learning of psychometric functions with multinomial logistic models Ji Hyun Bak, Jonathan Pillow P81 Inferring low-dimensional network dynamics with variational latent Gaussian process Yuan Zaho, Il Memming Park P82 Computational investigation of energy landscapes in the resting state subcortical brain network Jiyoung Kang, Hae-Jeong Park P83 Local repulsive interaction between retinal ganglion cells can generate a consistent spatial periodicity of orientation map Jaeson Jang, Se-Bum Paik P84 Phase duration of bistable perception reveals intrinsic time scale of perceptual decision under noisy condition Woochul Choi, Se-Bum Paik P85 Feedforward convergence between retina and primary visual cortex can determine the structure of orientation map Changju Lee, Jaeson Jang, Se-Bum Paik P86 Computational method classifying neural network activity patterns for imaging data Min Song, Hyeonsu Lee, Se-Bum Paik P87 Symmetry of spike-timing-dependent-plasticity kernels regulates volatility of memory Youngjin Park, Woochul Choi, Se-Bum Paik P88 Effects of time-periodic coupling strength on the first-spike latency dynamics of a scale-free network of stochastic Hodgkin-Huxley neurons Ergin Yilmaz, Veli Baysal, Mahmut Ozer P89 Spectral properties of spiking responses in V1 and V4 change within the trial and are highly relevant for behavioral performance Veronika Koren, Klaus Obermayer P90 Methods for building accurate models of individual neurons Daniel Saska, Thomas Nowotny P91 A full size mathematical model of the early olfactory system of honeybees Ho Ka Chan, Alan Diamond, Thomas Nowotny P92 Stimulation-induced tuning of ongoing oscillations in spiking neural networks Christoph S. Herrmann, Micah M. Murray, Silvio Ionta, Axel Hutt, Jérémie Lefebvre P93 Decision-specific sequences of neural activity in balanced random networks driven by structured sensory input Philipp Weidel, Renato Duarte, Abigail Morrison P94 Modulation of tuning induced by abrupt reduction of SST cell activity Jung H. Lee, Ramakrishnan Iyer, Stefan Mihalas P95 The functional role of VIP cell activation during locomotion Jung H. Lee, Ramakrishnan Iyer, Christof Koch, Stefan Mihalas P96 Stochastic inference with spiking neural networks Mihai A. Petrovici, Luziwei Leng, Oliver Breitwieser, David Stöckel, Ilja Bytschok, Roman Martel, Johannes Bill, Johannes Schemmel, Karlheinz Meier P97 Modeling orientation-selective electrical stimulation with retinal prostheses Timothy B. Esler, Anthony N. Burkitt, David B. Grayden, Robert R. Kerr, Bahman Tahayori, Hamish Meffin P98 Ion channel noise can explain firing correlation in auditory nerves Bahar Moezzi, Nicolangelo Iannella, Mark D. McDonnell P99 Limits of temporal encoding of thalamocortical inputs in a neocortical microcircuit Max Nolte, Michael W. Reimann, Eilif Muller, Henry Markram P100 On the representation of arm reaching movements: a computational model Antonio Parziale, Rosa Senatore, Angelo Marcelli P101 A computational model for investigating the role of cerebellum in acquisition and retention of motor behavior Rosa Senatore, Antonio Parziale, Angelo Marcelli P102 The emergence of semantic categories from a large-scale brain network of semantic knowledge K. Skiker, M. Maouene P103 Multiscale modeling of M1 multitarget pharmacotherapy for dystonia Samuel A. Neymotin, Salvador Dura-Bernal, Alexandra Seidenstein, Peter Lakatos, Terence D. Sanger, William W. Lytton P104 Effect of network size on computational capacity Salvador Dura-Bernal, Rosemary J. Menzies, Campbell McLauchlan, Sacha J. van Albada, David J. Kedziora, Samuel Neymotin, William W. Lytton, Cliff C. Kerr P105 NetPyNE: a Python package for NEURON to facilitate development and parallel simulation of biological neuronal networks Salvador Dura-Bernal, Benjamin A. Suter, Samuel A. Neymotin, Cliff C. Kerr, Adrian Quintana, Padraig Gleeson, Gordon M. G. Shepherd, William W. Lytton P107 Inter-areal and inter-regional inhomogeneity in co-axial anisotropy of Cortical Point Spread in human visual areas Juhyoung Ryu, Sang-Hun Lee P108 Two bayesian quanta of uncertainty explain the temporal dynamics of cortical activity in the non-sensory areas during bistable perception Joonwon Lee, Sang-Hun Lee P109 Optimal and suboptimal integration of sensory and value information in perceptual decision making Hyang Jung Lee, Sang-Hun Lee P110 A Bayesian algorithm for phoneme Perception and its neural implementation Daeseob Lim, Sang-Hun Lee P111 Complexity of EEG signals is reduced during unconsciousness induced by ketamine and propofol Jisung Wang, Heonsoo Lee P112 Self-organized criticality of neural avalanche in a neural model on complex networks Nam Jung, Le Anh Quang, Seung Eun Maeng, Tae Ho Lee, Jae Woo Lee P113 Dynamic alterations in connection topology of the hippocampal network during ictal-like epileptiform activity in an in vitro rat model Chang-hyun Park, Sora Ahn, Jangsup Moon, Yun Seo Choi, Juhee Kim, Sang Beom Jun, Seungjun Lee, Hyang Woon Lee P114 Computational model to replicate seizure suppression effect by electrical stimulation Sora Ahn, Sumin Jo, Eunji Jun, Suin Yu, Hyang Woon Lee, Sang Beom Jun, Seungjun Lee P115 Identifying excitatory and inhibitory synapses in neuronal networks from spike trains using sorted local transfer entropy Felix Goetze, Pik-Yin Lai P116 Neural network model for obstacle avoidance based on neuromorphic computational model of boundary vector cell and head direction cell Seonghyun Kim, Jeehyun Kwag P117 Dynamic gating of spike pattern propagation by Hebbian and anti-Hebbian spike timing-dependent plasticity in excitatory feedforward network model Hyun Jae Jang, Jeehyun Kwag P118 Inferring characteristics of input correlations of cells exhibiting up-down state transitions in the rat striatum Marko Filipović, Ramon Reig, Ad Aertsen, Gilad Silberberg, Arvind Kumar P119 Graph properties of the functional connected brain under the influence of Alzheimer’s disease Claudia Bachmann, Simone Buttler, Heidi Jacobs, Kim Dillen, Gereon R. Fink, Juraj Kukolja, Abigail Morrison P120 Learning sparse representations in the olfactory bulb Daniel Kepple, Hamza Giaffar, Dima Rinberg, Steven Shea, Alex Koulakov P121 Functional classification of homologous basal-ganglia networks Jyotika Bahuguna,Tom Tetzlaff, Abigail Morrison, Arvind Kumar, Jeanette Hellgren Kotaleski P122 Short term memory based on multistability Tim Kunze, Andre Peterson, Thomas Knösche P123 A physiologically plausible, computationally efficient model and simulation software for mammalian motor units Minjung Kim, Hojeong Kim P125 Decoding laser-induced somatosensory information from EEG Ji Sung Park, Ji Won Yeon, Sung-Phil Kim P126 Phase synchronization of alpha activity for EEG-based personal authentication Jae-Hwan Kang, Chungho Lee, Sung-Phil Kim P129 Investigating phase-lags in sEEG data using spatially distributed time delays in a large-scale brain network model Andreas Spiegler, Spase Petkoski, Matias J. Palva, Viktor K. Jirsa P130 Epileptic seizures in the unfolding of a codimension-3 singularity Maria L. Saggio, Silvan F. Siep, Andreas Spiegler, William C. Stacey, Christophe Bernard, Viktor K. Jirsa P131 Incremental dimensional exploratory reasoning under multi-dimensional environment Oh-hyeon Choung, Yong Jeong P132 A low-cost model of eye movements and memory in personal visual cognition Yong-il Lee, Jaeseung Jeong P133 Complex network analysis of structural connectome of autism spectrum disorder patients Su Hyun Kim, Mir Jeong, Jaeseung Jeong P134 Cognitive motives and the neural correlates underlying human social information transmission, gossip Jeungmin Lee, Jaehyung Kwon, Jerald D. Kralik, Jaeseung Jeong P135 EEG hyperscanning detects neural oscillation for the social interaction during the economic decision-making Jaehwan Jahng, Dong-Uk Hwang, Jaeseung Jeong P136 Detecting purchase decision based on hyperfrontality of the EEG Jae-Hyung Kwon, Sang-Min Park, Jaeseung Jeong P137 Vulnerability-based critical neurons, synapses, and pathways in the Caenorhabditis elegans connectome Seongkyun Kim, Hyoungkyu Kim, Jerald D. Kralik, Jaeseung Jeong P138 Motif analysis reveals functionally asymmetrical neurons in C. elegans Pyeong Soo Kim, Seongkyun Kim, Hyoungkyu Kim, Jaeseung Jeong P139 Computational approach to preference-based serial decision dynamics: do temporal discounting and working memory affect it? Sangsup Yoon, Jaehyung Kwon, Sewoong Lim, Jaeseung Jeong P141 Social stress induced neural network reconfiguration affects decision making and learning in zebrafish Choongseok Park, Thomas Miller, Katie Clements, Sungwoo Ahn, Eoon Hye Ji, Fadi A. Issa P142 Descriptive, generative, and hybrid approaches for neural connectivity inference from neural activity data JeongHun Baek, Shigeyuki Oba, Junichiro Yoshimoto, Kenji Doya, Shin Ishii P145 Divergent-convergent synaptic connectivities accelerate coding in multilayered sensory systems Thiago S. Mosqueiro, Martin F. Strube-Bloss, Brian Smith, Ramon Huerta P146 Swinging networks Michal Hadrava, Jaroslav Hlinka P147 Inferring dynamically relevant motifs from oscillatory stimuli: challenges, pitfalls, and solutions Hannah Bos, Moritz Helias P148 Spatiotemporal mapping of brain network dynamics during cognitive tasks using magnetoencephalography and deep learning Charles M. Welzig, Zachary J. Harper P149 Multiscale complexity analysis for the segmentation of MRI images Won Sup Kim, In-Seob Shin, Hyeon-Man Baek, Seung Kee Han P150 A neuro-computational model of emotional attention René Richter, Julien Vitay, Frederick Beuth, Fred H. Hamker P151 Multi-site delayed feedback stimulation in parkinsonian networks Kelly Toppin, Yixin Guo P152 Bistability in Hodgkin–Huxley-type equations Tatiana Kameneva, Hamish Meffin, Anthony N. Burkitt, David B. Grayden P153 Phase changes in postsynaptic spiking due to synaptic connectivity and short term plasticity: mathematical analysis of frequency dependency Mark D. McDonnell, Bruce P. Graham P154 Quantifying resilience patterns in brain networks: the importance of directionality Penelope J. Kale, Leonardo L. Gollo P155 Dynamics of rate-model networks with separate excitatory and inhibitory populations Merav Stern, L. F. Abbott P156 A model for multi-stable dynamics in action recognition modulated by integration of silhouette and shading cues Leonid A. Fedorov, Martin A. Giese P157 Spiking model for the interaction between action recognition and action execution Mohammad Hovaidi Ardestani, Martin Giese P158 Surprise-modulated belief update: how to learn within changing environments? Mohammad Javad Faraji, Kerstin Preuschoff, Wulfram Gerstner P159 A fast, stochastic and adaptive model of auditory nerve responses to cochlear implant stimulation Margriet J. van Gendt, Jeroen J. Briaire, Randy K. Kalkman, Johan H. M. Frijns P160 Quantitative comparison of graph theoretical measures of simulated and empirical functional brain networks Won Hee Lee, Sophia Frangou P161 Determining discriminative properties of fMRI signals in schizophrenia using highly comparative time-series analysis Ben D. Fulcher, Patricia H. P. Tran, Alex Fornito P162 Emergence of narrowband LFP oscillations from completely asynchronous activity during seizures and high-frequency oscillations Stephen V. Gliske, William C. Stacey, Eugene Lim, Katherine A. Holman, Christian G. Fink P163 Neuronal diversity in structure and function: cross-validation of anatomical and physiological classification of retinal ganglion cells in the mouse Jinseop S. Kim, Shang Mu, Kevin L. Briggman, H. Sebastian Seung, the EyeWirers P164 Analysis and modelling of transient firing rate changes in area MT in response to rapid stimulus feature changes Detlef Wegener, Lisa Bohnenkamp, Udo A. Ernst P165 Step-wise model fitting accounting for high-resolution spatial measurements: construction of a layer V pyramidal cell model with reduced morphology Tuomo Mäki-Marttunen, Geir Halnes, Anna Devor, Christoph Metzner, Anders M. Dale, Ole A. Andreassen, Gaute T. Einevoll P166 Contributions of schizophrenia-associated genes to neuron firing and cardiac pacemaking: a polygenic modeling approach Tuomo Mäki-Marttunen, Glenn T. Lines, Andy Edwards, Aslak Tveito, Anders M. Dale, Gaute T. Einevoll, Ole A. Andreassen P167 Local field potentials in a 4 × 4 mm2 multi-layered network model Espen Hagen, Johanna Senk, Sacha J. van Albada, Markus Diesmann P168 A spiking network model explains multi-scale properties of cortical dynamics Maximilian Schmidt, Rembrandt Bakker, Kelly Shen, Gleb Bezgin, Claus-Christian Hilgetag, Markus Diesmann, Sacha Jennifer van Albada P169 Using joint weight-delay spike-timing dependent plasticity to find polychronous neuronal groups Haoqi Sun, Olga Sourina, Guang-Bin Huang, Felix Klanner, Cornelia Denk P170 Tensor decomposition reveals RSNs in simulated resting state fMRI Katharina Glomb, Adrián Ponce-Alvarez, Matthieu Gilson, Petra Ritter, Gustavo Deco P171 Getting in the groove: testing a new model-based method for comparing task-evoked vs resting-state activity in fMRI data on music listening Matthieu Gilson, Maria AG Witek, Eric F. Clarke, Mads Hansen, Mikkel Wallentin, Gustavo Deco, Morten L. Kringelbach, Peter Vuust P172 STochastic engine for pathway simulation (STEPS) on massively parallel processors Guido Klingbeil, Erik De Schutter P173 Toolkit support for complex parallel spatial stochastic reaction–diffusion simulation in STEPS Weiliang Chen, Erik De Schutter P174 Modeling the generation and propagation of Purkinje cell dendritic spikes caused by parallel fiber synaptic input Yunliang Zang, Erik De Schutter P175 Dendritic morphology determines how dendrites are organized into functional subunits Sungho Hong, Akira Takashima, Erik De Schutter P176 A model of Ca2+/calmodulin-dependent protein kinase II activity in long term depression at Purkinje cells Criseida Zamora, Andrew R. Gallimore, Erik De Schutter P177 Reward-modulated learning of population-encoded vectors for insect-like navigation in embodied agents Dennis Goldschmidt, Poramate Manoonpong, Sakyasingha Dasgupta P178 Data-driven neural models part II: connectivity patterns of human seizures Philippa J. Karoly, Dean R. Freestone, Daniel Soundry, Levin Kuhlmann, Liam Paninski, Mark Cook P179 Data-driven neural models part I: state and parameter estimation Dean R. Freestone, Philippa J. Karoly, Daniel Soundry, Levin Kuhlmann, Mark Cook P180 Spectral and spatial information processing in human auditory streaming Jaejin Lee, Yonatan I. Fishman, Yale E. Cohen P181 A tuning curve for the global effects of local perturbations in neural activity: Mapping the systems-level susceptibility of the brain Leonardo L. Gollo, James A. Roberts, Luca Cocchi P182 Diverse homeostatic responses to visual deprivation mediated by neural ensembles Yann Sweeney, Claudia Clopath P183 Opto-EEG: a novel method for investigating functional connectome in mouse brain based on optogenetics and high density electroencephalography Soohyun Lee, Woo-Sung Jung, Jee Hyun Choi P184 Biphasic responses of frontal gamma network to repetitive sleep deprivation during REM sleep Bowon Kim, Youngsoo Kim, Eunjin Hwang, Jee Hyun Choi P185 Brain-state correlate and cortical connectivity for frontal gamma oscillations in top-down fashion assessed by auditory steady-state response Younginha Jung, Eunjin Hwang, Yoon-Kyu Song, Jee Hyun Choi P186 Neural field model of localized orientation selective activation in V1 James Rankin, Frédéric Chavane P187 An oscillatory network model of Head direction and Grid cells using locomotor inputs Karthik Soman, Vignesh Muralidharan, V. Srinivasa Chakravarthy P188 A computational model of hippocampus inspired by the functional architecture of basal ganglia Karthik Soman, Vignesh Muralidharan, V. Srinivasa Chakravarthy P189 A computational architecture to model the microanatomy of the striatum and its functional properties Sabyasachi Shivkumar, Vignesh Muralidharan, V. Srinivasa Chakravarthy P190 A scalable cortico-basal ganglia model to understand the neural dynamics of targeted reaching Vignesh Muralidharan, Alekhya Mandali, B. Pragathi Priyadharsini, Hima Mehta, V. Srinivasa Chakravarthy P191 Emergence of radial orientation selectivity from synaptic plasticity Catherine E. Davey, David B. Grayden, Anthony N. Burkitt P192 How do hidden units shape effective connections between neurons? Braden A. W. Brinkman, Tyler Kekona, Fred Rieke, Eric Shea-Brown, Michael Buice P193 Characterization of neural firing in the presence of astrocyte-synapse signaling Maurizio De Pittà, Hugues Berry, Nicolas Brunel P194 Metastability of spatiotemporal patterns in a large-scale network model of brain dynamics James A. Roberts, Leonardo L. Gollo, Michael Breakspear P195 Comparison of three methods to quantify detection and discrimination capacity estimated from neural population recordings Gary Marsat, Jordan Drew, Phillip D. Chapman, Kevin C. Daly, Samual P. Bradley P196 Quantifying the constraints for independent evoked and spontaneous NMDA receptor mediated synaptic transmission at individual synapses Sat Byul Seo, Jianzhong Su, Ege T. Kavalali, Justin Blackwell P199 Gamma oscillation via adaptive exponential integrate-and-fire neurons LieJune Shiau, Laure Buhry, Kanishka Basnayake P200 Visual face representations during memory retrieval compared to perception Sue-Hyun Lee, Brandon A. Levy, Chris I. Baker P201 Top-down modulation of sequential activity within packets modeled using avalanche dynamics Timothée Leleu, Kazuyuki Aihara Q28 An auto-encoder network realizes sparse features under the influence of desynchronized vascular dynamics Ryan T. Philips, Karishma Chhabria, V. Srinivasa Chakravarthy 25th Annual Computational Neuroscience Meeting: CNS*2016 Seogwipo City, Jeju-do, South Korea 2-7 July 2016 http://www.cnsorg.org/cns-2016-jejuissue-copyright-statement© The Author(s) 2016 ==== Body A1 Functional advantages of cell-type heterogeneity in neural circuits Tatyana O. Sharpee1 1Computational Neurobiology Laboratory, The Salk Institute for Biological Studies, San Diego, CA, USA Correspondence: Tatyana O. Sharpee - sharpee@snl.salk.edu BMC Neuroscience 2016, 17(Suppl 1):A1 Neural circuits are notorious for the complexity of their organization. Part of this complexity is related to the number of different cell types that work together to encode stimuli. I will discuss theoretical results that point to functional advantages of splitting neural populations into subtypes, both in feedforward and recurrent networks. These results outline a framework for categorizing neuronal types based on their functional properties. Such classification scheme could augment classification schemes based on molecular, anatomical, and electrophysiological properties. A2 Mesoscopic modeling of propagating waves in visual cortex Alain Destexhe1,2 1UNIC, CNRS, Gif sur Yvette, France; 2The European Institute for Theoretical Neuroscience (EITN), Paris, France Correspondence: Alain Destexhe - destexhe@unic.cnrs-gif.fr BMC Neuroscience 2016, 17(Suppl 1):A2 Propagating waves are large-scale phenomena widely seen in the nervous system, in both anesthetized and awake or sleeping states. Recently, the presence of propagating waves at the scale of microns–millimeters was demonstrated in the primary visual cortex (V1) of macaque monkey. Using a combination of voltage-sensitive dye (VSD) imaging in awake monkey V1 and model-based analysis, we showed that virtually every visual input is followed by a propagating wave (Muller et al., Nat Comm 2014). The wave was confined within V1, and was consistent and repeatable for a given input. Interestingly, two propagating waves always interact in a suppressive fashion, and sum sublinearly. This is in agreement with the general suppressive effect seen in other circumstances in V1 (Bair et al., J Neurosci 2003; Reynaud et al., J Neurosci 2012). To investigate possible mechanisms for this suppression we have designed mean-field models to directly integrate the VSD experiments. Because the VSD signal is primarily caused by the summed voltage of all membranes, it represents an ideal case for mean-field models. However, usual mean-field models are based on neuronal transfer functions such as the well-known sigmoid function, or functions estimated from very simple models. Any error in the transfer function may result in wrong predictions by the corresponding mean-field model. To palliate this caveat, we have obtained semi-analytic forms of the transfer function of more realistic neuron models. We found that the same mathematical template can capture the transfer function for models such as the integrate-and-fire (IF) model, the adaptive exponential (AdEx) model, up to Hodgkin–Huxley (HH) type models, all with conductance-based inputs. Using these transfer functions we have built “realistic” mean-field models for networks with two populations of neurons, the regular-spiking (RS) excitatory neurons, showing spike frequency adaptation, and the fast-spiking (FS) inhibitory neurons. This mean-field model can reproduce the propagating waves in V1, due to horizontal interactions, as shown previously using IF networks. This mean-field model also reproduced the suppressive interactions between propagating waves. The mechanism of suppression was based on the preferential recruitment of inhibitory cells over excitatory cells by afferent activity, which acted through the conductance-based shunting effect of the two waves onto one another. The suppression was negligible in networks with identical models for excitatory and inhibitory cells (such as IF networks). This suggests that the suppressive effect is a general phenomenon due to the higher excitability of inhibitory neurons in cortex, in line with previous models (Ozeki et al., Neuron 2009). Work done in collaboration with Yann Zerlaut (UNIC) for modeling, Sandrine Chemla and Frederic Chavane (CNRS, Marseille) for in vivo experiments. Supported by CNRS and the European Commission (Human Brain Project). A3 Dynamics and biomarkers of mental disorders Mitsuo Kawato1 1ATR Computational Neuroscience Laboratories, 2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0288, Japan Correspondence: Mitsuo Kawato - kawato@hip.atr.co.jp BMC Neuroscience 2016, 17(Suppl 1):A3 Current diagnoses of mental disorders are made in a categorical way, as exemplified by DSM-5, but many difficulties have been encountered in such categorical regimes: the high percentage of comorbidities, usage of the same drug for multiple disorders, the lack of any validated animal model, and the situation where no epoch-making drug has been developed in the past 30 years. NIMH started RDoC (research domain criterion) to overcome these problems [1], and some successful results have been obtained, including common genetic risk loci [2] and common neuroanatomical changes for multiple disorders [3] as well as psychosis biotypes [4]. In contrast to the currently dominant molecular biology approach, which basically assumes one-to-one mapping between genes and disorders, I postulate the following dynamics-based view of psychiatric disorders. Our brain is a nonlinear dynamical system that can generate spontaneous spatiotemporal activities. The dynamical system is characterized by multiple stable attractors, only one of which corresponds to a healthy or typically developed state. The others are pathological states. The most promising research approach within the above dynamical view is to combine resting-state functional magnetic resonance imaging, machine learning, big data, and sophisticated neurofeedback. Yahata et al. developed an ASD biomarker using only 16/9730 functional connections, and it did not generalize to MDD or ADHD but moderately to schizophrenia [5]. Yamashita’s regression model of working memory ability from functional connections [6] generalized to schizophrenia and reproduced the severity of working-memory deficits of four psychiatric disorders (in preparation). With the further development of machine learning algorithms and accumulation of reliable datasets, we hope to obtain a comprehensive landscape of many psychiatric and neurodevelopmental disorders. Guided by this full-spectrum structure, a tailor-made neurofeedback therapy should be optimized for each patient [7]. ReferencesInsel T, Cuthbert B, Garvey M., et al. Research domain criteria (RDoC): toward a new classification framework for research on mental disorders. Am J Psychiatry. 2010;167:748–51. Cross-disorder group of the psychiatric genomics consortium: identification of risk loci with shared effects on five major psychiatric disorders: a genome-wide analysis. Lancet. 2013;381:1371–9. Goodkind M, et al. Identification of a common neurobiological substrate for mental illness. JAMA Psychiatry. 2015;72:305–15. Clementz BA, et al. Identification of distinct psychosis biotypes using brain-based biomarkers. Am J Psychiatry. 2016;173:373–84. Yahata N, Morimoto J, Hashimoto R, Lisi G, Shibata K, Kawakubo Y, Kuwabara H, Kuroda M, Yamada T, Megumi F, Imamizu H, Nanez JE, Takahashi H, Okamoto Y, Kasai K, Kato N, Sasaki Y, Watanabe T, Kawato M: A small number of abnormal brain connections predicts adult autism spectrum disorder. Nature Commun. 2016;7:11254. doi:10.1038/ncomms11254. Yamashita M, Kawato M, Imamizu H. Predicting learning plateau of working memory from whole-brain intrinsic network connectivity patterns. Sci Rep. 2015;5(7622). doi:10.1038/srep07622. ATR Brain Information Communication Research Laboratory Group. DecNef Project. Available at http://www.cns.atr.jp/decnefpro/ (2016). F1 Precise recruitment of spiking output at theta frequencies requires dendritic h-channels in multi-compartment models of oriens-lacunosum/moleculare hippocampal interneurons Vladislav Sekulić1,2, Frances K. Skinner1,2,3 1Krembil Research Institute, University Health Network, Toronto, Ontario, Canada, M5T 2S8; 2Department of Physiology, University of Toronto, Toronto, Ontario, Canada, M5S 1A8; 3 Department of Medicine (Neurology), University of Toronto, Toronto, Ontario, Canada, M5T 2S8 Correspondence: Vladislav Sekulić - vlad.sekulic@utoronto.ca BMC Neuroscience 2016, 17(Suppl 1):F1 The theta rhythm (4–12 Hz) is a prominent network oscillation observed in the mammalian hippocampus and is correlated with spatial navigation and mnemonic processing. Inhibitory interneurons of the hippocampus fire action potentials at specific phases of the theta rhythm, pointing to distinct functional roles of interneurons in shaping this rhythmic activity. One hippocampal interneuron type, the oriens-lacunosum/moleculare (O-LM) cell, provides direct feedback inhibition and regulation of pyramidal cell activity in the CA1 region. O-LM cells express the hyperpolarization-activated, mixed-cation current (Ih) and, in vitro, demonstrate spontaneous firing at theta that is impaired upon blockade of Ih. Work using dynamic clamp has shown that in the presence of frequency-modulated artificial synaptic inputs, O-LM cells exhibit a spiking resonance at theta frequencies that is not dependent on Ih [1]. However, due to the somatic injection limitation of dynamic clamp, the study could not examine the potential contributions of putative dendritic Ih or the integration of dendritically-located synaptic inputs. To overcome this, we have used a database of previously developed multi-compartment computational models of O-LM cells [2]. We situated our OLM cell models in an in vivo-like context by injecting Poisson-based synaptic background activities throughout their dendritic arbors. Excitatory and inhibitory synaptic weights were tuned to produce similar baseline activity prior to modulation of the inhibitory synaptic process at various frequencies (2–30 Hz). We found that models with dendritic inputs expressed enhanced resonant firing at theta frequencies compared to models with somatic inputs. We then performed detailed analyses on the outputs of the models with dendritic inputs to further elucidate these results with respect to Ih distributions. The ability of the models to be recruited at the modulated input frequencies was quantified using the rotation number, or average number of spikes across all input cycles. Models with somatodendritic Ih were recruited at >50 % of the input cycles for a wider range of theta frequencies (3–9 Hz) compared to models with somatic Ih only (3–4 Hz). Models with somatodendritic Ih also exhibited a wider range of theta frequencies for which phase-locked output (vector strength >0.75) was observed (4–12 Hz), compared to models with somatic Ih (3–5 Hz). Finally, the phase of firing of models with somatodendritic Ih given 8–10 Hz modulated input was delayed 180–230° relative to the time of release from inhibitory synaptic input. O-LM cells receive phasic inhibitory inputs at theta frequencies from a subpopulation of parvalbumin-positive GABAergic interneurons in the medial septum (MS) timed to the peak of hippocampal theta, as measured in the stratum pyramidale layer [3]. Furthermore, O-LM cells fire at the trough of hippocampal pyramidal layer theta in vivo [4], an approximate 180˚ phase delay from the MS inputs, corresponding to the phase delay in our models with somatodendritic Ih. Our results suggest that, given dendritic synaptic inputs, O-LM cells require somatodendritic Ih channel expression to be precisely recruited during the trough of hippocampal theta activity. Our strategy of leveraging model databases that encompass experimental cell type-specificity and variability allowed us to reveal critical biophysical factors that contribute to neuronal function within in vivo-like contexts. Acknowledgements: Supported by NSERC of Canada, an Ontario Graduate Scholarship, and the SciNet HPC Consortium. ReferencesKispersky TJ, Fernandez FR, Economo MN, White JA. Spike resonance properties in hippocampal O-LM cells are dependent on refractory dynamics. J Neurosci. 2012;32(11):3637–51. Sekulić V, Lawrence JJ, Skinner FK. Using multi-compartment ensemble modeling as an investigative tool of spatially distributed biophysical balances: application to hippocampal oriens-lacunosum/moleculare (O-LM) cells. PLOS One. 2014;9(10):e106567. Borhegyi Z, Varga V, Szilágyi, Fabo D, Freund TF. Phase segregation of medial septal GABAergic neurons during hippocampal theta activity. J Neurosci. 2004;24(39):8470–9. Varga C, Golshani P, Soltesz I. Frequency-invariant temporal ordering of interneuronal discharges during hippocampal oscillations in awake mice. Proc Natl Acad Sci USA. 2012;109(40):E2726–34. F2 Kernel methods in reconstruction of current sources from extracellular potentials for single cells and the whole brains Daniel K. Wójcik1, Chaitanya Chintaluri1, Dorottya Cserpán2, Zoltán Somogyvári2 1Department of Neurophysiology, Nencki Institute of Experimental Biology, Warsaw, Poland; 2Department of Theory, Wigner Research Centre for Physics of the Hungarian Academy of Sciences, Budapest, H-1121, Hungary Correspondence: Daniel K. Wójcik - d.wojcik@nencki.gov.pl BMC Neuroscience 2016, 17(Suppl 1):F2 Extracellular recordings of electric potential, with a century old history, remain a popular tool for investigations of brain activity on all scales, from single neurons, through populations, to the whole brains, in animals and humans, in vitro and in vivo [1]. The specific information available in the recording depends on the physical settings of the system (brain + electrode). Smaller electrodes are usually more selective and are used to capture local information (spikes from single cells or LFP from populations) while larger electrodes are used for subdural recordings (on the cortex, ECoG), on the scalp (EEG) but also as depth electrodes in humans (called SEEG). The advantages of extracellular electric potential are the ease of recording and its stability. Its problem is interpretation: since electric field is long range one can observe neural activity several millimeters from its source [2–4]. As a consequence every recording reflects activity of many cells, populations and regions, depending on which level we focus. One way to overcome this problem is to reconstruct the distribution of current sources (CSD) underlying the measurement [5], typically done to identify activity on systems level from multiple LFP on regular grids [6]. We recently proposed a kernel-based method of CSD estimation from multiple LFP recordings from arbitrarily placed probes (i.e. not necessarily on a grid) which we called kernel Current Source Density method (kCSD) [7]. In this overview we present the original proposition as well as two recent developments, skCSD (single cell kCSD) and kESI (kernel Electrophysiological Source Imaging). skCSD assumes that we know which part of the recorded signal comes from a given cell and we have access to the morphology of the cell. This could be achieved by patching a cell, driving it externally while recording the potential on a multielectrode array, injecting a dye, and reconstructing the morphology. In this case we know that the sources must be located on the cell and this information can be successfully used in estimation. In kESI we consider simultaneous recordings with subdural ECoG (strip and grid electrodes) and with depth electrodes (SEEG). Such recordings are taken on some epileptic patients prepared for surgical removal of epileptogenic zone. When MR scan of the patient head is taken and the positions of the electrodes are known as well as the brain’s shape, the idea of kCSD can be used to bound the possible distribution of sources facilitating localization of the foci. Acknowledgements: Polish Ministry for Science and Higher Education (grant 2948/7.PR/2013/2), Hungarian Scientific Research Fund (Grant OTKA K113147), National Science Centre, Poland (Grant 2015/17/B/ST7/04123). ReferencesBuzsáki G, Anastassiou CA, Koch C. The origin of extracellular fields and currents—EEG, ECoG, LFP and spikes. Nat Rev Neurosci. 2012;13:407–20. Hunt MJ, Falinska M, Łęski S, Wójcik DK, Kasicki S. Differential effects produced by ketamine on oscillatory activity recorded in the rat hippocampus, dorsal striatum and nucleus accumbens. J Psychopharmacol. 2011;25:808–21. Lindén H, Tetzlaff T, Potjans TC, Pettersen KH, Gruen S, Diesmann M, Einevoll GT. Modeling the spatial reach of the LFP. Neuron. 2011;72:859–72.. Łęski S, Lindén H, Tetzlaff T, Pettersen KH, Einevoll GT. Frequency dependence of signal power and spatial reach of the local field potential. PLoS Comput Biol. 2013;9:e1003137. Wójcik DK. Current source density (CSD) analysis. In: Jaeger D, Jung R, editors. Encyclopedia of computational neuroscience. SpringerReference. Berlin: Springer; 2013. Mitzdorf U. Current source-density method and application in cat cerebral cortex: investigation of evoked potentials and EEG phenomena. Physiol Rev. 1985;65:37–100. Potworowski J, Jakuczun W, Łęski S, Wójcik DK. Kernel current source density method. Neural Comput. 2012;24:541–75. F3 The synchronized periods depend on intracellular transcriptional repression mechanisms in circadian clocks Jae Kyoung Kim1, Zachary P. Kilpatrick2, Matthew R. Bennett3, Kresimir Josić2,4 1Department of Mathematical Sciences, KAIST, Daejoen 34141, Republic of Korea; 2Department of Mathematics, University of Houston, Houston, TX 77004, USA; 3Department of Biochemistry and Cell Biology and Institute of Biosciences and Bioengineering, Rice University, Houston, TX 77005, USA; 4Department of Biology and Biochemistry, University of Houston, Houston, TX 77004, USA Correspondence: Jae Kyoung Kim - jaekkim@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):F2 In mammals, circadian (~24 h) rhythms are mainly regulated by a master circadian clock located in the suprachiasmatic nucleus (SCN) [1]. The SCN consists of ~20,000 neurons, each of which generates own rhythms via intracellular transcriptional negative feedback loop involving PER-CRY and BMAL1-CLOCK. These individual rhythms of each neuron are synchronized through intercellular coupling via neurotransmitters including VIP [2]. In this talk, I will discuss that the synchronized periods via coupling signal strongly depend on the mechanism of intracellular transcription repression [3–4]. Specifically, using mathematical modeling and phase response curve analysis, we find that the synchronized period of SCN stays close to the population mean of cells’ intrinsic periods (~24 h) if transcriptional repression occurs via protein sequestration. However, the synchronized period is far from the population mean when repression occurs via Hill-type regulation (e.g. phosphorylation-based repression). These results reveal the novel relationship between two major functions of the SCN-intracellular rhythm generation and intercellular synchronization of rhythms. Furthermore, this relationship provides an explanation for why the protein sequestration is commonly used in circadian clocks of multicellular organisms, which have a coupled master clock, but not in unicellular organisms [4]. Acknowledgements: This work was funded by the National Institutes of Health, through the joint National Science Foundation/National Institute of General Medical Sciences Mathematical Biology Program grant No. R01GM104974 (to M.R.B. and K.J.), National Science Foundation grants Nos. DMS-1311755 (to Z.P.K.) and DMS-1122094 (to K.J.), the Robert A. Welch Foundation grant No. C-1729 (to M.R.B.), National Science Foundation grant No. DMS-0931642 to the Mathematical Biosciences Institute (to J.K.K.), KAIST Research Allowance Grant G04150020 (to J.K.K) and the TJ Park Science Fellowship of POSCO TJ Park Foundation G01160001 (to J.K.K). ReferencesDibner C, Schibler U, Albrecht U. The mammalian circadian timing system: organization and coordination of central and peripheral clocks. Annu Rev Physiol. 2010;72:517–49. Welsh DK, Takahashi JS, Kay SA. Suprachiasmatic nucleus: cell autonomy and network properties. Annu Rev Physiol. 2010;72:551. Kim JK, Kilpatrick ZP, Bennett MR, Josić K. Molecular mechanisms that regulate the coupled period of the mammalian circadian clock. Biophys J. 2014;106(9):2071–81. Kim JK. Protein sequestration vs Hill-type repression in circadian clock models (in revision). O1 Assessing irregularity and coordination of spiking-bursting rhythms in central pattern generators Irene Elices1, David Arroyo1, Rafael Levi1,2, Francisco B. Rodriguez1, Pablo Varona1 1Grupo de Neurocomputación Biológica, Dpto. de Ingeniería Informática, Escuela Politécnica Superior, Universidad Autónoma de Madrid, Spain; 2Department of Biological Sciences, University of Southern California, CA, USA Correspondence: Irene Elices - irene.elices@uam.es BMC Neuroscience 2016, 17(Suppl 1):O1 Found in all nervous systems, central pattern generators (CPGs) are neural circuits that produce flexible rhythmic motor patterns. Their robust and highly coordinated spatio-temporal activity is generated in the absence of rhythmic input. Several invertebrate CPGs are among the best known neural circuits, as their neurons and connections have been identified and mapped. The crustacean pyloric CPG is one of these flagship neural networks [1, 2]. Experimental and computational studies of CPGs typically examine their rhythmic output in periodic spiking-bursting regimes. Aiming to understand the fast rhythm negotiation of CPG neurons, here we present experimental and theoretical analyses of the pyloric CPG activity in situations where irregular yet coordinated rhythms are produced. In particular, we focus our study in the context of two sources of rhythm irregularity: intrinsic damage in the preparation, and irregularity induced by ethanol. The analysis of non-periodic regimes can unveil important properties of the robust dynamics controlling rhythm coordination in this system. Adult male and female shore crabs (Carcinus maenas) were used for the experimental recordings. The isolated stomatrogastric ganglion was kept in Carcinus maenas saline. Membrane potentials were recorded intracellularly from the LP and PD cells, two mutually inhibitory neurons that form a half-center oscillator in the pyloric CPG. Extracellular electrodes allowed monitoring the overall CPG rhythm. Conductance-based models of the pyloric CPG neurons and their associated graded synapses as described in [3, 4] were also used in this dual experimental and theoretical study. Irregularity and coordination of the CPG rhythms were analyzed using measures characterizing the cells’ instantaneous waveform, period, duty cycle, plateau, hyperpolarization and temporal structure of the spiking activity, as well as measures describing instantaneous phases among neurons in the irregular rhythms and their variability. Our results illustrate the strong robustness of the circuit to keep LP/PD phase relationships in intrinsic and induced irregularity conditions while allowing a large variety of burst waveforms, durations and hyperpolarization periods in these neurons. In spite of being electrically coupled to the pacemaker cell of the circuit, the PD neurons showed a wide flexibility to participate with larger burst durations in the CPG rhythm (and larger increase in variability), while the LP neuron was more restricted in sustaining long bursts in the conditions analyzed. The conductance-based models were used to explain the role of asymmetry in the dynamics of the neurons and synapses to shape the irregular activity observed experimentally. Taking into account the overall experimental and model analyses, we discuss the presence of preserved relationships in the non-periodic but coordinated bursting activity of the pyloric CPG, and their role in the fast rhythm negotiating properties of this circuit. Acknowledgements: We acknowledge support from MINECO DPI2015-65833-P, TIN2014-54580-R, TIN-2012-30883 and ONRG grant N62909-14-1-N279. ReferencesMarder E, Calabrese RL. Principles of rhythmic motor pattern generation. Physiol Rev. 1996;76:687–717. Selverston AI, Rabinovich MI, Abarbanel HDI, Elson R, Szücs A, Pinto RD, Huerta R, Varona P. Reliable circuits from irregular neurons: a dynamical approach to understanding central pattern generators. J Physiol. 2000;94:357–74. Latorre R, Rodríguez FB, Varona P. Neural signatures: multiple coding in spiking-bursting cells. Biol Cybern. 2006;95:169–83. Elices I, Varona P. Closed-loop control of a minimal central pattern generator network. Neurocomputing. 2015;170:55–62. O2 Regulation of top-down processing by cortically-projecting parvalbumin positive neurons in basal forebrain Eunjin Hwang1, Bowon Kim1,2, Hio-Been Han1,3, Tae Kim4, James T. McKenna5, Ritchie E. Brown5, Robert W. McCarley5, Jee Hyun Choi1,2 1Center for Neuroscience, Korea Institute of Science and Technology, Hwarang-ro 14-gil 5, Seongbuk-gu, Seoul 02792, South Korea; 2Department of Neuroscience, University of Science and Technology, 217 Gajeong-ro, Yuseong-gu, Daejon 34113, South Korea; 3Department of Psychology, Yonsei University, 50 Yonsei-ro, Seodaemun-gu, Seoul 03722, South Korea; 4Department of Psychiatry, Kyung Hee University Hospital at Gangdong, 892, Dongnam-ro, Gangdong-gu, Seoul 05278, South Korea; 5Department of Psychiatry, Veterans Administration Boston Healthcare System and Harvard Medical School, Brockton, MA 02301, USA Correspondence: Jee Hyun Choi - jeechoi@kist.re.kr BMC Neuroscience 2016, 17(Suppl 1):O2 Particular behaviors are associated with different spatio-temporal patterns of cortical EEG oscillations. A recent study suggests that the cortically-projecting, parvalbumin-positive (PV+) inhibitory neurons in the basal forebrain (BF) play an important role in the state-dependent control of cortical oscillations, especially ~40 Hz gamma oscillations [1]. However, the cortical topography of the gamma oscillations which are controlled by BF PV+ neurons and their relationship to behavior are unknown. Thus, in this study, we investigated the spatio-temporal patterns and the functional role of the cortical oscillations induced or entrained by BF PV+ neurons by combining optogenetic stimulation of BF PV+ neurons with high-density EEG [2, 3] in channelrhodopsin-2 (ChR2) transduced PV-cre mice. First, we recorded the spatio-temporal responses in the cortex with respect to the stimulation of BF PV+ neurons at various frequencies. The topographic response patterns were distinctively different depending on the stimulation frequencies, and most importantly, stimulation of BF PV+ neurons at 40 Hz (gamma band frequency) induced a preferential enhancement of gamma band oscillations in prefrontal cortex (PFC) with a statistically significant increase in intracortical connectivity within PFC. Second, optogenetic stimulation of BF PV+ neurons was applied while the mice were exposed to auditory stimuli (AS) at 40 Hz. The time delay between optogenetic stimulation and AS was tested and the phase response to the AS was characterized. We found that the phase responses to the click sound in PFC were modulated by the optogenetic stimulation of BF PV+ neurons. More specifically, the advanced activation of BF PV+ neurons by π/2 (6.25 ms) with respect to AS sharpened the phase response to AS in PFC, while the anti-phasic activation (π, 12.5 ms) blunted the phase response. Interestingly, like PFC, the primary auditory cortex (A1) also showed sharpened phase response for the π/2 advanced optogenetic BF PV+ neuron activation during AS. Considering that no direct influence of BF PV+ neurons on A1 was apparent in the response to stimulation of BF PV+ neurons alone, the sharpened phase response curve of A1 suggests a top-down influence of the PFC. This result implies that the BF PV+ neurons may participate in regulating the top-down influence that PFC exerts on primary sensory cortices during attentive behaviors, and supports the idea that the modulating activities of BF PV+ neurons might be a potential target for restoring top-down cognitive functions as well as abnormal frontal gamma oscillations associated with psychiatric disorders. Acknowledgements: This research was supported by the Department of Veterans Affairs, the Korean National Research Council of Science & Technology (No. CRC-15-04-KIST), NIMH R01 MH039683 and Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2015R1D1A1A01059119). The contents of this report do not represent the views of the US Department of Veterans Affairs or the United States government. ReferencesKim T, et al. Cortically projecting basal forebrain parvalbumin neurons regulate cortical gamma band oscillations. Proc Natl Acad Sci. 2015;112(11):3535–40. Choi JH, et al. High resolution electroencephalography in freely moving mice. J Neurophysiol .2010;104(3):1825–34. Lee M, et al. High-density EEG recordings of the freely moving mice using polyimide-based microelectrode. J Vis Exp. 2011;47. http://www.jove.com/details.php?id=2562. doi:10.3791/2562. O3 Modeling auditory stream segregation, build-up and bistability James Rankin1, Pamela Osborn Popp1, John Rinzel1,2 1Center for Neural Science, New York University, New York 10003, NY; 2Courant Institute of Mathematical Sciences, New York University, New York 10012, NY Correspondence: James Rankin - james.rankin@nyu.edu BMC Neuroscience 2016, 17(Suppl 1):O3 With neuromechanistic modelling and psychoacoustic experiments we study the perceptual dynamics of auditory streaming (cocktail party problem). The stimulus is a sequence of two interleaved tones, A and B in a repeating triplet pattern: ABA_ABA_ (‘_’ is a silent gap). Initially, subjects hear a single integrated pattern, but after some seconds they hear segregated A_A_A_ and _B___B__ streams (build-up of streaming segregation). For long presentations, build-up is followed by irregular alternations between integrated and segregated (auditory bistability). We recently presented [1] the first neuromechanistic model of auditory bistability; it incorporates common competition mechanisms of mutual inhibition, slow adaptation and noise [2]. Our competition network is formulated to reside downstream of primary auditory cortex (A1). Neural responses in macaque A1 to triplet sequences [3] encode stimulus features and provide the inputs to our network (Fig. 1A). In our model recurrent excitation with an NMDA-like timescale links responses across gaps between tones and between triplets. It captures the dynamics of perceptual alternations and the stimulus feature dependence of percept durations. To account for build-up we incorporate early adaptation of A1 responses [3] (Fig. 1B, upper). Early responses in A1 are broadly tuned and do not reflect the frequency difference between the tones; later responses show a clear tonotopic dependence. This adaptation biases the initial percept towards integration, but occurs faster (~0.5 s) than the gradual build-up process (~5–10 s). The low initial probability of segregation gradually builds up to the stable probability of later bistable alternations (Fig. 1B, lower). During build-up, a pause in presentation may cause partial reset to integrated [4]. Our extended model shows this behavior assuming that after a pause A1 responses recover on the timescale of early adaptation. Moreover, the modeling results agree with our psychoacoustic experiments (compare filled and open circles in Fig. 1B, lower).Fig. 1 A Model schematic: tone inputs IA and IB elicit pulsatile responses in A1, which are pooled as inputs to a three-population competition network. Central unit AB encodes integrated, peripheral units A and B encode segregated. Mutual inhibition between units and recurrent excitation are incorporated with adaptation and noise. B A1 inputs show early initial adaptation, also if a pause is present. Build-up function shows proportion segregated increasing over time, here shown for three tone-frequency differences, DF, with no pause (dashed) or with a pause (solid curves). Time-snapshots from model (filled circles) agree with data (empty circles with SEM error bars, N = 8) Conclusions For the first time, we offer an explanation of the discrepancy in the timescales of early A1 responses and the more gradual build-up process. Recovery of A1 responses can explain resetting for stimulus pauses. Our model offers, to date, the most complete account of the early and late dynamics for auditory streaming in the triplet paradigm. ReferencesRankin J, Sussman E, Rinzel J. Neuromechanistic model of auditory bistability. PLoS Comput Biol. 2015;11:e1004555. Shpiro A, Moreno-Bote R, Rubin N, Rinzel J. Balance between noise and adaptation in competition models of perceptual bistability. J Comp Neurosci. 2009;27:37–54. Micheyl C, Tian B, Carlyon R, Rauschecker J. Perceptual organization of tone sequences in the auditory cortex of awake macaques. Neuron. 2005;48:139–48. Beauvois MW, Meddis R. Time decay of auditory stream biasing. Percept Psychophys. 1997;59:81–6. O4 Strong competition between tonotopic neural ensembles explains pitch-related dynamics of auditory cortex evoked fields Alejandro Tabas1, André Rupp2,†, Emili Balaguer-Ballester1,3,† 1Faculty of Science and Technology, Bournemouth University, Bournemouth, England, UK; 2Heidelberg University, Baden-Württemberg, Germany; 3Bernstein Center for Computational Neuroscience, Heidelberg-Mannheim, Baden-Württemberg, Germany Correspondence: Alejandro Tabas - atabas@bournemouth.ac.uk †Equal contribution BMC Neuroscience 2016, 17(Suppl 1):O4 Auditory evoked fields (AEFs) observed in MEG experiments systematically present a transient deflection known as the N100 m, elicited around 100 ms after the tone onset in the antero-lateral Heschl’s Gyrus. The exact N100m’s latency is correlated with the perceived pitch of a wide range of stimulus [1, 2], suggesting that the transient component reflects the processing of pitch in auditory cortex. However, the biophysical substrate of such precise relationship remains an enigma. Existing models of pitch, focused on perceptual phenomena, did not explain the mechanism generating cortical evoked fields during pitch processing in biophysical detail. In this work, we introduce a model of interacting neural ensembles describing, for the first time to our knowledge, how cortical pitch processing gives rise to observed human neuromagnetic responses and why its latency strongly correlates with pitch. To provide a realistic cortical input, we used a recent model of the auditory periphery and realistic subcortical processing stages. Subcortical processing was based on a delay-and-multiply operation carried out in cochlear nucleus and inferior colliculus [3], resulting in realistic patterns of neural activation in response to the stimulus periodicities. Subcortical activation is transformed into a tonotopic receptive-field-like representation [4] by a novel cortical circuit composed by functional blocks characterised by a best frequency. Each block consist of an excitatory and an inhibitory population, modelled using mean-field approximations [5]. Blocks interact with each other through local AMPA- and NMDA-driven excitation and GABA-driven global inhibition [5]. The excitation-inhibition competition of the cortical model describes a general pitch processing mechanism that explains the N100m deflection as a transient state in the cortical dynamics. The deflection is rapidly triggered by a rise in the activity elicited by the subcortical input, peaks after the inhibition overcomes the input, and stabilises when model dynamics reach equilibrium, around 100 ms after onset. As a direct consequence of the connectivity structure among blocks, the time necessary for the system to reach equilibrium depends on the encoded pitch of the tone. The model quantitatively predicts observed latencies of the N100m in agreement with available empirical data [1, 2] in a series of stimuli (see Fig. 2), suggesting that the mechanism potentially accounts for the N100 m dynamics.Fig. 2 N100 m predictions in comparison with available data [1, 2] for a range of pure tones (A) and HCTs (B) ReferencesSeither-Preisler A, Patterson R, Krumbholz K, Seither S, Lütkenhöner B. Evidence of pitch processing in the N100 m component of the auditory evoked field. Hear Res. 2006;213(1–2):88–98. Roberts TP, Ferrari P, Stufflebeam SM, Poeppel D. Latency of the auditory evoked neuromagnetic field components: stimulus dependence and insights toward perception. J Clin Neurophysiol. 2000;17(2):114–29. Meddis R, O’Mard LP. Virtual pitch in a computational physiological model. J Acoust Soc Am. 2006;6:3861–9. Balaguer-Ballester E, Clark, N. Understanding pitch perception as a hierarchical process with top-down modulation. PLoS Comput Biol. 2009;5(3):e1000301. Wong K-F, Wang X-J. A recurrent network mechanism of time integration in perceptual decisions. J Neurosci. 2006;26(4):1314–28. O5 A simple model of retinal response to multi-electrode stimulation Matias I. Maturana1,2, David B. Grayden2,3, Shaun L. Cloherty4, Tatiana Kameneva2, Michael R. Ibbotson1,5, Hamish Meffin1,5 1National Vision Research Institute, Australian College of Optometry, 3053, Australia; 2NeuroEngineering Laboratory, Dept. Electrical & Electronic Eng., University of Melbourne, 3010, Australia; 3Centre for Neural Engineering, University of Melbourne, 3010, Australia; 4Department of Physiology, Monash University, 3800, Australia; 5ARC Centre of Excellence for Integrative Brain Function, Department Optometry and Vision Sciences, University of Melbourne, 3010, Australia Correspondence: Hamish Meffin - hmeffin@unimelb.edu.au BMC Neuroscience 2016, 17(Suppl 1):O5 Retinal implants can restore vision to patients suffering photoreceptor loss by stimulating surviving retinal ganglion cells (RGCs) via an array of microelectrodes implanted within the eye [1]. However, the acuity offered by existing devices is low, limiting the benefits to patients. Improvements may come by increasing the number of electrodes in new devices and providing patterned vision, which necessitates stimulation using multiple electrodes simultaneously. However, simultaneous stimulation poses a number of problems due to cross-talk between electrodes and uncertainty regarding the resulting activation pattern. Here, we present a model and methods for estimating the responses of RGCs to simultaneous electrical stimulation. Whole cell in vitro patch clamp recordings were obtained from 25 RGCs with various morphological types in rat retina. The retinae were placed onto an array of 20 stimulating electrodes. Biphasic current pulses with 500 µs phase duration and 50 µs interphase gap were applied simultaneously to all electrodes at a frequency of 10 Hz, with the amplitude of current on each electrode sampled independently from a Gaussian distribution. A linear-nonlinear model was fit to the responses of each RGC using spike-triggered covariance analyses on 80 % of the recorded data. The analysis revealed a single significant principle component corresponding to the electrical receptive field for each cell, with the second largest principle component having negligible effect on the neural response (Fig. 3a). This indicates that interactions between electrodes are approximately linear in their influence on the cells’ responses.Fig. 3 a Spike triggered covariance showing the full set of stimuli (black dots) projected onto the first two principle components. Stimuli causing a spike formed two clusters: net cathodic first pulses (blue) and net anodic first pulse (red). b Electrical receptive fields superimposed on the electrode array are shown for the cathodic first (blue) and anodic first clusters (red) Furthermore, the spike-triggered ensemble showed two clusters (red and blue in Fig. 3a) corresponding to stimulation that had a net effect that was either anodic first or cathodic first. The electrical receptive fields for both anodic first and cathodic first stimulation were highly similar (Fig. 3b). They consisted of a small number (1–4) of electrodes that were close to the cell body (green dot). The remaining 20 % of data were used to validate the model. The average model prediction root-mean-square error was 7 % over the 25 cells. The accuracy of the model indicates that the linear-nonlinear model is appropriate to describe the responses of RGCs to electrical stimulation. Acknowledgements: This research was supported by the Australian Research Council (ARC). MI, HM, and SC acknowledge support through the Centre of Excellence for Integrative Brain Function (CE140100007), TK through ARC Discovery Early Career Researcher Award (DE120102210) and HM and TK through the ARC Discovery Projects funding scheme (DP140104533). ReferenceHadjinicolaou AE, Meffin H, Maturana M, Cloherty SL, Ibbotson MR. Prosthetic vision: devices, patient outcomes and retinal research. Clin Exp Optom. 2015;98(5):395–410. O6 Noise correlations in V4 area correlate with behavioral performance in visual discrimination task Veronika Koren1,2, Timm Lochmann1,2, Valentin Dragoi3, Klaus Obermayer1,2 1Institute of Software Engineering and Theoretical Computer Science, Technische Universitaet Berlin, Berlin, 10587, Germany; 2 Bernstein Center for Computational Neuroscience Berlin, Humboldt-Universitaet zu Berlin, Berlin, 10115, Germany; 3Department of Neurobiology and Anatomy, University of Texas-Houston Medical School, Houston, TX 77030, USA Correspondence: Veronika Koren - veronika.koren@bccn-berlin.de BMC Neuroscience 2016, 17(Suppl 1):O6 Linking sensory coding and behavior is a fundamental question in neuroscience. We have addressed this issue in behaving monkey visual cortex (areas V1 and V4) while animals were trained to perform a visual discrimination task in which two successive images were either rotated with respect to each other or were the same. We hypothesized that the animal’s performance in the visual discrimination task depends on the quality of stimulus coding in visual cortex. We tested this hypothesis by investigating the functional relevance of neuronal correlations in areas V1 and V4 in relation to behavioral performance. We measured two types of correlations: noise (spike count) correlations and correlations in spike timing. Surprisingly, both methods showed that correct responses are associated with significantly higher correlations in V4, but not V1, during the delay period between the two stimuli. This suggests that pair-wise interactions during the spontaneous activity preceding the arrival of the stimulus sets the stage for subsequent stimulus processing and importantly influences behavioral performance. Experiments were conducted in 2 adult monkeys that were previously trained for the task. After 300 ms of fixation, the target stimulus, consisting of a naturalistic stimulus, is shown for 300 ms, and after a random delay period (500–1200 ms), a test stimulus is shown for 300 ms. The test can either be identical to the target stimulus (match) or rotated with respect to the target (non-match). Monkey responded by pressing a button and was rewarded for a correct response with fruit juice. Two linear arrays with 16 recording channels each were used to record population activity in areas V1 and V4. The difficulty of the task is calibrated individually to have 70 % correct responses on average. The analysis is conducted on non-match condition, comparing activity in trials with correct responses with trials where the monkey responded incorrectly. Noise correlations were assessed as pair-wise correlations of spike counts (method 1) and of spike timing (method 2). For method 1, z-scores of spike counts of binned spike trains are computed in individual trials. r_sc is computed as Pearson correlation coefficient of z-scores in all available trials, balanced across correct/incorrect condition. For the method 2, cross-correlograms were computed, from which the cross-correlograms from shuffled trials are subtracted. Resulting function was summed around zero lag and normalized with sum of autocorrelograms [1]. While firing rates of single units or of the population did not significantly change for correct and incorrect responses, noise correlations during the delay period were significantly higher in V4 pairs, computed with both r_sc method (p = 0.0005 in monkey 1, sign-rank test) and with r_ccg method (p = 0.0001 and p = 0.0280 in monkey 1 and 2, respectively, 50 ms integration window). This result is robust to changes in the length of the bin (method 1) and to the length of the summation window (method 2). In agreement with [2], we confirm the importance of spontaneous activity preceding the stimulus on performance and suggest that higher correlations in V4 might be beneficial for successful read-out and reliable transmission of the information downstream. ReferencesBair W, Zohary E, Newsome WT. Correlated firing in macaque visual area MT: time scales and relationship to behavior. J Neurosci. 2001; 21(5):1676–97. Gutnisky DA, Beaman CB, Lew SE, Dragoi V. Spontaneous fluctuations in visual cortical responses influence population coding accuracy. Cereb Cortex. 2016;1–19. Cohen MR, Maunsell JH. Attention improves performance primarily by reducing interneuronal correlations. Nat Neurosci. 2009;12(12):1594–1600. Nienborg HR, Cohen MR, Cumming BG. Decision-related activity in sensory neurons: correlations among neurons and with behavior. Annu Rev Neurosci. 2012;35:463–83. O7 Input-location dependent gain modulation in cerebellar nucleus neurons Maria Psarrou1, Maria Schilstra1, Neil Davey1, Benjamin Torben-Nielsen1, Volker Steuber1 Centre for Computer Science and Informatics Research, University of Hertfordshire, Hatfield, AL10 9AB, UK Correspondence: Maria Psarrou - m.psarrou@herts.ac.uk BMC Neuroscience 2016, 17(Suppl 1):O7 Gain modulation is a brain-wide principle of neuronal computation that describes how neurons integrate inputs from different presynaptic sources. A gain change is a multiplicative operation that is defined as a change in the sensitivity (or slope of the response amplitude) of a neuron to one set of inputs (driving input) which results from the activity of a second set of inputs (modulatory input) [1, 2]. Different cellular and network mechanisms have been proposed to underlie gain modulation [2–4]. It is well established that input features such as synaptic noise and plasticity can contribute to multiplicative gain changes [2–4]. However, the effect of neuronal morphology on gain modulation is relatively unexplored. Neuronal inputs to the soma and dendrites are integrated in a different manner: whilst dendritic saturation can introduce a strong non-linear relationship between dendritic excitation and somatic depolarization, the relationship between somatic excitation and depolarization is more linear. The non-linear integration of dendritic inputs can enhance the multiplicative effect of shunting inhibition in the presence of noise [3]. Neurons in the cerebellar nuclei (CN) provide the main gateway from the cerebellum to the rest of the brain. Understanding how inhibitory inputs from cerebellar Purkinje cells interact with excitatory inputs from mossy fibres to control output from the CN is at the center of understanding cerebellar computation. In the present study, we investigated the effect of inhibitory modulatory input on CN neuronal output when the excitatory driving input was delivered at different locations in the CN neuron. We used a morphologically realistic conductance based CN neuron model [5] and examined the change in output gain in the presence of distributed inhibitory input under two conditions: (a) when the excitatory input was confined to one compartment (the soma or a dendritic compartment) and, (b), when the excitatory input was distributed across particular dendritic regions at different distances from the soma. For both of these conditions, our results show that the arithmetic operation performed by inhibitory synaptic input depends on the location of the excitatory synaptic input. In the presence of distal dendritic excitatory inputs, the inhibitory input has a multiplicative effect on the CN neuronal output. In contrast, excitatory inputs at the soma or proximal dendrites close to the soma undergo additive operations in the presence of inhibitory input. Moreover, the amount of the multiplicative gain change correlates with the distance of the excitatory inputs from the soma, with increasing distances from the soma resulting in increased gain changes and decreased additive shifts along the input axis. These results indicate that the location of synaptic inputs affects in a systematic way whether the input undergoes a multiplicative or additive operation. ReferencesSalinas E, Sejnowski TJ. Gain modulation in the central nervous system: where behavior, neurophysiology, and computation meet. Neuroscientist. 2001;7(5):430–40. Silver RA. Neuronal arithmetic. Nat Rev Neurosci. 2010;11(7):474–89. Prescott SA, De Koninck Y. Gain control of firing rate by shunting inhibition: roles of synaptic noise and dendritic saturation. Proc Natl Acad Sci USA. 2003;100(4):2076–81. Rothman J, Cathala L, Steuber V, Silver RA. Synaptic depression enables neuronal gain control. Nature. 2009;475:1015–18. Steuber V, Schultheiss NW, Silver RA, De Schutter E, Jaeger D. Determinants of synaptic integration and heterogeneity in rebound firing explored with data-driven models of deep cerebellar nucleus cells. J Comput Neurosci. 2011;30(3):633–58. O8 Analytic solution of cable energy function for cortical axons and dendrites Huiwen Ju1, Jiao Yu2, Michael L. Hines3, Liang Chen4 and Yuguo Yu1 1School of Life Science and the Collaborative Innovation Center for Brain Science, Fudan University, Shanghai, 200438, China; 2Linyi Hospital of Traditional Chinese Medicine, 211 Jiefang Road, Lanshan, Linyi, Shandong Province, 276000, China; 3Department of Neuroscience, Yale University School of Medicine, New Haven, CT 06520, USA; 4Department of Neurosurgery, Huashan Hospital, Shanghai Medical College, Fudan University, Shanghai, China Correspondence: Yuguo Yu - yuyuguo@fudan.edu.cn BMC Neuroscience 2016, 17(Suppl 1):O8 Accurate estimation of action potential (AP)-related metabolic cost is essential for understanding energetic constraints on brain connections and signaling processes. Most previous energy estimates of the AP were obtained using the Na+-counting method [1, 2], which seriously limits accurate assessment of metabolic cost of ionic currents that underlie AP generation. Moreover, the effects of axonal geometry and ion channel distribution on energy consumption related to AP propagation have not been systematically investigated. To address these issues, we return to the cable theory [3] that underlies our HH-type cortical axon model [4], which was constructed based on experimental measurements. Based on the cable equation that describes how ion currents flow along the cable as well as analysis of the electrochemical energy in the equivalent circuit, we derived the electrochemical energy function for the cable model, ∂2E∂x∂t=INaV-VNa+IKV-VK+ILV-VL-12πaia∂V∂x=gNamaxm3hVx,t-VNa2+gKmaxn4Vx,t-VK2+gLVx,t-VL2+Ga∂V∂x2 where gNamax (in a range of 50–650 mS/cm2), gKmax (5–100 mS/cm2), and gL = 0.033 mS/cm2 are the maximal sodium, maximal potassium, and leak conductance per unit membrane area, respectively; and VNa = 60, VK = −90 VL = −70 mV are the reversal potentials of the sodium, potassium, and leak channels, respectively. The gate variables m, h, and n are dimensionless activation and inactivation variables, which describe the activation and inactivation processes of the sodium and potassium channels [4]. This equation describes the AP-related energy consumption rate per unit membrane area (cm2/s) at any axonal distance and any time. The individual terms on the right-hand side of the equation represent the contributions of the sodium, potassium, leak, and axial currents, respectively. Then we employed the cable energy function to calculate energy consumption for unbranched axons and axons with several degrees of branching (branching level, BL). Calculations based on this function distinguish between the contributions of each item toward total energy consumption. Our analytical approach predicts an inhomogeneous distribution of metabolic cost along an axon with either uniformly or nonuniformly distributed ion channels. The results show that the Na+-counting method severely underestimates energy cost in the cable model by 20–70 %. AP propagation along axons that differ in length may require over 15 % more energy per unit of axon area than that required by a point model. However, actual energy cost can vary greatly depending on axonal branching complexity, ion channel density distributions, and AP conduction states. We also infer that the metabolic rate (i.e. energy consumption rate) of cortical axonal branches as a function of spatial volume exhibits a 3/4 power law relationship. Acknowledgements: Dr. Yu thanks for the support from the National Natural Science Foundation of China (31271170, 31571070), Shanghai program of Professor of Special Appointment (Eastern Scholar SHH1140004). ReferencesAlle H, Roth A, Geiger JR. Energy-efficient action potentials in hippocampal mossy fibers. Science. 2009;325(5946):1405–8. Carter BC, Bean BP. Sodium entry during action potentials of mammalian neurons: incomplete inactivation and reduced metabolic efficiency in fast-spiking neurons. Neuron. 2009;64(6):898–909. Rall W. Cable theory for dendritic neurons. In: Methods in neuronal modeling. MIT Press; 1989. p. 9–92. Yu Y, Hill AP, McCormick DA. Warm body temperature facilitates energy efficient cortical action potentials. PLoS Comput Biol. 2012;8(4):e1002456. O9 C. elegans interactome: interactive visualization of Caenorhabditis elegans worm neuronal network Jimin Kim1, Will Leahy2, Eli Shlizerman1,3 1Department of Applied Mathematics, University of Washington, Seattle, WA 98195, USA; 2Amazon.com Inc., Seattle, WA 98108, USA; 3Department of Electrical Engineering, University of Washington, Seattle, WA 98195, USA Correspondence: Eli Shlizerman - shlizee@uw.edu BMC Neuroscience 2016, 17(Suppl 1):O9 Modeling neuronal systems involves incorporating the two layers: a static map of neural connections (connectome), and biophysical processes that describe neural responses and interactions. Such a model is called the ‘dynome’ of a neuronal system as it integrates a dynamical system with the static connectome. Being closer to reproducing the activity of a neuronal system, investigation of the dynome has more potential to reveal neuronal pathways of the network than the static connectome [1]. However, since the two layers of the dynome are considered simultaneously, novel tools have to be developed for the dynome studies. Here we present a visualization methodology, called `interactome’, that allows to explore the dynome of a neuronal system interactively and in real-time, by viewing the dynamics overlaid on a graph representation of the connectome. We apply our methodology to the nervous system of Caenorhabditis elegans (C. elegans) worm, which connectome is almost fully resolved [2], and a computational model of neural dynamics and interactions (gap and synaptic) based on biophysical experimental findings was recently introduced [3]. Integrated together, C. elegans dynome defines a unique set of neural dynamics of the worm. To visualize the dynome, we propose a dynamic force-directed graph layout of the connectome. The layout is implemented using D3 visualization platform [4], and is designed to communicate with an integrator of the dynome. The two-way communication protocol between the layout and the integrator allows for stimulating (injecting current) into any subset of neurons at any time point (Fig. 4B). It also allows for simultaneously viewing the response of the network on top of the layout visualized by resizing graph nodes (neurons) according to their voltage. In addition, we support structural changes in the connectome, such as ablation of neurons and connections.Fig. 4 A Visualization of C. elegans dynome, B communication diagram between the dynome and the layout, C snapshots of visualization of C. elegans during the PLM/AVB excitations (forward crawling) Our visualization and communication protocols thereby display the stimulated network in an interactive manner and permit to explore different regimes that the stimulations induce. Indeed, with the interactome we are able to recreate various experimental scenarios, such as stimulation of forward crawling (PLM/AVB neurons and/or ablation of AVB) and show that its visualization assists in identifying patterns of neurons in the stimulated network. As connectomes and dynomes of additional neuronal systems are being resolved, the interactome will enable exploring their functionality and inference to its underlying neural pathways [5]. ReferencesKopell NJ, Gritton HJ, Whittingon MA, Kramer MA. Beyond the connectome: the dynome. Neuron. 2014;83(6):1319–28. Varshney LR, Chen BL, Paniagua E, Hall DH, Chkolvski DB. Structural properties of the caenorhabditis elegans neuronal network. PLoS Comput Biol. 2011;7(2):e1001066. Kunert J, Shlizerman E, Kutz JN. Low-dimensional functionality of complex network dynamics: neurosensory integration in the Caenorhabditis elegans connectome. Phys Rev E. 2014;89(5):052805. Bostock M, Ogievetsky V, Heer J. D3 data-driven documents. IEEE. 2011;17(12):2301–9. Kim J, Leahy W, Shlizerman E. C. elegans interactome: interactive visualization of Caenorhabditis elegans worm neuronal network. 2016 (in submission). O10 Is the model any good? Objective criteria for computational neuroscience model selection Justas Birgiolas1, Richard C. Gerkin1, Sharon M. Crook1,2 1School of Life Science, Arizona State University, Tempe, AZ 85287, USA; 2School of Mathematical and Statistical Sciences, Arizona State University, Tempe, AZ, 85287, USA Correspondence: Justas Birgiolas - justas@asu.edu BMC Neuroscience 2016, 17(Suppl 1):O10 Objectively evaluating and selecting computational models of biological neurons is an ongoing challenge in the field. Models vary in morphological detail, channel mechanisms, and synaptic transmission implementations. We present the results of an automated method for evaluating computational models against property values obtained from published cell electrophysiology studies. Seven published deterministic models of olfactory bulb mitral cells were selected from ModelDB [1] and simulated using NEURON’s Python interface [2]. Passive and spike properties in response to step current stimulation pulses were computed using the NeuronUnit [3] package and compared to their respective, experimentally obtained means of olfactory bulb mitral cell properties found in the NeuroElectro database [4]. Results reveal that across all models, the resting potential and input resistance property means deviated the most from their experimentally measured means (Rinputttest p = 0.02, VrestWilcoxon-test p = 0.01). The time constant, spike half-width, spike amplitude, and spike threshold properties, in the order of decreasing average deviation, matched well with experimental data (p > 0.05) (Fig. 5 top).Fig. 5 The average deviations of models and cell electrophysiology properties as measured in multiples of the 95 % CI bounds of experimental data means. Dashed line represents 1 CI bound threshold. Top rows show average deviations across all models for each cell property. Bottom rows show deviations across all cell properties for each model In three models, the property deviations were, on average, outside the 95 % CI of the experimental means (Fig. 5 bottom), but these averages were not significant (t test p > 0.05). All other models were within the 95 % CI, while the model of Chen et al. had the lowest deviation [5]. Overall, the majority of these olfactory bulb mitral cell models display some properties that are not significantly different from their experimental means. However, the resting potential and input resistance properties significantly differ from the experimental values. We demonstrate that NeuronUnit provides an objective method for evaluating the fitness of computational neuroscience cell models against publicly available data. Acknowledgements: The work of JB, RG, and SMC was supported in part by R01MH1006674 from the National Institutes of Health. ReferencesHines ML, Morse T, Migliore M, Carnevale NT, Shepherd GM. ModelDB: a database to support computational neuroscience. J Comput Neurosci. 2004;17(1):7–11. Hines M, Davison AP, Muller E. NEURON and Python. Front Neuroinform. 2009;3:1. Omar C, Aldrich J, Gerkin RC. Collaborative infrastructure for test-driven scientific model validation. In: Companion proceedings of the 36th international conference on software engineering. ACM; 2014. p. 524–7. Tripathy SJ, Savitskaya J, Burton SD, Urban NN, Gerkin RC. NeuroElectro: a window to the world’s neuron electrophysiology data. Front Neuroinform. 2014;8. Chen WR, Shen GY, Shepherd GM, Hines ML, Midtgaard J. Multiple modes of action potential initiation and propagation in mitral cell primary dendrite. J Neurophysiol. 2002;88(5):2755–64. O11 Cooperation and competition of gamma oscillation mechanisms Atthaphon Viriyopase1,2,3, Raoul-Martin Memmesheimer1,3,4, and Stan Gielen1,2 1Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen (Medical Centre), The Netherlands; 2Department for Biophysics, Faculty of Science, Radboud University Nijmegen, The Netherlands; 3Department for Neuroinformatics, Faculty of Science, Radboud University Nijmegen, The Netherlands; 4Center for Theoretical Neuroscience, Columbia University, New York, NY, USA Correspondence: Atthaphon Viriyopase - a.viriyopase@science.ru.nl BMC Neuroscience 2016, 17(Suppl 1):O11 Two major mechanisms that underlie gamma oscillations are InterNeuronal Gamma (“ING”), which is related to tonic excitation of reciprocally coupled inhibitory interneurons (I-cells), and Pyramidal InternNeuron Gamma (“PING”), which is mediated by coupled populations of excitatory pyramidal cells (E-cells) and I-cells. ING and PING are thought to serve different biological functions. Using computer simulations and analytical methods, we [1] therefore investigate which mechanism (ING or PING) will dominate the dynamics of a network when ING and PING interact and how the dominant mechanism may switch. We find that ING and PING oscillations compete: The mechanism generating the higher oscillation frequency “wins”. It determines the frequency of the network oscillations and suppresses the other mechanism. The network oscillation frequency (green lines corresponding to the network topology given in Fig. 6C) corresponding to the network with type-I-phase-response-curve interneurons and type-II-phase-response-curve interneurons is plotted in Fig. 6D, E, respectively. We explain our simulation results by a theoretical model that allows a full theoretical analysis.Fig. 6 Oscillations in full and reduced networks of reciprocally coupled pyramidal cells and interneurons. A, B Illustrate topologies of reduced networks that generate “pure” ING and “pure” PING, respectively, while C highlights the topology of a “full” network that could in principle generate either ING or PING oscillations or mixtures of both. D, E Frequency of pure ING-rhythm generated by the reduced network in A (blue line), pure PING-rhythm generated by the reduced network in b (red line), and rhythms generated by the full network in C (green line) as a function of mean current to I-cells I0,I and as function of mean current to E-cells I0,E, respectively. D Results for networks with type-I interneurons while E shows results for networks with type-II interneurons. Pyramidal cells are modeled as type-I Hodgkin–Huxley neurons Our study suggests experimental approaches to decide whether oscillatory activity in networks of interacting excitatory and inhibitory neurons is dominated by ING or PING oscillations and whether the participating interneurons belong to class I or II. Consider as an example networks with type-I interneurons where the external drive to the E-cells, I0,E, is kept constant while the external drive to the I-cells, I0,I, is varied. For both ING and PING dominated oscillations the frequency of the rhythm increases when I0,I increases (cf. Fig. 6D). Observing such an increase does therefore not allow to determine the underlying mechanism. However, the absolute value of the first derivative of the frequency with respect to I0,I allows a distinction, as it is much smaller for PING than for ING (cf. Fig. 6D). In networks with type-II interneurons, the non-monotonic dependence near the ING-PING transition may be a characteristic hallmark to detect the oscillation character (and the interneuron type): Decrease (increase) of the frequency when increasing I0,E indicates ING (PING), cf. Fig. 6E. These theoretical predictions are in line with experimental evidence [2]. ReferencesViriyopase A, Memmesheimer RM, Gielen S. Cooperation and competition of gamma oscillation mechanisms. J Neurophysiol. 2016. Craig MT, McBain CJ. Fast gamma oscillations are generated intrinsically in CA1 without the involvement of fast-spiking basket cells. J Neurosci. 2015;35(8):3616–24. O12 A discrete structure of the brain waves Yuri Dabaghian1,2, Justin DeVito1, Luca Perotti3 1Department of Neurology Pediatrics, Baylor College of Medicine, Houston, TX 77030, USA; 2Department of Computational and Applied Mathematics, Rice University, Houston, TX, 77005, USA; 3Physics Department, Texas Southern University, 3100 Cleburne St, Houston, TX 77004, USA Correspondence: Yuri Dabaghian - dabaghian@rice.edu BMC Neuroscience 2016, 17(Suppl 1):O12 A physiological interpretation of the biological rhythms, e.g., of the local field potentials (LFP) depends on the mathematical and computational approaches used for its analysis. Most existing mathematical methods of the LFP studies are based on braking the signal into a combination of simpler components, e.g., into sinusoidal harmonics of Fourier analysis or into wavelets of the Wavelet Analysis. However, a common feature of all these methods is that their prime components are presumed from the onset, and the goal of the subsequent analysis reduces to identifying the combination that best reproduces the original signal. We propose a fundamentally new method, based on a number of deep theorems of complex function theory, in which the prime components of the signal are not presumed a priori, but discovered empirically [1]. Moreover, the new method is more flexible and more sensitive to the signal’s structure than the standard Fourier method. Applying this method reveals a fundamentally new structure in the hippocampal LFP signals in rats in mice. In particular, our results suggest that the LFP oscillations consist of a superposition of a small, discrete set of frequency modulated oscillatory processes, which we call “oscillons”. Since these structures are discovered empirically, we hypothesize that they may capture the signal’s actual physical structure, i.e., the pattern of synchronous activity in neuronal ensembles. Proving this hypothesis will help enormously to advance a principal, theoretical understanding of the neuronal synchronization mechanisms. We anticipate that it will reveal new information about the structure of the LFP and other biological oscillations, which should provide insights into the underlying physiological phenomena and the organization of brains states that are currently poorly understood, e.g., sleep and epilepsy. Acknowledgements: The work was supported by the NSF 1422438 grant and by the Houston Bioinformatics Endowment Fund. ReferencePerotti L, DeVito J, Bessis D, Dabaghian Y, Dabaghian Y, Brandt VL, Frank LM. Discrete spectra of brain rhythms (in submisison). O13 Direction-specific silencing of the Drosophila gaze stabilization system Anmo J. Kim1,†, Lisa M. Fenk1,†, Cheng Lyu1, Gaby Maimon1 1Laboratory of Integrative Brain Function, The Rockefeller University, New York, NY 10065, USA Correspondence: Anmo J. Kim - anmo.kim@gmail.com † Authors contributed equally BMC Neuroscience 2016, 17(Suppl 1):O13 Many animals, including insects and humans, stabilize the visual image projected onto their retina by following a rotating landscape with their head or eyes. This stabilization reflex, also called the optomotor response, can pose a problem, however, when the animal intends to change its gaze. To resolve this paradox, von Holst and Mittelstaedt proposed that a copy of the motor command, or efference copy, could be routed into the visual system to transiently silence this stabilization reflex when an animal changes its gaze [1]. Consistent with this idea, we recently demonstrated that a single identified neuron associated with the optomotor response receives silencing motor-related inputs during rapid flight turns, or saccades, in tethered, flying Drosophila [2]. Here, we expand on these results by comprehensively recording from a group of optomotor-mediating visual neurons in the fly visual system: three horizontal system (HS) and six vertical system (VS) cells. We found that the amplitude of motor-related inputs to each HS and VS cell correlates strongly with the strength of each cell’s visual sensitivity to rotational motion stimuli around the primary turn axis, but not to the other axes (Fig. 7). These results support the idea that flies send rotation-axis-specific efference copies to the visual system during saccades—silencing the stabilization reflex only for a specific axis, but leaving the others intact. This is important because saccades consist of stereotyped banked turns, which involve body rotations around all three primary axes of rotation. If the gaze stabilization system is impaired for only one of these axes, then the fly is expected to attempt to maintain gaze stability, through a combination of head and body movements, for the other two. This prediction is consistent with behavioral measurements of head and body kinematics during saccades in freely flying blow flies [3]. Together, these studies provide an integrative model of how efference copies counteract a specific aspect of visual feedback signals to tightly control the gaze stabilization system.Fig. 7 The amplitudes of saccade-related potentials (SRPs) to HS and VS cells are strongly correlated with each cell’s visual sensitivity to rightward yaw motion stimuli. A Experimental apparatus. B Maximal-intensity z-projections of the lobula plate to visualize HS- or VS-cell neurites that are marked by a GAL4 enhancer trap line. C, D The amplitude of saccade-related potentials (SRPs) were inversely correlated with visual responses, when measured under rightward yaw motion stimuli, but not under clockwise roll motion stimuli. Each sample point corresponds to each cell type. Error bars indicate SEM Referencesvon Holst E, Mittelstaedt H. The principle of reafference. Naturwissenschaften.1950;37:464–76. Kim AJ, Fitzgerald JK, Maimon G. Cellular evidence for efference copy in Drosophila visuomotor processing. Nat Neurosci. 2015;18:1247–55. Schilstra C, van Hateren JH. Stabilizing gaze in flying blowflies. Nature. 1998;395:654. O14 What does the fruit fly think about values? A model of olfactory associative learning Chang Zhao1, Yves Widmer2, Simon Sprecher2, Walter Senn1 1Department of Physiology, University of Bern, Bern, 3012, Switzerland; 2Department of Biology, University of Fribourg, Fribourg, 1700, Switzerland Correspondence: Chang Zhao - zhao@pyl.unibe.ch BMC Neuroscience 2016, 17(Suppl 1):O14 Associative learning in the fruit fly olfactory system has been studied from the molecular to the behavior level [1, 2]. Fruit flies are able to associate conditional stimuli such as odor with unconditional aversive stimuli such as electrical shocks, or appetitive stimuli such as sugar or water. The mushroom body in the fruit fly brain is considered to be crucial for olfactory learning [1, 2]. The behavioral experiments show that the learning can not be explained simply by an additive Hebbian (i.e. correlation-based) learning rule. Instead, it depends on the timing between the conditional and unconditional stimulus presentation. Yarali and colleagues suggested a dynamic model on the molecular level to explain event timing in associative learning [3]. Here, we present new experiments together with a simple phenomenological model for learning that shows that associative olfactory learning in the fruit fly represents value learning that is incompatible with Hebbian learning. In our model, the information of the conditional odor stimulus is conveyed by Kenyon cells from the projection neurons to the mushroom output neurons; the information of the unconditional shock stimulus is represented by dopaminergic neurons to the mushroom output neurons through direct or indirect pathways. The mushroom body output neurons encode the internal value (v) of the odor (o) by synaptic weights (w) that conveys the odor information, v = w∙o. The synaptic strength is updated according to the value learning rule, Δw = η(s − v)õ, where s represents the (internal) strength of the shock stimulus, õ represents the synaptic odor trace, and η is the learning rate. The value associated with the odor determines the probability of escaping from that odor. This simple model reproduces the behavioral data and shows that olfactory conditioning in the fruit fly is in fact value learning. In contrast to the prediction of Hebbian learning, the escape probability for repeated odor-shock pairings is much lower than the escape probability for a single pairing with a correspondingly stronger shock. ReferencesAso Y, Sitaraman D, Ichinose T, Kaun KR, Vogt K, Belliart-Gurin G, Plaais PY, Robie AA, Yamagata N, Schnaitmann C, Rowell WJ, Johnston RM, Ngo TB, Chen N, Korff W, Nitabach MN, Heberlein U, Preat T, Branson KM, Tanimoto H, Rubin GM: Mushroom body output neurons encode valence and guide memory-based action selection in Drosophila. ELife. 2014;3:e04580. Heisenberg M. Mushroom body memoir: from maps to models. Nat Rev Neurosci. 2003;4:266–75. Yarali A, Nehrkorn J, Tanimoto H, Herz AVM. Event timing in associative learning: from biochemical reaction dynamics to behavioural observations. PLoS One. 2012;7(3):e32885. O15 Effects of ionic diffusion on power spectra of local field potentials (LFP) Geir Halnes1, Tuomo Mäki-Marttunen2, Daniel Keller3, Klas H. Pettersen4,5,Ole A. Andreassen2, Gaute T. Einevoll1,6 1Department of Mathematical Sciences and Technology, Norwegian University of Life Sciences, Ås, Norway; 2NORMENT, Institute of Clinical Medicine, University of Oslo, Oslo, Norway; 3The Blue Brain Project, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland; 4Letten Centre and Glialab, Department of Molecular Medicine, Instotute of Basic Medical Sciences, University of Oslo, Oslo, Norway; 5Centre for Molecular Medicine Norway, University of Oslo, Oslo, Norway; 6Department of Physics, University of Oslo, Oslo, Norway Correspondence: Geir Halnes - geir.halnes@nmbu.no BMC Neuroscience 2016, 17(Suppl 1):O15 The local field potential (LFP) in the extracellular space (ECS) of the brain, is a standard measure of population activity in neural tissue. Computational models that simulate the relationship between the LFP and its underlying neurophysiological processes are commonly used in the interpretation such measurements. Standard methods, such as volume conductor theory [1], assume that ionic diffusion in the ECS has negligible impact on the LFP. This assumption could be challenged during endured periods of intense neural signalling, under which local ion concentrations in the ECS can change by several millimolars. Such concentration changes are indeed often accompanied by shifts in the ECS potential, which may be partially evoked by diffusive currents [2]. However, it is hitherto unclear whether putative diffusion-generated potential shifts are too slow to be picked up in LFP recordings, which typically use electrode systems with cut-off frequencies at ~0.1 Hz. To explore possible effects of diffusion on the LFP, we developed a hybrid simulation framework: (1) The NEURON simulator was used to compute the ionic output currents from a small population of cortical layer-5 pyramidal neurons [3]. The neural model was tuned so that simulations over ~100 s of biological time led to shifts in ECS concentrations by a few millimolars, similar to what has been seen in experiments [2]. (2) In parallel, a novel electrodiffusive simulation framework [4] was used to compute the resulting dynamics of the potential and ion concentrations in the ECS, accounting for the effect of electrical migration as well as diffusion. To explore the relative role of diffusion, we compared simulations where ECS diffusion was absent with simulations where ECS diffusion was included. Our key findings were: (i) ECS diffusion shifted the local potential by up to ~0.2 mV. (ii) The power spectral density (PSD) of the diffusion-evoked potential shifts followed a 1/f2 power law. (iii) Diffusion effects dominated the PSD of the ECS potential for frequencies up to ~10 Hz (Fig. 8). We conclude that for large, but physiologically realistic ECS concentration gradients, diffusion could affect the ECS potential well within the frequency range considered in recordings of the LFP.Fig. 8 Power spectrum of ECS potential in a simulation including ECS diffusion (blue line) and a simulation without ECS diffusion (red line). Units for frequency and power are Hz and mV2/Hz, respectively ReferencesHolt G, Koch C. Electrical interactions via the extracellular potential near cell bodies. J Comput Neurosci. 1999;6:169–84. Dietzel I, Heinemann U, Lux H. Relations between slow extracellular potential changes, glial potassium buffering, and electrolyte and cellular volume changes during neuronal hyperactivity in cat. Glia. 1989;2:25–44. Hay E, Hill S, Schürmann F, Markram H, Segev I. Models of neocortical layer 5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS Comput Biol. 2011;7(7):e1002107. Halnes G, Østby I, Pettersen KH, Omholt SW, Einevoll GT: Electrodiffusive model for astrocytic and neuronal ion concentration dynamics. PLoS Comput Biol. 2013;9(12):e1003386. O16 Large-scale cortical models towards understanding relationship between brain structure abnormalities and cognitive deficits Yasunori Yamada1 1IBM Research - Tokyo, Japan Correspondence: Yasunori Yamada - ysnr@jp.ibm.com BMC Neuroscience 2016, 17(Suppl 1):O16 Brain connectivity studies have revealed fundamental properties of normal brain network organization [1]. In parallel, they have reported structural connectivity abnormalities in brain diseases such as Alzheimer’s disease (AD) [1, 2]. However, how these structural abnormalities affect information processing and cognitive functions involved in brain diseases is still poorly understood. To deepen our understanding of this causal link, I developed two large-scale cortical models with normal and abnormal structural connectivity of diffusion tensor imaging on aging APOE-4 non-carriers and carriers in the USC Multimodal Connectivity Database [2, 3]. The possession of the APOE-4 allele is one of the major risk factors in developing later AD, and it has known abnormalities in structural connectivity characterized by lower network communication efficiency in terms of local interconnectivity and balance of integration and interconnectivity [2]. The two cortical models share other parameters and consist of 2.4 million spiking neurons and 4.8 billion synaptic connections. First, I demonstrate the biological relevance of the models by confirming that they reproduce normal patterns of cortical spontaneous activities in terms of the following distinctive properties observed in vivo [4]: low firing rates of individual neurons that approximate log-normal distributions, irregular spike trains following a Poisson distribution, a network balance between excitation and inhibition, and greater depolarization of the average membrane potentials. Next, to investigate how the difference in structural connectivity affects cortical information processing, I compare cortical response properties to an input during spontaneous activity between the cortical models. The results show that the cortical model with the abnormal structural connectivity decreased the degree of cortical response as well as the number of cortical regions responding to the input (Fig. 9), suggesting that the structural connectivity abnormality observed in APOE-4 carriers might reduce cortical information propagation and lead to negative effects in information integration. Indeed, imaging studies support this suggestion by reporting structural abnormality with lower network communication efficiency observed in the structural connectivity of both APOE-4 carriers and AD patients [1, 2]. This computational approach allowing for manipulations and detailed analyses that are difficult or impossible in human studies can help to provide a causal understanding of how cognitive deficits in patients with brain diseases are associated with their underlying structural abnormalities.Fig. 9 Responses to input to the left V1 in the two cortical models with normal/abnormal structural connectivity. A Average firing rates. B–D Cortical regions and cortical areas that significantly responded to the input Acknowledgements: This research was partially supported by the Japan Science and Technology Agency (JST) under the Strategic Promotion of Innovative Research and Development Program. ReferencesStam CJ. Modern network science of neurological disorders. Nat Rev Neurosci. 2014;15(10):683–695. Brown JA, Terashima KH, Burggren AC, Ercoli LM, Miller KJ, Small GW, Bookheimer SY. Brain network local interconnectivity loss in aging APOE-4 allele carriers. Proc Natl Acad Sci USA. 2011;108(51):20760–5. Brown JA, Rudie JD, Bandrowski A, van Horn JD, Bookheimer SY. The UCLA multimodal connectivity database: a web-based platform for brain connectivity matrix sharing and analysis. Front Neuroinform. 2012;6(28). Ikegaya Y, Sasaki T, Ishikawa D, Honma N, Tao K, Takahashi N, Minamisawa G, Ujita S, Matsuki N. Interpyramid spike transmission stabilizes the sparseness of recurrent network activity. Cereb Cortex. 2013;23(2):293–304. O17 Spatial coarse-graining the brain: origin of minicolumns Moira L. Steyn-Ross1, D. Alistair Steyn-Ross1 1School of Engineering, University of Waikato, Hamilton 3240, New Zealand Correspondence: Moira L. Steyn-Ross - msr@waikato.ac.nz BMC Neuroscience 2016, 17(Suppl 1):O17 The seminal experiments of Mountcastle [1] over 60 years ago established the existence of cortical minicolumns: vertical column-like arrays of approximately 80–120 neurons aligned perpendicular to the pial surface, penetrating all six cortical layers. Minicolumns have been proposed as the fundamental unit for cortical organisation. Minicolumn formation is thought to rely on gene expression and thalamic activity, but exactly why neurons cluster into columns of diameter 30–50 μm containing approximately 100 neurons is not known. In this presentation we describe a mechanism for the formation of minicolumns via gap-junction diffusion-mediated coupling in a network of spiking neurons. We use our recently developed method of cortical “reblocking” (spatial coarse-graining) [2] to derive neuronal dynamics equations at different spatial scales. We are able to show that for sufficiently strong gap-junction coupling, there exists a minimum block size over which neural activity is expected to be coherent. This coherence region has cross-sectional area of order (40–60 μm)2, consistent with the areal extent of a minicolumn. Our scheme regrids a 2D continuum of spiking neurons using a spatial rescaling theory, established in the 1980s, that systematically eliminates high-wave-number modes [3]. The rescaled neural equations describe the bulk dynamics of a larger block of neurons giving “true” (rather than mean-field) population activity, encapsulating the inherent dynamics of a continuum of spiking neurons stimulated by incoming signals from neighbors, and buffeted by ion-channel and synaptic noise. Our method relies on a perturbative expansion. In order for this coarse-graining expansion to converge, we require not only a sufficiently strong level of inhibitory gap-junction coupling, but also a sufficiently large blocking ratio B. The latter condition establishes a lower bound for the smallest “cortical block”: the smallest group of neurons that can respond to input as a collective and cooperative unit. We find that this minimum block-size ratio lies between 4 and 6. In order to relate this 2D geometric result to the 3D extent of a 3-mm-thick layered cortex, we project the cortex onto a horizontal surface and count the number of neurons contained within each l × l grid micro-cell. Setting l ≈ 10 μm and assuming an average of one interneuron per grid cell, a blocking ratio at the mid-value B = 5 implies that the side-length of a coherent “macro-cell” will be L = Bl = 50 μm containing ~25 inhibitory plus 100 excitatory neurons (assuming an i to e abundance ratio of 1:4) in cross-sectional area L2. Thus the minicolumn volume will contain roughly 125 neurons. We argue that this is the smallest diffusively-coupled population size that can support cooperative dynamics, providing a natural mechanism defining the functional extent of a minicolumn. We propose that minicolumns might form in the developing brain as follows: Inhibitory neurons migrate horizontally from the ganglionic eminence to form a dense gap-junction coupled substrate that permeates all layers of the cortex [4]. Progenitor excitatory cells ascend vertically from the ventricular zone, migrating through the inhibitory substrate of the cortical plate. Thalamic input provides low-level stimulus to activate spiking activity throughout the network. Inhibitory diffusive coupling allows a “coarse graining” such that neurons within a particular areal extent respond collectively to the same input. The minimum block size prescribed by the coarse graining imposes constraints on minicolumn geometry, leading to the spontaneous emergence of cylindrical columns of coherent activity, each column centered on an ascending chain of excitatory neurons and separated from neighboring chains by an annular surround of inhibition. This smallest aggregate is preferentially activated during early brain development, and activity-based plasticity then leads to the formation of tangible structural columns. ReferencesMountcastle VB. Modality and topographic properties of single neurons of cat’s somatic sensory cortex. J Neurophysiol. 1957;20(4):408–34. Steyn-Ross ML, Steyn-Ross DA. From individual spiking neurons to population behavior: Systematic elimination of short-wavelength spatial modes. Phys Rev E. 2016;93(2):022402. Steyn-Ross ML, Gardiner CW. Adiabatic elimination in stochastic systems III. Phys Rev A. 1984;29(5):2834–44. Jones EG. Microcolumns in the cerebral cortex. Proc Natl Acad Sci USA. 2000;97(10):5019–21. O18 Modeling large-scale cortical networks with laminar structure Jorge F. Mejias1, John D. Murray2, Henry Kennedy3, and Xiao-Jing Wang1,4 1Center for Neural Science, New York University, New York, NY, 10003, USA; 2Department of Psychiatry, Yale School of Medicine, New Haven, CT, 06511, USA; 3INSERM U846, Stem Cell and Brain Research Institute, Bron Cedex, France; 4NYU-ECNU Institute of Brain and Cognitive Science, NYU Shanghai, Shanghai, China Correspondence: Jorge F. Mejias - jorge.f.mejias@gmail.com BMC Neuroscience 2016, 17(Suppl 1):O18 Visual cortical areas in the macaque are organized according to an anatomical hierarchy, which is defined by specific patterns of anatomical projections in the feedforward and feedback directions [1, 2]. Recent macaque studies also suggest that signals ascending through the visual hierarchy are associated with gamma rhythms, and top-down signals with alpha/low beta rhythms [3–5]. It is not clear, however, how oscillations presumably originating at local populations can give rise to such frequency-specific large-scale interactions in a mechanistic way, or the role that anatomical projections patterns might have in this. To address this question, we build a large-scale cortical network model with laminar structure, grounding our model on a recently obtained anatomical connectivity matrix with weighted directed inter-areal projections and information about their laminar origin. The model involves several spatial scales—local or intra-laminar microcircuit, inter-laminar circuits, inter-areal interactions and large-scale cortical network—and a wide range of temporal scales—from slow alpha oscillations to gamma rhythms. At any given level, the model is constrained anatomically and then tested against electrophysiological observations, which provides useful information on the mechanisms modulating the oscillatory activity at different scales. As we ascend through the local to the inter-laminar and inter-areal levels, the model allows us to explore the sensory-driven enhancement of gamma rhythms, the inter-laminar phase-amplitude coupling, the relationship between alpha waves and local inhibition, and the frequency-specific inter-areal interactions in the feedforward and feedback directions [3, 4], revealing a possible link with the predictive coding framework. When we embed our modeling framework into the anatomical connectivity matrix of 30 areas (which includes novel areas not present in previous studies [2, 6]), the model gives insight into the mechanisms of large-scale communication across the cortex, accounts for an anatomical and functional segregation of FF and FB interactions, and predicts the emergence of functional hierarchies, which recent studies have found in macaque [4] and human [5]. Interestingly, the functional hierarchies observed experimentally are highly dynamic, with areas moving across the hierarchy depending on the behavioral context [4]. In this regard, our model provides a strong prediction: we propose that these hierarchical jumps are triggered by laminar-specific modulations of input into cortical areas, suggesting a strong link between hierarchy dynamics and context-dependent computations driven by specific inputs. ReferencesFelleman DJ, Van Essen DC. Distributed hierarchical processing in the primate cerebral cortex. Cereb Cortex. 1991;1(1):1–47. Markov NT, Vezoli J, Chameau P, Falchier A, Quilodran R, Huissoud C, Lamy C, Misery P, Giroud P, Ullman S, et al. Anatomy of hierarchy: feedforward and feedback pathways in macaque visual cortex. J Comp Neurol. 2014;522:225–259. van Kerkoerle T, Self MW, Dagnino B, Gariel-Mathis MA, Poort J, van der Togt C, Roelfsema PR. Alpha and gamma oscillations characterize feedback and feedforward processing in monkey visual cortex. Proc Natl Acad Sci USA. 2014;111;14332–41. Bastos AM, Vezoli J, Bosman CA, Schoffelen JM, Oostenveld R, Dowdall JR, De Weerd P, Kennedy H, Fries P. Visual areas exert feedforward and feedback influences through distinct frequency channels. Neuron. 2015;85:390–401. Michalareas G, Vezoli J, van Pelt S, Schoffelen JM, Kennedy H, Fries. Alpha–beta and gamma rhythms subserve feedback and feedforward influences among human visual cortical areas. Neuron. 2016;89:384–97. Chaudhuri R, Knoblauch K, Gariel MA, Kennedy H, Wang XJ. A large-scale circuit mechanism for hierarchical dynamical processing in the primate cortex. Neuron. 2015;88:419–31. O19 Information filtering by partial synchronous spikes in a neural population Alexandra Kruscha1,2, Jan Grewe3,4, Jan Benda3,4 and Benjamin Lindner1,2 1Bernstein Center for Computational Neuroscience, Berlin, 10115, Germany; 2Institute for Physics, Humboldt-Universität zu Berlin, Berlin, 12489, Germany; 3Institue for Neurobiology, Eberhardt Karls Universität Tübingen, Germany; 4Bernstein Center for Computational Neuroscience, Munich, Germany Correspondence: Alexandra Kruscha - alexandra.kruscha@bccn-berlin.de BMC Neuroscience 2016, 17(Suppl 1):O19 Synchronous firing of neurons is a prominent feature in many brain areas. Here, we are interested in the information transmission by the synchronous spiking output of a noisy neuronal population, which receives a common time-dependent sensory stimulus. Earlier experimental [1] and theoretical [2] work revealed that synchronous spikes encode preferentially fast (high-frequency) components of the stimulus, i.e. synchrony can act as an information filter. In these studies a rather strict measure of synchrony was used: the entire population has to fire within a short time window. Here, we generalize the definition of the synchronous output, for which only a certain fraction γ of the population needs to be active simultaneously—a setup that seems to be of more biological relevance. We characterize the information transfer in dependence of this fraction and the population size, by the spectral coherence function between the stimulus and the partial synchronous output. We present two different analytical approaches to derive this frequency-resolved measure (one that is more suited for small population sizes, while the second one is applicable to larger populations). We show that there is a critical synchrony fraction, namely the probability at which a single neuron spikes within the predefined time window, which maximizes the information transmission of the synchronous output. At this value, the partial synchronous output acts as a low-pass filter, whereas deviations from this critical fraction lead to a more and more pronounced band-pass filtering effect. We confirm our analytical findings by numerical simulations for the leaky integrate-and-fire neuron. We also show that these findings are supported by experimental recordungs of P-Units electroreceptors of weakly electric fish, where the filtering effect of the synchronous output occurs in real neurons as well. Acknowledgement: This work was supported by Bundesministerium für Bildung und Forschung Grant 01GQ1001A and DFG Grant 609788-L1 1046/2-1. ReferencesMiddleton JW, Longtin A, Benda J, Maler L. Postsynaptic receptive field size and spike threshold determine encoding of high-frequency information via sensitivity to synchronous presynaptic activity. J Neurophysiol. 2009;101:1160–70. Sharafi N, Benda J, Lindner B. Information filtering by synchronous spikes in a neural population. J Comp Neurosc. 2013;34:285–301. O20 Decoding context-dependent olfactory valence in Drosophila Laurent Badel1, Kazumi Ohta1, Yoshiko Tsuchimoto1, Hokto Kazama1 1RIKEN Brain Science Institute, 2-1 Hirosawa, Wako, 351-0198, Japan Correspondence: Laurent Badel - laurent@brain.riken.jp BMC Neuroscience 2016, 17(Suppl 1):O20 Many animals rely on olfactory cues to make perceptual decisions and navigate the environment. In the brain, odorant molecules are sensed by olfactory receptor neurons (ORNs), which convey olfactory information to the central brain in the form of sequences of action potentials. In many organisms, axons of ORNs expressing the same olfactory receptor converge to one or a few glomeruli in the first central region (the antennal lobe in insects and the olfactory bulb in fish and mammals) where they make contact with their postsynaptic targets. Therefore, each glomerulus can be considered as a processing unit that relays information from a specific type of receptor. Because different odorants recruit different sets of glomeruli, and most glomeruli respond to a wide array of odors, olfactory information at this stage of processing is contained in spatiotemporal patterns of glomerular activity. How these patterns are decoded by the brain to guide odor-evoked behavior, however, remains largely unknown. In Drosophila, attraction and aversion to specific odors have been linked to the activation of one or a few glomeruli (reviewed in [1]) in the antennal lobe (AL). These observations suggest a “labeled-line” coding strategy, in which individual glomeruli convey signals of specific ethological relevance, and their activation triggers the execution of hard-wired behavioral programs. However, because these studies used few odorants, and a small fraction of glomeruli were tested, it is unclear how the results generalize to broader odor sets, and whether similar conclusions hold for each of the ~50 glomeruli of the fly AL. Moreover, how compound signals from multiple glomeruli are integrated is poorly understood. Here, we combine optical imaging, behavioral and statistical techniques to address these questions systematically. Using two-photon imaging, we monitor Ca2+ activity in the AL in response to 84 odors. We next screen behavioral responses to the same odorants. Comparing these data allows us to formulate a decoding model describing how olfactory behavior is determined by glomerular activity patterns in a quantitative manner. We find that a weighted sum of normalized glomerular responses recapitulates the observed behavior and predicts responses to novel odors, suggesting that odor valence is not determined solely by the activity a few privileged glomeruli. This conclusion is supported by genetic silencing and optogenetic activation of individual ORN types, which are found to evoke modest biases in behavior in agreement with model predictions. Finally, we test the model prediction that the relative valence of a pair of odors depends on the identity of other odors presented in the same experiment. We find that the relative valence indeed changes, and may even switch, suggesting that perceptual decisions can be modulated by the olfactory context. Surprisingly, our model correctly captured both the direction and the magnitude of the observed changes. These results indicate that the valence of olfactory stimuli is decoded from AL activity by pooling contributions over a large number of glomeruli, and highlight the ability of the olfactory system to adapt to the statistics of its environment, similarly to the visual and auditory systems. ReferenceLi Q, Liberles SD. Aversion and attraction through olfaction. Curr Biol. 2015;25(3):R120–9. P1 Neural network as a scale-free network: the role of a hub B. Kahng1 1Department of Physics and Astronomy, Seoul National University, 08826, Korea Correspondence: B. Kahng - bkahng@snu.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P1 Recently, increasing attention has been drawn to human neuroscience in network science communities. This is because recent fMRI and anatomical experiments have revealed that neural networks of normal human brain are scale-free networks. Thus, accumulated knowledges in a broad range of network sciences can be naturally applied to neural networks to understand functions and properties of normal and disordered human brain networks. Particularly, the degree exponent value of the human neural network constructed from the fMRI data turned out to be approximately two. This value has particularly important meaning in scale-free networks, because the number of connections to neighbors of a hub becomes largest and thus functional role of the hub becomes extremely important. In this talk, we present the role of the hub in pattern recognition and dynamical problems in association with neuroscience. P2 Hemodynamic responses to emotions and decisions using near-infrared spectroscopy optical imaging Nicoladie D. Tam1 1Department of Biological Sciences, University of North Texas, Denton, TX 76203, USA Correspondence: Nicoladie D. Tam - nicoladie.tam@unt.edu BMC Neuroscience 2016, 17(Suppl 1):P2 This study focuses on the relationship between the emotional response, decision and the hemodynamic responses in the prefrontal cortex. This is based on the computational emotional model that hypothesizes the emotional response is proportional to the discrepancy between the expectancy and the actuality. Previous studies had shown that emotional responses are related to decisions [1, 2]. Specifically, the emotional responses of happy [3], sad [4], angry [5], jealous [6] emotions are proportional to the discrepancy between what one wants and what one gets [1, 3–7]. Methods Human subjects are asked to perform the classical behavioral economic experiment called Ultimatum Game (UG) [8]. This experimental paradigm elicits the interrelationship between decision and emotion in human subjects [3–6]. The hemodynamic responses of the prefrontal cortex were recorded while the subjects performed the UG experiment. Results The results showed that the hemodynamic response, which corresponds to the neural activation and deactivation based on the metabolic activities of the neural tissues, are proportional to the emotional intensity and the discrepancy between the expectancy and the actuality. This validates the hypothesis of the proposed emotional theory [9–11] that the intensity of emotion is proportional to the disparity between the expected and the actual outcomes. These responses are also related to the fairness perception [7], with respect to the survival functions [9, 10] similar to the responses established for happy [1] emotion, and for fairness [12] experimentally. This is consistent with the computational relationship between decision and fairness [13]. ReferencesTam ND. Quantification of happy emotion: dependence on decisions. Psychol Behav Sci. 2014;3(2):68–74. Tam ND. Rational decision-making process choosing fairness over monetary gain as decision criteria. Psychol Behav Sci. 2014;3(6–1):16–23. Tam ND. Quantification of happy emotion: Proportionality relationship to gain/loss. Psychol Behav Sci. 2014;3(2):60–7. Tam ND: Quantitative assessment of sad emotion. Psychol Behav Sci 2015, 4(2):36-43. Tam DN. Computation in emotional processing: quantitative confirmation of proportionality hypothesis for angry unhappy emotional intensity to perceived loss. Cogn Comput. 2011;3(2):394–415. Tam ND, Smith KM. Cognitive computation of jealous emotion. Psychol Behav Sci. 2014;3(6–1):1–7. Tam ND. Quantification of fairness perception by including other-regarding concerns using a relativistic fairness-equity model. Adv Soc Sci Research J. 2014;1(4):159–69. von Neumann J, Morgenstern O, Rubinstein A. Theory of games and economic behavior. Princeton: Princeton University Press; 1953. Tam D. EMOTION-I model: A biologically-based theoretical framework for deriving emotional context of sensation in autonomous control systems. Open Cybern Syst J. 2007;1:28–46. Tam D. EMOTION-II model: a theoretical framework for happy emotion as a self-assessment measure indicating the degree-of-fit (congruency) between the expectancy in subjective and objective realities in autonomous control systems. Open Cybern Syst J. 2007;1:47–60. Tam ND. EMOTION-III model. A theoretical framework for social empathic emotions in autonomous control systems. Open Cybern Syst J. 2016 (in press). Tam ND: Quantification of fairness bias in relation to decisions using a relativistic fairness-equity model. Adv in Soc Sci Research J 2014, 1(4):169-178. Tam ND. A decision-making phase-space model for fairness assessment. Psychol Behav Sci. 2014;3(6–1):8–15. P3 Phase space analysis of hemodynamic responses to intentional movement directions using functional near-infrared spectroscopy (fNIRS) optical imaging technique Nicoladie D. Tam1, Luca Pollonini2, George Zouridakis3 1Department of Biological Sciences, University of North Texas, Denton, TX 76203, USA; 2College of Technology, the University of Houston, TX, 77204, USA; 3Departments of Engineering Technology, Computer Science, and Electrical and Computer Engineering, University of Houston, Houston, TX, 77204, USA Correspondence: Nicoladie D. Tam - nicoladie.tam@unt.edu BMC Neuroscience 2016, 17(Suppl 1):P3 We aim to extract the intentional movement directions of the hemodynamic signals recorded from noninvasive optical imaging technique, such that a brain-computer-interface (BCI) can be built to control a wheelchair based on the optical signals recorded from the brain. Real-time detection of neurodynamic signals can be obtained using functional near-infrared spectroscopy (fNIRS), which detects both oxy-hemoglobin (oxy-Hb) and deoxy-hemoglobin (deoxy-Hb) levels in the underlying neural tissues. In addition to the advantage of real-time monitoring of hemodynamic signals using fNIRS over fMRI (functional magnetic resonance imaging), fNIRS also can detect brain signals of human subjects in motion without any movement artifacts. Previous studies had shown that hemodynamic responses are correlated with the movement directions based on the temporal profiles of the oxy-Hb and deoxy-Hb levels [1–5]. In this study, we will apply a phase space analysis to the hemodynamic response to decode the movement directions instead of using the temporal analysis in the previous studies. Methods In order to decode the movement directions, human subjects were asked to execute two different orthogonal directional movements in the front-back and right-left directions while the optical hemodynamic responses were recorded in the motor cortex of the dominant hemisphere. We aim to decode the intentional movement directions without a priori any assumption on how arm movement directions are correlated with the hemodynamic signals. Therefore, we used the phase space analysis to determine how the trajectories of oxy-Hb and deoxy-Hb are related to each other during these arm movements. Results The results show that there are subpopulations of cortical neurons that are task-related to the intentional movement directions. Specifically, using phase space analysis of the oxy-Hb and deoxy-Hb levels, opposite movement direction is represented by the different hysteresis of the trajectories in opposite direction in the phase space. Since oxy-Hb represents the oxygen delivery and deoxy-Hb represents the oxygen extraction by the underlying brain tissues, the phase space analysis provides a means to differentiate the movement direction by the ratio between oxygen delivery and oxygen extraction. In other words, the oxygen demands in the subpopulation of neurons in the underlying tissue differ depending on the movement direction. This also corresponds to the opposite patterns of neural activation and deactivation during execution of opposite movement directions. Thus, phase space analysis can be used as an analytical tool to differentiate different movement directions based on the trajectory of the hysteresis with respect to the hemodynamic variables. ReferencesTam ND, Zouridakis G. Optical imaging of motor cortical activation using functional near-infrared spectroscopy. BMC Neurosci. 2012;13(Suppl 1):P27. Tam ND, Zouridakis G. Optical imaging of motor cortical hemodynamic response to directional arm movements using near-infrared spectroscopy. Int J Biol Eng. 2013;3(2):11–17. Tam ND, Zouridakis G. Decoding of movement direction using optical imaging of motor cortex. BMC Neurosci. 2013; P380. Tam ND, Zouridakis G. Temporal decoupling of oxy- and deoxy-hemoglobin hemodynamic responses detected by functional near-infrared spectroscopy (fNIRS). J Biomed Eng Med Imaging. 2014;1(2):18–28. Tam ND, Zouridakis G. Decoding movement direction from motor cortex recordings using near-infrared spectroscopy. In: Infrared spectroscopy: theory, developments and applications. Hauppauge: Nova Science; 2014. P4 Modeling jamming avoidance of weakly electric fish Jaehyun Soh1, DaeEun Kim1 1Biological Cybernetics, School of Electrical and Electronic Engineering, Yonsei University, Shinchon, Seoul, 120-749, South Korea Correspondence: DaeEun Kim - daeeun@yonsei.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P4 Weakly electric fish use electric field generated by the electric organ in the tail of the fish. They detect objects by sensing the electric field with electroreceptors on the fish’s body surface. Obstacles in the vicinity of the fish distort the electric field generated by the fish and the fish detect this distortion to recognize environmental situations. Generally, weakly electric fish produce species-dependent electric organ discharge (EOD) signals. Frequency bands of the fish’s signals include a variety of frequencies, 50–600 Hz or higher than 800 Hz. The EOD signals can be disturbed by similar frequency signals emitted by neighboring weakly electric fish. They change their EOD frequencies to avoid jamming signals when they detect the interference of signals. This is called jamming avoidance response (JAR). Electroreceptors of the fish read other electric fish’s EOD while they sense their own EOD. Therefore, when two weakly electric fish are close enough and they sense similar frequencies, their sensing ability by EOD is impaired because of signal jamming [1, 2]. The fish lowers its EOD frequency in response to the jamming signals when a slightly higher frequency of signals are detected and otherwise, raises its EOD. This response is shown in Fig. 10. The fish shift their EOD frequency almost immediately without trial and error.Fig. 10 Jamming avoidance response The method of how to avoid jamming has been studied for a long time, but the corresponding neural mechanisms have not been revealed yet so far. The JAR of Eigenmannia can be analyzed by Lissajous graphs which consist of amplitude modulations and differential phase modulations. Relative intensity of signals at each skin can show that the signal frequency is higher than its own signal frequency or lower [3]. We suggest an algorithm of jamming avoidance for EOD signals, especially for wave-type fish. We explore the diagram of amplitude modulation versus phase modulation, and analyze the shape over the graph. The phase differences or amplitude differences will contribute to the estimation of the signal jamming situation. From that, the jammed signal frequency can be detected and so it can guide the jamming avoidance response. It can provide a special measure to predict the jamming avoidance response. However, what type of neural structure is available in weakly electric fish is an open question. We need further study on this subject. Acknowledgements: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MEST) (No. 2014R1A2A1A11053839). ReferencesHeiligenberg W. Electrolocation of objects in the electric fish eigenmannia (rhamphichthyidae, gymnotoidei). J Comp Physiol. 1973;87(2):137–64. Heiligenberg W. Principles of electrolocation and jamming avoidance in electric fish. Berlin: Springer; 1977. Heiligenberg W. Neural nets in electric fish. Cambridge: MIT Press; 1991. P5 Synergy and redundancy of retinal ganglion cells in prediction Minsu Yoo1, S. E. Palmer1,2 1Committee on Computational Neuroscience, University of Chicago, Chicago, IL, USA; 2Department of Organismal Biology and Anatomy, University of Chicago, Chicago, IL, USA Correspondence: Minsu Yoo - minsu@uchicago.edu BMC Neuroscience 2016, 17(Suppl 1):P5 Recent work has shown that retina ganglion cells (RGC) of salamanders predict future sensory information [1]. It has also been shown that these RGC’s carry significant information about the future state of their own population firing patterns [2]. From the perspective of downstream neurons in the visual system that do not have independent access to the visual scene, the correlations in the RGC firing, itself, may be important for predicting the future visual input. In this work, we explore the structure of the generalized correlation in firing patterns in the RGC, with a particular focus on coding efficiency. From the perspective of efficient neural coding, we might expect neurons to code for their own future state independently (decorrelation across cells), and to have very little predictive information extending forward in time (decorrelation in the time domain). In this work, we quantify whether neurons in the retina code for their own future input independently, redundantly, or synergistically, and how long these correlations persist in time. We use published extracellular multi-electrode data from the salamander retina in response to repeated presentations of a natural movie [1]. We find significant mutual information in the population firing that is almost entirely independent except at very short time delays, where the code is weakly redundant (Fig. 11). We also find that the information persists to delays of up to a few 100 ms. In addition, we find that individual neurons vary widely in the amount of predictive information they carry about the future population firing state. This heterogeneity may contribute to the diversity of predictive information we find across groups in this experiment.Fig. 11 Predictive information in the retinal response is coded for independently. Red the mutual information between the binary population firing patterns at times t and t + Δt, for 1000 randomly selected groups of 5 cells from our 31-cell population. Time is binned in 16.67 ms bins, and the (rare) occurrence of two spikes in a bin is recorded as a ‘1’. Blue the sum of the mutual information between a single cell response at time t and the future response of the group at time t + Δt. Error bars indicate the standard error of the mean across groups. All information quantities are corrected for finite-size effects using quadratic extrapolation [3] The results in this study may provide useful information for building a model of the RGC population that can explain why redundant coding is only observed at short delays, or what makes one RGC more predictive than another. Building this type of model will illustrate how the retina represents the future. ReferencesPalmer SE, Marre O, Berry MJ, Bialek W. Predictive information in a sensory population. Proc. Natl. Acad. Sci. 2015;112:6908–13. Salisbury J, Palmer SE. Optimal prediction and natural scene statistics in the retina. ArXiv150700125 Q-Bio [Internet]. 2015 [cited 2016 Feb 25]; Available from: http://arxiv.org/abs/1507.00125. Panzeri S, Senatore R, Montemurro MA, Petersen RS. Correcting for the sampling bias problem in spike train information measures. J. Neurophysiol. 2007;98:1064–72. P6 A neural field model with a third dimension representing cortical depth Viviana Culmone1, Ingo Bojak1 1School of Psychology, University of Reading, Reading, Berkshire, RG1 6AY, UK Correspondence: Viviana Culmone - v.culmone@pgr.reading.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P6 Neural field models (NFMs) characterize the average properties of neural ensembles as a continuous excitable medium. So far, NFMs have largely ignored the extension of the dendritic tree, and its influence on the neural dynamics [1]. As shown in Fig. 12A, we implement a 3D-NFM, including the dendritic extent through the cortical layers, starting from a well-known 2D-NFM [2]. We transform the equation for the average membrane potential he for the point-like soma in the 2D-NFM [2] to a full cable equation form (added parts in bold):Fig. 12 A The 3D-NFM adds a dendritic dimension to the 2D one [1]. One single macrocolumn has inhibitory (I) and excitatory (E) subpopulations. B (Top) Discretization of the dendrite. (Bottom) Equilibrium membrane potential along the dendrite for two different synaptic inputs. C PSDs of he for the 2D- and 3D-NFM. Increasing the synaptic input recovers the lost alpha rhythm τe∂he(x,z,t)∂t=-he(x,z,t)-her+λ2∂2he(x,z,t)∂z2+fsyn∑kψke(he)Ike(x,z,t) The 3D-NFM is modeled considering the dendritic tree as a single linear cable. Figure 12B shows the resulting resting potential along the extended dendrite for synaptic input in two different locations. Naively keeping the parameters of the 2D-NFM for the 3D-NFM results in a power spectral density (PSD) without an alpha rhythm resonance, see Fig. 12C. However, increasing the synaptic input by a factor fsyn can compensate for the dispersion along the dendrite and recovers the peak in the alpha band. We study the influence of varying the distribution of synaptic inputs along the dendritic (vertical) dimension and of changing the (horizontal) area of the simulated cortical patch. We also provide an outlook on how to compare our results with local field potential recordings from real cortical tissues. We expect that 3D-NFMs will be used widely in the future for describing such experimental data, and that the methods used to extend the specific 2D-NFM used here [2] will generalize to other 2D-NFMs. ReferencesSpruston N. Pyramidal neurons: dendritic structure and synaptic integration. Nat Rev Neurosci. 2008;9:206–221. Bojak I, Liley DTJ. Modeling the effects of anesthesia on the electroencephalogram. Phys Rev E. 2005;71:041902. P7 Network analysis of a probabilistic connectivity model of the Xenopus tadpole spinal cord Andrea Ferrario1, Robert Merrison-Hort1, Roman Borisyuk1 1School of Computing and Mathematics, Plymouth University, Plymouth, PL4 8AA, United Kingdom Correspondence: Andrea Ferrario - andrea.ferrario@plymouth.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P7 Our previous results [1, 2] describe a computational anatomical model of the Xenopus tadpole spinal cord which includes about 1400 neurons of seven types allocated on two sides of the body. This model is based on a developmental approach, where axon growth is simulated and synapses are created (with some probability) when axons cross dendrites. A physiological model of spiking neurons with the generated connectivity of about 85,000 synapses produces a very reliable swimming pattern of anti-phase oscillations in response to simulated sensory input [2]. Using the developmental model we generate 100 different sets of synaptic connections (“connectomes”), and use this information to create a generalized probabilistic model. The probabilistic model provides a new way to easily generate tadpole connectomes and, remarkably, these connectomes produce similar simulated physiological behavior to those generated using the more complex developmental approach (e.g. they swim when stimulated). Studying these generated connectivity graphs allows us to analyze the structure of connectivity in a typical tadpole spinal cord. Many complex neuronal networks have been found to have “small world” properties, including those in the nematode worm C. elegans [3, 6], cat and macaque cortex and the human brain [4]. Small world networks are classified between regular and random networks, and are characterized by a high value of the clustering coefficient C and a relatively small value of the average path length L, when compared with Erdős-Rényi and degree matched graphs of a similar size. We used graph theory tools to calculate the strongly connected component of each network, which was then used to measure C and L. For the degree-matched network, these computations have been based on finding the probabilistic generating function [5]. By comparing these measures with those of degree matched random graphs, we found that tadpole’s network can be considered a small world graph. This is also true for the sub-graph consisting only of neurons on one side of the body, which displays properties very similar to those of the C. elegans network. Another important subgraph, comprising only the two main neuron types in the central pattern generator (CPG) network also shows small world properties, but is less similar to the C. elegans network. Our approach allows us to study the general properties of the architecture of the tadpole spinal cord, even though in reality the actual network varies from individual to individual (unlike in C. elegans). This allows us to develop ideas about the organizing principles of the network, as well as to make predictions about the network’s functionality that can be tested first in computer simulations and later in real animal experiments. In this work we combine several graph theory techniques in a novel way to analyze the structure of a complex neuronal network where not all biological details are known. We believe that this approach can be applied widely to analyze other animals’ nervous systems. ReferencesBorisiuk R, al Azad AK, Conte D, Roberts A, Soffe SR. A developmental approach to predicting neuronal connectivity from small biological datasets: a gradient-based neuron growth model. PloS One. 2014;9(2):e89461. Roberts A, Conte A, Hull M, Merrison-Hort R, al Azad AK, Buhl E, Borisyuk R, Soffe SR. Can simple rules control development of a pioneer vertebrate neuronal network generating behavior? J Neurosci. 2014;34(2):608–21. Watts DJ, Strogatz SH. Collective dynamics of ‘small-world’ networks. Nature. 1998;440–2. Kaiser M. A tutorial in connectome analysis: topological and spatial features of brain networks. NeuroImage. 2011;892–907. Newman MEJ, Strogatz SH, Watts DJ. Random graphs with arbitrary degree distribution and their applications. Phys. Rev. 2001;E64:026118. Vershney LR, Chen BL, Paniagua E, Hall DH Chklovskii DB. Structural properties of the Caenorhabditis elegans neuronal network. PloS Comput Biol. 2011;7(2):e1001066. P8 The recognition dynamics in the brain Chang Sub Kim1 1Department of Physics, Chonnam National University, Gwangju, 61186, Republic of Korea Correspondence: Chang Sub Kim - cskim@jnu.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P8 Over the years an extensive research endeavor has been given to understanding the brain’s cognitive function in a unified principle and to providing a formulation of the corresponding computational scheme of the brain [1]. The explored free-energy principle (FEP) claims that the brain’s operation on perception, learning, and action rests on brain’s internal mechanism of trying to avoid aberrant events encountering in its habitable environment. The theoretical measure for this biological process has been suggested to be the informational free-energy (IFE). The computational actualization of the FEP is carried out via the gradient descent method (GDM) in machine learning theory. The information content of the cognitive processes is encoded in the biophysical matter as spatiotemporal patterns of the neuronal correlates of the external causes. Therefore, any realistic attempt to account for the brain function must conform to the physics laws and the underlying principles. Notwithstanding the grand simplicity, however, the FEP framework embraces some extra-physical constructs. Two major such extra-physical constructs are the generalized motions, which are non-Newtonian objects, and the GDM in executing the brain’s computational mechanism of perception and active inference. The GDM is useful in finding mathematical solutions in the optimal problems, but not derived from a physics principle. In this work, we cast the FEP in the brain science into the framework of the principle of least action (PLA) in physics [2]. The goal is to remove the extra-physical constructs embedded in the FEP and to reformulate the GDM within the standard mechanics arena. Previously, we suggested setting up the minimization scheme of the IFE in the Lagrange mechanics formalism [3] which contained only primitive results. In the present formulation we specify the IFE as the information-theoretic Lagrangian and thus formally define the informational action (IA) as time-integral of the IFE. Then, the PLA prescribes that the viable brain minimizes the IA when encountering uninhabitable events by selecting an optimal path among all possible dynamical configurations in the brain’s neuronal network. Specifically, the minimization yields the mechanistic equations of motion of the brain states, which are inverting algorithms of sensory inputs to infer their external causes. The obtained Hamilton–Jacobi–Bellman-type equation prescribes the brain’s recognition dynamics which do not require the extra-physical concept of higher order motions. Finally, a neurobiological implementation of the algorithm is presented which complies with the hierarchical, operative structure of the brain. In doing so, we adopt the local field potential and the local concentration of ions in the Hodgkin–Huxley model as the effective brain states [4]. Thus, the brain’s recognition dynamics is operatively implemented in a neuro-centric picture. We hope that our formulation, conveying a wealth of structure as an interpretive and mechanistic description of explaining how the brain’s cognitive function may operate, will provide with a helpful guidance for future simulation. ReferencesFriston K. The free-energy principle: a unified brain theory? Nat Reivew Neurosci. 2010;11:127–38. Landau LP. Classical mechanics. 2nd ed. NewYork: Springer; 1998. Kim CS. The adaptive dynamics of brains: Lagrangian formulation. Front Neurosci Conf Abstr Neuroinform. 2010. doi:10.3389/conf.fnins.2010.13.00046. Hodgkin A, Huxley A. A quantitative description of membrane current and its application to conduction and excitation in nerve. J Physiol. 1952;117:500–44. P9 Multivariate spike train analysis using a positive definite kernel Taro Tezuka1 1Faculty of Library, Information and Media Science, University of Tsukuba, Tsukuba, 305-0821, Japan Correspondence: Taro Tezuka - tezuka@slis.tsukuba.ac.jp BMC Neuroscience 2016, 17(Suppl 1):P9 Multivariate spike trains, obtained by recording multiple neurons simultaneously, is a key to uncovering information representation in the brain [1]. Other expressions used to refer to the same type of data include “multi-neuron spike train” [2] and “parallel spike train’” [3]. One approach to analyze spike trains is to use kernel methods, which are known to be among the most powerful machine learning methods. Kernel methods rely on defining a symmetric positive-definite kernel suited to the given data. This work proposes a general way of extending kernels on univariate (or single-unit) spike trains to multivariate spike trains. In this work, the mixture kernel, which naturally extends a kernel defined on univariate spike trains, is proposed and evaluated. There are many univariate spike train kernels proposed [4–9], and the mixture kernel is applicable to any of these kernels. Considered abstractly, a multivariate spike train is a set of time points at which different types of events occurred. In other words, it is a sample taken from a marked point process. The method proposed in this paper is therefore applicable to other data with the same structure. The mixture kernel is defined as a linear combination of symmetric positive-definite kernels on the components of the target data structure, in this case univariate spike trains. The name “mixture kernel” derives from the common use of the word “mixture” to indicate a linear combination in physics and machine learning, for example in Gaussian mixture models. One can prove that the mixture kernel is symmetric positive-definite if coefficient matrix of the mixture is a symmetric positive-semidefinite matrix. The performance of the mixture kernel was evaluated by kernel ridge regression for estimating the value of the parameter for generating synthetic spike train data, and also the stimulus given to the animal as the spike trains were recorded. For synthetic data, multivariate spike trains were generated using homogenous Poisson processes. For real data, the pvc-3 data set [2] in the CRCNS (Collaborative Research in Computational Neuroscience) data sharing website was used, which is a 10-unit multivariate spike trains recorded from the primary visual cortex of a cat. Acknowledgement: This work was supported in part by JSPS KAKENHI Grant Numbers 21700121, 25280110, and 25540159. ReferencesGerstner W, Kistler WM, Naud R, Paninski L. Neuronal dynamics. Cambridge: Cambridge University Press; 2014. Blanche T. Multi-neuron recordings in primary visual cortex, CRCNS.org; 2009. Grun S, Rotter S. Analysis of parallel spike trains. Berlin: Springer; 2010. Paiva A, Park IM, Principe JC. A reproducing kernel Hilbert space framework for spike train signal processing, Neural Comput. 2009;21(2):424–49. Park IM, Seth S, Rao M, Principe JC. Strictly positive definite spike train kernels for point process divergences. Neural Comput. 2012;24:2223–50. Park IM, Seth S, Paiva A, Li L, Principe JC. Kernel methods on spike train space for neuroscience: a tutorial. Signal Process Mag. 2013;30(4):149–60. Li L, Park IM, Brockmeier AJ, Chen B, Seth S, Francis JT, Sanchez JC, Principe JC. Adaptive inverse control of neural spatiotemporal spike patterns with a reproducing kernel Hilbert space (RKHS) framework. IEEE Trans Neural Syst Rehabil Eng. 2013;21(4):532–43. Shpigelman L, Singer Y, Paz R, Vaadia E. Spikernels: embedding spiking neurons in inner product spaces. Adv Neural Inf Process Syst. 2003;15:125–32. Eichhorn J, Tolias A, Zien A, Kuss M, Rasmussen CE, Weston J, Logothetis N, Scholkopf B. Prediction on spike data using kernel algorithms. Adv Neural Inf Process Syst. 2004;16:1367–74. P10 Synchronization of burst periods may govern slow brain dynamics during general anesthesia Pangyu Joo1 1Physics, POSTECH, Pohang, 37673, Republic of Korea Correspondence: Pangyu Joo - pangyu32@postech.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P10 Researchers have utilized electroencephalogram (EEG) as an important key to study brain dynamics in general anesthesia. Representative features of EEG in deep anesthesia are slow wave oscillation and burst suppression [1], and they have so different characteristics that they seem to have different origins. Here, we propose that the two feature may be a different aspect of same phenomenon and show that the slow oscillation could arise from partial synchronization of bursting periods. To model the synchronization of burst periods, modified version of Ching’s model of burst suppression [2] is used. 20 pyramidal neurons and 20 fast spiking neurons are divided into 10 areas composed of 2 pyramidal and 2 fast spiking neurons so that each area exhibit burst suppression behavior independently. Then, all the pyramidal neurons are all to all connected and the connection strength modulates the amount of synchronization of burst periods. The action potentials of pyramidal neurons are substituted by 1 when the action potential larger than 0, and all other case 0. Then they are averaged over the neurons and convoluted with 50 ms square function to see the collective activity of the neurons. As shown in Fig. 13A, At high level of ATP recovery rate (JATP > 1), there are no suppression period so that slow oscillation does not appear regardless of synchronization. At low level of ATP recovery rate (JATP = 0.5), we can observe that the slow oscillation appears with increasing amplitude and finally become burst suppression as relative connection strength increases (Fig. 13B). When the ATP recovery rate is 0, then the pyramidal neurons do not fire at all. These results suggest that the burst period synchronization model could explain some important features of EEG during general anesthesia: the increasing slow oscillation amplitude as anesthesia deepen, significantly high activity in bursting period, and the peak max phase amplitude coupling in deep anesthesia.Fig. 13 A The convoluted signal with different ATP recovery rates (JATP) and relative connection strengths (C). B Standard deviation of the convoluted signals ReferencesPurdon PL, Pierce ET, Mukamel EA, et al. Electroencephalogram signatures of loss and recovery of consciousness from propofol. PNAS. 2013;110(12):E1142–51. Ching S, Purdon PL, Vijayan S, Kopell NJ, Brown EN. A neurophysiological–metabolic model for burst suppression. PNAS. 2012;109(8):3095–100. P11 The ionic basis of heterogeneity affects stochastic synchrony Young-Ah Rho1,4, Shawn D. Burton2,3, G. Bard Ermentrout1,3, Jaeseung Jeong4, Nathaniel N. Urban2,3 1Department of Mathematics, University of Pittsburgh, Pittsburgh, PA, USA 15260; 2Department of Biological Sciences, Carnegie Mellon University, Pittsburgh, PA, USA 15213; 3Center for the Neural Basis of Cognition, Pittsburgh, PA, USA 15213; 4Department of Bio and Brain Engineering/Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea 34141 Correspondence: Young-Ah Rho - yarho75@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P11 Synchronization in neural oscillations is a prominent feature of neural activity and thought to play an important role in neural coding. Theoretical and experimental studies have described several mechanisms for synchronization based on coupling strength and correlated noise input. In the olfactory systems, recurrent and lateral inhibition mediated by dendrodendritic mitral cell–granule cell synapses are critical for synchronization, and intrinsic biophysical heterogeneity reduce the ability to synchronize. In our previous study, a simple phase model was used to examine how physiological heterogeneity in biophysical properties and firing rates across neurons affects correlation-induced synchronization (stochastic synchrony). It has showed that heterogeneity in the firing rates and in the shapes of the phase response curves (PRCs) reduced output synchrony. In this study, we extend the previous phase model to a conductance based model to examine how the density of specific ion channels in mitral cells impacts on stochastic synchrony. A recent study revealed that mitral cells are highly heterogeneous in the expression of the sag current, a hyperpolarization-activated inward current (Angelo, 2011). The variability in the sag contributes to the diversity of mitral cells and thus we wanted to know how this variability influences synchronization. Mitral cell oscillations and bursting are also regulated by an inactivating potassium current (IA). Based on these ion channels, we examined the effect of changing the current densities (gA, gH) on diversity of PRCs and of synchrony. In order to identify oscillatory patterns of bursting and repetitive spiking across gA and gH to the model, two parameter bifurcation analysis was performed in the presence and absence of noise. Increasing gH alone reduces the region of bursting, but does not completely eliminate bursting, and PRCs changed much more with respect to gA than gH. Focusing on varying gA, we next examined a role of gA density and firing rate in stochastic synchrony by introducing the fluctuating correlated input resembling the shared presynaptic drives. We found that heterogeneity in A-type current mainly influenced on stochastic synchrony as we predicted in PRCs investigated theoretically, and diversity in firing rate alone didn’t account for it. In addition, heterogeneous population with respect to gA, given decent amount of gA density, showed better stochastic synchrony than homogeneous population in same firing rate. P12 Circular statistics of noise in spike trains with a periodic component Petr Marsalek1,2 1Institute of Pathological Physiology, First Faculty of Medicine, Charles University in Prague, 128 53, Czech Republic; 2Czech Technical University in Prague, Zikova 1903/4, 166 36, Czech Republic Correspondence: Petr Marsalek - petr.marsalek@lf1.cuni.cz BMC Neuroscience 2016, 17(Suppl 1):P12 Introduction We estimate parameters of the inter-spike interval distributions in binaural neurons of the mammalian sound localization neural circuit, neurons of the lateral and medial superior olive [1]. We present equivalent descriptions of spike time probabilities using both standard and circular statistics. We show that the difference between sine function and beta density in the circular domain is negligible. Results Estimation of the spike train probability density function parameters is presented in relation to harmonic and complex sound input. The resulting densities are expressed analytically with the use of harmonic and Bessel functions. Parameter fits are verified by numerical simulations of spike trains (Fig. 14). Fig. 14 Comparison of circular probability density functions of sine and beta density. A Beta density with parameters a = b = 3.3818, matches closely that of the sine function, used as a probability density function (PDF). Beta density with parameters a = b = 3 solid line, is matched by sine function y = 1.05 − 1.1 cos(2π x/1.1). B Cumulative distribution function (CDF) is shown for these densities together with the difference between the two CDFs multiplied by 100 to visualize the comparison of the two distributions. C For testing different vector strengths we use uniform distributions with pre-set vector strengths (ρ = 0.8, 0.5 and 0.08) Conclusions We use analytical techniques, where it is possible. We calculate the one-to-one correspondence of vector strength parameters and parameters of circular distributions used for description of data. We show here introductory figure of our paper with the two representative circular densities. We also use experimental data [2, 3] and simulated data to compare them with these theoretical distributions. Acknowledgements: Supported by the PRVOUK program no. 205024 at the Charles University in Prague. I acknowledge contributions to the analytical computations by Ondrej Pokora and simulation in Matlab by Peter G. Toth. ReferencesBures Z, Marsalek P. On the precision of neural computation with interaural level differences in the lateral superior olive. Brain Res. 2013;1536:16–26. Joris P, Carney L, Smith P, Yin T. Enhancement of neural synchronization in the anteroventral cochlear nucleus. I. Responses to tones at the characteristic frequency. J Neurophysiol. 1994;71(3):1022–36. Joris P, Smith P, Yin T. Enhancement of neural synchronization in the anteroventral cochlear nucleus. II. Responses in the tuning curve tail. J Neurophysiol. 1994;71(3):1037–51. P14 Representations of directions in EEG-BCI using Gaussian readouts Hoon-Hee Kim1,2, Seok-hyun Moon3, Do-won Lee3, Sung-beom Lee3, Ji-yong Lee3, Jaeseung Jeong1,2 1Department of Bio and Brain Engineering and 2Program of Brain and Cognitive Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea, 34141; 3Korea Science Academy of KAIST, Busan, South Korea, 10547 Correspondence: Jaeseung Jeong - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P14 EEG (electroencephalography) is one of most useful neuroimaging technology and best options for BCI (Brain-Computer Interface) because EEG has portable size, wireless and well-wearing design in any situations. The key objective of BCI is physical control of machine such as cursor movement in screen and robot movement [1, 2]. In previously study, the motor imagery had used for represent of direction to movement [1, 2]. For example, the left hand imagery mapping to move the left, the right hand imagery mapping to move the right and both hand imagery mapping to move the forward. In this study, however, we considered only brain signals when a subject thinks directions to movements not motor imageries. We designed the recurrent neural networks which consist of 300–10,000 artificial linear neurons using Echo State Networks paradigm [3]. We also recorded EEG signals using Emotiv EPOC+ which has 16 channels (AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4 and two of reference). All raw data of channels were normalized and then used inputs to recurrent neural networks. For representation of directions, we had built Gaussian readouts which has preferred directions and fitted the Gaussian functions (Fig. 15). The firing rate of readout were high when the subject thought preferred direction. However, when the subject thought not preferred direction, the firing rate of readout slightly low down. For implement these readouts, all of neuros in recurrent neural networks had linearly connected to all readouts and weights of these connections were trained by linear learning rules. In result, we considered 5 healthy subjects and recorded EEG signals for each directions. The readouts were showed well Gaussian fitted direction preference. In this study, we considered only two dimensions but many situations of BCI has three dimensional space. Therefore, our study which using Gaussian readouts should be extended to three dimensional version.Fig. 15 Design of recurrent neural networks and readouts ReferencesChae Y, Jeong J, Jo S. Toward brain-actuated humanoid robots: asynchronous direct control using an EEG-based BCI. IEEE Trans Robot. 2012;28(5):1131–44. LaFleur K, Cassady K, Doud A, Shades K, Rogin E, He B. Quadcopter control in three-dimensional space using a noninvasive motor imagery-based brain–computer interface. J Neural Eng. 2013;10(4):046003. Jaeger H, Haas H. Harnessing nonlinearity predicting chaotic systems and saving energy in wireless communication. Science. 2004;304(5667):78–80. P15 Action selection and reinforcement learning in basal ganglia during reaching movements Yaroslav I. Molkov1, Khaldoun Hamade2, Wondimu Teka3, William H. Barnett1, Taegyo Kim2, Sergey Markin2, Ilya A. Rybak2 1Department of Mathematics and Statistics, Georgia State University, Atlanta, GA 30303, USA; 2Department of Neurobiology and Anatomy, Drexel University, Philadelphia, PA 19129, USA; 3Department of Mathematical Sciences, Indiana University – Purdue University, Indianapolis, IN 46202, USA Correspondence: Yaroslav I. Molkov - ymolkov@gsu.edu BMC Neuroscience 2016, 17(Suppl 1):P15 The basal ganglia (BG) comprise a number of interconnected nuclei that are collectively involved in a wide range of motor and cognitive behaviors. The commonly accepted theory is that the BG play a pivotal role in action selection and reinforcement learning facilitated by the activity of dopaminergic neurons of substantia nigra pars compacta (SNc). These dopaminergic neurons encode prediction errors when reward outcomes exceed or fall below anticipated values. The BG gate appropriate behaviors from multiple moto-cortical command candidates arriving at the striatum (BG’s input nuclei) but suppress competing inappropriate behaviors. The selected motor action is realized when the internal segment of the globus pallidus (GPi) (BG’s output nuclei) disinhibits thalamic neurons corresponding to the gated behavior. The BG network performs motor command selection through the facilitation of the appropriate behavior via the “direct” striatonigral (GO) pathway and inhibition of competing behaviors by the “indirect” striatopallidal (NOGO) pathway. Several modeling studies have showed plausibility of the above concept in simplified cases, e.g. for binary action selection in response to a binary cue. However, in these previous models, the possible actions/behaviors were represented in an abstract way, and did not have a detailed implementation as specific neuronal patterns actuating the muscular-skeletal apparatus. To address these details, the motor system in the present study includes a 2D-biomechanical arm model in the horizontal plane to simulate realistic reaching movements. The arm consists of two segments (upper arm and forearm) and has two joints (shoulder and elbow) controlled by four monoarticular (flexor and extensor at each joint) and two bi-articular (shoulder and elbow flexor, and shoulder and elbow extensor) muscles. The neural component of the model includes the BG, the thalamus, the motor cortex, and spinal circuits. The low-level spinal circuitry contains six motoneurons (each controlling one muscle), and receives proprioceptor feedback from muscles. Cortical neurons provide inputs to the spinal network. Their activity is calculated by solving an inverse problem (inverting the internal model) based on the initial position of the arm, reaching distance and direction. In the model, reaching movements in different directions were used as a set of possible behaviors. We simulated movements in response to a sensory cue defining the target arm position. The cortex generated signals corresponding to the cue and all possible motor commands and delivered these signals to the BG. The resulting neuronal patterns in the motor cortex were calculated as a convolution of the thalamic activity and all possible motor commands. The function of BG was to establish the association between the cue and the appropriate action(s) by adjusting weights of plastic corticostriatal projections through reinforcement learning. The BG model contained an exploratory mechanism, operating through the subthalamic nucleus (STN) that allowed the model to constantly seek better cue-action associations that deliver larger rewards. Reinforcement learning relied on the SNc dopaminergic signal that measured trial-to-trial changes in the reward value, defined by performance errors. Using this model, we simulated several learning tasks in the conditions of different unexpected perturbations. When a perturbation was introduced, the model was capable of quickly switching away from pre-learned associations and learning novel cue-action associations. The analysis of the model reveals several features, that can have general importance for brain control of movements: (1) potentiation of the cue-NOGO projections is crucial for quick destruction of preexisting cue-action associations; (2) the synaptic scaling (the decay of the cortical-striatal synaptic weights in the absence of dopamine-mediated potentiation/depression) has a relatively short time-scale (10–20 trials); (3) quick learning is associated with a relatively poor accuracy of the resultant movement. We suggest that BG may be involved in a quick search for behavioral alternatives when the conditions change, but not in the learning of skilled movements that require good precision. P17 Axon guidance: modeling axonal growth in T-junction assay Csaba Forro1, Harald Dermutz1,László Demkó1, János Vörös1 1LBB, ETH Zürich, Zürich, 8051, Switzerland Correspondence: Csaba Forro - forro@biomed.ee.ethz.ch BMC Neuroscience 2016, 17(Suppl 1):P17 The current field of neuroscience investigates the brain at scales varying from the whole organ, to brain slices and down to the single cell level. The technological advances miniaturization of electrode arrays has enabled the investigation of neural networks comprising several neurons by recording electrical activity from every individual cell in the network. This level of complexity is key in the study of the core principles at play in the machinery of the brain. Indeed, it is the first layer of complexity above the single cell that is still tractable for the human scientist without needing to resort to a ‘Big Data’ approach. In light of this, we strive to create topologically well-defined neural networks, akin to mathematical directed graphs, as a model systems in order to study the basic mechanisms emerging in networks of increasing complexity and varying topology. This approach will also yield statistically sound and reproducible observations, something which is sought after in neuroscience [1]. The first step in realizing such a well-defined neural network is to reliably control the guidance of individual axons in order to connect the network of cells in a controlled way. For this purpose, we present a method consisting of obstacles forcing the axon to turn one way or the other. The setup is made of PolyDiMethylSiloxane (PDMS) which is microstructured by ways of state of the art photolithography procedures. Two tunnels of 5 µ height are patterned into a block of 100 µ thick PDMS and connected in the shape of a T-junction (Fig. 16). Primary cortical neurons are inserted via entry holes at the base of the tunnels. The entry angle of the bottom tunnel (“vertical part of the T”) into the junction is varied between 20° (steep entry) and 90° (vertical entry). We observe that the axons prefer to turn towards the smaller angle. We show how this observed angular selectivity in axon guidance can be explained by a simple model and how this principle can be used to create topologically well-defined neural networks (Fig. 16B).Fig. 16 A The T-junction assay with an entry angle of 20°. The axon is expected to prefer a right-turn at this angle. B A simple model is constructed where the direction of growth of the axon is proportional to area (red) it can explore ReferenceButton KS, et al. Power failure: why small sample size undermines the reliability of neuroscience. Nat Rev Neurosci. 2013;14(5):365–76. P19 Transient cell assembly networks encode persistent spatial memories Yuri Dabaghian1,2, Andrey Babichev1,2 1Department of Neurology Pediatrics, Baylor College of Medicine, Houston, TX 77030, USA; 2Department of Computational and Applied Mathematics, Rice University, Houston, TX, 77005, USA Correspondence: Yuri Dabaghian - dabaghian@rice.edu BMC Neuroscience 2016, 17(Suppl 1):P19 The reliability of our memories is nothing short of remarkable. Thousands of neurons die every day, synaptic connections appear and disappear, and the networks formed by these neurons constantly change due to various forms of synaptic plasticity. How can the brain develop a reliable representation of the world, learn and retain memories despite, or perhaps because of, such complex dynamics? Here we consider the specific case of spatial navigation in mammals, which is based on mental representations of their environments—cognitive maps—provided by the network of the hippocampal place cells—neurons that become active only in a particular region of the environment, known as their respective place fields. Experiments suggest that the hippocampal map is fundamentally topological, i.e., more similar to a subway map than to a topographical city map, and hence amenable to analysis by topological methods [1]. By simulating the animal’s exploratory movements through different environments we studied how stable topological features of space get represented by assemblies of simulated neurons operating under a wide range of conditions, including variations in the place cells’ firing rate, the size of the place fields, the number of cells in the population [2,3]. In this work, we use methods from Algebraic Topology to understand how the dynamic connections between hippocampal place cells influence the reliability of spatial learning. We find that although the hippocampal network is highly transient, the overall spatial map encoded by the place cells is stable. Acknowledgements: The work was supported by the NSF 1422438 grant and by the Houston Bioinformatics Endowment Fund. ReferencesDabaghian Y, Brandt VL, Frank LM. Reconceiving the hippocampal map as a topological template. eLife. 2014. doi:10.7554/eLife.03476. Dabaghian Y, Mémoli F, Frank L, Carlsson G. A topological paradigm for hippocampal spatial map formation using persistent homology. PLoS Comput Biol. 2012;8:e1002581. Arai M, Brandt V, Dabaghian Y. The effects of theta precession on spatial learning and simplicial complex dynamics in a topological model of the hippocampal spatial map. PLoS Comput Biol. 2014;10:e1003651. P20 Theory of population coupling and applications to describe high order correlations in large populations of interacting neurons Haiping Huang1 1RIKEN Brain Science Institute, Wako-shi, Saitama, Japan Correspondence: Haiping Huang - physhuang@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P20 Correlations among neurons spiking activities play a prominent role in deciphering the neural code. Various models were proposed to understand the pairwise correlations in the population activity. Modeling these correlations sheds light on the functional organization of the nervous system. In this study, we interpret correlations in terms of population coupling, a concept recently proposed to understand the multi-neuron firing patterns of the visual cortex of mouse and monkey [1]. We generalize the population coupling to its higher order (PC2), characterizing the relationship of pairwise firing with the population activity. We derive the practical dimensionality reduction method for extracting the low dimensional representation parameters, and test our method on different types of neural data, including ganglion cells in the salamander retina onto which a repeated natural movie was projected [2], and layer 2/3 as well as layer 5 cortical cells in the medial prefrontal cortex (MPC) of behaving rats [3]. For the retinal data, by considering the correlation between the pairwise firing activity and the global population activity, i.e., the second order population coupling, the three-cell correlation could be predicted partially (64.44 %), which suggests that PC2 acts as a key circuit variable for third order correlations. The interaction matrix revealed here may be related to the found overlapping modular structure of retinal neuron interactions [4]. In this structure, neurons interact locally with their adjacent neurons, and in particular this feature is scalable and applicable for larger networks. About 94.79 % of three-cell correlations are explained by PC2 in the MPC circuit. The PC2 matrix shows clear hubs’ structure in the cortical circuit. Some neuron interacts strongly with a large portion of neurons in the population, and such neurons may play a key role in shaping the collective spiking behavior during the working memory task. The hubs and non-local effects are consistent with findings reported in the original experimental paper [3]. Acknowledgements: We are grateful to Shigeyoshi Fujisawa and Michael J Berry for sharing us the cortical and retinal data, respectively. We also thank Hideaki Shimazaki and Taro Toyoizumi for stimulating discussions. This work was supported by the program for Brain Mapping by Integrated Neurotechnologies for Disease Studies (Brain/MINDS) from Japan Agency for Medical Research and development, AMED. ReferencesOkun M, Steinmetz NA, Cossell L, Iacaruso MF, Ko H, Bartho P, et al. Diverse coupling of neurons to populations in sensory cortex. Nature. 2015;521:511–15. Tkacik G, Marre O, Amodei D, Schneidman E, Bialek W, Berry II MJB. Searching for collective behavior in a large network of sensory neurons. PLoS Comput Biol. 2014;10:e1003408. Fujisawa S, Amarasingham A, Harrison MT, Buzsaki G. Behavior-dependent short-term assembly dynamics in the medial prefrontal cortex. Nat Neurosci. 2008;11:823–33. Ganmor E, Segev R, Schneidman E. The architecture of functional interaction networks in the retina. J Neurosci. 2011;31(8):3044–54. P21 Design of biologically-realistic simulations for motor control Sergio Verduzco-Flores1 1Computational Neuroscience Unit, Okinawa Institute of Science and Technology, Okinawa 1919-1, Japan Correspondence: Sergio Verduzco-Flores - sergio.verduzco@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P21 Several computational models of motor control, although apparently feasible, fail when simulated in 3-dimensional space with redundant manipulators [1, 2]. Moreover, it has become apparent that the details of musculoskeletal simulations, such as the muscle model used, can fundamentally affect the conclusions of a computational study [3]. There would be great benefits from being able to test theories involving motor control within a simulation framework that brings realism in the musculoskeletal model, and in the networks that control movements. In particular, it would be desirable to have: (1) a musculoskeletal model considered to be research-grade within the biomechanics community, (2) afferent information provided by standard models of the spindle afferent and the Golgi tendon organ, (3) muscle stimulation provided by a spiking neural network that follows the basic known properties of the spinal cord, and (4) a cerebellar network as part of adaptive learning. Creating this type of model is only now becoming practical, not only due to faster computers, but due to properly validated musculoskeletal models and simulation platforms from the biomechanics community, as well as mature software and simulations techniques from the computational neuroscience community. We show how these can be harnessed in order to create simulations that are grounded both by physics and by neural implementation. This pairing of computational neuroscience and biomechanics is sure to bring further insights into the workings of the central nervous system. ReferencesGielen S. Review of models for the generation of multi-joint movements in 3D. In: Sternad D, editor. Progress in motor control. New-York: Springer; 2009. Verduzco-Flores SO, O’Reilly RC. How the credit assignment problems in motor control could be solved after the cerebellum predicts increases in error. Front Comput Neurosci. 2015;9:39. Gribble PL, Ostry DJ, Sanguineti V, Laboissière R. Are complex control signals required for human arm movement? J Neurophysiol. 1998;79:1409–24. P22 Towards understanding the functional impact of the behavioural variability of neurons Filipa Dos Santos1, Peter Andras1 1School of Computing and Mathematics, Keele University, Newcastle-under-Lyme, ST5 5BG, UK Correspondence: Filipa Dos Santos - f.d.s.brandao@keele.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P22 The same neuron may play different functional roles in the neural circuits to which it belongs. For example, neurons in the Tritonia pedal ganglia may participate in variable phases of the swim motor rhythms [1]. While such neuronal functional variability is likely to play a major role the delivery of the functionality of neural systems, it is difficult to study it in most nervous systems. We work on the pyloric rhythm network of the crustacean stomatogastric ganglion (STG) [2]. Typically network models of the STG treat neurons of the same functional type as a single model neuron (e.g. PD neurons), assuming the same conductance parameters for these neurons and implying their synchronous firing [3, 4]. However, simultaneous recording of PD neurons shows differences between the timings of spikes of these neurons. This may indicate functional variability of these neurons. Here we modelled separately the two PD neurons of the STG in a multi-neuron model of the pyloric network. Our neuron models comply with known correlations between conductance parameters of ionic currents. Our results reproduce the experimental finding of increasing spike time distance between spikes originating from the two model PD neurons during their synchronised burst phase. The PD neuron with the larger calcium conductance generates its spikes before the other PD neuron. Larger potassium conductance values in the follower neuron imply longer delays between spikes, see Fig. 17.Fig. 17 The time distances between the first and second spikes of the simulated PD neurons as a function of the gK and gCaT conductances of the neuron with variable conductances. A first spikes. B Second spikes. The PD neuron with fixed conductances had gK = 1.5768 μS and gCaT = 0.0225 μS Neuromodulators change the conductance parameters of neurons and maintain the ratios of these parameters [5]. Our results show that such changes may shift the individual contribution of two PD neurons to the PD-phase of the pyloric rhythm altering their functionality within this rhythm. Our work paves the way towards an accessible experimental and computational framework for the analysis of the mechanisms and impact of functional variability of neurons within the neural circuits to which they belong. ReferencesHill ES, Vasireddi SK, Bruno AM, Wang J, Frost WN. Variable neuronal participation in stereotypic motor programs. PLoS One. 2012;7:1–11. Bucher D, Johnson CD, Marder E. Neuronal morphology and neuropil structure in the stomatogastric ganglion. J Comp Neurol. 2007;501:185–205. Soto-Treviño C, Rabbah P, Marder E, Nadim F. Computational model of electrically coupled, intrinsically distinct pacemaker neurons. J Neurophysiol. 2005;94:590–604. Golowasch J, Casey M, Abbott LF, Marder E. Network stability from activity-dependent regulation of neuronal conductances. Neural Comput. 1999;11:1079–96. Khorkova O, Golowasch J. Neuromodulators, not activity, control coordinated expression of ionic currents. J Neurosci. 2007;27:8709–18. P23 Different oscillatory dynamics underlying gamma entrainment deficits in schizophrenia Christoph Metzner1, Achim Schweikard2, Bartosz Zurowski3 1Science and Technology Research Institute, University of Hertfordshire, Hatfield, United Kingdom; 2Institute for Robotics and Cognitive Systems, University of Luebeck, Luebeck, Germany; 3Department of Psychiatry, University of Luebeck, Schleswig–Holstein, Luebeck, Germany Correspondence: Filipa Dos Santos - c.metzner@herts.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P23 In recent years, a significant amount of biomarkers and endophenotypic signatures of psychiatric illnesses have been identified, however, only a very limited number of computational models in support thereof have been described so far [1]. Furthermore, the few existing computational models typically only investigate one possible mechanism in isolation, disregarding the potential multifactoriality of the network behaviour [2]. Here we describe a computational instantiation of an endophenotypic finding for schizophrenia, an impairment in gamma entrainment in auditory click paradigms [3]. We used a model of primary auditory cortex from Beeman [4] and simulated a click entrainment paradigm with stimulation at 40 Hz, to investigate gamma entrainment deficits, and at 30 Hz as a control condition. We explored the multifactoriality by performing an extensive parameter search (approx. 4000 simulations). We focused on synaptic and connectivity parameters of the fast spiking inhibitory interneurons in the model (i.e. number and strength of and, GABAergic decay times at I-to-E and I-to-I connections, independently). We performed a time–frequency analysis of simulated EEG signals and extracted the power in the 40 Hz and the 30 Hz band, respectively. Using the power in the 40 Hz band for 40 Hz stimulation we identified regions in the parameter space showing strong reductions in gamma entrainment. For these we calculated cycle-averaged EEG signals and spike time histograms of both network populations, in order to explore the dynamics underlying the reduction in gamma power. We find three regions in the parameter space which show strong reductions in gamma power. These three regions, however, have very different parameter settings and show very different oscillatory dynamics. The first, which produces the strongest reduction, is characterised by a strong prolongation of decay times at I-to-E synapses and strong and numerous I-to-E connections. Cycle-averaged spike histograms show a broadening of distributions which indicate that the overall synchrony is reduced, leading to the strong reduction in gamma power. However, this parameter setting also produced a strong reduction of power in the 30 Hz control condition, which is not seen experimentally. The second region, is characterized by prolonged I-to-I decay times together with numerous and strong I-to-I connectivity. Here, a second peak appears in the cycle-average spike histogram of the excitatory population, which leads to a loss of synchrony and thus a reduction in gamma power. The third parameter region, is also characterized by prolonged I-to-I decay times. Moreover, it is associated with a reduction in I-to-I connection numbers and strengths together with strong I-to-E connections. Here, we found that in every second cycle, the spike histogram of the inhibitory neurons showed two peaks, one at the beginning and one in the middle of the cycle. This second peak then inhibited the excitatory neurons’ response to the next stimulation. Hence, the EEG signal showed beat-skipping, i.e. every second gamma peak was suppressed, resulting in a decrease in gamma power. Performing an extensive parameter search in an in silico instantiation of an endophenotypic finding for schizophrenia, we have identified distinct regions of the parameter space that give rise to analogous network level behaviour found in schizophrenic patients using electrophysiology [3]. However, the oscillatory dynamics underlying this behaviour substantially differ across regions. These regions might correspond to different subtypes of schizophrenic patients and hence, subtypes of what might have different targets for alleviating the deficits because of their differences in underlying dynamics. ReferencesSiekmeier P. Computational modeling of psychiatric illnesses via well-defined neurophysiological and neurocognitive biomarkers. Neurosci Biobehav Rev. 2015;57:365–80. Pavão, R, Tort ABL, Amaral OB. Multifactoriality in psychiatric disorders: a computational study of schizophrenia. Schizophrenia Bull. 2015;41(4):980–88. Kwon JS, O’Donnell BF, Wallenstein GV, Greene RW, Hirayasu Y, Nestor PG, Hasselmo ME, Potts GF, Shenton ME, McCarley RW. Gamma frequency-range abnormalities to auditory stimulation in schizophrenia. Arch Gen Psychiatry. 1999;56(11):1001–5. Beeman D. A modeling study of cortical waves in primary auditory cortex. BMC Neurosci. 2013;14(Suppl. 1):P23. P24 Memory recall and spike frequency adaptation James P. Roach1, Leonard M. Sander2,3, Michal R. Zochowski2,3,4 1Neuroscience Graduate Program, University of Michigan, Ann Arbor, MI 48109, USA; 2Center for the Study of Complex Systems, University of Michigan, Ann Arbor, MI 48109, USA; 3Department of Physics, University of Michigan, Ann Arbor, MI 48109, USA; 4Biophysics Program, University of Michigan, Ann Arbor, MI 48109, USA Correspondence: James P. Roach - roachjp@umich.edu BMC Neuroscience 2016, 17(Suppl 1):P24 In the brain, representations of the external world are encoded by patterns of neural activity. It is critical that representations be stable, but still easily moved between. This phenomenon has been modeled at the network level as auto associative memory. In auto associative network models, such as the Hopfield network, representations, or memories, are stored within synaptic weights and form stable fixed points, or attractors [1]. Spike frequency adaptation (SFA) provides a biologically plausible mechanism for switching between stabile fixed points in the Hopfield network. In the present work we show that for low levels of SFA networks will stabilize in a representation that corresponds to the nearest memory activity space, regardless of strength. In networks with higher levels of SFA only the pattern corresponding to the strongest memory, or a global minimum in activity space. The effects of SFA are similar to fast, or thermodynamic noise, but also allows for deterministic destabilization of memories leading to periodic activation of memories through time. We argue that control of SFA level is a universal mechanism for network-wide attractor selectivity. SFA is tightly regulated by the neurotransmitter acetylcholine (ACh) and can be changed on behaviorally relevant timescales. To support this claim we demonstrate that SFA controls selectivity of spatial attractors in a biophysical model of cholinergic modulation in cortical networks [2, 3]. This model produces localized bumps of firing. A region with enhanced recurrent excitation acts as an attractor for the bump location and selectivity for these regions is quickly diminishes as SFA level increases [3]. When multiple spatial attractors of varying strengths are stored in a network moderate increases SFA level will lead to the weak attractors being destabilized and activity localizing within the strongest attractor. This effect is qualitatively similar to the effects of SFA in the Hopfield network. These results indicate that ACh controls memory recall and perception within the cortex by regulation of SFA and explain the important role cholinergic modulation plays in cognitive functions such as attention and memory consolidation [4]. Acknowledgements: JPR was supported by an NSF Graduate Research Fel- lowship Program under Grant No. DGE 1256260 and a UM Rackham Merit Fellowship. MRZ and LMS were supported by NSF PoLS 1058034. ReferencesHopfield JJ. Neural networks and physical systems with emergent collective computational abilities. PNAS. 1982;79: 2554–8. Stiefel KM, Gutkin BS, Sejnowski TJ. The effects of cholinergic neuromodulation on neuronal phase-response curves of modeled cortical neurons. J Comp Neurosci. 2008;26:289–301. Roach JP, Ben-Jacob E, Sander LM, Zochowski MR. Formation and dynamics of waves in a cortical model of cholinergic modulation. PLoS Comput Biol. 2015;11(8): e1004449. Hasselmo ME, Sarter M. Modes and models of forebrain cholinergic neuromodulation of cognition. Neuropsychopharmacology. 2011;36:52–73. P25 Stability of neural networks and memory consolidation preferentially occur near criticality Quinton M. Skilling1, Nicolette Ognjanovski2, Sara J. Aton2, Michal Zochowski1,3 1Biophysics Program, University of Michigan, Ann Arbor, MI 48109 USA; 2Department of Molecular, Cellular, and Developmental Biology, University of Michigan, Ann Arbor, MI, 48109 USA; 3Department of Physics, University of Michigan, Ann Arbor, MI 48109 USA Correspondence: Michal Zochowski - michalz@umich.edu BMC Neuroscience 2016, 17(Suppl 1):P25 Dynamic neural representations underlie cognitive processing and are an outcome of complex interactions of network structural properties and cellular dynamics. We have developed a new framework to study dynamics of network representations during rapid memory formation in the hippocampus in response to contextual fear conditioning (CFC) [1]. Experimentally, this memory paradigm is achieved by exposing mice to foot shocks while in a novel environment and later testing for behavioral responses when reintroduced to that environment. We employ the average minimum distance (AMD) functional connectivity algorithm to spiking data recorded before, during, and after CFC using implanted stereotrodes. Comparing changes in functional connectivity using cosine similarity, we find that stable functional representations correlate well with animal performance in learning. Using extensive computer simulations, we show that the most robust changes compared to baseline occur when the system resides near criticality. We attribute these results to emergence of long-range correlations during the initial process of memory formation. Furthermore, we have developed a generic model using a generalized Hopfield framework to link formation of novel memory representation to functional stability changes. The network initially stores a single representation, which is to exemplify biologically already stored (old) memories, and is then presented a new representation by freezing a randomly chosen fraction of nodes from a novel pattern. We show that imposing fractional input of the new representation may partially stabilize this representation near the phase transition (critical) point. We further show that invoking synaptic plasticity rules may fully stabilize this new representation only when the dynamics of the network reside near criticality. Taken together these results show, for the first time, that only when the network is at criticality can it stabilize novel memory representations, the dynamical regime which also yields an increase of network stability. Furthermore, our results match well experimental data observed from CFC experiments. ReferenceOgnjanovski N, Maruyama D, Lashner N, Zochowski M, Aton SJ. CA1 hippocampal network activity changes during sleep-dependent memory consolidation. Front Syst Neurosci. 2014;8:61. P26 Stochastic oscillation in self-organized critical states of small systems: sensitive resting state in neural systems Sheng-Jun Wang1,2, Guang Ouyang2, Jing Guang3, Mingsha Zhang3, K. Y. Michael Wong4, Changsong Zhou2,5,6 1Department of Physics, Shaanxi Normal University, Xi’An City, ShaanXi Province, China; 2Department of Physics and Centre for Nonlinear Studies, Institute of Computational and Theoretical Studies, Hong Kong Baptist University, Kowloon Tong, Hong Kong; 3State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing, China; 4Department of Physics, Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong; 5Beijing Computational Science Research Center, Beijing 100084, People’s Republic of China; 6Research Centre, HKBU Institute of Research and Continuing Education, Shenzhen, China Correspondence: Changsong Zhou - cszhou@hkbu.edu.hk BMC Neuroscience 2016, 17(Suppl 1):P26 Self-organized critical states (SOCs) and stochastic oscillations (SOs) are simultaneously observed in neural systems [1], which appears to be theoretically contradictory since SOCs are characterized by scale-free avalanche sizes but oscillations indicate typical scales. Here, we show that SOs can emerge in SOCs of small size systems due to temporal correlation between large avalanches at the finite-size cutoff, resulting from the accumulation-release process in SOCs. In contrast, the critical branching process without accumulation-release dynamics cannot exhibit oscillations. The reconciliation of SOCs and SOs is demonstrated both in the sandpile model and robustly in biologically plausible neuronal networks. The oscillations can be suppressed if external inputs eliminate the prominent slow accumulation process, providing a potential explanation of the widely studied Berger effect or event-related desynchronization in neural response. The features of neural oscillations and suppression are confirmed during task processing in monkey eye-movement experiments. Our results suggest that finite-size, columnar neural circuits may play an important role in generating neural oscillations around the critical states, potentially enabling functional advantages of both SOCs and oscillations for sensitive response to transient stimuli. The results have been published in [2]. Acknowledgements: This work was partially supported by Hong Kong Baptist University Strategic Development Fund, NSFCRGC Joint Research Scheme HKUST/NSFC/12-13/01 (or N-HKUST 606/12), RGC (Grants No. 604512, No. 605813, and No. 12302914), NSFC (Grants No.11275027, No. 11328501, and No. 11305098), and the Fundamental Research Funds for the Central Universities (Grant No. GK201302008). ReferencesGireesh E, Plenz D, Neuronal avalanches organize as nested theta-and beta/gamma-oscillations during development of cortical layer 2/3. Proc Natl Acad Sci USA. 2008;105:7576–81. Wang SJ, Ouyang G, Guang J, Zhang MS, Wong KYM, Zhou CS. Stochastic oscillation in self-organized critical states of small systems: sensitive resting state in neural systems. Phys Rev Lett. 2016;116:018101. P27 Neurofield: a C++ library for fast simulation of 2D neural field models Peter A. Robinson1,2, Paula Sanz-Leon1,2, Peter M. Drysdale1,2, Felix Fung1,2, Romesh G. Abeysuriya3, Chris J. Rennie1,2, Xuelong Zhao1,2 1School of Physics, University of Sydney, Sydney, New South Wales, 2006, Australia; 2Center for Integrative Brain Function, University of Sydney, Sydney, New South Wales, 2006, Australia Correspondence: Paula Sanz-Leon - paula.sanz-leon@sydney.edu.au BMC Neuroscience 2016, 17(Suppl 1):P27 Neural field theory [1] has addressed numerous questions regarding brain dynamics and its interactions across many scales, becoming a highly flexible and unified framework for the study and prediction experimental observables of the electrical activity of the brain. These include EEG spectra [2, 3], evoked response potentials, age-related changes to the physiology of the brain [4], epileptic seizures [5, 6], and synaptic plasticity phenomena [7]. However, numerical simulations of neural field models are not widely available despite their extreme usefulness in cases where analytic solutions are less tractable. This work introduces the features of NeuroField, a research-ready library applicable to simulate a wide range of neural field based systems involving multiple structures (e.g., cortex, cortex and thalamic nuclei, and basal ganglia). The link between a given neural field model, its mathematical representation (i.e., a delay-partial differential equations system with spatial periodic boundary conditions) and its computational implementation is described. The resulting computational model has the capability to represent from spatially extended to neural-mass-like systems, and it has been extensively validated against analytical solutions and against experiment [1–10]. To illustrate its flexibility, a range of simulations modeling a variety of arousal-, sleep- and epilepsy-state phenomena is presented [8, 9]. NeuroField has been written using object-oriented programming in C++ and is bundled together with MATLAB routines for quantitative offline analysis, such as spectral and dynamic spectral analysis. ReferencesRobinson PA, Rennie CJ, Wright JJ. Propagation and stability of waves of electrical activity in the cortex. Phys Rev E. 1997;56:826–40. Robinson PA, Rennie CJ, Wright JJ, Bahramali H, Gordon E, Rowe D. Prediction of electroencephalographic spectra from neurophysiology. Phys Rev E. 2001;63:021903. Robinson PA, Rennie CJ, Rowe DL, O’Connor SC. Estimation of multiscale neurophysiologic parameters by electroencephalographic means. Hum Brain Mapp. 2004;23:53–72. van Albada SJ, Kerr CC, Chiang AKI, Rennie CJ, Robinson PA. Neurophysiological changes with age probed by inverse modeling of EEG spectra. Clin Neurophysiol. 2010;121:21–38. Robinson PA, Rennie CJ, Rowe DL. Dynamics of large-scale brain activity in normal arousal states and epileptic seizures. Phys Rev E. 2002; 65:041924. Breakspear M, Roberts JA, Terry JR, Rodrigues S, Mahant N, Robinson PA. A unifying explanation of primary generalized seizures through nonlinear brain modeling and bifurcation analysis. Cereb Cortex. 2006;16:1296–1313. Fung PK, Haber AL, Robinson PA. Neural field theory of large-scale synaptic plasticity in the cerebral cortex. J Theor Biol. 2013; 318:44–57. Abeysuriya RG, Rennie CJ, Robinson PA. Physiologically based arousal state estimation and dynamics. J Neurosci Methods. 2015; 253:55–69. Robinson, PA, Postnova, S, Abeysuriya, RG, Kim, JK, Roberts, JA, McKenzie-Sell L, Karanjai, A, Kerr, CC, Fung, F, Anderson, R, Breakspear, MJ, Drysdale, PM, Fulcher, BD, Phillips, AKJ, Rennie, CJ, Yin G. Chapter 5: a multiscale “working brain” model. In: Validating neurocomputational models of of neurological and psychiatric disorders. Paris: Springer; 2015. O’Connor SC, Robinson PA. Spatially uniform and nonuniform analysis of electroencephalographic dynamics, with application to the topography of the alpha rhythm. Phys Rev E. 2004;70:110–9. P28 Action-based grounding: Beyond encoding/decoding in neural code Yoonsuck Choe1, Huei-Fang Yang2 1Department of Computer Science & Engineering, Texas A&M University, College Station, TX, 77845, USA; 2Research Center for Information Technology Innovation, Academia Sinica, Taipei, Taiwan Correspondence: Yoonsuck Choe - choe@tamu.edu BMC Neuroscience 2016, 17(Suppl 1):P28 How can we decode the neural activation patterns (Fig. 18A)? This is a key question in neuroscience. We as scientists have the luxury of controlling the stimulus, based on which we can find the meaning of the spikes (Fig. 18C-right). However, as shown in Fig. 18A (and C-left), the problem seems intractable from the point of view of the brain itself since neurons deeply embedded in the brain do not have direct access to the stimulus. In [1] and related work, we showed that the decoding problem seems intractable only because we left out the motor system from the picture. Figure 18D shows how motor action can help processes deeply embedded in the brain can understand the meaning of the spikes by generating motor behavior and observing the resulting change in the neural spikes. Here, a key principle is to generate motion that keeps the neural spike pattern invariant over time (Fig. 18E), which allows the following to coincide (1) the property of the motion (diagonal movement) and (2) the encoded property of the input (45° orientation). Using reinforcement learning, we showed that the invariance criterion leads to near optimal state-action mapping for synthetic and natural image inputs (Fig. 18F, G), where the encoded property of the input is mapped to congruent motor action. Furthermore, we showed that the receptive fields can be learned simultaneously with the state-action mapping (Fig. 18H). The main lesson we learned is that the encoding/decoding framework in neural code can lead to a dead end unless the problem is posed from the perspective of the brain itself; and the motor system can play an important role in the shaping of the sensory/perceptual primitives (also see [2]).Fig. 18 Concept (A–E) and simulation results (F–H). A Four activities without any clear meaning. b Activities in A are V1 response to oriented lines. C Comparison of brain’s view of spikes (left; apparently intractable) and scientist’s view of spikes (right; decoding possible). D Visuomotor agent set up. E Invariance principle. F Ideal state(s)-action(a) mapping R(s, a) (a), learned R(s, a) (b: synthetic input), learned R(s, a) (c: natural input). G Input (a), initial gaze trajectory (b), and learned gaze trajectory (c). H Learned state-action mapping (a: unordered; b: reordered rows), and learned receptive fields (c: unordered; d: reordered as b) [1] ReferencesChoe Y, Yang HF, Misra N. Motor system’s role in grounding, receptive field development, and shape recognition. In: 7th IEEE international conference on development and learning (ICDL 2008). IEEE. p. 67–72. Salinas E. How behavioral constraints may determine optimal sensory representations. PLoS Biol. 2006;4(12):e387. P29 Neural computation in a dynamical system with multiple time scales Yuanyuan Mi1,†, Xiaohan Lin1,†, Si Wu1 1State Key Lab of Cognitive Neuroscience & Learning, IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing 100875, China Correspondence: Si Wu - wusi@bnu.edu.cn † Y.M. and X.L. contributed equally to this work BMC Neuroscience 2016, 17(Suppl 1):P29 The brain performs computation by updating its internal states in response to external inputs. Neurons, synapses, and the circuits are the fundamental units for implementing brain functions. At the single neuron level, a neuron integrates synaptic inputs and generates spikes if its membrane potential crosses the threshold. At the synapse level, neurons interact with each other to enhance or depress their responses. At the network level, the topology of neuronal connection pattern shapes the overall population activity. These fundamental computation units of different levels encompass rich short-term dynamics, for example, spike-frequency adaptation (SFA) at single neurons [1], short-term facilitation (STF) and depression (STD) at neuronal synapses [2]. These dynamical features typically expand a broad range of time scale and exhibit large diversity in different brain regions. Although they play a vital part in the rise of various brain functions, it remains unclear what is the computational benefit for the brain to have such variability in short-term dynamics. In this study, we propose that one benefit for having multiple dynamical features with varied time scales is that the brain can fully exploit the advantages of these features to implement which are otherwise contradictory computational tasks. To demonstrate this idea, we consider STF, SFA and STD with increasing time constants in the dynamics of a CANN. The potential brain regions with these parameter values are the sensory cortex, where the neuronal synapses are known to be STD-dominating. We show that the network is able to implement three seemingly contradictory computations, which are persistent activity, adaptation and anticipative tracking (see Fig. 19). Simply state, the role of STF is to hold persistent activity in the absence of external drive, the role of SFA is to support anticipative tracking for a moving input, and the role of STD is to eventually suppress neural activity for a static or transient input. Notably, the time constants of SFA and STD can be swapped with each other, since SFA and STD have the similar effects on the network dynamics. Nevertheless, we need to include both of them, since a single negative feedback modulation is unable to achieve both anticipative tracking and plateau decay concurrently. The implementation of each individual computational task based on a single dynamical feature has been studied previously. Here, our contribution is on revealing that these tasks can be realized concurrently in a single neural circuit by combined dynamical features with coordinated time scales. We hope that this study will shed light on our understanding of how the brain orchestrates its rich dynamics at various levels to realize abundant cognitive functions.Fig. 19 Networks implement different computations. A Persistent activity; network can sustain activity after removing stimulus. B Adaptation; network activity attenuates to background level given continuous stimulus. C Anticipative tracking; D network response leads moving stimulus in a certain speed ReferenceBenda J, Herz AVM. A universal model for spike-frequency adaptation. Neural Comput. 2003;15(11):2523–64. Markram H, Wang Y, Tsodyks M. Differential signaling via the same axon of neocortical pyramidal neurons. Proc Natl Acad Sci. 1998;95(9):5323–28. P30 Maximum entropy models for 3D layouts of orientation selectivity Joscha Liedtke1,2, Manuel Schottdorf1,2, Fred Wolf1,2 1Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany; 2Bernstein Center for Computational Neuroscience, Göttingen, Germany Correspondence: Joscha Liedtke - joscha@nld.ds.mpg.de, Manuel Schottdorf - manuel@nld.ds.mpg.de BMC Neuroscience 2016, 17(Suppl 1):P30 The neocortex is composed of 6 different layers. In the primary visual cortex (V1), the functional architecture of basic stimulus selectivity is experimentally found to be similar across these layers [1]. The organization in functional columns justifies the use of cortical models describing only two-dimensional layers and disregarding functional organization in the third dimension. Here we show theoretically that already small deviations from an exact columnar organization can lead to non-trivial three-dimensional functional structures (see Fig. 20). Previously, two-dimensional orientation domains were modeled by Gaussian random fields, the maximum entropy ensemble, allowing for an exact calculation of pinwheel densities [2]. Pinwheels are points surrounded by neurons preferring all possible orientations and these points generalize to pinwheel strings in three dimensions. We extend the previous two-dimensional model characterized by its typical scale of orientation domains to a three-dimensional model by keeping the typical scale in each layer and introducing a columnar correlation length. We dissect in detail the three-dimensional functional architecture for flat geometries and for curved gyri-like geometries with different columnar correlation lengths. The model is analyzed analytically complemented by numerical simulations to obtain solutions for its intrinsic statistical parameters. We find that (i) pinwheel strings are generally curved, (ii) for large curvatures closed loops and reconnecting pinwheel strings appear and (iii) for small columnar correlation lengths a novel transition to a rodent-like interspersed organization emerges.Fig. 20 A Three-dimensional orientation domains with columnar correlation length of Λ. B String singularities of orientation domains in A. Typical scale of cats Λ ≈ 1 mm This theory extends the work of [2, 3] by adding a columnar dimension and supplements the work of [4] by a rigorous statistical treatment of the three-dimensional functional architecture of V1. Furthermore, the theory sheds light on the required precision of experimental techniques for probing the fine structure of the columnar organization in V1. ReferencesHubel DN, Wiesel TN. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. J Physiol. 1962;160:106–54. Schnabel M, Kaschube M, Löwel S, Wolf F. Random waves in the brain: symmetries and defect generation in the visual cortex. Eur Phys J Spec Topics. 2007;145(1):137–57. Wolf F, Geisel T. Spontaneous pinwheel annihilation during visual development. Nature. 1998;395:73–8. Tanaka S, Moon CH, Fukuda M, Kim SG: Three-dimensional visual feature representation in the primary visual cortex. Neural Networks 2011, 24(10):1022-1035. P31 A behavioral assay for probing computations underlying curiosity in rodents Yoriko Yamamura1, Jeffery R. Wickens1 1Neurobiology Research Unit, Okinawa Institute of Science and Technology, Onna-son, Okinawa, 904-0412, Japan Correspondence: Yoriko Yamamura - yoriko@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P31 Curiosity in humans appears to follow an inverted U-shaped function of unpredictability: stimuli that are neither too predictable nor too unpredictable evoke the greatest interest [1]. Rewarding moderate sensory unpredictability is an effective strategy for reinforcing explorations that improve our predictive models of the world [1, 2]. However, the computations and neural circuits underlying this unpredictability-dependence of curiosity remain largely unknown. A rodent model of curiosity would be useful for elucidating its underlying neural circuitry, because more specific manipulation techniques are available than in humans. It has been shown that mice prefer unpredictable sounds to predictable ones when the sounds are paired with light [3]. However, frequency of stimulus presentation was a potential confound in this study. Furthermore, a more systematic sampling of stimulus unpredictability is necessary to determine whether a rodent analogue of the U-shaped curve indeed exists. We have devised an operant conditioning paradigm building on [3], using sensory stimuli as “reward” to quantify the rewardingness of various levels of sensory predictability for rats. Rats (Long Evans, male) are placed in a soundproofed chamber with two nosepoke holes. A combination of sound and light stimuli is presented whenever the rat pokes the active hole; no stimulus is associated with the inactive hole (counterbalanced across subjects). We hypothesize that reward is also a U-shaped function of stimulus unpredictability in rats, and that this is due to a Bayesian precision weighting placing more importance on deviations from reliabile predictions. This departs from previous learning-based accounts [2]. There are five experimental conditions, systematically varied in unpredictability of the sound stimuli (as quantified by entropy H), and a control condition, in which a nosepoke in neither hole has any consequence (Fig. 21). Specifically, the sound stimuli are random sequences of two possible 125-ms sound snippets of equal value to the rat, with their frequencies of occurrence varied across conditions to vary H. Each sequence contains eight such snippets. Across all conditions, the light stimulus simply remains on while the sound is being played; it is added to enhance the rats’ responding to auditory stimuli [3]. We predict that the rats’ active nosepoke responses will be maximally increased at intermediate H (Fig. 21).Fig. 21 Schematic of the sound stimuli used in all conditions, and the predicted reward for each In preliminary experiments for conditions 0 and 2 (N = 3 each; three sessions), rats preferred the active hole to the inactive, replicating the earlier results in mice [3]. Moreover, as hypothesized, rats responded more to the active hole in condition 2 (mean = 15.0, SD = 5.32) than in condition 0 (mean = 11.3, SD = 4.05); t(22) = 1.91, p = 0.0345 (one-tailed t test). We note that in mice, most across-condition differences did not emerge until around session 7 [3]. The proposed assay quantifies the rewardingness of sensory unpredictability in rats. By systematically varying the entropy of the sound sequence, we can probe the computations behind the putative unpredictability-driven reward. The assay can furthermore be used to study the effect of pharmacological or genetic manipulations on unpredictability-driven reward, in order to validate mechanistic implementations of such computations. ReferencesKidd C, Hayden BY. The psychology and neuroscience of curiosity. Neuron. 2015;88:449–60. Oudeyer PY, Kaplan F. What is intrinsic motivation? A typology of computational approaches. Front Neurorobot. 2007;1:6. Olsen CM, Winder DG. Stimulus dynamics increase the self-administration of compound visual and auditory stimuli. Neurosci Lett. 2012;511:8–11. P32 Using statistical sampling to balance error function contributions to optimization of conductance-based models Timothy Rumbell1, Julia Ramsey2, Amy Reyes2, Danel Draguljić2, Patrick R. Hof3, Jennifer Luebke4, Christina M. Weaver2 1Computational Biology Center, IBM Research, Thomas J. Watson Research Center, Yorktown Heights, NY 10598; 2Department of Mathematics, Franklin and Marshall College, Lancaster, PA 17604; 3Fishberg Department of Neuroscience and Friedman Brain Institute, Icahn School of Medicine at Mount Sinai, New York, NY 10029; 4Department of Anatomy and Neurobiology, Boston University School of Medicine, Boston, MA 02118 Correspondence: Christina M. Weaver - christina.weaver@fandm.edu BMC Neuroscience 2016, 17(Suppl 1):P32 Recently we developed a three-stage optimization method for fitting conductance-based models to data [1]. The method makes novel use of Latin hypercube sampling (LHS), a statistical space-filling design, to determine appropriate weights automatically for various error functions that quantify the difference between empirical target and model output. The method uses differential evolution to fit parameters active in the subthreshold and suprathreshold regimes (below and above action potential threshold). We have applied the method to spatially extended models of layer 3 pyramidal neurons from the prefrontal cortex of adult rhesus monkeys, in which in vitro action potential firing rates are significantly higher in aged versus young animals [2]. Here we validate our optimization method by testing its ability to recover parameters used to generate synthetic target data. Results from the validation fit the voltage traces of the synthetic target data almost exactly (Fig. 22A–C), whether fitting a model with 4 ion channels (10 parameters), or 8 ion channels (23 parameters). The optimized parameter values are either identical to, or nearby, the original target values (Fig. 22D–F), except for a few parameters that were not well constrained by the simulated protocols. Further, our LHS-based scheme for weighting error functions is significantly more efficient at recovering target parameter values than by weighting all error functions equally, or by choosing weights manually. We are now using the method to fit models to data from several young, middle-aged, and aged monkeys. Adding new conductances to the model, and allowing altered channel kinetics in the axon initial segment versus the soma, improves the quality of the model fits to data. We use published results from empirical studies of layer 3 neocortical pyramidal neurons to determine whether the optimized parameter sets are biologically plausible.Fig. 22 A–C Membrane potential of the synthetic target (black), and of randomly chosen members of the final population (colors, overlaid almost exactly), from three validation studies. Optimized 10 and 23 parameters in A–C respectively. D–F Parameter values used to generate synthetic data (black lines), and mean ± standard deviation of values recovered in the searches (colored circles), normalized to the range used in the optimization ReferencesRumbell T, Draguljić D, Luebke J, Hof P, Weaver CM. Prediction of ion channel parameter differences between groups of young and aged pyramidal neurons using multi-stage compartmental model optimization. BMC Neurosci. 2015;16(Suppl. 1):P282. Chang YM, Rosene DL, Killiany RJ, Mangiamele LA, Luebke JI. Increased action potential firing rates of layer 2/3 pyramidal cells in the prefrontal cortex are significantly related to cognitive performance in aged monkeys. Cereb Cortex. 2005;15(4):409–18. P33 Exploration and implementation of a self-growing and self-organizing neuron network building algorithm Hu He1, Xu Yang2, Hailin Ma1, Zhiheng Xu1, Yuzhe Wang1 1Institute of Microelectronics, Tsinghua University, Beijing, 100081, China; 2School of Software, Beijing Institute of Technology, Beijing, 100083, China Correspondence: Xu Yang - yangxu@tsinghua.edu.cn BMC Neuroscience 2016, 17(Suppl 1):P33 In this work, an algorithm to build self-growing and self-organizing neuron network according to external signals is presented, in attempt to build neuron network with high intelligence. This algorithm takes a bionic way to build complex neuron network. We begin with very simple external signals to provoke neurons. In order to propagate the signals, neurons will seek to connect to each other, thus building neuron networks. Those generated networks will be verified and optimized, and be treated as seeds to build more complex networks. Then we repeat this process, use more complex external signals, and build more complex neuron networks. A parallel processing method is presented, to enhance the computation efficiency of the presented algorithm, and to help build large scale of neuron network with reasonable time. The result shows that, neuron network built by our algorithm can self-grow and self-organize as the complexity of the input external signals increase. And with the screening mechanism, neuron network that can identify different input external signals is built successfully (Fig. 23).Fig. 23 Neuron network generated by our algorithm Acknowledgements: This work is supported by the Core Electronic Devices, High-End General Purpose Processor, and Fundamental System Software of China under Grant No. 2012ZX01034-001-002, the National Natural Science Foundation of China under Grant No. 61502032, Tsinghua National Laboratory for Information Science and Technology (TNList), and Samsung Tsinghua Joint Laboratory. P34 Disrupted resting state brain network in obese subjects: a data-driven graph theory analysis Kwangyeol Baek1,2, Laurel S. Morris1, Prantik Kundu3, Valerie Voon1 1Department of Psychiatry, University of Cambridge, Cambridge, CB2 0QQ, United Kingdom; 2Department of Biomedical Engineering, Ulsan National Institute of Science and Technology, Ulsan, South Korea; 3Departments of Radiology and Psychiatry, Icahn School of Medicine at Mount Sinai, New York City, 10029, USA Correspondence: Kwangyeol Baek - kb567@cam.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P34 The efficient organization and communication of brain networks underlies cognitive processing, and disruption in resting state brain network has been implicated in various neuropsychiatric conditions including addiction disorder. However, few studies have focused on whole-brain networks in the maladaptive consumption of natural rewards in obesity and binge-eating disorder (BED). Here we use a novel multi-echo resting state functional MRI (rsfMRI) technique along with a data-driven graph theory approach to assess global and regional network characteristics in obesity and BED. We collected multi-echo rsfMRI scans from 40 obese subjects (including 20 BED patients) and 40 healthy controls, and used multi-echo independent component analysis (ME-ICA) to remove non-BOLD noise. We estimated the normalized correlation across mean rsfMRI signals in 90 brain regions of the Automated Anatomical Labeling atlas, and computed global and regional network metrics in the binarized connectivity matrix with density threshold of 5–25 %. In addition, we confirmed the observed alterations in network metrics using the Harvard-Oxford atlas which was parcellated into 470 even-sized regions. Obese subjects exhibited significantly reduced global and local efficiency as well as decreased modularity in the whole-brain network compared to healthy controls (Fig. 24). Both BED patients and the obese subjects without BED exhibited the same alteration of network metrics compared with healthy controls, but two obese groups did not differ from each other. In regional network metrics, bilateral putamen, thalamus and right pallidum exhibited profoundly decreased nodal degree and efficiency in obese subjects, and left superior frontal gyrus showed decreased nodal betweeness in obese subjects (all p < 0.05, Bonferroni correction). Network-based statistics revealed a cortico-striatal/cortico-thalamic network with significantly decreased functional connectivity which consisted of bilateral putamen, pallidum, thalamus, primary motor cortex, primary somatosensory cortex, supplementary motor area, paracentral lobule, superior parietal lobule, superior temporal cortex and left amygdala. Interestingly, when examining the same network properties but using only single-echo rsfMRI data analysis without ME-ICA, we find no significant differences between groups.Fig. 24 A Disrupted resting state brain network in obese subjects. B Global network properties network-based statistics Therefore, using data-driven graph theory analysis of multi-echo rsfMRI data, we highlight more subtle impairments in cortico-striatal/cortico-thalamic networks in obesity that have previously been associated with substance addictions. We emphasize global impairments in network efficiency in obesity with disrupted local network organization closer to random networks. Mathematically capturing brain network alterations in obesity provides novel insights into potential biomarkers and therapeutic targets. P35 Dynamics of cooperative excitatory and inhibitory plasticity Everton J. Agnes1, Tim P. Vogels1 1Centre for Neural Circuits and Behaviour, University of Oxford, Oxford, OX1 3SR, UK Correspondence: Everton J. Agnes - everton.agnes@cncb.ox.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P35 Neurons receive balanced excitatory and inhibitory inputs, a phenomenon thought to be essential for a variety of computations [1–3]. Inhibitory synaptic plasticity is an obvious candidate for imposing this balanced input regime [2,4], leaving excitatory synapses available to learn patterns and memories. Recent experimental work seems to agree with that notion of collaborative excitatory and inhibitory plasticity [4], but recent models do not take direct interactions into consideration. Instead, learning rules are usually tuned to indirectly but constructively interact via the firing-rates they elicit [3,5]. Without proper parameter tuning, this can be problematic because excitatory and inhibitory synaptic plasticity models may have different homeostatic set points, making synaptic weights fluctuate wildly (Fig. 25A, B; green lines). Here we present a hybrid model of inhibitory synaptic plasticity that combines the simplicity of spike-based models with the addition of a excitatory/inhibitory input dependence. It captures recent experimental findings showing that changes at inhibitory synapses are strongly correlated with the balance between excitation and inhibition and that inhibitory synapses do not change when excitatory input is blocked [4]. Essentially, our model is a symmetric spike-timing-dependent plasticity (STDP) rule in which the learning-rate is controlled by excitatory and inhibitory activities—a spike-timing- and current-dependent plasticity (STCDP) model. Balance is maintained, but the learning rule does not impose fixed-point attractor dynamics to post-synaptic neurons, because there is no change in inhibitory synapses once the total input is balanced. Inhibitory synapses change depending on excitatory synapses, which means that plasticity depends on at least three synaptic participants (trisynaptic) instead of only two (bisynaptic). We show that when combined with an excitatory synaptic plasticity model, both excitatory and inhibitory weights converge to stable values, as the firing-rate reaches the fixed-point imposed by the excitatory learning rule (Fig. 25B; yellow lines). More importantly, the learning rule allows efficient and stable learning of new weights when the balance is disrupted, opening the door for effective and stable learning of arbitrary synaptic patterns.Fig. 25 A Schematics representing the neuronal network. A group of 2000 excitatory neurons and 500 inhibitory neurons are recurrently connected with sparse connectivity and the excitatory neurons receive random input from an external pool of neurons. B Excitatory neurons’ mean firing-rate (top), mean excitatory weight onto excitatory neurons (middle) and mean inhibitory weight onto excitatory neurons (connections marked as plastic in A). Simulation of the neuronal network with a spike-based inhibitory learning rule is represented by green lines (STDP) while simulation with our novel spike-timing- and current-dependent learning rule is shown in yellow (STCDP). The dashed lines represent the fixed points imposed by the excitatory (high) and inhibitory (low) learning rules. The low fixed point only exists for the inhibitory STDP model (simulation represented by the green lines) Acknowledgements: This work was partially funded by the Brazilian agency CNPq (Grant Agreement Number 235144/2014-2) and the Sir Henry Dale Fellowship (Grant Agreement WT100000). ReferencesDenève S, Machens CK. Efficient codes and balanced networks. Nat Neurosci. 2016;19:375–85. Vogels TP, Froemke RC, Doyon N, Gilson M, Haas JS, Liu R, Maffei A, Miller P, Wierenga CJ, Woodin MA, et al. Inhibitory synaptic plasticity: spike timing-dependence and putative network function. Front Neural Circuits. 2013;7:119. Zenke F, Agnes EJ, Gerstner W. Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nat Commun. 2015;6:6922. D’amour JA, Froemke RC. Inhibitory and excitatory spike-timing-dependent plasticity in the auditory cortex. Neuron. 2015;86:514–28. Sprekeler H, Clopath C, Vogels TP. Interactions of excitatory and inhibitory synaptic plasticity. Front Comp. P36 Frequency-dependent oscillatory signal gating in feed-forward networks of integrate-and-fire neurons William F. Podlaski1, Tim P Vogels1 1Centre for Neural Circuits and Behaviour, University of Oxford, Oxford, UK Correspondence: William F. Podlaski - william.podlaski@cncb.ox.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P36 Neural oscillations—the periodic synchronisation of neuronal spiking—is a common feature of brain activity, with several hypothesised functions relating to information flow, attention and brain state [1]. Previous experimental work has shown that oscillatory activity correlates with moments of heightened attention, and that communication between different brain areas is often marked by an increase in oscillatory coherence between the regions [2]. Theoretical and modelling work has helped to explore the mechanisms behind neuronal oscillations, and some of their effects on neural coding and signal propagation [3]. Recently, theoretical studies have explored how resonance might affect signal processing [4, 5] and how information can be propagated along different pathways according to oscillatory phase and frequency [6]. We expand this work here by studying how resonance at the single neuron level might be used for frequency-dependent gating of information flow in neuronal networks. We show that in feed-forward spiking network simulations background oscillations can synchronise or desynchronise the spikes of a propagated signal, changing its content and emphasis from rate code to synfire code or vice versa. Such a mechanism can modulate information flow without rewiring the signal pathways themselves, allowing to select for specific downstream readout targets. Building on this idea, we can create entire pathways that can be selectively (in-)activated by different background oscillatory frequencies without changing the connectivity of the network. We hypothesise that neuronal resonance, combined with resonance in synapses and network motifs, can allow for precise oscillatory gating of information in cortex. Building on previous studies of resonance and oscillatory signal propagation [4,5,6] we propose a plausible mechanism for how fast and precise frequency-dependent gating might be achieved in the brain. Acknowledgements: Research was supported by a Sir Henry Dale Royal Society and Wellcome Trust Research Fellowship (WT100000). ReferencesBuzsáki G. Rhythms of the brain. Oxford: Oxford University Press; 2011. Engel AK, Fries P, Singer W. Dynamic predictions: oscillations and synchrony in top-down processing. Nat Rev Neurosci. 2001;2(10):704–16. Wang XJ. Neurophysiological and computational principles of cortical rhythms in cognition. Physiol Rev. 2010;90:1195–1268. Richardson MJE, Brunel N, Hakim V. From subthreshold to firing-rate resonance. J Neurophysiol. 2003;89:2538–54. Hahn G, Bujan AF, Frégnac Y, Aertsen A, Kumar A. Communication through resonance in spiking neuronal networks. PLoS Comput Biol. 2014;10(8):e1003811. Akam T, Kullmann DM. Oscillatory multiplexing of population codes for selective communication in the mammalian brain. Nat Rev Neurosci. 2014;15(2):111–22. P37 Phenomenological neural model for adaptation of neurons in area IT Martin Giese1, Pradeep Kuravi2, Rufin Vogels2 1Section Computational Sensomotorics, CIN & HIH, Department of Cognitive Neurology, University Clinic Tübingen, Germany; 2Lab. Neuro en Psychofysiologie, Dept. Neuroscience, KU Leuven, Belgium Correspondence: Martin Giese - martin.giese@uni-tuebingen.de BMC Neuroscience 2016, 17(Suppl 1):P37 For repeated stimulation neurons in higher-level visual cortex show adaptation effects. Such effects likely influence repetition suppression paradigms in fMRI studies and the formation of high-level after-effects, e.g. for faces [1]. A variety of theoretical explanations has been discussed, which are difficult to distinguish without detailed electrophysiological data [2]. Meanwhile, detailed physiological experiments on the adaptation of shape-selective neurons in inferotemporal cortex (area IT) have provided constraints that help to narrow down possible neural processes. We propose a neurodynamical model that reproduces a number of these experimental observations by biophysically plausible neural circuits. Our model uses the mean-field limit and consists of a neural field of shape-selective dynamic linear-threshold neurons that are augmented several adaptation processes: (i) spike-rate adaptation; (ii) an input fatigue adaptation process, modeling adaptation in earlier hierarchy levels and of afferent synapses; (iii) a firing-rate fatigue adaptation process that models adaptation dependent on the output firing rates of the neurons. The model with a common parameter set is compared to results from several studies about adaptation in area IT. The model reproduces the following experimentally observed effects: (i) shape of the typical PSTHs of IT neurons; (ii) temporal decay for repeated stimulation of the same neurons with many repetitions of the same stimulus [3] (Fig. 26A); (iii) dependence of adaptation on efficient and ineffective adaptor stimuli, which stimulate the neuron strongly or only moderately [4] (Fig. 26B); (iv) dependence of the strength of the adaptation effect on the duration of the adaptor (Fig. 26C). A mean field model with several additional adaptive processes can account for the observed experimental effects, where all introduced processes were necessary to account for the results. Especially the observed dependence on the effectivity of the adaptor cannot be reproduced without an appropriate mixture if an input fatigue and a firing-rate fatigue mechanism. This suggests that adaptation in IT neurons is significantly influenced by several biophysical processes with different spatial and temporal scales.Fig. 26 Simulation results. A Decay of neural activity for multiple repetitions of the same stimulus. B Experiment adapting with effective and ineffective stimuli. C Dependence of the PSTH on adaptor duration and unadapted response (black) Acknowledgements: Supported by EC Fp7-PEOPLE-2011-ITN PITN-GA-011-290011 (ABC), FP7-ICT-2013-FET-F/604102 (HBP), FP7-ICT-2013-10/611909 (Koroibot), BMBF, FKZ: 01GQ1002A, DFG GI 305/4-1 + KA 1258/15-1. ReferencesLeopold DA, O’Toole AJ, Vetter T, Blanz V: Prototype-referenced shape encoding revealed by high-level aftereffects. Nat Neurosci. 2001;4(1):89–94. Grill-Spector K, Henson R, Martin A. Repetition and the brain: neural models of stimulus-specific effects. Trends Cogn Sci. 2006;10(1):14–23. Sawamura H, Orban GA, Vogels R. Selectivity of neuronal adaptation does not match response selectivity: a single-cell study of the FMRI adaptation paradigm. Neuron. 2006;49(2):307–18. De Baene W, Vogels R. Effects of adaptation on the stimulus selectivity of macaque inferior temporal spiking activity and local field potentials. Cereb Cortex. 2010;20(9):2145–65. P38 ICGenealogy: towards a common topology of neuronal ion channel function and genealogy in model and experiment Alexander Seeholzer1,†, William Podlaski2,†, Rajnish Ranjan3, Tim Vogels2 1Laboratory of Computational Neuroscience, EPF Lausanne, Switzerland; 2Centre for Neural Circuits and Behaviour, University of Oxford, UK; 3The Blue Brain Project, EPF Lausanne, Switzerland Correspondence: Alexander Seeholzer - alex.seeholzer@epfl.ch † These authors contributed equally to this work. BMC Neuroscience 2016, 17(Suppl 1):P38 Ion channels are fundamental constituents determining the function of single neurons and neuronal circuits. To understand their complex interactions, the field of computational modeling has proven essential: since its emergence, thousands of ion channel models have been created and published as part of detailed neuronal simulations [1]. Faced with this large variety of models, it is difficult to determine how particular models relate to each other, to the interpretability of simulations and, importantly, to experimental data. Here, we present a framework within which we analyzed a pilot set of 2378 voltage- or calcium-dependent published ion channel models for the NEURON simulator [1]. We extracted annotated metadata from all associated publications, helping identify their use in simulations (e.g. the animal type, neuron type or area of compartmental models) and the provenance of ion channel models as they were derived from other published work. This categorical and relational metadata is combined with quantitative evaluations of all channel models: individual channels are characterized by their responses to voltage clamp protocols. With subsequent cluster analysis, we extract topologies of ion channel similarity and genealogy, identifying redundancy and groups of common channel kinetics. The result of this large-scale assay of published work is freely accessible through interactive visualizations (see Fig. 27A) on the Ion Channel Genealogy (ICG) web-resource [2], providing a tool for model discovery and comparison. Bridging the gap between model and experiment, our resource allows classifying new channel models and experimental current traces within the topology of all models currently in the database (see Fig. 27B, C). The ICG framework thus allows for quantitative comparison of ion channel kinetics, experimental and model alike, aimed to facilitate field-wide standardization of experimentally-constrained modeling.Fig. 27 A Visualizations available on the web-resource [2] for model browsing. B Schematic of upload and evaluation. Both experimental current traces and mod files can be uploaded to our servers, where they are scored and compared to all models currently in the database. C Exemplary result of automated comparison: Current traces (recorded from “Ramp” and “Activation” voltage clamp protocols) of the uploaded model (red) together with mean (1st, 2nd, 3rd, 4th) and individual (gray) traces of the four most similar clusters of channel models in the database Acknowledgements: Research was supported by a Sir Henry Dale Royal Society & Wellcome Trust Research Fellowship (WT100000). A.S. was supported by the Swiss National Science Foundation (200020_147200). R.R. was supported by the EPFL Blue Brain Project Fund and the ETH Board funding to the Blue Brain Project. ReferencesHines ML, Morse T, Migliore M, Carnevale NT, Shepherd GM. ModelDB. A database to support computational neuroscience. J Comput Neurosci. 2004;17:7–11. ICGenealogy Project Website. http://icg.neurotheory.ox.ac.uk. P39 Temporal input discrimination from the interaction between dynamic synapses and neural subthreshold oscillations Joaquin J. Torres1, Fabiano Baroni2, Roberto Latorre3, Pablo Varona3 1Departamento de Electromagnetismo y Física de la Materia, and Institute “Carlos I” for Theoretical and Computational Physics, University of Granada, Granada, Spain; 2School of Psychological Sciences, Faculty of Biomedical and Psychological Sciences, Monash University, Australia; 3Grupo de Neurocomputación Biológica, Dpto. de Ingeniería Informática, Escuela Politécnica Superior, Universidad Autónoma de Madrid, Spain Correspondence: Pablo Varona - pablo.varona@uam.es BMC Neuroscience 2016, 17(Suppl 1):P39 Neuronal subthreshold oscillations underlie key mechanisms of information discrimination in single cells while dynamic synapses provide channel-specific input modulation. Previous studies have shown that intrinsic neuronal properties, in particular subthreshold oscillations, constitute a biophysical mechanism for the emergence of non-trivial single-cell input/output preferences (e.g., preference towards decelerating vs. accelerating input trains of the same average rate) [1, 2]. It has also been shown that short-term synaptic dynamics, in the form of short-term depression and/or short-term facilitation, can provide a channel-specific mechanism for the enhancement of the post-synaptic effects of temporally specific input sequences [3, 4]. While intrinsic oscillations and synaptic dynamics are typically studied independently, it is reasonable to hypothesize that their interplay can lead to more selective and complex temporal input processing. Here, we extend and refine our previous computational study on the interaction between subthreshold oscillations and synaptic depression [5]. In particular, we investigated whether, and under which conditions, the combination of intrinsic subthreshold oscillations and short-term synaptic dynamics can act synergistically to enable the emergence of robust and channel-specific selectivity in neuronal input–output transformations. We calculated analytically the voltage trajectories and spike output of generalized integrate-and-fire (GIF) model neurons in response to temporally distinct trains of input EPSPs. In particular, we considered triplets of input EPSPs in a range that covers intrinsic and synaptic time scales, and analyzed the model output as intrinsic and synaptic parameters were varied. Our results show that intrinsic and synaptic dynamics interact in a complex manner for the emergence of specific input–output transformations. In particular, precise non-trivial preferences emerge from synergistic intrinsic and synaptic preferences, while broader selectivity is observed for mismatched intrinsic and synaptic dynamics. We discuss the conditions for robustness of the observed input/output relationships. We conclude that the interaction of intrinsic and synaptic properties can enable the biophysical implementation of complex and channel-specific mechanisms for the emergence of selective neuronal responses. We further interpret our results in the light of experimental evidence describing distinct short-term synaptic dynamics in different afferents converging onto the same neuron, as in the case of parallel and climbing fiber inputs to cerebellar Purkinje cells, and advance specific hypotheses that link heterogeneous synaptic dynamics of distinct pathways onto the same post-synaptic target to their distinct computational function. We also discuss the impact of single-channel/single-neuron temporal input discrimination in the context of information processing based on heterogeneous elements. Acknowledgements: We acknowledge support from MINECO FIS2013-43201-P, DPI2015-65833-P, TIN-2012-30883 and ONRG Grant N62909-14-1-N279. ReferencesBaroni F, Varona P. Subthreshold oscillations and neuronal input–output relationships. Neurocomputing. 2007;70:1611–14. Baroni F, Torres JJ, Varona P. History-dependent excitability as a single-cell substrate of transient memory for information discrimination. PLoS One. 2010;5:e15023. O’Donnell C, Nolan MF. Tuning of synaptic responses: An organizing principle for optimization of neural circuits. Trends Neurosci. 2011;34:51–60. Torres JJ, Kappen HJ. Emerging phenomena in neural networks with dynamic synapses and their computational implications. Front Comp Neurosci. 2013;7. Latorre R, Torres JJ, Varona P. Interplay between subthreshold oscillations and depressing synapses in single neurons. PLoS One. 2016;11:e0145830. P40 Different roles for transient and sustained activity during active visual processing Bart Gips1,†, Eric Lowet1,2,†, Mark J Roberts1,2, Peter de Weerd2, Ole Jensen1, Jan van der Eerden1 1Radboud University, Donders Institute for Brain, Cognition and Behaviour, 6525 EN Nijmegen, The Netherlands; 2Faculty of Psychology and Neuroscience, Maastricht University, 6200 MD Maastricht, the Netherlands Correspondence: Bart Gips - bart.gips@donders.ru.nl † Authors have made equal contribution BMC Neuroscience 2016, 17(Suppl 1):P40 Neural activity in awake primate early visual cortex exhibits transients with intervals of 250-300 ms. Experimental work by us and others has shown that these transients are related to microsaccadic eye movements [1, 2]. These short transients are followed by periods of steady activity that last until the next microsaccade (Fig. 28A).Fig. 28 A Time–frequency representation of local field potential (LFP) locked to a microsaccade (MS) recorded in primate V1. B Time–frequency representation of simulated LFP. C Schematic representation of the model network illustrating input (injection current), recurrent connection pattern and output (spike trains). D The input to the neurons is best reflected in the simulated spike trains (output) during phase I, quantified by mutual information (MI). E Recurrent connection pattern is best reflected in the output during phase II We found that computational models of excitatory-inhibitory spiking networks organized in a structure of columns and hypercolumns, are able to represent relevant stimulus information when subjected to 3–4 Hz saccade-like transients. The simulated networks expressed evoked responses with power in the alpha–beta band (~8–25 Hz) as well as gamma rhythmic activity (~25–80 Hz) similar to in vivo local field recordings in monkey V1 (Fig. 28A, B). We show that in phase I, the model produces large-scale spatial synchrony and pronounced alpha–beta power. In phase II the model exhibits narrow-band gamma oscillations with spatially local synchrony. The activity in the model network (rate and timing coding) in phase I mainly reflects feedforward input (Fig. 28C, D), whereas, the network activity in phase II was dominated by recurrent connections (Fig. 28C, E). The model network activity closely matches that found in experiments. The simulation results suggest that transient phase (phase I) allows for resetting the network and rapid feedfoward processing of novel information, whereas detailed processing and contextualization by recurrent activity take place in the period of steady gamma activity (phase II). Therefore we arrived at hypotheses on the functional interpretation of phases I and II that can be possibly tested in an experimental setup. First, because of the reset of network activity by a microsaccade, phase I is the optimal time window to switch information flow among competing networks through a top-down signal. This indicates that signals related to visual attention are most likely to occur just after a saccade. Second, the increased efficacy of recurrent connections during phase II indicate that contextualization operations such as figure-ground segregation [3] and contour completion occur in the steady phase ~100 ms after the onset of a (micro)saccade. ReferencesLowet E, Roberts MJ, Bosman CA, Fries P, de Weerd P. Areas V1 and V2 show microsaccade-related 3–4 Hz covariation in gamma power and frequency. Eur J Neurosci. 2015. Martinez-Conde S, Otero-Millan J, Macknik SL. The impact of microsaccades on vision: towards a unified theory of saccadic function. Nat Rev Neurosci. 2013;14:83–96. Self MW, van Kerkoerle T, Supèr H, Roelfsema PR. Distinct roles of the cortical layers of area V1 in figure-ground segregation. Curr Biol. 2013:1–9. P41 Scale-free functional networks of 2D Ising model are highly robust against structural defects: neuroscience implications Abdorreza Goodarzinick1, Mohammad D. Niry1,2, Alireza Valizadeh1,3 1Department of Physics, Institute for Advanced Studies in Basic Sciences (IASBS), Zanjan 45137-66731, Iran; 2Center for Research in Climate Change and Global Warming (CRCC), Institute for Advanced Studies in Basic Sciences (IASBS), Zanjan 45137-66731, Iran; 3School of Cognitive Sciences, Institute for Research in Fundamental Sciences (IPM), Tehran - Iran Correspondence: Abdorreza Goodarzinick - a.goodarzinick@iasbs.ac.ir BMC Neuroscience 2016, 17(Suppl 1):P41 In recent years, several experimental observations have confirmed the emergence of self-organized criticality (SOC) in the brain at different scales [1]. At large scale, functional brain networks obtained from fMRI data have shown that node-degree distributions and probability of finding a link versus distance are indicative of scale-free and small-world networks regardless of the tasks in which the subjects were involved [2]. At small scale, the study of neuronal avalanches in networks of living neurons revealed power-law behavior in both spatial and temporal scales [3]. It is also shown that functional networks of the brain are strikingly similar to those derived from the 2D Ising model at critical temperature [4] and the 2D abelian sandpile model [5]. The importance to see whether brain network’s scaling properties associated with healthy conditions are altered under various pathologies and how structural defects of a system at criticality can affect its functional connectivity motivated us to study robustness of functional networks of 2D Ising model at critical point against elimination of structural sites. The results showed that the statistics of the functional network indicative of criticality (evident in healthy brain controls), such as power-law behavior and small-worldness remained robust against random elimination of structural sites up to percolation limit (see Fig. 29). The resulting functional network maintained its key properties orders of magnitude higher than those of the same system poised in a super-critical or sub-critical state. These results can show that self-organized critical behavior, besides having unique advantages like fasciliation of alteration of functional patterns, optimization of information transfer and maximization of correlation length, shows striking robustness against structural deficits. Taking into account brain’s long-range anatomical connections and compensatory mechanisms like neuroplasticity, if the results of this study are generalizable to the brain, they may help to explain the delay in clinical diagnosis of multiple neurodegenerative diseases in which possible deficit in functional connectivity among brain regions contribute to the cognitive dysfunctions.Fig. 29 Relevant parameters of functional network of 2D Ising model at critical point versus fraction of defect to the structural cells. A Power-law exponent of degree-distribution, B small-worldness measure, C average degree ReferencesChialvo DR. Emergent complex neural dynamics. Nat Phys. 2010;6:744–50. Eguíluz VM, Chialvo DR, Cecchi GA, Baliki M, Apkarian AV. Scale-free brain functional networks. Phys Rev Let.t 2005;94:018102. Beggs J, Plenz D. Neuronal avalanches in neocortical circuits. J Neurosci. 2003;23:11167–77. Fraiman D, Balenzuela P, Foss J, Chialvo D. Ising-like dynamics in large-scale functional brain networks. Phys Rev E. 2009;79:061922. Zarepour M, Niry MD, Valizadeh A. Functional scale-free networks in the two-dimensional Abelian sandpile model. Phys Rev E. 2015;92:012822. P42 High frequency neuron can facilitate propagation of signal in neural networks Aref Pariz1, Shervin S Parsi1, Alireza Valizadeh1,2 1Department of Physics, Institute for advanced studies in basic sciences, Zanjan, Iran; 2School of Cognitive Sciences, Institute for Studies in Theoretical Physics and Mathematics, Niavaran, Tehran, Iran Correspondence: Aref Pariz - a.pariz@iasbs.ac.ir BMC Neuroscience 2016, 17(Suppl 1):P42 Signal transmission is of interest from both fundamental and clinical perspective and has been well studied in nonlinear science and complex networks [1, 2]. In particular, in nervous systems, cognitive processing involves signal propagation through multiple brain regions and the activation of large numbers of specific neurons [3–6]. In information propagation through brain regions, each part, known as generator, activated locally as information comes to it from neighboring generators. Although the problem is well studied in the context of complex networks, our focus here is on the effect of the intrinsic dynamical properties of the reciprocal generators on the propagation of signal. In this study we explored the propagation of information in a chain of neurons and networks. As signal propagate through the chain of networks, the firing rate of networks show a fluctuation as host network (the network which receive signal). Here the response is the amplitude of fast Fourier transform of firing rates of each network. If the host network has sufficiently higher intrinsic firing rate than others, signal can transfer with higher amplitude, otherwise, other networks will not get affected. As a result of propagation of signal, for the former case, all networks will show a peak in frequency domain at exactly the same frequency as input signal (Fig. 30A), but with different amplitude which show the efficacy of transmitted information. Also the same result can obtain by a chain of single LIF neurons (Fig. 30B). As phase response curve of the chain and it response to signal show, if the host neuron has higher firing rate (call it leader neuron), the propagation of information will be enhanced. But this higher firing rate has a limit which after that the whole chain will act asynchronously and results the loss of information was aimed to propagate.Fig. 30 Inhomogeneity of input current on host network, increases the response of network. A, B Response of networks of neurons and chain of neurons, for different inhomogeneity on host network and host neuron, respectively ReferencesLiang X, Liu Z, Li B. Weak signal transmission in complex networks and its application in detecting connectivity. Phys Rev E. 2010;80:046102. Perc M. Stochastic resonance on weakly paced scale-free networks. Phys Rev E. 2008;78:036105. Abeles M. Corticonics: neural circuits of the cerebral cortex. Cambridge: Cambridge UP; 1991. Aertsen A, Diesmann M, Gewaltig MO. Propagation of synchronous spiking activity in feedforward neural networks. J Physiol. 1996;90:243–247. van Rossum MC, Turrigiano GG, Nelson SB. Fast propagation of firing rates through layered networks of noisy neurons. J Neurosci. 2002;22:1956–66. Vogels TP, Abbott LF. Signal propagation and logic gating in networks of integrate-and-fire neurons. J Neurosci. 2005;25(46):10786–95. P43 Investigating the effect of Alzheimer’s disease related amyloidopathy on gamma oscillations in the CA1 region of the hippocampus Julia M. Warburton1, Lucia Marucci2, Francesco Tamagnini3,4, Jon Brown3,4, Krasimira Tsaneva-Atanasova5 1Bristol Centre for Complexity Sciences, University of Bristol, Bristol, BS8 1TR, UK; 2Department of Engineering Mathematics, University of Bristol, Bristol, BS8 1UB, UK; 3School of Physiology and Pharmacology, University of Bristol, Bristol, BS8 1TD, UK; 4Medical School, University of Exeter, Exeter, EX4 4PE, UK; 5Department of Mathematics, University of Exeter, Exeter, EX4 4QF, UK Correspondence: Julia M. Warburton - julia.warburton@bristol.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P43 Alzheimer’s disease (AD) is the main form of dementia and is characterised clinically by cognitive decline and impairments to memory function. One of the key histopathological features of AD thought to cause this neurodegeneration is the abnormal aggregation of the protein amyloid-β (Aβ) [1]. Transgenic mouse models that overexpress Aβ are used to investigate the potential functional consequences of this amyloidopathy in AD. In this study we use in vitro electrophysiology data recorded from PDAPP transgenic mice (a mouse model of amyloidopathy) and their wild-type littermates to parameterise a hippocampal network model [2]. The aim of the study is to investigate how amyloidopathy alters gamma frequency oscillations within the hippocampus, which is one of the regions first affected in AD. We use a synaptically connected network of excitatory pyramidal neurons and inhibitory interneurons to simulate the gamma frequency activity [3]. Each cell is described by a single-compartment Hodgkin–Huxley type equation, with the properties of the voltage-gated channels fit to the intrinsic properties measured experimentally, which included stimulated firing frequency data and the associated action potentials from CA1 pyramidal neurons and three-types of CA1 interneuron. Network activity is either driven deterministically via a direct stimulus, such as a step pulse or a theta wave, or via a stochastic input. We perform power spectral density analysis to analyse the oscillatory activity. Our model focuses on gamma frequency oscillations, which lie in the 30–100 Hz range, because of the associations with attention, sensory processing and potentially of most relevance to AD, with learning and memory. It has been shown that within the hippocampus gamma oscillations enable cross-talk between distributed cell assemblies, with low frequency gamma associated with coupling between the CA1 and the CA3 region and fast frequency gamma associated with coupling between the CA1 and the medial entorhinal cortex [4]. EEG measurements from AD mouse models have identified network hypersynchrony alongside decreased gamma activity, with the role of interneurons in this process highlighted. [5]. By incorporating the pyramidal neuron and interneuron data in our model we aim to learn more about which parameters are most significant in these effects and to further understanding of the effects of amyloidopathy on oscillatory activity. Acknowledgements: This work was supported by funding from the EPSRC. ReferencesHardy J, Selkoe DJ. The amyloid hypothesis of Alzheimer’s disease: progress and problems on the road to therapeutics. Science. 2002;297:353–6. Kerrigan TL, Brown JT, Randall TL. Characterization of altered intrinsic excitability in hippocampal CA1 pyramidal cells of the Aβ-overproducing PDAPP mouse. Neuropharmacology. 2014;79:515–24. Kopell NJ, Borgers C, Pervouchine D, Maerba P, Tort A. Gamma and theta rhythms in biophysical models of hippocampal circuits. In: Hippocampal microcircuits: a computational modeler’s resource book, chap 15; p. 423–57. Colgin LL, Denninger T, Fyhn M, Hafting T, Bonnevie T, Jensen O, Moser M-B, Moser EI. Frequency of gamma oscillations routes flow of information in the hippocampus. Nature. 2009;462:353–7. Verret L, et al. Inhibitory interneuron deficit links altered network activity and cognitive dysfunction in Alzheimer model. Cell. 2012;149:708–21. P44 Long-tailed distributions of inhibitory and excitatory weights in a balanced network with eSTDP and iSTDP Florence I. Kleberg1, Jochen Triesch1 1Frankfurt Institute for Advanced Studies, Frankfurt am Main, Hessen, Germany, 60438 Correspondence: Florence I. Kleberg - kleberg@fias.uni-frankfurt.de BMC Neuroscience 2016, 17(Suppl 1):P44 The strengths of excitatory synapses in cortex and hippocampus have been shown to follow a rightward-skewed or long-tailed distribution [1,2]. Such distributions can be achieved in recurrent balanced networks [3, 4], after synaptic modification by spike-timing dependent plasticity (STDP) [5] and synaptic scaling [6]. Recently, long-tailed distributions have also been observed for inhibitory synapses in cultured cortical neurons [7], confirming early findings in hippocampal slices [8]. However, the conditions and plasticity mechanisms necessary for achieving long-tailed distributions of inhibitory synapses are unknown. Furthermore, different forms of inhibitory STDP have been reported, but their effect on the distribution of inhibitory synaptic efficacies are largely unknown [9-11]. Here we investigate how plasticity in the inhibitory synapses in a self-organised recurrent neural network (SORN [12]) with leaky integrate-and-fire neurons can lead to long-tailed distributions of synaptic weights. We examine different inhibitory STDP (iSTDP) rules and characterize the conditions under which right-skewed shapes of inhibitory synaptic weight distributions are obtained while a balance between excitation and inhibition is maintained. While the ratio of long-term potentiation to long-term depression in iSTDP affects the shape of the distribution, a variety of window shapes for iSTDP can each achieve long-tailed distributions of inhibitory weights. We find that a precise balance of excitation and inhibition can be achieved with a strongly right-skewed distribution of inhibitory weights. Our results suggest that long-tailed distributions of inhibitory weights could be a ubiquitous feature of neural circuits that employ different plasticity mechanism. ReferencesBekkers, JM, Stevens, CF. NMDA and non-NMDA receptors are co-localized at individual excitatory synapses in cultured rat hippocampus. Nature. 1989;341:230–3. Loewenstein Y, Kuras A, Rumpel S. Multiplicative dynamics underlie the emergence of the log-normal distribution of spine sizes in the neocortex in vivo. J Neurosci. 2011;31(26):9481–8. Effenberger F, Jost J, Levina A. Self-organization in balanced state networks by STDP and homeostatic plasticity. PLoS Comput Biol. 2015;11(9):e1004420. Miner D, Triesch J. Plasticity-driven self-organization under topological constraints accounts for non-random features of cortical synaptic wiring. PLoS Comput Biol. 2016;12(2):e1004759. Bi G, Poo M. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci. 1998;18(24):10464–72. Turrigiano GG, Leslie KR, Desai NS, Rutherford LC, Nelson SB. Activity-dependent scaling of quantal amplitude in neocortical neurons. Nature. 1998;391(6670):892–6. Rubinski A, Ziv NE. Remodeling and tenacity of inhibitory synapses: relationships with network activity and neighboring excitatory synapses. PLoS Comput Biol. 2015;11(11):e1004632. Miles R. Variation in strength of inhibitory synapses in the CA3 region of guinea-pig hippocampus in vitro. J Physiol. 1990;431:659–76. Woodin MA, Ganguly K, Poo M. Coincident pre-and postsynaptic activity modifies GABAergic synapses by postsynaptic changes in Cl− transporter activity. Neuron. 2003;39(5):807–20. Haas JS, Nowotny T, Abarbanel HDI. Spike-timing-dependent plasticity of inhibitory synapses in the entorhinal cortex. J Neurophysiol. 2006;96(6):3305–13. D’Amour JA, Froemke RC. Inhibitory and excitatory spike-timing-dependent plasticity in the auditory cortex. Neuron. 2015;86(2):514–28. Lazar A, Pipa G, Triesch J. SORN: a self-organizing recurrent neural network. Front Comp Neurosci. 2009;3:23. P45 Simulation of EMG recording from hand muscle due to TMS of motor cortex Bahar Moezzi1, Nicolangelo Iannella1,4, Natalie Schaworonkow2, Lukas Plogmacher2, Mitchell R. Goldsworthy3, Brenton Hordacre3, Mark D. McDonnell1, Michael C. Ridding3, Jochen Triesch2 1Computational and Theoretical Neuroscience Laboratory, School of Information Technology and Mathematical Sciences, University of South Australia, Australia; 2Frankfurt Institute for Advanced Studies, Goethe-Universität, Germany; 3Robinson Research Institute, School of Medicine, University of Adelaide, Australia; 4School of Mathematical Sciences, University of Nottingham, UK Correspondence: Bahar Moezzi - bahar.moezzi@unisa.edu.au BMC Neuroscience 2016, 17(Suppl 1):P45 Single pulse transcranial magnetic stimulation (TMS) is a technique which (at moderate intensities) activates corticomotor neuronal output cells transynaptically and evokes a complex descending volley in the corticospinal tract. Rusu et al. developed a computational model of TMS induced I-waves that reproduced observed epidural recordings in conscious humans [1]. In humans, epidural responses can be recorded in anaesthetized subjects during surgery or conscious subjects with electrodes implanted for the treatment of chronic pain. Such opportunities are uncommon and invasive. The effects of TMS can be non-invasively studied using surface electromyography (EMG) recordings from the hand first dorsal interosseous (FDI) muscle. We simulated the surface EMG signal due to TMS of motor cortex in the hand FDI muscle. Our model comprises a population of cortical layer 2/3 cells, which drive layer 5 cortico-motoneuronal cells with excitatory and inhibitory synaptic inputs as in [1]. The layer 5 cells in turn project to a pool of motoneurons, which are modeled as an inhomogeneous population of integrate-and-fire neurons to simulate motor unit recruitment and rate coding. The input to motoneurons from cortical layer 5 consists of TMS-induced spikes and baseline firing. We modeled baseline firing with a Poisson drive to layer 2/3 cells. Hermite-Rodriguez functions were used to simulate motor unit action potential shape. The EMG signal was obtained from the summation of motor unit action potentials of active motor units. Parameters were tuned to simulate recordings from the FDI muscle. Our simulated EMG signals match experimental surface EMG recordings due to TMS of motor cortex in the hand FDI muscle in shape, size and time scale both at rest and during voluntary contraction (see Fig. 31). The simulated EMG traces exhibit cortical silent periods (CSP) that lie within the biological range.Fig. 31 Comparison of simulated and experimental EMG during A rest, B 10 % maximum voluntary contraction ReferenceRusu CV, Murakami M, Ziemann U, Triesch J. A model of TMS-induced I-waves in motor cortex. Brain Stimul. 2014;7(3):401–14. P46 Structure and dynamics of axon network formed in primary cell culture Martin Zapotocky1,2, Daniel Smit1,2,3, Coralie Fouquet3, Alain Trembleau3 1Institute of Physiology of the Czech Academy of Sciences, Prague, Czech Republic; 2Institute of Biophysics and Informatics, First Faculty of Medicine, Charles University in Prague, Czech Republic; 3IBPS, Neuroscience Paris Seine, CNRS UMR8246, Inserm U1130, UPMC UM 119, Université Pierre et Marie Curie, Paris, France Correspondence: Martin Zapotocky - zapotocky@biomed.cas.cz BMC Neuroscience 2016, 17(Suppl 1):P46 Axons growing in vivo or in culture may adhere to each other and form a connected network, which subsequently guides the paths of newly arriving axons. We investigated the development of such a network formed by growing axons in primary cell culture. Olfactory epithelium explants from mouse embryos (day 13–14) were cultured on laminin substrate for 2 days and then recorded using DIC or phase contrast videomicroscopy for up to 24 h. The growing axons established a dense network within which large fascicles of axons were progressively formed. Within the recorded time period, the network remained stable, with limited further gowth of the axons but with ongoing rearrangement in the network structure. Based on segmentation of the recorded images, we determined the principal network characteristics (including the total length, the total number of vertices, and the network anisotropy) and their evolution in time. This quantitative characterization permitted an analysis of the mechanisms of the observed network coarsening. We relate the network dynamics to the elementary processes of zippering, during which two axons or axon fascicles progressively adhere to each other [1]. We compare the structural features of the network (such as the distribution of vertex angles) with those reported in an electron microscopy investigation of a plexus of sensory neurites in Xenopus embryo [2]. We show that both our ex vivo study and the in vivo study of Ref. [2] support a similar underlying mechanism of the formation of the axon network. Acknowledgements: Work supported by GAČR 14-16755S, GAUK 396213, MŠMT 7AMB12FR002, NIH 1RO1DCO12441 and ANR 2010-BLAN-1401-01. ReferencesSmit D, Fouquet C, Pincet F, Trembleau A, Zapotocky M. Axon zippering in neuronal cell culture and its biophysical modeling. BMC Neurosci. 2015;16(Suppl. 1):P298. Roberts A, Taylor JSH. A scanning electron microscope study of the development of a peripheral sensory neurite network. J Embryol Exp Morph. 1982;69:237–50. P47 Efficient signal processing and sampling in random networks that generate variability Sakyasingha Dasgupta1,2, Isao Nishikawa3, Kazuyuki Aihara3, Taro Toyoizumi2 1IBM Research - Tokyo, Tokyo, Japan; 2RIKEN Brain Science Institute, Tokyo, Japan; 3The University of Tokyo, Tokyo, Japan Correspondence: Sakyasingha Dasgupta - sdasgup@jp.ibm.com BMC Neuroscience 2016, 17(Suppl 1):P47 The source of cortical variability and its influence on signal processing remain an open question. We address the latter, by studying two types of randomly connected networks of quadratic integrate-and-fire neurons with balanced excitation-inhibition that produce irregular spontaneous activity patterns (Fig. 32A): (a) a deterministic network with strong synaptic interactions that actively generates variability by chaotic dynamics (internal noise) and (b) a stochastic network that has weak synaptic interactions but receives noisy input (external noise), e.g. by stochastic vesicle releases. These networks of spiking neurons are analytically tractable in the limit of a large network-size and slow synaptic-time-constant. Despite the difference in their sources of variability, spontaneous (baseline) activity patterns of these two models are indistinguishable unless majority of neurons are simultaneously recorded. We characterize the network behavior with dynamic mean field analysis and reveal a single-parameter family that allows interpolation between the two networks, sharing nearly identical spontaneous activity (Fig. 32B). Despite the close similarity in the spontaneous activity, the two networks exhibit remarkably different sensitivity to external stimuli. Input to the former network reverberates internally and can be successfully read out over long time. Contrarily, input to the latter network rapidly decays and can be read out only for short time. This is also observed in the significant changes in the spiking probability of evoked responses across this family (Fig. 32C). The difference between the two networks is further enhanced if input synapses undergo activity-dependent plasticity, producing significant difference in the ability to decode external input from neural activity. We show that, this difference naturally leads to distinct performance of the two networks to integrate spatio-temporally distinct signals from multiple sources. Unlike its stochastic counterpart, the deterministic chaotic network activity can serve as a reservoir to perform near optimal Bayesian integration and Monte-Carlo sampling from the posterior distribution. We describe implications of the differences between deterministic and stochastic neural computation on population coding and neural plasticity.Fig. 32 A Schematic illustrations of the two balanced QIF networks models considered in the present study. The left network consists of strongly coupled neurons without noise, while the right network consists of weak coupling among neurons with noisy input. B Nearly identical rate autocorrelation functions in the two networks. The red line (C 0) represents the value of the autocorrelation at time 0 and cyan line (C ∞) is the value of auto-correlation function in the limit of large t. C Change in spiking probability for different network connectivity strengths (g~), after being stimulated by a brief input at time t = 0 P48 Modeling the effect of riluzole on bursting in respiratory neural networks Daniel T. Robb1, Nick Mellen2, and Natalia Toporikova3 1Department of Mathematics, Computer Science and Physics, Roanoke College, Salem, VA 24153, USA; 2Department of Pediatrics, University of Louisville, Louisville, KY 40208, USA; 3Department of Biology, Washington and Lee University, Lexington, VA 24450, USA Correspondence: Daniel T. Robb - robb@roanoke.edu BMC Neuroscience 2016, 17(Suppl 1):P48 To accommodate constantly changing environmental and metabolic demands, breathing should be able to vary flexibly within a range of frequencies. The respiratory neural network in the pre-Botzinger complex of the ventrolateral medulla controls and flexibly maintains the breathing rhythm, coordinating network-wide bursting to signal the inspiratory phase of the breath. The frequency of this rhythmic activity is controlled by a number of neuromodulators, the majority of which are excitatory. Therefore, the central pattern generator for rhythmic respiratory activity should possess two seemingly contradictory properties: it has to be able to change frequency in response to excitatory input, but it also has to preserve stable rhythmic activity under a wide range of conditions. A persistent sodium current (INaP) been identified as one of the key currents for generation of inspiratory activity [1]. It has been shown that some of the neurons in Pre-BotC possess an intrinsic bursting mechanism, which relies on inactivation of this current. Higher expression of INaP correlates with higher burst frequency of a single pacemaker neuron [2]. However, the INaP pacemaker mechanism can only function within very narrow ranges of external excitation—NaP dependent pacemaker tends to switch to tonic firing after a small increase in depolarizing current [3]. In this combined experimental and computational study, we tested the effect of the persistent sodium blocker Riluzole (RIL) in several different levels of continuous depolarization, simulated by application of K+. Whereas increased potassium increases the bursting frequency of the control network, in the presence of RIL the increased potassium does not alter the bursting frequency (Fig. 33). These findings indicate that INaP is responsible for flexible modulation of respiratory rhythm, but there is another mechanism, which can sustain rhythmic activity in its absence. We developed a computational model which incorporates a Calcium sensitive Non-specific cationic current (IcaN) in addition to INaP. Our simulations indicate that IcaN and INaP can maintain the rhythm in respiratory neurons in the presence of RIL, and are capable of providing stable oscillations in the presence of tonic excitation by K+.Fig. 33 Summary of experiment on the effect of riluzole on the dependence of burst frequency on potassium concentration. Without riluzole (left), the frequency increases steadily with increasing potassium concentration. With riluzole present (right), the frequency remains essentially constant with increasing potassium concentration ReferencesButera RJ Jr, Rinzel J, Smith JC. Models of respiratory rhythm generation in the pre-Bötzinger complex. I. Bursting pacemaker neurons. J Neurophysiol. 1999;82:382–97. Purvis LK, Smith JC, Koizumi H, Butera RJ. Intrinsic bursters increase the robustness of rhythm generation in an excitatory network. J Neurophysiol. 2007;97:1515–26. Del Negro CA, Morgado-Valle C, Hayes JA, Mackay DD, Pace RW, Crowder EA, Feldman JL. Sodium and calcium current-mediated pacemaker neurons and respiratory rhythm generation. J Neurosci Off J Soc Neurosci. 2005;25:446–53. P49 Mapping relaxation training using effective connectivity analysis Rongxiang Tang1, Yi-Yuan Tang2 1Department of Psychology, Washington University in St. Louis, St. Louis, MO 63130, USA; 2Department of Psychological Sciences, Texas Tech University, TX 79409, USA Correspondence: Yi-Yuan Tang - yiyuan.tang@ttu.edu BMC Neuroscience 2016, 17(Suppl 1):P49 Relaxation training (RT)is a behavioral therapy that has been applied in stress management, muscle relaxation and other health benefit. However, compared to short-term meditation training, previous studies did not show the significant differences in brain changes following same amount of RT [1,2]. One possible reason might derive from the insensitive correlation based routine functional connectivity method that could not reveal training-related changes in effective connectivity (directed information flow) among these distributed brain regions. Here, we applied a novel spectral dynamic causal modeling (spDCM) to resting state fMRI to characterize changes in effective connectivity. Twenty-three healthy college students were recruited through campus advertisements and received 4 weeks of RT (10 h in total), previously reported in our randomized studies [1, 2]. All neuroimaging data were collected using an Allegra 3-Telsa Siemens scanner and processed using the Data Processing Assistant for Resting-State fMRI, which is based on SPM and Resting-State fMRI Data Analysis Toolkit [3]. For each participant, the subsequent standard procedures included slice timing, motion correction, regression of WM/CSF signals, and spatial normalization [3]. Based on previous literature, we specified four regions of interest within default mode network (DMN)—medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC), and bilateral inferior parietal lobule (left IPL and right IPL), same coordinates as in previous spDCM studies [4]. A standard DCM analysis involves a specification of plausible models, which are then allows the model parameters to be estimated following Bayesian model selection. In both pre- and post-RT conditions, the procedure selected the fully connected model as the best model with a posterior probability of almost 1. The fully connected model had 24 parameters describing the extrinsic connections between nodes, the intrinsic (self-connections) within nodes and neuronal parameters describing the neuronal fluctuations within each node. We used Bayesian Parametric Average to quantify the differences between pre- and post-RT, and a classical multivariate test—canonical variate analysis to test for any significances in these differences [4]. Our results showed no significant differences in causal relationships among the above nodes following RT (all P > 0.05). Conclusions Four weeks of RT could not induce significant changes in effective connectivity among DMN nodes. Long-term RT effect on brain changes warrants further investigation. Acknowledgements: This work was supported by the Office of Naval Research. ReferencesTang YY, Holzel BK, Posner MI. The neuroscience of mindfulness meditation. Nat Rev Neurosci. 2015;16:213–25. Tang YY, Lu Q, Geng X, Stein EA, Yang Y, Posner MI. Short-term meditation induces white matter changes in the anterior cingulate. Proc Natl Acad Sci USA. 2010;107:15649–52. Tang YY, Tang R, Posner MI. Brief meditation training induces smoking reduction. Proc Natl Acad Sci USA. 2013;110:13971–75. Razi A, Kahan J, Rees G, Friston KJ. Construct validation of a DCM for resting state fMRI. Neuroimage. 2015;106:1–14. P50 Modeling neuron oscillation of implicit sequence learning Guangsheng Liang1, Seth A. Kiser2,3, James H. Howard, Jr.3, Yi-Yuan Tang1 1Department of Psychological Sciences, Texas Tech University, TX 79409, USA; 2The Department of Veteran Affairs, District of Columbia VA Medical Center, Washington, DC 20420, USA; 3Department of Psychology, The Catholic University of America, Washington, DC 20064, USA Correspondence: Yi-Yuan Tang - yiyuan.tang@ttu.edu BMC Neuroscience 2016, 17(Suppl 1):P50 Implicit learning (IL) occurs without goal-directed intent or conscious awareness but has important influences on our everyday functioning and overall health such as environmental adaptation, developing habits and aversions. Most of IL studies used event-related potentials (ERPs) to study brain response by taking the grand average of all event-related brain signals. How neuron oscillation (EEG frequency band) involves in IL remains unknown. Moreover, ERP analysis requires brain signals that are not only time locked, but also phase locked to the event, therefore the information with phase locked signals are missed and not presented in potentials. To address this issue, we applied time–frequency analysis and cluster-based permutation test in this study. Fifteen healthy participants were recruited to perform three sessions of triplets learning task (TLT), an IL task commonly used in the field [1]. Three successive cues were presented and participants were asked to observe the first two cues and only respond to the third cue (target) by pressing corresponding keys. During the task, EEG signals were recorded. Cluster based permutation on alpha and theta band is used to deal with family-wise error rate and in the same time, help to find out difference occurred in specific time range along with spatial information among different triplet types. Base on the behavioral result, overall learning occurs in session1, while triplet-specific learning takes place in session2. We find significant difference in both alpha (8–13 Hz) and theta (4–8 Hz) frequency band. For alpha band, power modulation shows significant difference between high versus low frequency triplet group in session2 in the frontal cortex. For theta band, theta power shows significant difference between session1 and session3 in the frontal cortex. It started from as early as target onset until the end of the trial in high frequency triplet group. However, in the low frequency triplet group, the power differential occurs later, from around 1000 ms till the end of the next trial. Conclusions Behavioral result showed that the brain learned the regularity of sequence implicitly. Alpha power modulation indicated that the brain allocated resource in attention among two different triplet types. Theta power modulation showed the difference of memory processing and retrieval among two different triplet types. Our results indicated that participants did not find the regularity of the triplet types till the end of the study, but the brain in fact reacts to these two different triplet types differently. Acknowledgements: This work was supported by the Office of Naval Research. ReferenceHoward JH, Howard DV, Dennis N, Kelly AJ. Implicit learning of predictive relationships in three-element visual sequences by young and old adults. J Exp Psychol Learn Mem Cogn. 2008, 34: 1139–57. P51 The role of cerebellar short-term synaptic plasticity in the pathology and medication of downbeat nystagmus Julia Goncharenko1, Neil Davey1, Maria Schilstra1, Volker Steuber1 1Centre for Computer Science and Informatics Research, University of Hertfordshire, Hatfield, AL10 9EJ, UK Correspondence: Julia Goncharenko - i.goncharenko@herts.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P51 Downbeat nystagmus (DBN) is a common eye fixation disorder that is linked to cerebellar pathology. DBN patients are treated with 4-aminopyridine (4-AP), a K channel blocker, but the underlying mechanism is unclear. DBN is associated with an increased activity of floccular target neurons (FTNs) in the vestibular nuclei. It was previously believed that the reason for the increased activity of FTNs in DBN is a pathological decrease in the spike rate of their inhibitory Purkinje cell inputs, and that the effect of 4-AP in treating DBN could be mediated by an increased Purkinje cell activity, which would restore the inhibition of FTNs and bring their activity back to normal [1]. This assumption, however, has been questioned by in vitro recordings of Purkinje cells from tottering (tg/tg) mice, a mouse model of DBN. It was shown that therapeutic concentrations of 4-AP did not increase the spike rate of the Purkinje cells, but that they restored the regularity of their spiking, which is impaired in tg/tg mice [2]. Prompted by these experiments, Glasauer and colleagues performed computer simulations to investigate the effect of the regularity of Purkinje cell spiking on the activity of FTNs [3]. Using a conductance based FTN model, they found that changes in the regularity of the Purkinje cell input only affected the FTN spike rate when the input was synchronized. In this case, increasing the regularity of the Purkinje cell spiking resulted in larger gaps in the inhibitory input to the FTN and an increased FTN spike rate. These results predict that the increased irregularity in the Purkinje cell activity in DBN should lead to a decreased activity of the FTNs, rather than the increased activity that is found in experiments, and they are therefore unable to explain the therapeutic effect of 4-AP. However, the model by Glasauer and colleagues does not take short-term depression (STD) at the Purkinje cell—FTN synapses into account. We hypothesized that this absence of STD could explain the apparent contradiction between the experimental [2] and computational [3] results. To study the role of STD in the pathology and 4-AP treatment of DBN, we used a morphologically realistic conductance based model of a cerebellar nucleus (CN) neuron [4, 5] as an FTN model to simulate the effect of irregular versus regular Purkinje cell input. The coefficients of variation of the irregular and regular Purkinje cell spike trains during DBN and after 4-AP treatment, respectively, were taken from recordings from wild-type and tg/tg mice [6], which served as a model system for DBN. We presented the FTN model with synchronized and unsynchronized input and found that, for both conditions, irregular (DBN) input trains resulted in higher FTN spike rates than regular (4-AP) ones. In the presence of unsynchronized Purkinje cell input, the acceleration of the FTN spike output during simulated DBN and the deceleration during simulated 4-AP treatment depended on STD at the Purkinje cell synapses. Our results provide a potential explanation for the pathology and 4-AP treatment of pathological nystagmus. ReferencesGlasauer S, Kalla R, Buttner U, Strupp M, Brandt T. 4-aminopyridine restores visual ocular motor function in upbeat nystagmus. J Neurol Neurosurg Psychiatry. 2005;76:451–3. Alvina K, Khodakhah K. The therapeutic mode of action of 4-aminopyridine in cerebellar ataxia. J Neurosci. 2010;30:7258–68. Glasauer S, Rössert C, Strupp M. The role of regularity and synchrony of cerebellar Purkinje cells for pathological nystagmus. Ann NY Acad Sci. 2011;1233:162–7. Steuber V, Schultheiss NV, Silver RA, de Schutter E, Jaeger D. Determinants of synaptic integration and heterogeneity in rebound firing explored with data-driven models of deep cerebellar nucleus cells. J Comp Neurosci. 2011;30:633–58. Luthman J, Hoebeek FE, Maex R, Davey N, Adams R, de Zeeuw CI, Steuber V. STD-dependent and independent encoding of input irregularity as spike rate in a computational model of a cerebellar nucleus neuron. Cerebellum. 2011;10:667–82. Hoebeek FE, Stahl JS, van Alphen AM, Schonewille M, Luo C, Rutteman M, van den Maagdenberg AM, Molenaar PC, Goossens HH, Frens MA, et al. Increased noise level of Purkinje cell activities minimizes impact of their modulation during sensorimotor control. Neuron 2005, 45(6):953–965. P52 Nonlinear response of noisy neurons Sergej O. Voronenko1,2, Benjamin Lindner1,2 1Department of Physics, Humboldt University, Berlin, 10099, Germany; 2Bernstein Center for Computational Neuroscience, Berlin, 10115, Germany Correspondence: Sergej O. Voronenko - sergej@physik.hu-berlin.de BMC Neuroscience 2016, 17(Suppl 1):P52 In many neuronal systems that exhibit high trial-to-trial variability the time-dependent firing rate is thought to be the main information channel for time-dependent signals. However, for nerve cells with low intrinsic noise and highly oscillatory activity synchronization, mode locking and frequency locking seem to be of major importance. Here, we present an extension to the linear response theory [1, 2] for the leaky integrate-and-fire neuron model to second order and demonstrate how the time-dependent firing rate can exhibit features that are reminiscent of mode-locking and frequency-locking. Although our theory allows to predict the response to general weak time-dependent signals, the second-order effects are best demonstrated using cosine signals as in Fig. 34A. We consider a leaky integrate-and-fire model for which the subthreshold voltage, Fig. 34B, is subject to the signal and to Gaussian white noise. Whenever the voltage hits the threshold, it is reset to zero and a spike time is recorded in the raster plot, Fig. 34C. The firing rate can be obtained numerically by averaging over the spike trains or via a perturbation approach similar to the weakly nonlinear analysis in [3]. We find that the firing rate can exhibit pronounced nonlinear behavior as can be seen from the excitation of a harmonic oscillation in Fig. 34D. Further effects that are not shown in Fig. 34 but are revealed by our analysis are a signal-dependent change of the mean firing rate and a pronounced nonlinear response to the sum of two cosine signals.Fig. 34 Nonlinear modulation of the firing rate by a cosine signal. A Signal, B subthreshold voltage, C rasterplot, D The time-dependent firing rate (red, noisy trace) is significantly different from the linear theory (dashed line) but is accurately described by the second-order response (solid line) Summary and conclusions Here we demonstrate that the time-dependent firing rate (equivalent to the instantaneous population rate for neurons driven by a common stimulus) can exhibit pronounced nonlinearities even for weak signal amplitudes. The linear theory does not only give quantitatively wrong predictions but also fails to capture the timing of the modulation peaks. Hence, our theory has not only implications for sinusoidal stimulation that is commonly used to study dynamic properties of nerve cells but also demonstrates the relevance of the nonlinear response for the encoding of complex time-dependent signals. Acknowledgements: This work was supported by the BMBF (FKZ: 01GQ1001A) and the DFG (research training group GRK1589/2). ReferencesBrunel N, Chance FS, Fourcaud N, Abbott LF. Effects of synaptic noise and filtering on the frequency response of spiking neurons. PRL. 2001;86(10):2186–9. Lindner B, Schimansky-Geier L. Transmission of noise coded versus additive signals through a neuronal ensemble. PRL. 2001;86(14):2934–7. Brunel N, Hakim V. Fast global oscillations in networks of integrate-and-fire neurons with low firing rates. Neural Comput. 1999;11(7):1621–71. P53 Behavioral embedding suggests multiple chaotic dimensions underlie C. elegans locomotion Tosif Ahamed1, Greg Stephens1,2 1Biological Physics Theory Unit, Okinawa Institute of Science and Technology, Okinawa 904-0495, Japan; 2Department of Physics and Astronomy, Vrije Universiteit Amsterdam Correspondence: Tosif Ahamed - tosif.ahamed@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P53 Behavior is the primary output of an organism; genetic and neural circuits, no matter how complex, seek to optimize this output. A quantitative understanding of behavior is therefore crucial to our understanding of biological processes. A key characteristic of natural behavior is variability; even the most stereotyped movements such as reaching to a target, which are similar in aggregate, can vary substantially from trial to trial. In motor control such variability is often ascribed to noise in the sensorimotor control circuit. On the other hand, deterministic dynamical systems can generate variability intrinsically when operating in a chaotic regime. Differentiating between the two is important as they generate separate mechanistic predictions about how variability is generated in the brain. Here, we use tools from nonlinear dynamics to understand behavioral variability in the movement of C. elegans. We reconstruct a 6-dimensional phase space by developing a novel extension of multivariate singular systems analysis [1] and applying it to a low-dimensional but complete representation of worm postures obtained from videos of freely foraging worms [2]. At a coarse level, the reconstructed phase space naturally separates into three stereotyped behaviors: forward locomotion, reversals and turns (Fig. 35A, B). However, there is also substantial variability at finer scales, which is reflected in positive maximal Lyapunov exponents (MLE) [3] within trajectories corresponding to each individual behavior (Fig. 35C). The MLEs calculated this way differ significantly from MLEs calculated from a random shuffle of the data that preserves its linear structure (or power spectrum). This implies that the positive MLEs, which indicate sensitive dependence to initial conditions in C. elegans behavior are a result of nonlinear structure present in the dynamics. These results are strengthened by the fact that we observe little inter-animal variability in the estimated values of MLE, additionally the values also agree well with the estimated time scales of the three behaviors. Based on these observations we propose that C. elegans behavior might be driven by the activity of multiple coupled chaotic attractors. We expect our analysis will also be relevant in understanding global neural dynamics, recently imaged in freely-moving worms [4].Fig. 35 Phase space portrait and divergence of nearby trajectories. A The top panel shows the orthogonal relationship between the forward and reversal behaviors, while the bottom panel shows the transition from reversal to an omega turn in the phase space. To aid visualization color coding is done by radial distance from the origin. B Escape response visualized in the phase planes. When the worm is hit with a laser impulse, it makes a reversal, followed by an omega turn and then resumes forward crawling. Color map encodes time in frames. C Divergence curves for the three different attractors. Y-axis shows the exponential of the divergence between neighboring trajectories plotted on a semilog scale on the y axis, each curve corresponds to a single worm (n = 12). λ L is estimated by calculating the slope of the linear region. Boxplots show the range of λ L obtained from different animals ReferencesRead PL. Phase portrait reconstruction using multivariate singular systems analysis. Phys D. 1993;69(3):353–65. Stephens GJ, Johnson-Kerner, B, Bialek W, Ryu WS. Dimensionality and dynamics in the behavior of C. elegans. PLoS Comput Biol. 2008;4(4):e1000028. Kantz H. A robust method to estimate the maximal Lyapunov exponent of a time series. Phys Lett A. 1994;185(1):77–87. Nguyen JP, Shipley FB, Linder AN, Plummer GS, Liu M, Setru SU, Shaevitz JW, Leifer AM. Whole-brain calcium imaging with cellular resolution in freely behaving Caenorhabditis elegans. PNAS. 2015;201507110. P54 Fast and scalable spike sorting for large and dense multi-electrodes recordings Pierre Yger1, Baptiste Lefebvre1, Giulia Lia Beatrice Spampinato1, Elric Esposito, Marcel Stimberg et Olivier Marre1 1Institut de la Vision, INSERM UMRS 968, CNRS UMR 7210, Paris Correspondence: Pierre Yger - pierre.yger@inserm.com BMC Neuroscience 2016, 17(Suppl 1):P54 Understanding how assemblies of neurons encode information requires recording of large populations of cells in the brain. In recent years, multi-electrode arrays and large silicon probes have been developed to record simultaneously from thousands of electrodes packed with a high density. However, these new devices challenge the classical way to do spike sorting. First, the large number of electrodes preclude approaches based on manual clustering. Even automatic approaches need to be fast enough to handle the amount of extracellular data. Second, the density of the electrodes is high enough so that a single spike will be detected on many electrodes. So the different channels must be processed simultaneously. Third, within a large and dense array of electrodes, overlapping spikes are rather the rule than the exception, and it is known that classical clustering methods cannot easily capture the synchronous occurrence of two spikes from two different cells [1]. Here we developed a new software to solve all these aforementioned issues, based on a highly automated algorithm to extract spikes from extracellular data, and show that this algorithm reached near optimal performance both in vitro and in vivo. The algorithm is composed of two main steps: (1) a “template-finding” phase to extract the cell templates, i.e. the pattern of activity evoked over many electrodes when one neuron fires an action potential; (2) a “template-matching” phase where the templates are matched onto the raw data to find the location of the spikes. The manual intervention by the user is reduced to the minimal, and the time spent on manual curation did not scale with the number of electrodes. For the template-finding phase, we start by detecting all the possible times in the raw data that could contain a spike. Spikes are then clustered into groups using a density-based clustering derived from [1], and we then extract the template corresponding to each group. In the fitting phase, we match the templates onto the raw data with a method that allows amplitude variation for each template [2]. The algorithm is written in Python and is entirely parallelized such that it can handle large amount of data. It also provides a graphical user interface so that the output of the algorithm can be checked, and to refine the sorting. We tested our algorithm with large-scale data from in vitro and in vivo recordings, from 32 and up to 4225 electrodes. In all cases, we estimated its performance on data with ground truth, i.e. cases where the solution to the sorting problem is at least partially known. The performance was always close to the maximal expected performance. Therefore, our method appears as a general solution to sort spikes from large-scale extracellular recordings. ReferencesEinevoll GT, et al. Towards reliable spike-train recordings from thousands of neurons with multielectrodes. Curr Opin Neurobiol. 2012;22:11–17. Rodriguez A, et al. Clustering by fast search and find of density peaks. Science. 2014;344(6191):1492–96. P55 Sufficient sampling rates for fast hand motion tracking Hansol Choi1, Min-Ho Song2 1Bernstein Center Freiburg, Institute of Biology III, University of Freiburg, Germany, 79100; 2fourMs group, Dept. Musicology, University of Oslo, Norway, 0371 Correspondence: Min-Ho Song - minho.song@imv.uio.no BMC Neuroscience 2016, 17(Suppl 1):P55 When tracking fine motor behaviors in human body parts, passive marker-based tracking is one of the best-suited methods not only because of its high spatial precision and temporal resolution, but also allowing high degrees-of-freedom [1]. However, the passive marker approach suffers from identity confusion problem (Fig. 36A) between the markers. As the speed of motion increases, sufficient sampling rate is required to avoid the problem. In a recent study [2], we reported that the problem still occurs even with the sampling rate significantly higher than the Nyquist sampling rate. The study suggested a sampling rate criterion to avoid identity problem for the worst-case condition.Fig. 36 Experimental design and result. A Marker confusion. Grey dots are markers. d 1, d 2 are the distances between markers, t s is the sampling latency, v is speed of marker. Green lines show the markers, which identified as same. Left correct identification right example of marker confusion. B Experimental set up. Red dots are keys to press by the thumb and the little finger during repeats. C The probabilities of continuous marker identification In this poster, the confusion problem is tested in more realistic human motor control behavior. Grids of 3 × 3 markers with different distances (1, 1.5 and 2 cm) were attached to a skilled piano player’s right hand (Fig. 36B). The experimental task was repeated right-hand alternative keystrokes between D#5 and D#7 (two octave) with a tempo of 176 bpm for 10 s. This is an excerpt from Liszt’s La Campanella, which requires fast horizontal jump of the right-hand. These motions were recorded with 7 optical motion capture cameras (Qualisys Ltd. Oqus 400) changing the sampling rates from 50 to 200 Hz. The maximum frequency components of these hand movements were lower than 8 Hz. The probability of successful tracking is measured by counting the number of successful repetition of the center marker (Fig. 36C). Estimated required sampling rates for successful tracking (where the probabilities reach 100 %) were 101, 137z, and 181 Hz (fitted to piecewise linear functions by expectation maximization). The theoretically predicted values are 176, 235, and 353 Hz [2]. We found that the required sampling rates are lower than the theoretical criterion. This is because the theoretical prediction was developed to avoid the worst case where marker trajectories overlap from perfect periodic motion; not realistic for human movement, which has variability. Our results show that in practical situations involving human movements, the sampling criterion can be weakened considerably. But, it should be note that a motion slower than 10 Hz still requires more than 100 Hz, which far exceeds the Nyquist sampling rate. ReferencesGuerra-Filho G. Optical motion capture: theory and implementation. J Theor Appl Inf. 2005;12(2):61–8.9 Song M-H, Godøy RI. How fast is your body motion? Determining a sufficient frame rate for an optical motion tracking system using passive markers. PLoS One (in press). P56 Linear readout of object manifolds SueYeon Chung1, Dan D. Lee2, Haim Sompolinsky1,3 1Center for Brain Science, Harvard University, Cambridge, MA 02138, USA; 2Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104, USA; 3Edmond and Lily Safra Center for Brain Sciences, Hebrew University, Jerusalem 91904, Israel Correspondence: SueYeon Chung - schung@fas.harvard.edu BMC Neuroscience 2016, 17(Suppl 1):P56 Objects are represented in sensory systems by continuous manifolds due to sensitivity of neuronal responses to changes in physical features such as location, orientation, and intensity [1]. It has been hypothesized that object identity can be decoded from high level representations, by simple downstream readout networks. What makes certain sensory representations better suited for invariant decoding of objects by downstream networks? We generalize Gardner’s statistical mechanical analysis of points [2, 3] and establish a replica theory of linear classification of manifolds synthesizing statistical and geometric properties of high dimensional signals. We show how changes in the dimensionality, size, and shape of the object manifolds affect the capacity and the distribution of configurations in downstream perceptrons (Fig. 37).Fig. 37 Theoretical predictions (lines) and numerical simulation (markers) are shown. A1 Classification of line segments. (Solid) lines embedded in the margin, (dotted) lines touching the margin, (striped) interior lines .A2 Capacity α = P/N of a network N = 200 as a function of R (line length) with margins κ = 0, 0.5. A3 Fraction of configurations at capacity with κ = 0. (red) lines in the margin, (blue) touching the margin, (black) interior lines. B1 D2 balls, B2 capacity α = P/N for κ = 0 for large D = 50 and R ∝ D−1/2 as a function of RD. (Blue solid) αD(0, R) compared with α0(RD) (red square). (Inset) capacity α at κ = 0 for 0.35 ≤ R ≤ 20 and D = 20: (blue) theoretical α compared with approximate form (1 + R−2)/D (red dashed). C1 2D L1 balls. C2 Fraction of configurations as a function of radius R at capacity with κ = 0. (red) entire manifold embedded, (blue) touching margin at a single vertex, (gray) touching with two corners (one side), (purple) interior manifold Our analysis shows how linear separability of the manifolds depends intimately upon the dimensionality, size and shape of the the manifolds. These properties are expected to differ at different stages in the sensory hierarchy. Thus, the present work enables systematic analysis of the degree to which this reformatting enhances the capacity for object classification in different sensory processing stages. The present work lays the groundwork for a computational theory of neuronal processing of objects in the presence of variability, providing quantitative measures for assessing the properties of representations in biological and artificial neural networks. ReferencesDiCarlo JJ, Cox DD. Untangling invariant object recognition. Trends Cogn Sci. 2007;11(8):333–41. Gardner E. Maximum storage capacity in neural networks. EPL (Europhys Lett). 1987;4(4):481 Abbott LF, Kepler TB. Universality in the space of interactions for network models. J Phys A Math Gen. 1989;22(12):2031. P57 Differentiating models of intrinsic bursting and rhythm generation of the respiratory pre-Bötzinger complex using phase response curves Ryan S. Phillips1,2, Jeffrey Smith1 1NINDS, NIH, Bethesda, MD 20892, USA; 2Department of Physics, University of New Hampshire, Durham, NH, 03824, USA Correspondence: Ryan S. Phillips - Ryan.Phillips@nih.gov BMC Neuroscience 2016, 17(Suppl 1):P57 The pre-Bötzinger complex (PBC) is an essential rhythmogenic brainstem nucleus located in the ventrolateral medulla. Rhythmic output from the PBC is relayed through premotor and motor neurons to the diaphragm and intercostal muscles to drive the active inspiratory phase of respiration. The specific biophysical mechanisms responsible for generating rhythmic bursting and network synchronization are not well understood and remain a highly controversial topic within the field. A wide variety of biophysical mechanisms have been proposed to explain the origins of intrinsic bursting and rhythmogenesis including persistent sodium currents [1, 2], calcium-activated nonspecific cation channels [2, 3], inositol trisphosphate (IP3) signaling [2], and synaptic mechanisms [4]. Computational simulations of these models produce similar patterns of bursting and network synchronization compared to each other and experimental recordings despite having different underlying mechanisms. In this theoretical study we demonstrate a method to differentiate between biophysically distinct models using phase response curves (PRCs). PRCs characterize the change in phase of an oscillator as a function of the timing of a perturbation to the system. Depolarizing and hyperpolarizing perturbations were generated by incorporating the light sensitive channels, Channelrhodopsin-2 and Archaerhodopsin, respectively, into our conductance based PBC models. A library of model PBC neurons was generated by varying the conductance of each ion channel over equally spaced intervals. PRCs were then calculated for each neuron capable of producing rhythmic bursting. Preliminary results found that in general depolarizing perturbations produced qualitatively similar PRCs and could both advance or delay the next cycle. Conversely, hyperpolarizing perturbations produced qualitatively distinct PRCs depending on the combination of conductance magnitudes. In conclusion, these PRCs provide a method for differentiating models of intrinsic bursting and rhythm generation based on underlying biophysical mechanisms and provide a means for interpreting experimentally derived PRCs from the PBC. ReferencesButera RJ, Rinzel J, Smith JC. Models of respiratory rhythm generation in the pre-Bötzinger complex. I. Bursting pacemaker neurons. J Neurophysiol. 1999;82:382–97. Jasinski PE, Molkov YI, Shevtsova NA, Smith JC, Rybak IA. Sodium and calcium mechanisms of rhythmic bursting in excitatory neural networks of the pre-Bötzinger complex: a computational modelling study. Eur J Neurosci. 2013;37:212–30. Rubin JE, Hayes JA, Mendenhall JL, Del Negro CA. Calcium-activated nonspecific cation current and synaptic depression promote network-dependent burst oscillations. Proc Natl Acad Sci USA. 2009;106:2939–44. Guerrier C, Hayes JA, Fortin G, Holcman D. Robust network oscillations during mammalian respiratory rhythm generation driven by synaptic dynamics. Proc Natl Acad Sci. 2015;201421997. P58 The effect of inhibitory cell network interactions during theta rhythms on extracellular field potentials in CA1 hippocampus Alexandra Pierri Chatzikalymniou1,2, Katie Ferguson1,3, Frances K. Skinner1,2,4 1Krembil Research Institute, University Health Network, Toronto, ON, Canada; 2Department of Physiology, University of Toronto, Toronto ON, Canada; 3Department of Neuroscience, Yale School of Medicine, New Haven, CT, 06520, USA; 4Department of Medicine (Neurology), University of Toronto, Toronto ON, Canada Correspondence: Alexandra Pierri Chatzikalymniou - alexandra.chatzikalymniou@mail.utoronto.ca BMC Neuroscience 2016, 17(Suppl 1):P58 Oscillatory local field potentials (LFPs) are extracellularly recorded potentials with frequencies of up to ~500 Hz. They are associated with a number of physiological functions in health and disease and complement the information obtained by analysis of spikes. Because multiple neuronal processes contribute to the LFP, the signal is inherently ambiguous and more difficult to interpret than spikes [1]. However, the biophysical origin of LFPs is well understood in the framework of volume conductor theory [4]. Using “LFPy” [3], a python package that implements this framework, we construct a pyramidal cell model in CA1 hippocampus which generates extracellular potentials. Our pyramidal cell model receives inhibitory synaptic input from four different types of CA1 interneuron populations. These interneuron models are taken from a previous, experimentally constrained inhibitory network model developed to understand spontaneous theta (4–12 Hz) rhythms as expressed in an intact hippocampus preparation [2]. We investigate the contribution of the different inhibitory cell type interactions to the extracellular potential. In our current model we placed a virtual electrode probe along the vertical axis of the pyramidal cell to record its output in a layer dependent manner. We identified distinct regimes where specific interneuron cell type interactions distinctively affect the polarity, amplitude and frequency of the LFP signal (Fig. 38). We also distinguish between regimes where synaptic connection strengths preserve the extracellular potential frequency versus those that lead to lag or abolishment of the extracellular rhythm. In this way, our model helps us understand the cellular contributions to extracellular field patterns that arise in experimental recordings as a function of biologically relevant network states when the efficacy of inhibitory connections dynamically varies.Fig. 38 Example of the spatial attenuation of the extracellular potential signal for a particular set of inhibitory connections. The temporal traces at two electrode locations are represented with blue and green dots accordingly. Average over the absolute maximum extracellular potential amplitudes is shown in 2D space. According to the schematic the rate of the extracellular signal spatial attenuation generated by the pyramidal cell is approximately 400 μ Acknowledgements: Supported by NSERC Canada, U. of T. Fellowship P.S.L., and the SciNet HPC Consortium. ReferencesBuzsáki G, Anastassiou CA, Koch C. The origin of extracellular fields and currents—EEG, ECoG, LFP and spikes. Nat Rev Neurosci. 2012;13:407–20. Ferguson KA, Huh CYL, Amilhon B, Williams S, Skinner FK. Network models provide insight into how oriens-lacunosum-moleculare (OLM) and bistratified cell (BSC) interactions influence local CA1 theta rhythms. Front Syst Neurosci. 2015;9:110 Lindén H, Hagen E, Leski S, Norheim ES, Pettersen KH, Einevoll GT. LFPy: a tool for biophysical simulation of extracellular potentials generated by detailed model neurons. Front Neuroinform. 2014;7:41. Rall W, Shepherd GM. Theoretical reconstruction of field potentials and dendrodendritic synaptic interactions in olfactory bulb. J Neurophysiol. 1968;31:884–915. P59 Expansion recoding through sparse sampling in the cerebellar input layer speeds learning N. Alex Cayco Gajic1, Claudia Clopath2, R. Angus Silver1 1Department of Neuroscience, Physiology and Pharmacology, University College London, London, UK; 2Department of Bioengineering, Imperial College London, London, UK Correspondence: N. Alex Cayco Gajic - natasha.gajic@ucl.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P59 Feed-forward networks often have many more output neurons than input neurons. This is thought to enable them to project neuronal activity patterns into a higher dimensional space. When such expansion recoding is combined with sparse representations it provides a powerful way to increase the separation between activity patterns [1–4]. In the input layer of the cerebellar cortex, granule cells (GCs) integrate sensorimotor input from the less numerous mossy fibre afferents (MFs), with each GC sampling from only 2 to 7 local MFs. Recent work has revealed that this sparse sampling of input by GCs provides an optimal tradeoff between information transmission and sparsification over a range of activity levels [4]. Moreover, theories of cerebellar function have linked expansion recoding under sparse regimes to pattern separation and associative learning [5, 6]. However, the relationship between the feedforward excitatory synaptic connectivity and the learning performance is poorly understood. To investigate how the number of MF inputs per GC affects the performance we simulated a model of an 80µ ball of the MF-GC feedforward layer with either random or clustered MF activation. The connectivity profile of the network was constrained with recent anatomical data [4]. MF stimulation was modeled as random binary patterns with varying levels of population activity and correlation, and GCs as high-threshold rectified linear units. We then measured the speed at which granule cell population activity can be used to classify random patterns via backpropagation learning as the number of synaptic inputs was varied. We found that the largest speedup of learning in the GC activity (compared to learning based on the MF inputs) occurred when each GC received only a few synaptic inputs. We probed this result by analyzing the eigenvalues of the covariance matrix of the population-level activity, finding that sparse sampling of MF inputs allows GCs to both expand and decorrelate MF activity patterns. Interestingly, this feature is robustly preserved even in the presence of clustered inputs. In summary, we find that sparse sampling combined with sparsification of activity allows GCs to optimize both pattern expansion and pattern decorrelation. Acknowledgement: This research is funded by the Wellcome Trust. ReferencesLaurent G. Olfactory network dynamics and the coding of multidimensional signals. Nat Rev Neurosci. 2001;3:884–95. Olshausen BA, Field DJ. Sparse coding of sensory inputs. Curr Opin Neurobiol. 2004;14:481–7. Billings G, Piasini E, Lorincz A, Nusser Z, Silver A. Network structure within the cerebellar input layer enables lossless sparse encoding. Neuron. 2014;83:960–74. Babadi B, Sompolinksy H. Sparseness and expansion in sensory representations. Neuron. 2014;83:1–14. Marr D. A theory of cerebellar cortex. J Physiol. 1969;202:437–70. Tyrell T, Wilshaw D. Cerebellar cortex: its simulation and the relevance of Marr’s theory. Philos Trans R Soc B. 1992;336:239–57. P60 A set of curated cortical models at multiple scales on Open Source Brain Padraig Gleeson1, Boris Marin1, Sadra Sadeh1, Adrian Quintana1, Matteo Cantarelli2, Salvador Dura-Bernal3, William W. Lytton3, Andrew Davison4, R. Angus Silver1 1Department of Neuroscience, Physiology and Pharmacology, University College London, London, UK; 2Metacell LLC, San Diego, CA, USA; 3State University of New York Downstate Medical Center, Brooklyn, NY, USA; 4Neuroinformatics group Unité de Neurosciences, Information et Complexité, CNRS, Gif sur Yvette, France Correspondence: Padraig Gleeson - p.gleeson@ucl.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P60 Computational models of spiking cortical networks are implemented using a variety of approaches from large scale models with simplified point neurons and anatomically inspired connectivity, to networks on smaller scales with morphologically and biophysically detailed neurons. In between these scales many published models have used intermediate representations of neurons (e.g. conductance based with one compartment or abstract morphologies). These studies, and the associated modelling scripts, provide many potential starting points for experimental and theoretical neuroscientists wishing to use biologically constrained cortical models in their investigations. In addition, there are an increasing number of public neuroinformatics resources which are providing structured experimental data on the electrophysiology, connectivity and morphology of cortical neurons. While these modelling and experimental resources should lead to a proliferation in well constrained cortical models there remain a number of practical and technical barriers to more widespread development and use of such models among researchers. The Open Source Brain (OSB) initiative (http://www.opensourcebrain.org) is a resource for collaborative development of models in computational neuroscience. Sharing of models in standardised representations such as NeuroML 2 [1] and PyNN is encouraged and actively supported on the site. Conversion of cell and network models to NeuroML allows them to be visualised and analysed in 3D in a standard web browser through the OSB website. We have recently added a feature to allow simulations to be executed on our servers (e.g. by conversion to NEURON) and the results displayed within the browser. We have been actively converting published cortical models to NeuroML format and making these available on OSB. These range from point neuron models [2, 3], to abstract [4] and detailed [5] multicompartmental models. We are working to develop these and others into a curated set of cortical models in a common format which can be used as the basis for new models. We have also developed frameworks for importing resources from neuroinformatics datasets such as the Allen Institute Cell Types database (http://celltypes.brain-map.org) and NeuroMorpho.org (http://neuromorpho.org). We have greatly improved compatibility between PyNN and NeuroML, allowing the modeller freedom to choose between procedural (Python) and declarative (XML) model specification. We have also extended a model optimisation framework (https://github.com/NeuralEnsemble/NeuroTune) facilitating generation of new NeuroML models from electrophysiological data. All of this work is aimed at making existing cortical models easier to access, visualise and simulate, simplifying development of new models based on these prototypes, and ensuring the latest experimental datasets can be used to constrain and validate complex models of cortical function. Acknowledgements: This work has been primarily funded by the Wellcome Trust (101445/095667). ReferencesCannon RC, Gleeson P, Crook S, Ganapathy G., Marin B, Piasini E, Silver RA. LEMS: a language for expressing complex biological models in concise and hierarchical form and its use in underpinning NeuroML2. Front Neuroinform. 2014;8:79. Izhikevich E. Simple model of spiking neurons. IEEE Trans Neural Netw. 2003;14(6):1569–72. Brunel N. Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. J Comput Neurosci. 2000;8(3):183–208. Traub RD, Contreras D, Cunningham MO, et al. Single-column thalamocortical network model exhibiting gamma oscillations, sleep spindles, and epileptogenic bursts. J Neurophysiol. 2005;93(4):2194–2232. Markram H, Muller E, Ramaswamy S, et al. Reconstruction and simulation of neocortical microcircuitry. Cell. 2015;163(2):456–92. P61 A synaptic story of dynamical information encoding in neural adaptation Luozheng Li1, Wenhao Zhang1, Yuanyuan Mi1, Dahui Wang1,2, Si Wu1 1State Key Laboratory of Cognitive Neuroscience & Learning, IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing 100875, China; 2School of System Science, Beijing Normal University, Beijing 100875, China Correspondence: Si Wu - wusi@bnu.edu.cn BMC Neuroscience 2016, 17(Suppl 1):P61 Adaptation refers to the general phenomenon that a neural system dynamically adjusts its response property according to the statistics of external inputs [1]. In response to a prolonged constant stimulation, neuronal firing rates always first increase dramatically at the onset of the stimulation; and afterwards, they decrease rapidly to a low level close to background activity (see Fig. 39A). This attenuation of neural activity seems to be contradictory to our experience that we can still sense the stimulus after the neural system is adapted [2]. Thus, it prompts a question: where is the stimulus information encoded during the adaptation? Here, we investigate a computational model in which the neural system employs a dynamical encoding strategy during the neural adaptation: at the early stage of the adaptation, the stimulus information is mainly encoded in the strong independent firings; and as time goes on, the information is shifted into the weak but concerted responses of neurons (see Fig. 39B). We find that short-term plasticity [3], a general feature of synapses, provides a natural mechanism to achieve this goal. Furthermore, we demonstrate that with balanced excitatory and inhibitory inputs, this correlation-based information can be read out efficiently. The implications of this study on our understanding of neural information encoding are discussed.Fig. 39 Firing rates, synaptic efficacy and cross-correlation change during the adaptation. A The time course of firing rates and the averaged synaptic efficacy of the network during the adaptation. ux is temporally enhanced during the adaptation due to the STF, but in the long term, strong STD drives the synaptic efficacy to background level. Stimulation is during 0–1500 ms. B The enhancement of cross-correlation between neurons during the adaptation Conclusions We have explored a dynamical encoding strategy in neural adaptation. By constructing a computational model, we show that this can be achieved through varying the information encoder during the adaptation, that is, at the early stage of the adaptation, the stimulus information is mainly encoded in the strong and independent firings of neurons; and as time goes on, the stimulus information is shifted into the weak but concerted responses of neurons. This shift of information encoder can be naturally implemented via STP, a general feature of synapses. ReferencesWark B, Lundstrom B N, Fairhall A. Sensory adaptation. Curr Opin Neurobiol. 2007;17(4):423–9. Christopher deCharms R, Merzenich MM. Primary cortical representation of sounds by the coordination of action-potential timing. Nature. 1996;381:13. Markram H, Wang Y, Tsodyks M. Differential signaling via the same axon of neocortical pyramidal neurons. Proc Natl Acad Sci. 1998;95(9):5323–28. P62 Physical modeling of rule-observant rodent behavior Youngjo Song1, Sol Park1,2, Ilhwan Choi2, Jaeseung Jeong1, Hee-sup Shin2 1Bio and Brain Engineering, KAIST, Daejeon, 34141, Republic of Korea; 2Center for Cognition and Sociality, IBS, Daejeon, 34047, Republic of Korea Correspondence: Jaeseung Jeong - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P62 There is training room for a mouse. In the room, there are two lights which is used as a cue for reward, and there are two reward zone. If a mouse go left reward zone when the left light cue turns on, the mouse gets reward, and if a mouse go right reward zone when the right light cue turns on, the mouse gets reward. The reward is given by brain stimulation from the electrode implanted in MFB. A pair of mice trained individually in the room in order to make them understand the meaning of two light cues. After individual training, the two mice released in the same training room at the same time. In this experiment, 15 out of 19 pairs showed tendency to separate their own reward zone. In order to explain the rodent behavior, we made a computational rodent model. This model is based on Rescorla–Wagner Model. We defined a success rate as probability that a mouse is in the correct reward zone when any cue is given. In each trial, the model mouse learns the left success rate and the right success rate by reinforcement learning. In this model, we assume that success rates for the left cue and right cue are independent and we eliminate social interaction between two model mice. Results The simulation result is given in below figures. Figure 40A is a graph of success rate of a pair of model mice which shows rule between them. You can see the right que success rate of mouse1 and the left que success rate of mouse2 converge to 1, but the left que success rate of mouse1 and the right que success rate of mouse2 converge below 0.4. It means that mouse1 tends to move only when the right cue is given, and mouse2 tends to move only when the left cue is given. Our model mice, however, shows this rule with less probability than actual behavior result. Figure 40B, C 43 % of model mice showed rule, but 79 % of actual mice showed rule in behavior experiment.Fig. 40 A Success rate of two model rats which shows rule between them (blue dots represent the left cue success rate of model rat1, orange plus represent the right cue success rate of model rat1, yellow cross represent the left cue success rate of model rat2, and purple line represent the right cue success rate of model rat2). B Simulation result (600 iterations). C Behavior experiment result (19 pairs) Conclusion If we assume the mouse as a simple independent creature which doesn’t have social element, they show rules with lower probability than the actual mouse in behavior experiment. It means that we can’t regard the mouse as a non-social creature. Therefore, we have to add the social factors such as empathy or cooperation to simulate actual rodent behavior. ReferenceGlimcher PW, Fehr E. Neuroeconomics, 2nd ed. Academic Press. P64 Predictive coding in area V4 and prefrontal cortex explains dynamic discrimination of partially occluded shapes Hannah Choi1,2,3, Anitha Pasupathy2,3, Eric Shea-Brown1,3 1Department of Applied Mathematics, University of Washington, Seattle, WA 98195, USA; 2Department of Biological Structure, University of Washington, Seattle, WA 98195, USA; 3UW Institute for Neuroengineering, University of Washington, Seattle, WA 98195, USA Correspondence: Hannah Choi - hannahch@uw.edu BMC Neuroscience 2016, 17(Suppl 1):P64 The visual system recognizes objects in natural scenes without difficulty, even when most objects are partially occluded. The neural basis of this capacity is unknown. Recent results from primate area V4, an intermediate stage in the shape processing pathway, suggest that feedback from higher cortices may be important for the emergence of V4 shape selective signals [1] when animals are engaged in discriminating partially occluded shapes. Here we implement predictive coding, which has been previously applied to explain responses in early visual areas [2], to investigate possible mechanisms underlying robust discrimination of partially occluded shapes in V4. We propose that higher cortical areas such as prefrontal cortex (PFC) make predictions about V4 activities; when these PFC signals are relayed via feedback to V4, they can reproduce the delayed peak of V4 responses observed in experiments. With a model (Fig. 41A) composed of PFC and V4 units that are selective for different input features, we capture response characteristics of V4 and PFC measured in experiments, by combining feed-forward sensory inputs and feedback predictions to maximize the posterior probability of the responses. We found that inclusion of the feedback predictions results in stronger shape-selective responses across a range of occlusion levels (Fig. 41B), thus maintaining robust discrimination of partially occluded shapes (Fig. 41C).Fig. 41 A Schematic of the V4-PFC network model. B Optimal representation of the shape-selective V4 responses as a function of occlusion level. C Neuronal responses with a noise projected onto the test shape-selective (unit 1)/non-selective (unit 2) V4 response plane, before (top) and after (bottom) the feedback inputs from PFC. Feedback inputs move the responses away from the unity line, improving shape discriminability under occlusion Acknowledgements: This research was supported by the Washington Research Foundation Innovation Postdoctoral Fellowship in Neuroengineering (HC), National Science Foundation CRCNS Grant IIS1309725 (AP), and NEI Grant R01EY018839 (AP). ReferencesKosai Y, El-Shamayleh Y, Fyall AM, Pasupathy A. The role of visual area V4 in the discrimination of partially occluded shapes. J Neurosci. 2014;34:8570–84. Rao RPN, Ballard DH. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neurosci. 1999;2:79–87. P65 Stability of FORCE learning on spiking and rate-based networks Dongsung Huh1, Terrence J. Sejnowski1,2 1The Salk Institute for Biological Studies, La Jolla, CA 92037 USA; 2Division of Biological Sciences, University of California at San Diego, La Jolla, CA 92095 USA Correspondence: Dongsung Huh - huh@salk.edu BMC Neuroscience 2016, 17(Suppl 1):P65 Neurons in the brain often exhibit complex activity patterns, with fluctuations on time scales of several seconds. The generation of complex patterns is critical for directing movements, and is likely to be involved in processing time-varying input (such as speech). However, it is not yet understood how networks of spiking neurons, with time constants of only a few milliseconds, could exhibit such slow dynamics. This should be contrasted with rate-based neural networks, which can be easily trained to generate arbitrary complex activity patterns in a reservoir-based manner [1] by an iterative training method (FORCE learning [2]). So far, however, FORCE learning has not led to successful training of spiking neural networks. Here, we analyze the stability of the networks that result from such learning schemes. For linear rate-based networks, we can analytically predict the full dynamic property of the networks. As the network’s recurrent connectivity reaches the “edge of chaos”, the neuronal activity exhibits a broad distribution of phase, providing appropriate basis for generating the fluctuations. For weaker recurrent connectivity, however, the phase distribution becomes much narrower. In this case, the trained network exhibits highly non-normal structure, which becomes unstable even under small perturbations. Our analysis also illuminates the source of instability in training spiking networks, which is mainly due to the rectified nature of the neuronal output. In numerical simulations, rectified-linear rate networks exhibit narrow phase distribution, even with strong recurrent connectivity near the edge of chaos. Moreover, introducing spiking-dynamics further reduces the width of the distribution, leading to highly unstable network dynamics. Our result reveals the limitation of the reservoir-based approaches, and may lead to more stable, alternative training methods. Acknowledgements: Supported by HHMI. ReferencesJaeger H, Haas H. Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science. 2004;304:78–80. Sussillo D, Abbott LF. Generating coherent patterns of activity from chaotic neural networks. Neuron. 2009;63(4):544–57. P66 Stabilising STDP in striatal neurons for reliable fast state recognition in noisy environments Simon M. Vogt1, Arvind Kumar2,3, Robert Schmidt1,2 1BrainLinks-BrainTools, Cluster of Excellence, University of Freiburg, Germany; 2Faculty of Biology and Bernstein Center Freiburg, University of Freiburg, Germany; 3Department of Computational Biology, Royal Institute of Technology Stockholm, Sweden Correspondence: Simon M. Vogt - simonsunimail@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P66 The brain must be able to quickly identify environmental states based on sensory inputs and select appropriate actions. To obtain neurons that respond selectively to states with short response latencies, a neural plasticity rule is required. Spike timing dependent plasticity (STDP) is known to find the earliest predictors of spatiotemporal spike patterns without supervision. While this feature of STDP is often seen as a hindrance when aiming to reproduce an exact target output spike train, we exploit it to generate short-latency, reliable pattern recognition. There are, however, some difficulties in using STDP in real-world continuous scenarios. In models, presynaptic firing rates are often assumed to be stationary or have constant correlation for simplification, and STDP rule parameters must be closely fitted to form a stable fixed point in the postsynaptic firing rate and normalise postsynaptic activity. Any deviation from these requirements can cause the postsynaptic neuron to become quiet before it is able to form strong selectivity, or exhibit a runaway effect that makes the neuron responsive to a large set of inputs. Soft-bound STDP with a weight-dependent attractor has been suggested as a means for stabilising postsynaptic activity, but this actively hinders separation of spatiotemporal patterns from background noise: as during learning the synaptic weight moves away from the attractor, noisy input becomes more likely to undo these weight changes. Furthermore, activity-dependent scaling also aims to keep the postsynaptic neuron active indefinitely, so neurons may lose any learned selectivity when enduring long periods of background noise. Interpolations between different types of STDP have been suggested, but a problem of balance between premature selectivity and longtime noise robustness remains. We solve this dilemma by including a slow continuous potentiation in our model, which depends on metabolic cost of maintaining strong synapses and slowly vanishes as neurons become selective. It is independent of pre- or postsynaptic activity and can recover silent neurons if a metabolic cost function allows it to. Together with negative-integral STDP as taken from experimental data, it stabilises the activity of untrained neurons for a much wider range of rule parameters and heterogeneous input activity. Our model maintains a unimodal weight distribution while the postsynaptic neuron has not yet become selective, but does not impair the formation of selectivity to spatiotemporal patterns. Selectivity is quickly achieved as soon as patterns are present, even after enduring long periods of noise. Connections that only present noise, represent only other patterns, or present only late parts of a trained pattern become ineffective as the neuron becomes selective, and may be pruned. Any selectivity is hence ensured to represent actual spatiotemporal spike patterns that were at some point present in the postsynaptic neuron’s inputs. This makes the process of training neurons to detect environmental states encoded as spatiotemporal patterns more robust to variations in input statistics and rule parameters, thus easing application in larger-scale networks. Our model of fast pattern detection may apply specifically to the striatum of the basal ganglia where fast reliable decisions need to be made within milliseconds. Unsupervised learning should coexist with rare dopaminergic reinforcement to continuously form new representations of environmental events and decide which of these events are behaviourally important and which can safely be ignored. Acknowledgements: Cluster of Excellence BrainLinks-BrainTools funded by German Research Foundation (DFG, Grant Number EXC 1086). P67 Electrodiffusion in one- and two-compartment neuron models for characterizing cellular effects of electrical stimulation Stephen Van Wert1, Steven J. Schiff1,2 1Center for Neural Engineering, Department of Engineering Science and Mechanics, The Pennsylvania State University, University Park, PA 16802, USA; 2Departments of Neurosurgery and Physics, The Pennsylvania State University, University Park, PA 16802, USA Correspondence: Stephen Van Wert - szv124@psu.edu BMC Neuroscience 2016, 17(Suppl 1):P67 Standard approaches for modeling the neuronal effects of electrical fields and currents (such as [1, 2]) apply transmembrane current to Hodgkin–Huxley membrane patches without regard to ion fluxes and conservation of ions inside and outside the cell. We propose cellular models that reflect polarization and preserve the biophysics of the spaces the neurons are embedded within. By including ion fluxes and maintaining conservation of mass and charge, the gradients of ionic concentrations both within and outside of the neuron can be accounted for. This requires characterizing the ionic fluxes with electrodiffusion, such that ionic charge gradients as well as ionic concentration gradients drive flux. This electrodiffusion mechanism, derived from Nernst-Planck flux equation, not only allows for more accurate modeling of physiological and pathophysiological conditions with substantial ionic and volume changes, but it also provides a means to model application of electrical stimulation. The model being developed here builds on recent one-compartment model development in [3] that extends the Hodgkin–Huxley formalism in several distinct ways, most notably in using conservation to track all ion fluxes and volume changes to determine the extra- and intra-cellular concentrations. The proposed model also extends the model in [3] to a two compartment model which allows for simulation of neuronal polarization with control applied in the direction of a soma-dendritic axis. We first characterize the resultant model dynamics in the absence of any control stimulus and compare these to the dynamics seen with standard diffusion. In particular, the dynamics are characterized through trajectories of dynamically evolving variables, with a focus on bifurcation structure at points where the dynamics transition from different states such as normal firing, seizure, or spreading depression. We then simulate the effects of applying excitatory or inhibitory control on these dynamics and optimize the dynamics of the neuron to be consistent with experimental evidence. Such a model gives very different results from the customary approach to modeling the effects of electrical stimulation, where the stimulation is applied internally or externally to a neuron without taking the nature of the charge carriers present into account. There are a variety of effects of excitatory and inhibitory stimulation observed now that could not be possible before, and we can describe the trajectories of the effects of such stimulation in ways that shed light on multiple experimental scenarios. This type of model development offers the ability to understand a wide variety of previously unexplained experimental observations for both excitatory and inhibitory stimulation. Doing so offers a platform for us to study electrical feedback control of neuronal systems and to offer model-based control strategies for pathological dynamics such as seizures and spreading depression. ReferencesPark E-H, Barreto E, Gluckman BJ, Schiff SJ, So P. A model of the effects of applied electric fields on neuronal synchronization. J Comput Neurosci. 2005;19: 53–70. Berzhanskaya J, Chernyy N, Gluckman BJ, Schiff SJ, Ascoli GA. Modulation of hippocampal rhythms by subthreshold electric fields and network topology. J Comput Neurosci. 2013;34(3):369–89. Wei Y, Ullah G, Schiff SJ. Unification of neuronal spikes, seizures, and spreading depression. J Neurosci. 2014;34(35):11733–43. P68 STDP improves speech recognition capabilities in spiking recurrent circuits parameterized via differential evolution Markov chain Monte Carlo Richard Veale1, Matthias Scheutz2 1National Institute for Physiological Sciences, Okazaki, Aichi, Japan; 2Department of Computer Science, Tufts University, Medford, MA, USA Correspondence: Richard Veale - richard@nips.ac.jp BMC Neuroscience 2016, 17(Suppl 1):P68 A major issue in using spiking neural circuits for pragmatic tasks such as speech recognition is how to parameterize them. Here, we apply a hybrid Differential Evolution/Markov Chain Monte Carlo (DE/MCMC) [1, 2] approach to estimate optimal parameters for a spiking neural circuit that is used for real time speech recognition [3] from raw auditory input using PSWEEP2 (rveale.com/software.php). To avoid the expensive training step, we use a surrogate measure of word recognition performance. Specifically, we maximize average within-word similarity in the neural circuit’s state space trajectory, while simultaneously minimizing between-word similarity. We executed the algorithm for 7000 generations (48 h of runtime) using 2016 cores of the super computer Big Red II at Indiana university. The average fitness increases significantly with successive generations. Panel a is a visualization of the first 3 principle components of the state space for the exemplars of each different word category (shown as different colors). The different categories move to take more distant trajectories through the state space with successive generations. We verify that the state-space separation is a good surrogate measure of word recognition performance by taking the set of circuits with the highest fitness from the first, middle, and last 100 generations and training readout neurons for the 7-word corpus. Word recognition performance increases from 19 to 71 to 85 % (Fig. 42).Fig. 42 A Fitness evolution generation 0–4000 (first 3 principle components). Each color is a different word class, each line a different utterance token of the word. B Change in state space trajectory from STDP adaptation Finally, we evaluated the performance benefit of adaptation to sensory stimuli via synaptic plasticity mechanisms, to make pragmatic use of our previous work investigating auditory habituation [4]. We take the most performant parameter points that were found during the parameter sweep, and test their fitness before and after exposure to 100 presentations of the word stimuli while a nearest-neighbor temporally asymmetric Hebbian plasticity model of spike timing dependent plasticity (STDP) is implemented in all excitatory synapses. Although sensitive to STDP model parameters, Panel B shows that word recognition performance can be improved by as much as 8 % by familiarizing a neural circuit to the type of sensory stimulus that it will be used to compute. This follows previous findings by Triesch et al. [5], who reported similar effects in non-spiking neural networks. ReferencesVeale R, Isa T, Yoshida M. Applying differential evolution MCMC to parameterize large-scale spiking neural simulations. In: IEEE conference on evolutionary computation (CEC). IEEE. p. 1620–27. Laloy E, Vrugt JA. High-dimensional posterior exploration of hydrologic models using multiple-try dream (zs) and high-performance computing, Water Resour Res. 2012;48(1). Veale R, Scheutz M. Neural circuits for any-time phrase recognition. In: Proceedings of the 34th annual conference of the Cognitive Science Society; 2012. p. 1072–7. Veale R, Scheutz M. Auditory habituation via spike-timing dependent. In: Proceedings of the international conference on development and learning and epigenetic robotics (ICDL), San Diego, CA; 2012. Lazar A, Pipa G, Triesch J. Fading memory and time series prediction in recurrent networks with different forms of plasticity. Neural Netw. 2007;20(3):312–22. P69 Bidirectional transformation between dominant cortical neural activities and phase difference distributions Sang Wan Lee1,2,3 1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea; 2Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea; 3KAIST Institute for Health Science and Technology, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea Correspondence: Sang Wan Lee - sangwan@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P69 The brain is a complex nonlinear dynamic system comprising multiple different types of subsystems. Each subsystem encodes different types of information, and its states are context-dependent. Unlike sensory or motor processing occurring at relatively early or terminal stages of functional hierarchy, cognitive processes, including learning, inference, and top-down attention, require interactions between brain’s multiple subsystems. The associated neural dynamics inevitably leave an imprint on neural activity patterns over a wide areas of cortex. Considerable progress has been made toward understanding such functional network dynamics. This includes the causal connectivity [1], its extension to distinguish causality from correlation within nonseparable weakly connected dynamic systems [2], and the integrated information theory to quantify the effect of a neuronal network connectivity on the increase in the amount of information above and beyond the capability of a single locally connected network [3, 4]. However, none of these methods is applicable to real-time analyses when the network size is large. Here I develop a simple and efficient computational framework for analyzing cortical dynamics both in time and space, arising from complex interactions between brain’s multiple subsystems. Accommodating the fact that both a covariance and a gram matrix can be computed by using a combination of a certain idempotent matrix with a data matrix, I derive a set of matrix operators F, with which one set of eigenvectors associated with a covariance matrix of data and of mean-corrected data can be transposed to another set of eigenvectors associated with a gram matrix, and vice versa (see Fig. 43). For example, suppose that a d-by-n data matrix is a set of time-series data recorded from multiple locations of cortices where d and n refers to the number of electrodes and time points, respectively, and d≪n. One can then perform a singular value decomposition (SVD) for the d-by-d covariance or gram matrix to obtain an associated eigenvector set, followed by applying matrix operator F′ to the eigenvector set to convert it to the eigenvector set of their counterparts. There is no need to perform SVD for n-by-n matrix whose computational load is high. If n≪d, then one could simply start with performing the SVD for the n-by-n matrix, followed by applying the matrix operator F to its eigenvector set. It is noted that the acquired d-by-1k eigenvectors and n-by-1k eigenvectors correspond to dominant cortical neural activities and phase difference distributions, respectively.Fig. 43 Computational framework for analyzing space–time cortical dynamics. T is an idempotent projection matrix Conclusion An efficient computational framework for analyzing cortical dynamics both in time and space is proposed, taking into account the relationship between a covariance and a gram matrix. For analyzing neural data acquired from multiple locations of cortices, the framework replaces the SVD with a simple matrix operator F so as to reduce a heavy computational load of performing SVD on large-size data matrices. In doing so, it allows efficient bidirectional transformation between dominant neural activities and phase difference distributions. Acknowledgements: I thank Barclay Lee and Dae-Hyun Kim for their assistance. This work was supported by the research fund of the KAIST (Korea Advanced Institute of Science and Technology) (Grant code: G04150045). References Seth AK. Causal connectivity of evolved neural networks during behavior. Network. 2005;16:35–54. Sugihara G, May R, Ye H, Hsieh C, Deyle E, Fogarty M, Munch S. Detecting causality in complex ecosystems. Science. 2012;338(6106):496–500. Balduzzi D, Tononi G. Qualia: the geometry of integrated information. PLoS Comput Biol. 2009;5(8):e1000462. Edlund JA, Chaumont N, Hintze A, Koch C, Tononi G, Adami C. Integrated information increases with fitness in the evolution of animats. PLoS Comput Biol. 2011;7(10):e1002236. P70 Maturation of sensory networks through homeostatic structural plasticity Júlia Gallinaro1, Stefan Rotter1 1Bernstein Center Freiburg and Faculty of Biology, University of Freiburg, Freiburg, Baden-Württember, 79194, Germany Correspondence: Simon M. Vogt - julia.gallinaro@bcf.uni-freiburg.de BMC Neuroscience 2016, 17(Suppl 1):P70 Neurons in the adult visual cortex of mice prefer to make synapses with neurons responding to similar visual features. As such a bias in connectivity is not observed at the time of eye opening, it has been proposed that the functional subnetworks are formed through rewiring of recurrent synaptic connections, induced by visual experience [1]. However, it is not clear according to which rules this structure develops. The emergence of feature specific wiring was recently demonstrated in a balanced network model with appropriate rules of functional synaptic plasticity [2]. In this model, however, connectivity was evaluated based on the strength of already existing synapses, and the structure of the network remained unchanged throughout the simulation. Referring to recent findings of homeostatic regulation of cortical activity in rodent visual cortex in vivo [3], we employ here a structural plasticity rule based on firing rate homeostasis described previously [4] for simulating network restructuring during sensory stimulation. We show that, next to other biologically meaningful properties, feature specific connectivity also emerges in a balanced network of changing structure (see Fig. 44), using a plasticity rule that does not depend on spike timing.Fig. 44 Network connectivity before and after sensory stimulation. A, B. Connectivity matrix, pre- and post-synaptic neurons are sorted according to their preferred orientation (PO) and subdivided into groups. C, D. Mean output connectivity plotted against the difference between pre and post PO Acknowledgements: Supported by the Erasmus Mundus Joint Doctoral program EuroSPIN, the German Federal Ministry of Education and Research, grant 01GQ0830, and the state of Baden-Württemberg through bwHPC. ReferencesKo H, Cossell L, Baragli C, Antolik J, Clopath C, Hofer SB, Mrsic-Flogel TD: The emergence of functional microcircuits in visual cortex. Nature. 2013;496:96–100. Sadeh S, Clopath C, Rotter S. Emergence of functionalspecificity in balanced networks with synapticplasticity. PLoS Comput Biol. 2015;11:e1004307. Bishop HI, Zito K. The downs and ups of sensorydeprivation: evidence for firing rate homeostasis in vivo. Neuron. 2013;80:247–9. van Ooyen A: Using theoreticalmodels to analyseneuraldevelopment. Nat Rev Neurosci. 2011;12:311–26. P71 Corticothalamic dynamics: structure, number of solutions and stability of steady-state solutions in the space of synaptic couplings Paula Sanz-Leon1,2, Peter A Robinson1,2 1School of Physics, University of Sydney, New South Wales, Australia; 2Center for Integrative Brain Function, University of Sydney, New South Wales, Australia Correspondence: Paula Sanz-Leon - paula.sanz-leon@sydney.edu.au BMC Neuroscience 2016, 17(Suppl 1):P71 The interconnections of a model of the corticothalamic system [1] define an 8-dimensional parameter space where specific combinations of dimensions correspond to one of the three loops of the system (e.g., intracortical, corticothalamic and intrathalamic). The form of the steady-state equation of the corticothalamic system imposes an odd number of solutions, which in terms of dynamics correspond to fixed points of the system. Here, the structure of regions with different number of solutions is systematically investigated within physiologically valid ranges of synaptic couplings representing different brain states [2, 3]. For instance, Fig. 45A, Bdisplay the regions where the steady state equation has one, three or five solutions for two 3-dimensional subsets of the full space. These results show how small changes in the connectivity can cause additional roots of the steady state equation to appear or vanish. More importantly, they illustrate the effect of intracortical feedback: for more than one solution to exist the total intracortical feedback needs to be negative (inhibitory). The occurrence of multiple roots happens for parameter values that characterize normal arousal states [3], indicating that the approach presented here has a potential to (i) quantify and predict the existence of additional (abnormal) arousal states and (ii) categorize subtle differences in states such as anesthesia, coma [4].Fig. 45 Three dimensional subsets of the 8D corticothalmic coupling space. A, B Regions with 1, 3 or 5 roots are enclosed by surfaces (blue, violet and yellow respectively). The sharp transition between zones of 1–3 roots along the vee axis (excitatory intracortical feedback) indicates the plane at which the total intracortical feedback (vee + vei) changes sign. The difference between A, B is the value of vei (inhibitory intracortical feedback). In a the probability of having multiple roots is lower than in B ReferencesRobinson PA, Rennie CJ, Wright JJ, Bourke PD. Steady states and global dynamics of electrical activity in the cerebral cortex. Phys Rev E. 1998;58:3557–71. Robinson PA, Rennie CJ, Wright JJ, Bahramali H, Gordon E, Rowe DL. Prediction of electroencephalographic spectra from neurophysiology. Phys Rev E. 2001; 63:021903. Abeysuriya RG, Rennie CJ, Robinson PA. Physiologically based arousal state estimation and dynamics. J Neurosci Methods. 2015;253:55–69. Steyn-Ross ML, Steyn-Ross DA, Sleigh JW, Wilcocks LC. Toward a theory of the general-anesthetic-induced phase transition of the cerebral cortex. I. A thermodynamics analogy. Phys Rev E. 2001;64:942–53. P72 Optogenetic versus electrical stimulation of the parkinsonian basal ganglia. Computational study Leonid L. Rubchinsky1,2, Chung Ching Cheung1, Shivakeshavan Ratnadurai-Giridharan1 1Department of Mathematical Sciences, Indiana University-Purdue University Indianapolis, Indianapolis, IN, USA; 2Stark Neurosciences Research Institute, Indiana University School of Medicine, Indianapolis, IN, USA Correspondence: Leonid L. Rubchinsky - lrubchin@iupui.edu BMC Neuroscience 2016, 17(Suppl 1):P72 Deep brain stimulation (DBS) is used as a therapeutic procedure to treat symptoms of several neurological and neuropsychiatric disorders. In particular it is used to treat motor symptoms of Parkinson’s disease (PD) by delivering high-frequency regular stimulation to subcortical targets. Hypokinetic symptoms of PD are associated with excessive oscillatory synchronized activity in the beta frequency band, and effective DBS is believed to suppress it. An alternative way to stimulate neural circuits is an emerging technology of optogenetics. It is an experimental technique and it is not clear if it eventually will be possible to implement it in clinical practice. However it is used as an experimental tool, and maybe, in time, it will be developed into safe therapeutic technique. The goal of his study is to explore how effective an optogenetic stimulation in comparison with electrical stimulation in their network effects on elevated synchronized oscillatory activity. We use a model for the basal ganglia activity [1], which was developed to reproduce experimentally observed beta-band activity patterns [2]. We introduce electrical stimulation as well as optogenetic stimulation of two types: excitatory via channelrhodopsin and inhibitory via halorodopsin. We explore the effect of different stimulation types on oscillatory synchronized dynamics and consider the efficacy of stimulation for different kind of network’s dynamics. All three modes of stimulation can decrease beta synchrony that is commonly associated with hypokinetic symptoms of Parkinson’s disease. Generally speaking, growing intensity of stimulation leads to larger suppression of the beta-band synchronized oscillatory activity. But the actions of different stimulation types on the beta activity may differ from each other. Electrical DBS and optogenetic excitation have somewhat similar effects on the network. Both of these stimulation types cause desynchronization and suppression of the beta-band bursting. As intensity of stimulation is growing, they synchronize the network at higher (non-beta) frequencies in a close to tonic spiking dynamics. Optogenetic inhibition effectively reduces spiking and bursting activity of the targeted neurons. We compare the stimulation modes in terms of the minimal effective current delivered to basal ganglia neurons in order to suppress beta activity below a threshold: the less stimulation current is needed to suppress the activity, the more efficacious stimulation is. We found that optogenetic inhibition usually requires less effective current than electrical DBS to achieve beta suppression. Optogenetic excitation, while as not efficacious as optogenetic inhibition, still usually requires less effective current than electrical DBS to suppress beta activity. Thus our results suggest that optogenetic stimulation may introduce less of effective currents to a neuron than conventional electrical DBS, but still achieve sufficient beta activity suppression. Optogenetics is presently not used in humans. However, it was implemented in the basal ganglia of non-human primates [3]. So we suppose our results may motivate further research into applicability of optogenetic technologies in humans. Optogenetic stimulation is also used as a research tool. Our results suggest that it may be more effective than electrical stimulation in control of synchronized oscillatory activity, because it does its job with less current injected into the neurons. Acknowledgements: The study was supported by ICTSI and the Indiana University Health – Indiana University School of Medicine Strategic Research Initiative. ReferencesPark C, Worth RM, Rubchinsky LL. Neural activity in Parkinsonian brain: the boundary between synchronized and nonsynchronized dynamics. Phys Rev E. 2011;83:042901. Park C, Worth RM, Rubchinsky LL. Fine temporal structure of beta oscillations synchronization in subthalamic nucleus in Parkinson’s disease. J Neurophysiol. 2010;103:2707–16. Galvan A, Hu X, Smith Y, Wichmann T. In vivo optogenetic control of striatal and thalamic neurons in non-human primates. PLoS One. 2012;7(11):e50808. P73 Exact spike-timing distribution reveals higher-order interactions of neurons Safura Rashid Shomali1, Majid Nili Ahmadabadi1,2, Hideaki Shimazaki3, S Nader Rasuli4,5 1School of Cognitive Sciences, Institute for Research in Fundamental Sciences (IPM), Tehran, 19395-5746, Iran; 2School of ECE, College of Engineering, University of Tehran, Tehran, 14155-6619, Iran; 3RIKEN Brain Science Institute, Wako, Saitama, 351-0198, Japan; 4Department of Physics, University of Guilan, Rasht, 41335-1914, Iran; 5School of Physics, Institute for Research in Fundamental Sciences (IPM), Tehran, 19395-5531, Iran Correspondence: Safura Rashid Shomali - safura@ipm.ir BMC Neuroscience 2016, 17(Suppl 1):P73 It has been suggested that variability in spike patterns of individual neuron is largely due to noisy fluctuations caused by asynchronous synaptic inputs balanced near the threshold regime [1–3]. In this regime, small fluctuations in synaptic inputs to a neuron do cause output spikes; because the membrane potential is maintained below but close enough to the threshold potential. To successfully transfer signals under such noisy conditions, it is proposed that a few relatively stronger synapses and/or an assembly of nearly synchronous ones form “signaling inputs” [4]. Thus one fundamental question is how such relatively strong signaling input modifies the spiking activity of a post-synaptic neuron which receives noisy background inputs balanced near the threshold regime. Nonetheless, analytical studies on the effect of the signaling input under such conditions are scarce even with the popular leaky integrate-and-fire (LIF) neuron model. Here we analytically study the impact of a specified signaling input on spike timing of the postsynaptic LIF neuron which receives noisy inputs at the threshold regime. To this end, we first revisit Fokker–Planck analysis of a first spike-timing distribution when the LIF neuron receives noisy synaptic inputs, but no signaling input, at the threshold regime. We then perform perturbation analysis to investigate how a signaling input modifies this first spike-timing distribution. Fortunately, we could solve all terms of perturbation analytically and find the exact first spike-timing distribution of the postsynaptic neuron; it is applicable to not only excitatory but also inhibitory input. This analytical solution allows us to describe the statistics of output spiking activity as a function of background noise, membrane dynamics, and signaling input’s timing and amplitude. The proposed analysis of signaling input provides a powerful framework for studying information transmission, neural correlation, and timing-dependent synaptic plasticity. Among them, we investigate the impact of common signaling inputs on population activities of postsynaptic neurons. Using mixture models based on our analytical first spike-timing distribution, we calculate the higher-order interactions [5] of postsynaptic neurons in different network architectures. Comparing these results with higher-order interactions, measured from experimental data in monkey V1 [6], we try to answer whether one can reveal network architecture, responsible for the ubiquitously observed sparse activities. ReferencesVreeswijk C, Sompolinsky H. Chaos in neuronal networks with balanced excitatory and inhibitory activity. Science. 1996;274: 1724–6. Renart A, De La Rocha J, Bartho P, Hollender L, Parga N, Reyes A, Harris KD. The asynchronous state in cortical circuits. Science. 2010;327:587–90. Tan AY, Chen Y, Scholl B, Seidemann E, Priebe NJ. Sensory stimulation shifts visual cortex from synchronous to asynchronous states. Nature. 2014;509:226–9. Teramae Jn, Tsubo Y, Fukai T. Optimal spike-based communication in excitable networks with strong-sparse and weak-dense links. Sci Rep. 2:485. Nakahara H, Amari S. Information-geometric measure for neural spikes. Neural Comput. 2002;14:2269–316. Ohiorhenuan IE, Victor JD. Information-geometric measure of 3-neuron firing patterns characterizes scale dependence in cortical networks. J Comput Neurosci. 2011;30:125–41. P74 Neural mechanism of visual perceptual learning using a multi-layered neural network Xiaochen Zhao1, Malte J. Rasch1 1State Key Lab of Cognitive Neuroscience and Learning, IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing 100875, China Correspondence: Malte J. Rasch - malte.rasch@bnu.edu.cn BMC Neuroscience 2016, 17(Suppl 1):P74 Recently, a study [1] has found by recording the activities of neurons in monkeys performing a contour detection task that the response properties of the primary visual cortex (V1) change continuously during perceptual learning. In particular, the figure-background contrast was continuously enhanced in the course of learning. However, the exact neural circuit mechanisms that causes the V1 responses to change during perceptual learning remain unclear. In order to understand how the underlying neural network needs to change, we here train a multi-layered neural network model to perform the contour detection task on the very same visual stimuli as in the experiments and investigate the network’s performance and the resulting synaptic weight structure. In this study, we first model the V1 representation of each visual stimulus by using a non-classical receptive field model (NCRF) which takes into account orientation selective inhibition [3]. We further assume that the higher visual areas (up to a decision unit) are hierarchically structured, read out the V1 activity, and learn to change their synaptic weights to optimally perform the contour detection task. AGREL (attention-gated reinforcement learning) algorithm [2], which considers feedback connections and biologically plausible local synaptic adjustments, is applied to train the network (see Fig. 46). We found that the multi-layered model trained with AGREL could replicate the behavioral performance increase in a contour detection task as observed in experiments. Moreover, learning the network model structure showed enhanced synaptic weights in the region of the detected contour. It further demonstrated that “predictive” feedback signals from higher layers facilitate the responses of V1 neurons to the contour and thus increased the figure-background contrast in V1 with improved behavioral performance. The results suggest that the experimental observed V1 response facilitation could be caused by selective synaptic strengthening of feed-forward and feed-back pathways.Fig. 46 V1 representations of the stimuli and simulation results of the model. A V1 representations of the stimuli by using NCRF model. B Model performance increases with perceptual training. C Synaptic weight changes after perceptual learning ReferencesYan Y, Rasch MJ, Chen M, et al. Perceptual training continuously refines neuronal population codes in primary visual cortex. Nat Neurosci. 2014;17(10):1380–7. Roelfsema PR, van Ooyen A. Attention-gated reinforcement learning of internal representations for classification. Neural Comput. 2005;17(10):2176–214. Zeng C, Li Y, Yang K, et al. Contour detection based on a non-classical receptive field model with butterfly-shaped inhibition subregions. Neurocomputing. 2011;74(10):1527–34. P75 Inferring collective spiking dynamics from mostly unobserved systems Jens Wilting1, Viola Priesemann1,2 1Max-Planck-Institute for Dynamics and Self-Organization, D-37077 Göttingen, Germany; 2Bernstein Center for Computational Neuroscience, University of Göttingen, D-37075 Göttingen, Germany Correspondence: Jens Wilting - jwilting@nld.ds.mpg.de BMC Neuroscience 2016, 17(Suppl 1):P75 What can we know about a high-dimensional dynamical system if we can only observe a very small part of it? This problem of spatial subsampling is common to almost every area of research where spatially extended, time evolving systems are investigated, and is particularly severe when assessing population spiking dynamics in neuroscience. Previous studies have shown that subsampling can lead to spurious results when assessing the dynamical state of spiking activity, in particular when discriminating whether neural networks operates at criticality [1, 2]. Here we present further insight why the distance to criticality is systematically overestimated, and introduce a novel estimator which for the first time allows to correctly infer the distance to criticality even under strong subsampling. Neuronal systems have been proposed to operate close to criticality, because in models criticality maximizes information processing capacities [e.g. 3]. Indeed, power-law distributions of the avalanche size, an indication of criticality, have been found for local field potentials from in vitro systems [1] to humans in vivo [4]. However, for neuronal systems criticality also comes with the risk of spontaneous runaway activity, which may lead to pathological states like epilepsy. Experiments indeed indicate that spiking activity in rats, cats, and monkeys is in a sub-critical regime, thereby keeping a safety-margin from criticality [5]. Quantifying the precise distance to criticality may help to shed light on how the brain maximizes its information processing capacities without risking runaway activity. In neural systems, critical dynamics is typically compared to dynamics from models that resemble branching processes [1]. Their dynamics are controlled by a single parameter, the expected number σ of postsynaptic spikes generated by one individual spike, showing either stationary dynamics (sub-critical, σ < 1) or transient growth (super-critical, σ > 1). For σ = 1 branching processes are critical and produce heavy tailed avalanche size distributions. We used a driven branching process, which allows to exactly match the model neuron firing rate to that observed in experiments for any σ. We propose a stochastic representation of subsampling and show that under subsampling established approaches to inferring σ are substantially biased. We derived a novel approach based on multistep regression [6], which for the first time allows to quantify the distance to criticality even under strong subsampling. Our method generalizes to auto-regressive processes with both additive and multiplicative noise, making it widely applicable in diverse fields of research. We validate our method by applying subsampling to simulated branching networks with invasion, and also to a network of integrate-and-fire neurons. We applied this method to spike recordings from awake macaque monkeys prefrontal cortex, cat visual cortex, and rat hippocampus. We found that neuronal population activity operates close to criticality, but in a subcritical regime with 0.94 < σ < 0.995. These results point at a novel universal organization principle: spiking dynamics in vivo is in a subcritical regime which does not yield maximum, but sufficient information processing capacity, and at the same time keeps a safety-margin from unstable supercritical states. ReferencesBeggs J, Plenz D. Neuronal avalanches in neocortical circuits. J Neurosci. 2003;23(35):11167–77. Priesemann V, Munk MHJ, Wibral M. Subsampling effects in neuronal avalanche distributions recorded in vivo. BMC Neurosci. 2009;10:40. Boedecker J, Obst O, Lizier JT, Mayer NM, Asada M. Information processing in echo state networks at the edge of chaos. Theory Biosci. 2012;131:205–13. Priesemann V, Valerrame M, Wibral M, Le Van Quyen M. Neuronal avalanches differ from wakefulness to deep sleep—evidence from intracranial depth recordings in humans. PloS Comput Biol. 2013;9(3):e1002985. Arviv O, Goldstein A, Shriki O. Near-critical dynamics in stimulus-evoked activity of the human brain and its relation to spontaneous resting-state activity. J Neurosci. 2015;35(41):13927–42. Priesemann V, et al. Spike avalanches in vivo suggest a driven, slightly subcritical brain state. Front Syst Neurosci. 2014;8:108. Wilting J, Priesemann V. Quantifying the distance to criticality under subsampling. BMC Neurosci. 2015;16:O3. P76 How to infer distributions in the brain from subsampled observations Anna Levina1, Viola Priesemann2 1IST Austria, Klosterneuburg, 3400, Austria; 2BCCN & MPI for Dynamics and Self-Organization, Göttingen, 37077, Germany Correspondence: Anna Levina - anna.levina@ist.ac.at BMC Neuroscience 2016, 17(Suppl 1):P76 Inferring the dynamics of a system from observations is a challenge, even if one can observe all system units or components. The same task becomes even more challenging if one can sample only a small fraction of the units at a time. As a prominent example, spiking activity in the brain can be accessed only for a very small fraction of all neurons in parallel. These limitations do not affect our ability to infer single neuron properties, but it influences our understanding of the global network dynamics or connectivity: Subsampling can hamper inferring whether a system shows scale-free topology or scale-free dynamics (criticality) [1, 2]. Criticality is a dynamical state that maximizes information processing capacity in models, and therefore is a favorable candidate state for brain function. Experimental approaches to test for criticality extract spatio-temporal clusters of spiking activity, called avalanches, and test whether they followed power laws. Avalanches can propagate over the entire system, thus observations are strongly affected by subsampling. We developed a formal ansatz to infer avalanche distributions in the full system from spatial subsampling using both analytical and numerical approaches. In the mathematical model subsampling from exponential distribution does not change the class of distribution, but only its parameters. In contrast, power law distributions, despite their alias “scale-free”, do not manifest as power laws under subsampling [2]. We study changes in distributions to derive “subsampling scaling” that allows to extrapolate the results from subsampling to a full system: P(s) = psubPsub(s/psub) where P(s) is the original distribution, Psub is the one under subsampling, and psub=NM is the probability to sample a unit, N—number of sampled units, M—system size. In the model with critical avalanches, subsampling scaling collapses distributions for any N (Fig. 47B). However, for subcritical models, no distribution collapse is observed (Fig. 47D). Thus we demonstrate that subsampling scaling allows to distinguish critical from non-critical systems. With the help of this novel method we studied dissociated cortical cultures. For these we artificially subsampled recordings by considering only fraction of all 60 electrodes. We find that in the first days subsampling scaling does not collapse distributions well, whereas mature cultures (~from day 21) allow for a good collapse, indicating development toward criticality (Fig. 47C, E).Fig. 47 Subsampling scaling in model and experiment. Left branching process model; right: experiments on developing cultures A Avalanche size counts f(s) from the full and the subsampled critical model; N: number of sampled neurons. B Under subsampling scaling, all f(s) collapse. C Collapse of subsampled avalanche-size distribution from the culture at the age of 21 days. D For subcritical models, the same scaling ansatz does not result in a collapse. E No collapse of f(s) from the culture at age 7 days Acknowledgements: AL received funding from the Marie Curie Actions (FP7/2007–2013) under REA Grant Agreement No. (291734). VP received funding from BMBF Bernstein 01GQ1005B. ReferencesStumpf, MPH, Wiuf C, May RM. Subnets of scale-free networks are not scale-free: sampling properties of networks. PNAS 2005;102(12):4221–24. Priesemann V, Munk MH, Wibral M. Subsampling effects in neuronal avalanche distributions recorded in vivo. BMC Neurosci. 2009;10(1):40. Uhlig M, Levina A, Geisel T, Herrmann JM. Critical dynamics in associative memory networks. Front Comp Neurosci. 2013;7. P77 Influences of embedding and estimation strategies on the inferred memory of single spiking neurons Lucas Rudelt1, Joseph T. Lizier2, Viola Priesemann1 1Department of Non-linear Dynamics, Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany; 2School of Civil Engineering, The University of Sydney, Sydney, NSW, Australia Correspondence: Lucas Rudelt - l.rudelt@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P77 Information theory provides a generic framework for studying statistical dependencies, and is widely used in neuroscience. However, a correct estimation of the involved quantities can be challenging. For example, a correct estimation of transfer entropy (TE), TE(X → Y) = I(X−, Y|Y−), or active information storage (AIS), AIS(X) = I(X−, X), requires past state variables X− and Y− that encode all information about the past that is relevant when predicting X or Y [1]. For a spiking neuron, states can be defined by transforming the spike train into a binary sequence of spike counts in sufficiently small, equally spaced bins with some bin size Δt (Fig. 48A). For neurons, however, it is unclear how many past bins a sufficient state variable typically comprises. In practice, past states have often been limited to only one time bin to reduce the complexity of estimation. This points at the main challenge one faces when estimating TE and AIS: A reliable estimation of probabilities from recorded data becomes more and more difficult with increasing complexity of the state variables, i.e. with considering more past bins. We used AIS for single spiking neurons to estimate (a) how much memory there is, (b) how long it reaches typically into the past, (c) the (non-)linear contributions of the memory. To this end, we first examined the performance of different estimators for a realistic model neuron [2] whose AIS can be directly computed (dashed line, Fig. 48B). Using constant external drive we simulated a recording of 12 h. In a model-free approach, probabilities were directly estimated from relative frequencies using the standard ‘plugin’ estimator or the ‘NSB’ estimator [3]. In addition, we fitted a generalized linear model (GLM) whose predictions constitute an estimator that is constrained to linear contributions. For all these estimation strategies, the number of past bins k and thus the time range was systematically varied (Fig. 48B). We then applied the same estimators to in vitro and in vivo recordings (Fig. 48c, d) of 3 h and 1 h duration.Fig. 48 Relative active information storage as a function of the time range of the past state for different estimators Considering a very small number k of past bins can lead to a substantial underestimation of AIS. Increasing the number of past bins, however, leads to severe positive bias of the model-free estimators when the complexity of the past state becomes too large. This manifests in the estimators exceeding the true AIS (Fig. 48B). Assuming a point process with linear contributions (GLM), in contrast, allows a robust estimation but does not capture non-linear effects. This approach can also be used to take the global past activity of the neural population into account, thereby unveiling redundancies in the activity of the single neuron with the population activity. While the model neuron intrinsically has only linear dependencies on its past, it is surprising that, in vitro, there also seem to be very little non-linear contributions and a lot of redundancy. In vivo, the non-linear contributions are more prominent and the memory is clearly non-redundant. We thus showed that appropriate embedding is necessary, otherwise AIS is underestimated and likewise, TE might be mis-estimated. Furthermore, our results suggest that in vivo the information processing is more evolved. ReferencesWibral M, Lizier J, Priesemann V. Bits from brains for biologically-inspired computing. Comput Intell. 2015;2:5. Pozzorini, C, Naud R, Mensi S, Gerstner W. Temporal whitening by power-law adaptation in neocortical neurons. Nat Neurosci. 2013;16(7):9428. Nemenman I, Shafee F, Bialek W. Entropy and inference, revisited. Adv Neural Inf Process Syst. 2002;14:471–478. P78 A nearest-neighbours based estimator for transfer entropy between spike trains Joseph T. Lizier1, Richard E. Spinney1, Mikail Rubinov2,3, Michael Wibral4, Viola Priesemann5,6 1Complex Systems Research Group, Faculty of Engineering & IT, The University of Sydney, NSW 2006, Australia; 2Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, VA 20147, USA; 3Department of Psychiatry, University of Cambridge, Cambridge, UK; 4MEG Unit, Brain Imaging Center, Goethe University, 60528 Frankfurt am Main, Germany; 5Department of Nonlinear Dynamics, Max Planck Institute for Dynamics & Self-Organization, Göttingen, Germany; 6Bernstein Center for Computational Neuroscience, Göttingen, Germany Correspondence: Joseph T. Lizier - joseph.lizier@sydney.edu.au BMC Neuroscience 2016, 17(Suppl 1):P78 The nature of a directed relationship (or lack thereof) between brain areas is a fundamental topic of inquiry in computational neuroscience [1]. A particular focus of such inquiry is in regards to the analysis of information flows in a network [2], and such investigations take place at all levels of analysis, from interregional connectivity in fMRI imaging data [3] down to directed relationships between spike trains at the neuronal level [4]. In all of the aforementioned studies, information theory provides the primary tool, transfer entropy (TE) [5], for analysis of such directed relationships. TE measures the predictive gain about state transitions in a target time-series from observing some source time-series. While the TE has been used extensively to analyse recordings from fMRI, MEG and EEG for example [1–3], fewer applications [4] have been made to spiking time-series. Although one can apply temporal binning on such time-series before measuring TE on the resultant binary time-series [4], it remains unclear: (a) how to set parameters for this approach (e.g. bin sizes), (b) whether an estimate can be achieved by avoiding temporal binning and instead working directly with continuous-valued time stamps of spikes, and (c) whether such an estimate would actually improve on binning approaches. Recent theoretical developments have pointed to how transfer entropy may be derived from continuous-valued time-stamps of spikes directly, using spike rates conditioned on previous spike histories [6]. Yet, it is not immediately obvious how an estimator for this form would be constructed, and indeed construction of such an estimator has previously defaulted to a binning or discretisation of time [7]. Here, we propose an estimator for this continuous-time point-process formulation of TE that remains in the continuous-time regime by harnessing a nearest-neighbours approach [8] to matching (rather than binning) inter-spike interval (ISI) histories and future spike-times. By retaining as much information about ISIs as possible, this estimator is expected to improve on properties of TE such as robustness to noise and undersampling, bias removal, and sensitivity, etc. We are currently implementing the proposed estimation algorithm in open-source code (i.e. contributing to JIDT [9] and TRENTOOL [10]), and evaluating the properties of the algorithm particularly in comparison to temporal binning approaches. ReferencesWibral M, Vicente R, Lizier JT, editors. Directed information measures in neuroscience. Berlin: Springer-Verlag; 2014. Vicente R, Wibral M, Lindner M, Pipa G. Transfer entropy—a model-free measure of effective connectivity for the neurosciences. J Comp Neurosci. 2011;30(1):45–67. Lizier JT, Heinzle J, Horstmann A, Haynes J-D, Prokopenko M. Multivariate information-theoretic measures reveal directed information structure and task relevant changes in fMRI connectivity. J Comp Neurosci. 2011;30(1):85–107. Ito S, Hansen ME, Heiland R, Lumsdaine A, Litke AM, Beggs JM. Extending transfer entropy improves identification of effective connectivity in a spiking cortical network model. PLoS One. 2011;6(11):e27431. Schreiber T. Measuring information transfer. Phys Rev Lett. 2000;85:461–4. Bossomaier T, Barnett L, Harré M, Lizier JT. An introduction to transfer entropy: information flow in complex systems. Berlin: Springer; 2016 (in press). Kim S, Putrino D, Ghosh S, Brown EN. A Granger causality measure for point process models of ensemble neural spiking activity. PLoS Comput Biol. 2011;7(3):e1001110. Kraskov A, Stögbauer H, Grassberger P. Estimating mutual information. Phys Rev E. 2004;69(6):066138. Lizier JT. JIDT: an information-theoretic toolkit for studying the dynamics of complex systems. Front Robot AI. 2014;1:11. Lindner M, Vicente R, Priesemann V, Wibral M. TRENTOOL: a Matlab open source toolbox to analyse information flow in time series data with transfer entropy. BMC Neurosci. 2011;12(1):119. P79 Active learning of psychometric functions with multinomial logistic models Ji Hyun Bak1, Jonathan Pillow2 1Department of Physics & Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton, NJ 08544, USA; 2Department of Psychology & Princeton Neuroscience Institute, Princeton University, Princeton, NJ 08544, USA Correspondence: Ji Hyun Bak - jhbak@princeton.edu BMC Neuroscience 2016, 17(Suppl 1):P79 As new technologies expand the capacity for making large-scale measurements of neural activity, there is a growing need for methods to rapidly characterize behavior and its dependence on stimuli. In typical experiments, an animal is presented with a stimulus on each trial and has to select a response among several options. Since such experiments are costly, a problem of practical importance is to learn the animal’s psychometric choice functions from a minimal amount of data. Here we show that one can achieve substantial speedups over traditional randomized designs via active learning, in which stimuli are selected adaptively on each trial according to an information-theoretic criterion, as shown in Fig. 49. Specifically, we model behavior with a multinomial logistic regression model, in which the probability of each choice given a stimulus depends on a set of linear weights. Our work extends previous work on this problem [1–3] in several important ways. First, we incorporate an explicit lapse rate to account for the fact that observers may occasionally make errors on “easy” trials due to lapses in concentration or memory [4]. Second, we develop an efficient method based on Markov Chain Monte Carlo (MCMC) sampling that is accurate in settings in which the log-likelihood is not concave, for example as in the presence of lapse rates. Third, we extend consideration for multiple-alternative responses, extending previous work for binary responses. We compare the performance of our sampling-based method to one based on a local (Laplace) approximation to the posterior [5], and show that failure to incorporate lapse rates can have deleterious effects on the accuracy of inferred parameters under both methods. We test our method on simulated data, as well as on an experimental dataset concerning the multiple-alternative choice behavior of monkeys [6], demonstrating that active sampling of the stimulus space facilitates the learning of the psychometric function significantly, as well as suggesting that the full range of the multi-dimensional stimulus could have been exploited more efficiently using our active learning framework. Finally, we discuss the comparative advantages and disadvantages of the different methods, and how one might adapt these algorithms to achieve best results.Fig. 49 Example of active learning, simulated with a three-alternatives model on 1D stimulus. After each observation, the psychometric functions are estimated based on the accumulated data, and the next stimulus is chosen to maximize the expected information gain. The estimated psychometric functions (solid lines) quickly approach the true functions (dashed lines) through the adaptive and optimal choice of stimuli Acknowledgements: We thank Anne Churchland for providing the data. JP was supported by the McKnight Foundation, Simons Global Brain Initiative, NSF CAREER Award IIS-1150186, and NIMH grant MH099611. JHB was supported by NSF grant PHY-1521553. ReferencesKontsevich LL, Tyler CW. Bayesian adaptive estimation of psychometric slope and threshold. Vis Res. 1999;39:2729–37. Zocchi SS, Atkinson AC. Optimum experimental designs for multinomial logistic models. Biometrics. 1999;55:437–44. DiMattina C. Fast adaptive estimation of multidimensional psychometric functions. J Vis. 2015;15:1–20. Kuss M, Jakel F, Wichmann FA. Bayesian inference for psychometric functions. J Vis. 2005;5:478–92. Lewi J, Butera R, Paninski L. Sequential optimal design of neurophysiology experiments. Neural Comput. 2009;21:619–87. Churchland AK, Kiani R, Shadlen MN. Decision-making with multiple alternatives. Nat Neurosci. 2008;11:693–702. P81 Inferring low-dimensional network dynamics with variational latent Gaussian process Yuan Zaho1,2, Il Memming Park1,3 1Department of Neurobiology and Behavior, Stony Brook University, Stony Brook, NY 11794, USA; 2Department of Applied Mathematics and Statistics, Stony Brook University, Stony Brook, NY 11794, USA; 3Institute for Advanced Computational Science, Stony Brook University, Stony Brook, NY 11794, USA Correspondence: Il Memming Park - memming.park@stonybrook.edu BMC Neuroscience 2016, 17(Suppl 1):P81 Surprisingly, large-scale population recordings often show signatures of low-dimensional dynamics, that is, variations in a small number of common factors explain most of the dependence among neurons [1–3]. This supports the idea that a large neuronal network is implementing necessary computations described by continuous low-dimensional nonlinear dynamics. Sufficient amount of redundancy in the population activity would allow us access to the internal computation process of interest even when we only observe a small subset of neurons. Thus, it is necessary to deduce the latent dynamics from neural time series in order to understand if and how neural systems operate in this regime. There are several latent variable models that aim at recovering the latent dynamics, howerver, they make inadequate assumptions in favor of fast inference [4]. Here we describe an approximate inference method that recovers the latent dynamics under a natural generative model with minimal assumptions. We implemented a probabilistic method to extract shared low-dimensional latent dynamics from multi-channel neural recordings (LFP and spike trains) to reveal how neural population encodes information, and how multiple functional neural populations dynamically interact with each other. Key assumptions of our model are: (1) each neural signal represent a noisy mixture of common latent dynamics, and (2) latent dynamics are independent and temporally smooth (with possibly different time scales). We use autoregressive generalized linear model driven by latent dynamics. Unlike most of the literature [5], we do not impose linear dynamics as a prior on the latent process, instead we use a general gaussian process prior which provides a flexible framework for imposing structure such as smoothness. However, as a result, the exact posterior inference is intractable, thus we developed a variational method to find a Gaussian approximation to the posterior [6]. Our inference algorithm is memory-efficient and fast: both linear in time using a low-rank approximation of the covariance. We compare our method on both simulated systems and real data from V1 driven by drifting gratings. For a population of 148 V1 neurons, 11.4 % of the variance was explained by a shared 4-dimensional latent process, while 10 % of the variance was explained by independent variability of each neuron. We recovered orientation dependent embedding that faithfully encode the stimulus drive on average, and the population-wide trial-to-trial modulation. In conclusion, we present an efficient and scalable method to recover underlying dynamics from noisy partial observations to study neural code and computation. ReferencesOkun M, Steinmetz NA, Cossell L, Iacaruso MF, Ko H, Barthó P, Moore T, Hofer SB, Mrsic-Flogel TD, Carandini M, Harris KD. Diverse coupling of neurons to populations in sensory cortex. Nature. 2015;521(7553):511–15. Goris RLT, Movshon JA, Simoncelli EP. Partitioning neuronal variability. Nat Neurosci. 2014;17(6):858–65. Luczak A, Bartho P, Harris KD. Gating of sensory input by spontaneous cortical activity. J Neurosci. 2013;33(4):1684–95. Yu BM, Cunningham JP, Santhanam G, Ryu SI, Shenoy KV, Sahani M. Gaussian-process factor analysis for low-dimensional single-trial analysis of neural population activity. J Neurophysiol. 2009;102(1):614–35. Paninski L, Ahmadian Y, Ferreira DGG, Koyama,S, Rahnama Rad K, Vidne M, Vogelstein J, Wu W. A new look at state-space models for neural data. Journal of Comput Neurosci. 2010;29(1–2):107–26. Blei DM, Kucukelbir A, McAuliffe JD. Variational inference: a review for statisticians. arXiv 2016 1601.00670 [http://arxiv.org/abs/1601.00670 arXiv] P82 Computational investigation of energy landscapes in the resting state subcortical brain network Jiyoung Kang1, Hae-Jeong Park2 1Graduate School of Life Science, University of Hyogo, 3-2-1 Koto, Kamigori, Ako, Hyogo, 678-1297, Japan; 2Department of Nuclear Medicine, Radiology and Psychiatry, Yonsei University College of Medicine, Department of Cognitive Science, Yonsei University, 50 Yonsei-ro, Sinchon-dong Seodaemoon-gu, Seoul, 120-752, Republic of Korea Correspondence: Hae-Jeong Park - parkhj@yuhs.ac BMC Neuroscience 2016, 17(Suppl 1):P82 Recently, energy landscapes of the resting state functional brain network have been researched using the pairwise maximum entropy model (MEM). This approach considers not only activities of nodes but also interactions among nodes in modeling brain networks and thus estimating energy landscapes of the brain state [1, 2]. From the energy landscape models, we can identify major stable states (local minima) and estimate transition rates among stable states. The brain networks of the resting states are known to be affected by brain diseases or treatments. In the pairwise MEM, such effects correspond to changes in the parameters of the baseline activities and pairwise interactions. In the present study, we investigated the energy landscape and its robustness of the subcortical human brain network that plays a central role in the human brain. The subcortical brain regions we examined were 15 regions of interests (ROIs); hippocampus, amygdala, caudate, putamen, pallidum, thalamus, nucleus accumbens, and brainstem. To construct a pairwise MEM for spontaneous interactions among subcortical brain regions, we used resting state fMRI (rs-fMRI) data of the human connectome projects, which contains 468 people’s data. The blood oxygen level-dependent signals in the ROI were first binarized to represent states (zero for inactive, one for active states) of the ROI, and thereby 215 brain states were considered. The parameters of the MEM were fitted to reproduce observed activation patterns of the rs-fMRI data. The constructed MEM showed high accuracy of fit (~92.6 %) and reliability (~99.9 %). We found symmetric properties for the left and right hemispheres, and confirmed estimated parameters grossly reflecting the known anatomical connectivity of the subcortical brain. We further investigated the robustness of the system by perturbing the global weight for interactions, parameters for baseline activities of ROI and parameters for interactions between pairs of all ROIs (1906 edges), one by one from the original MEM. Alteration of the energy landscapes after perturbation was measured with respect to the number of local minima. We found that the number of the local minima of the subcortical system without any perturbation is very high. This implies that the subcortical brain system is optimal in the sense of its largest coverage of local minima (maximal number of local minima). This result suggests that brain was built to have multiple stable states. We also found different categories of parameters that affect the energy landscape of the resting state. For example, small increase in the pairwise parameter between the caudate and putamen dramatically reduced the numbers of the local minima while reduction in this parameter did not change the energy landscape. In conclusion, MEM analysis of resting state functional network would be an important tool to understand principles of the brain organization and could be useful in researching brain disease. ReferencesWatanabe T, Hirose S, Wada H, Imai Y, Machida T, Shirouzu I, Konishi S, Miyashita Y, Masuda N. A pairwise maximum entropy model accurately describes resting-state human brain networks. Nat Commun. 2013;4:1370. Watanabe T, Hirose S, Wada H, Imai Y, Machida T, Shirouzu I, Konishi S, Miyashita Y, Masuda N. Energy landscapes of resting-state brain networks. Front Neuroinform. 2014;8:12. P83 Local repulsive interaction between retinal ganglion cells can generate a consistent spatial periodicity of orientation map Jaeson Jang1, Se-Bum Paik1,2 1Department of Bio and Brain engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea; 2Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea Correspondence: Jaeson Jang - jaesonjang@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P83 Orientation map in the primary visual cortex (V1) is of great interest among functional maps in the brain, but its developmental mechanism has been under debate. A recently suggested idea is that a moiré interference pattern between ON and OFF retinal ganglion cells (RGCs) can develop a quasi-periodic structure of orientation map [1] (Fig. 50A). In this model, the mosaics of ON and OFF RGCs that are in hexagonal lattice patterns generate a periodic interference pattern and induce a cortical orientation preference map. This model successfully explains the mechanism of map development, but two questions remain unanswered yet; (1) How does the hexagonal pattern of RGC mosaic develop and (2) how is the angle alignment (θ) between ON and OFF RGC mosaics (Fig. 50A) restricted to seed the consistent spatial periodicity of orientation map? Here, we suggest that a local repulsive interaction between the nearby cells is enough to develop hexagonal RGC mosaics and consistent alignment of ON and OFF mosaics.Fig. 50 Local repulsive interaction develops a consistent interference between mosaics. A Moiré pattern of RGC. B Developmental model of RGC mosaic with local repulsive interaction between nearby cells. C Developed cell mosaic. D Autocorrelation of developed mosaics. e Approach between ON and OFF mosaics induces a gradual reinforcement of heterotypic interaction. F Angle alignment between mosaics (θ) is limited to low angles as mosaics approach (*: p < 0.05, Ranksum test, error bar: SE) To validate this idea of developmental process of cell mosaic, we assumed a local repulsive force between the nearby cells as a function of distance between two cells (Fig. 50B), which induces a gradual shift of cell position. In our model simulations, we confirmed that this model could develop a hexagonal pattern in the monotypic RGC mosaic (Fig. 50C, D). Next, we examined how the angle alignment between ON and OFF mosaics can be achieved by homotypic (ON–ON or OFF–OFF) and heterotypic (ON–OFF) interaction between RGCs. We simulated the development of ON and OFF mosaics as we allow a heterotypic interaction and gradually reduce the distance between two mosaics (Fig. 50E). When two mosaics get closer enough, we observed that the angle alignment between ON and OFF mosaics was limited to low angles (Fig. 50F). Finally, we analyzed previously reported cat RGC mosaic data [2, 3] and concluded that the data suggests the existence of heterotypic repulsive interaction between ON and OFF mosaics, as our model predicted. Our results suggest that a local repulsive interaction between RGCs can develop a hexagonal pattern in mosaics and restrict the angle alignment between ON and OFF RGC mosaics to generate a constant spatial period of orientation map. This model may provide a complementary mechanism of the retinal origin of periodic functional maps in the brain. ReferencesPaik S-B, Ringach DL. Retinal origin of orientation maps in visual cortex. Nat Neurosci. 2011;14:919–925. Zhan XJ, Troy JB. Modeling cat retinal beta-cell arrays. Vis Neurosci. 2000;17:23–39. Wassle H, Boycott BB, Illing R-B. Morphology and mosaic of ON- and OFF-beta cells in the cat retina and some functional considerations. Proc R Soc B Biol Sci. 1981;212:177–95. P84 Phase duration of bistable perception reveals intrinsic time scale of perceptual decision under noisy condition Woochul Choi1,2, Se-Bum Paik1,2 1Department of Bio and Brain Engineering, KAIST, Daejeon 34141, Republic of Korea; 2Program of Brain and Cognitive Engineering, KAIST, Daejeon 34141, Republic of Korea Correspondence: Woochul Choi - choiwc1128@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P84 When we see an ambiguous visual stimulus such as Necker cube, our perceived state switches periodically between two possible interpretations. This phenomenon, called bistable perception, is considered important to the study of dynamic mechanism of sensory perception. In particular, the time duration of each perceptual state, termed “phase duration”, seems to be a crucial factor to understanding the temporal features of underlying neural activity during sensory perception under this condition, which has not been studied intensively yet. In this study, we assume that phase duration is intrinsically correlated with time delay in cognitive tasks, such as decision making from sensory information. Our hypothesis is that the periodic switching in bistable perception is a repeated process of decision making and reveals the time scale required for this decision task. To confirm our hypothesis, we performed a human psychophysics experiment using the “racetrack” type stimulus [1], which can induce a motion perception from both bistable illusion and real motion by varying the coherence parameter, c (Fig. 51A). We examined the relationship between the phase duration, τ, under illusory motion (c = 0) and the response delay for coherent motion with different degrees of ambiguity (c > 0, Fig. 51B). Our result showed that the response delay in the coherent motion detection task (c > 0) was highly correlated with the phase duration in the bistable illusory motion perception (c = 0) (Fig. 51C, N = 19, R = 0.61, p < 0.01, Pearson’s correlation coefficient). For a systematic analysis of subjects’ performance for these two tasks, we designed a theoretical model of simple double-well energy potential [2] (Fig. 51D). The model could successfully replicate the correlation between phase duration and response delay in each task, suggesting that bistable perception and perceptual decision making processes may share a common neural mechanism (Fig. 51C).Fig. 51 Correlation between bistable perception and perception under ambiguous signal. A Racetrack stimulus. Rotational motion can be either illusory or ambiguous depending on coherence. B Example response of racetrack. Perceived motion can be bistable (top) or follows actual motion with response time (bottom). C Subjects’ (black) and model’s (red) phase duration and response time are highly correlated. D Double-well energy model to describe behavior during bistable perception and perceptual decision making task Conclusions Our findings show that the phase duration of bistable perception is highly correlated with the response time of a cognitive task. Our simple model suggests that the bistable perception can be interpreted as perceptual decision making process under highly ambiguous condition and share similar temporal dynamics. ReferencesJain S. Performance characterization of Watson Ahumada motion detector using random dot rotary motion stimuli. PLoS One. 2009;4. Kornmeier J, Bach M. Ambiguous figures—what happens in the brain when perception changes but not the stimulus. Front Hum Neurosci. 2012;6:51. P85 Feedforward convergence between retina and primary visual cortex can determine the structure of orientation map Changju Lee1, Jaeson Jang1, Se-Bum Paik1,2 1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea; 2Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea Correspondence: Changju Lee - lcj110808@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P85 Orientation map in the primary visual cortex (V1) is one of the most studied functional maps in the brain. In higher mammals such as monkeys and cats, preferred orientation of each V1 neuron appears continuous and periodic across cortical space. On the other hand, in rodents, it appears completely discontinuous, forming a structure called salt-and-pepper map. However, the developmental mechanism of salt-and-pepper orientation maps remains unclear. Previously, a model study suggested that a moiré interference pattern between ON and OFF retinal ganglion cell (RGC) mosaics can seed a periodic orientation maps (Fig. 52A), and a salt-and-pepper map can also be developed when the spatial periodicity of interference pattern is very short [1]. However, our analysis suggests that the spatial periodicity of map, estimated from rat RGC mosaics data, is not small enough to generate a salt-and-pepper structure (Fig. 52B, C). To address this issue, here we suggest that feedforward convergent wiring between retina and V1 is a crucial factor that decides the structure of orientation map.Fig. 52 The simulation model for developmental mechanism of salt-and-pepper map by feedforward convergence between retina and V1. A Moiré interference between ON and OFF RGC mosaics. B Moiré interference with small and large alignment angles can generate various range of periodicity, S. C From rat RGC mosaics, both smooth and salt-and-pepper map can be developed. D Spatial distribution of preferred orientations of V1 cells by different convergence conditions; Smooth map model (high sampling ratio and large convergence range), Salt-and-pepper map model (low sampling ratio and short convergence range). E The structure of orientation map depending on convergence range and sampling ratio To find a convergence condition that develops salt-and-pepper map, we modulated two parameters in our simulations: (1) the convergence range of V1 cells and (2) the sampling ratio of RGCs within the range. The regularity of orientation map was estimated from the measurement of the preferred orientation difference between local neurons (Fig. 52D). We found that a salt-and-pepper map was developed with low sampling ratio and large convergence range, while a smooth map was developed when convergence range was relatively small and sampling ratio was high (Fig. 52E). To further analyze the map structure generated by our model, we compared the profile of correlation between local receptive fields structure in our simulated V1 map to the previous observation in animal experiment [2]. We confirmed that our salt-and-pepper map model well matched the statistics of observed experimental data. Conclusions Our result suggests that a salt-and-pepper map can be developed by sparse and long-range convergence in feedforward wiring, while smooth map can be developed by localized convergence. We suggest that the condition of feedforward convergence between retina and V1 is a critical factor to determine the structure of orientation map. ReferencesPaik S-B, Ringach DL. Retinal origin of orientation maps in visual cortex. Nat Neurosci. 2011;14:919–25. Bonin V, Histed MH, Yurgenson S, Reid RC. Local diversity and fine-scale organization of receptive fields in mouse visual cortex. J Neurosci. 2011;31:18506–21. P86 Computational method classifying neural network activity patterns for imaging data Min Song1,2, Hyeonsu Lee1, Se-Bum Paik1,2 1Department of Bio and Brain Engineering, KAIST, Daejeon 34141, Republic of Korea; 2Program of Brain and Cognitive Engineering, KAIST, Daejeon 34141, Republic of Korea Correspondence: Min Song - night@kaist.ac.kr, Hyeonsu Lee - hslee9305@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P86 In neural imaging data, various types of spatio-temporal activity patterns are observed which may reflect dynamic features of information processing in the brain [1, 2]. Classification of these patterns is required to analyze the brain activity further. However, development of analysis tool for spatio-temporal neural activity pattern has been regarded difficult because of highly complex connections between neurons and nonlinearity of activity patterns. In this study, we suggest a novel method of classifying activity patterns in two aspects: spatial geometry and temporal dynamics. We show that our method efficiently categorizes complicated spatio-temporal patterns in brain. First, we defined meaningful activity as salient distribution of highly activated parts. Its spatial feature could be described by size and peak amplitude (Fig. 53A), temporal feature by velocity and dispersion of activity (Fig. 53A). Thus, we designed geometric profile as two-dimensional profile of instantaneous neural activity, by measuring the topography of supra-threshold area with a shifting threshold. This profile contains the information of size, peak amplitude, and geometric contours of meaningful activity. With this method, we could readily estimate similarity or correlation of different activities in terms of size, peak amplitude, and amplitude contour (Fig. 53B).Fig. 53 A novel index effectively describes different neural activity patterns obtained from imaging. A Neural activity obtained from optical imaging could be analyzed with appearance and propagation. B Appearance index of four distinct sample. C Propagation index of straight trajectory (top) and curved trajectory (bottom). D Propagation Index of non-dispersive sample (top) and dispersive sample (bottom) Next, we defined propagation profile as a characteristic of temporal displacement of activity on each direction against time and angle axis. We measured trajectory and speed of activity using a normalized cross-correlation. This profile intuitively shows dominant trajectory, speed change and dispersion of the activity: how disperse to every direction. So we can compare activities: whether the activity is moving straight or curved trajectories, or whether the activity propagation is accelerating or not (Fig. 53C). Our new method can easily perform not only classification of overall dynamics in brain, but also a simplified description of complex patterns (Fig. 53D), that may be applicable to the analysis of various kinds of brain imaging data. P87 Symmetry of spike-timing-dependent-plasticity kernels regulates volatility of memory Youngjin Park1, Woochul Choi1,2, Se-Bum Paik1,2 1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea; 2Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea Correspondence: Youngjin Park - yodamaster@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P87 Synaptic plasticity is considered the core mechanism of learning and memory [1]. However, how plasticity can specifically modulate synaptic connections to generate short term or long term memory has not been understood completely. Here we introduce a theoretical model which suggests that a key mechanism of short term and long term memory can be implemented by a small difference in spike-timing-dependent-plasticity (STDP) rule. (Fig. 54A).Fig. 54 Different learning rules reproduce volatile/nonvolatile memory system. A Spike timing dependent plasticity. B Memory decaying properties of different learning rule. Poisson spikes are given for 1000 s to simulate decaying environment. C, D Multiple patterns was given to the system every 200 s. C Memory performance of each pattern in AS memory system. D Memory performance of each pattern in SS memory system To test our idea, we designed simulations using a model feedforward neural network where two types of synaptic plasticity are implemented; asymmetric STDP (AS) [2] and symmetric STDP (SS). We defined the memory as the ability of a system to retrieve a consistent response spike pattern when we repeatedly introduce an identical pattern of spikes, and then we examined the performance of the system in terms of memory sustainability and appendability. In our simulations, a network with AS showed performances similar to short-term memory while a network with SS showed long-term memory like properties. Memory in an AS Network decayed as a function of time, while memory in a SS network did not show a noticeable decay (Fig. 54B). Moreover, when a new input pattern was given to the network in addition to old memory, AS system replaced old memory with new memory pattern (Fig. 54C), while SS system maintained the old memory together with a newly trained memory (Fig. 54D). Based on our findings, we suggest a new memory system called hybrid memory that is capable of showing intermediate properties between a long-term memory and a short-term memory (Fig. 54B, hybrid). This model suggests that transition between short term and long term memory might not be discrete but gradual. Conclusions We have shown that our model network can implement different types of memory performance from the variation of plasticity, or learning rule. Our results imply that the various types of memory may be originated from a small difference in the shape of STDP kernel. ReferencesBliss TV, Collingridge GL. A synaptic model of memory: long-term potentiation in the hippocampus. Nature. 1993;361(6407): 31–9. Gütig R, Aharonov R, Rotter S, Sompolinsky H. Learning input correlations through nonlinear temporally asymmetric Hebbian plasticity. J Neurosci. 2003;23(9):3697–3714. P88 Effects of time-periodic coupling strength on the first-spike latency dynamics of a scale-free network of stochastic Hodgkin–Huxley neurons Ergin Yilmaz1, Veli Baysal1, Mahmut Ozer2 1Department of Biomedical Engineering, Bülent Ecevit University, Zonguldak 67100, Turkey; 2Department of Electrical and Electronics Engineering, Bülent Ecevit University, Zonguldak 67100, Turkey Correspondence: Ergin Yilmaz - erginyilmaz@yahoo.com BMC Neuroscience 2016, 17(Suppl 1):P88 It is still not known clearly which encoding mechanism that neurons utilize for coding of sensory information. One of the proposed encoding mechanism is latency coding which suppose that first-spike latency conveys much of the information about the stimulus. In this context, Pankratova et al. [1] studied the effects of noise on the first-spike latency dynamics of stochastic the Hodgkin–Huxley (HH) neuron, and obtained a bell-shaped dependence of mean response time of the neuron on noise intensity, emerging a phenomenon called “noise delayed decay” (NDD). Later, this finding have been extensively studied by using complex neuronal networks [2]. On the other hand, neurons exchange information via coupling at the special location called synapse. Thus, coupled neurons in networks play a decisive role on the phenomenon occurring in neuronal networks. Majority of the studies examining the NDD effect assume that coupling strength among coupled neurons is constant and the fact that synapses are plastic, that is, coupling strength among neurons can change with time, have been neglected. To present the effects of plasticity or time-varying-coupling strength Birzu et al. [3] studied the effects of time-periodic coupling strength (TPCS) on the firing dynamics of a globally coupled array of FHN neurons. Here, our aim is to present the effects of the frequency of TPCS on the NDD phenomenon in a scale-free network of HH neurons. We construct the network with N = 200 neurons modeled by a stochastic HH equation including ion channel noise, and average degree of kavg = 4. We consider that the coupling strength among coupled units changes with time-periodic fashion as proposed in [3]. To measure the mean latency and jitter of the network, the first-spike times of each neuron are recorded. For the comparison purpose, we give the constant coupling strength effect on the latency dynamics of scale-free network. Obtained result are depicted in Fig. 55.Fig. 55 The statistics of the first-spike occurrence times (amplitude of TPCS ε0 = 0.2, cell size S = 100 μm2, frequency of suprathreshold signal f = 20 Hz and amplitude of it A = 4μA/cm2). A Mean latency of the network, B jitter of the network Conclusions It is seen that mean latency and jitter of the first-spike times exhibit a damped sine wave dependence on the frequency of TPCS, indicating that TPCS can significantly increase or decrease the latency time which passes until sensing of the suprathreshold stimulus by each neuron at fixed intensity of channel noise (S = 100 μm2). The frequencies of TPCS greater than ω = 2 m s−1 are not significantly affect the mean latency and jitter of the network, as compared to the constant coupling strength. As a result, with finely tuned values of the frequency of TPCS input signal detection performance of the scale-free network can be prominently increased by mitigating the response time of the each neuron. ReferencesPankratova EV, Polovinkin AV, Mosekilde E. Resonant activation in a stochastic Hodgkin–Huxley model: interplay between noise and suprathres-hold driving effects. Eur Phys J B. 2005;45:391–7. Yilmaz E. Impacts of hybrid synapses on the noise-delayed decay in scale-free neural networks. Chaos Solitons Fractals. 2014;66:1–8. Birzu A, Krischer K. Resonance tongues in a system of globally coupled FitzHugh-Nagumo oscillators with time-periodic coupling strength. Chaos. 2010;20(4):043114. P89 Spectral properties of spiking responses in V1 and V4 change within the trial and are highly relevant for behavioral performance Veronika Koren1,2, Klaus Obermayer1,2 1Institute of Software Engineering and Theoretical Computer Science, Technische Universitaet Berlin, Berlin, 10587, Germany; 2Bernstein Center for Computational Neuroscience Berlin, Humboldt-Universitaet zu Berlin, Berlin, 10115, Germany Correspondence: Veronika Koren - veronika.koren@bccn-berlin.de BMC Neuroscience 2016, 17(Suppl 1):P89 Linking sensory coding and behavior is a fundamental question in neuroscience. We have addressed this issue in behaving monkey visual cortex (areas V1 and V4) while animals were trained to perform a visual discrimination task in which two successive images (target and test stimuli, with a delay period in between) were either rotated with respect to each other or were the same. We hypothesized that animal’s performance in the visual discrimination task depends on the quality of stimulus coding in visual cortex. We tested this hypothesis by investigating the power spectral density of spiking signal from single neurons (spectra) and of pairs of neurons (cross-spectra) in relation to correct and incorrect behavioral responses. Our analysis shows that spectral properties systematically change with behavioral performance. Correct responses are associated with significantly higher spectra during the delay period. Cross-spectra of correct responses are significantly lower during the target period but significantly higher afterwards (delay period and test period). Spectral properties of single neurons and even more of pair-wise interactions therefore change within the trial, presumably following functional demands of stimulus processing in different epochs of the trial. Interestingly, differential dynamics in visual cortex sustains successful versus unsuccessful behavioral performance. Preprocessing methods are used in order to avoid biases due to limited measurement time. The spike train is multiplied with Hanning window for low frequencies up to 22 Hz and with Slepian multitapers for frequencies between 24 and 140 Hz [1]. We use 300 ms window of sustained activity during stimulus periods and 500 ms window of activity during the delay period. Spectrum and cross spectrum are computed with Fast Fourier transform (Matlab, Mathworks). The cross-spectrum being a complex function, we consider its absolute value. Spectra are averaged in bins of 6 Hz. The variance and the covariance of the spiking signal are computed as sums over frequencies of auto and cross-spectra, respectively, up to the cut-off frequency (140 Hz). Auto spectra are significantly higher in correct compared to incorrect trials in most of frequency bands in both V1 and V4 areas (p < 0.05 in 21 and 22 out of 24 frequency bands in V1 and V4, respectively, one-tailed sign-rank test). Consistently, the variance is significantly higher for correct responses (p < 10−4 in V1 and p = 0.0014 in V4. Cross-spectra are lower in correct trials during target period (18 and 10 out of 24 frequency bands are significant in V1 and V4, respectively, no significant effect in remaining bands) but higher in delay period (11 and 20 bands are significant in V1 and V4, respectively, no effect in remaining bands) and test period (19 and 13 significant bands in V1 and V4, respectively, no effect in remaining bands). Consistently, the covariance is significantly lower for correct responses during target stimulus (p = 0.0003, p = 0.0002 in V1 and V4) and higher during the delay (p < 10−4 in V1 and V4) and the test stimulus (p < 10−4 in V1 and V4). Our results show that spectra and cross spectra change during behavioral task and that spectral information in visual cortex might be highly relevant for behavioral performance. References Fries P, Womelsdorf T, Oostenwald R, Desimone R. The effects of visual stimulation and selective visual attention on rhythmic neuronal synchronization in Macaque area V4. J Neurosci. 2008;28(18):4823–35. Womelsdorf T, Schoffelen JM, Oostenveld R, Singer W, Desimone R, Engel AK, Fries P. Modulation of neuronal interactions through neuronal synchronization. Science. 2007;316:1609–12. Appel W. Mathematics for physics & physicists. Oxfordshire: Princeton University Press; 2007. Gutnisky DA, Dragoi V. Adaptive coding of visual information in neural populations. Nature. 2008;452(7184):220–4. P90 Methods for building accurate models of individual neurons Daniel Saska1, Thomas Nowotny1 1School of Engineering and Informatics, Sussex Neuroscience, University of Sussex, Falmer, Brighton BN1 9QJ, UK Correspondence: Daniel Saska - research@saska.io BMC Neuroscience 2016, 17(Suppl 1):P90 Formulating predictive models of single neuron dynamics has become a challenge taken up by many researchers ever since Hodgkin and Huxley published their widely accepted phenomenological model of electrophysiological dynamics of the squid giant axon [1]. Advances include, amongst others, modelling complex cells (such as cells of the stomatogastric ganglion (STG) in lobster and crab [2]) or increasingly automated modelling methods [3]. However, for each problem solved, new ones emerge. One such problem has been pointed out by Golowasch et al. [4] who discovered that averaging multiple measurements from the same cell type can produce models that fail to reproduce the behaviour of the target cells. This issue does not only affect methods that rely on averaging to achieve better signal-to-noise ratio but more generally all methods that examine ion channels in separate preparations. This includes classical voltage clamp, in which different ionic conductances are measured in separate individual cells because many pharmacological blockers cannot be fully reversed. We here propose a different approach for parameter estimation aiming to build a model based on data from a single, individual cell. The proposed method consists of the consecutive use of a voltage clamp like protocol and parameter estimation in a current clamp mode. For the ‘voltage clamp protocol’ we use genetic algorithms (GA) to evolve a set of ‘highlighting’ voltage waveforms and specific observation windows, so that the resulting currents within the windows depend on a highlighted parameter but not so much on the values of other parameters. These parameter-specific waveforms are then applied to a live neuron (so far in simulation) and the resulting currents are observed and fitted with another GA, focusing on the highlighted parameters for each of the voltage waveforms. The resulting model is then transferred into current clamp, where the parameters are again estimated using a GA. The neuron models in the GA population are coupled to the observed cell to achieve a degree of synchronization and so smooth the error landscape. The coupling is reduced adiabatically until the model neurons and experimental cell remain synchronized with (virtually) no coupling. We found that combining voltage and current clamp works particularly well since the fitness landscape in voltage clamp has few local minima but is fairly shallow whereas the opposite is true for the current clamp mode. We can hence find approximate parameter values from arbitrary initial guesses in our ‘voltage clamp’ mode and once the parameters are in the right area, they can be refined in current clamp. Our method can produce accurate models of cells in the crab STG for the cases of one, two, three and four-spike bursters. Table 1 shows the resulting parameter values for the example of a one-spike burster cell illustrated in Fig. 56.Table 1 Estimated parameters for one-spike burster STG cell fPar C gNa ENa gKd EKd gA EA gCa Ca0 Caf Cat gKCa EKCa gleak Eleak Min 0.1 0 0 0 −100 0 −100 0 0.01 14 20 0 −100 0 −100 Max 10 800 100 200 0 75 0 5 0.1 16 250 300 0 1 0 Real 0.628 50 50 100 −80 5 −80 4 0.05 14.96 200 250 −80 0.01 −50 Estim. 0.613 62.07 39.05 95.77 −76.09 4.622 −88.59 3.912 0.0438 15.59 194.7 246.2 −80.22 0.0106 −50.15 Fig. 56 One-spike burster and estimated model as in Table 1 ReferencesHodgkin AL, Huxley AF. A quantitative description of membrane current and its application to conduction and excitation in nerve. J Physiol. 1952;117(4):500–44. Liu Z, Golowasch J, Marder E, Abbott LF. A model neuron with activity-dependent conductances regulated by multiple calcium sensors. J Neurosci. 1998;18(7):2309–20. Willms AR. NEUROFIT: software for fitting Hodgkin–Huxley models to voltage-clamp data. J Neurosci Methods. 2002;121(2):139–50. Golowasch J, Goldman MS, Abbott LF, Marder E. Failure of averaging in the construction of a conductance-based neuron model. J Neurophysiol. 2002;87(2):1129–31. P91 A full size mathematical model of the early olfactory system of honeybees Ho Ka Chan1, Alan Diamond1, Thomas Nowotny1 1School of Engineering and Informatics, University of Sussex, Falmer, Brighton, BN1 9QJ, UK Correspondence: Ho Ka Chan - hc338@sussex.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P91 Experimental measurements often can only provide limited data from animals’ sensory systems. As a result, data driven models are similarly limited. However, in order to make biologically relevant predictions, it is important to consider inputs representative of the full sensory input space. Here we present a full size model of the early olfactory system of honeybees that extrapolates inputs from the limited subset of available experimental observations. Our model comprise olfactory receptor neurons (ORNs), local neurons (LNs) and projection neurons (PNs) organized in 160 glomeruli. The ORN response patterns are generated using a set of ordinary differential equations describing the binding and activation of receptors as in [2]. The parameters for these processes are chosen to match the statistical distribution of experimental observed quantities in [3, 4] as well as the statistics of asymptotic responses to time-invariant odours at high concentration observed in calcium imaging of glomeruli with bath-applied Ca dyes [5]. To generate the PN responses, we considered a network in which PNs and LNs both receive excitatory input from ORNs in the same glomerulus and inhibitory input from LNs in all glomeruli. The connectivity between PNs and LNs is based on the correlation between the activities of their respective glomeruli as in [5]. A rate model, derived using the leaky integrate-and-fire model with the assumption of constant input, is used to determine the input–output relationship. We tested our ORN model with continuous stimuli and short pulses. The average normalized ORN responses to a chemical stimulus are qualitatively similar to that of biological ORNs measured by electro-antennogram recordings [4] as shown in Fig. 57, except that the time scale of response latency is a little smaller in the model. This can be explained by the lack of temporal filtering of input conductance to output spiking in our rate model. The responses of PNs driven by ORN activity can be compared to calcium imaging data with back-filled PNs [6], which confirms that the responses of our model PNs can replicate key features of those of biological PNs. With appropriate data, our model can be generalized to the early olfactory systems of other insects. It hence provides a possible basis for future numerical studies of olfactory processing in insects.Fig. 57 Experimental ORN responses to stimulus measured by electro-antennogram recordings in [5] (top, black line) is qualitative similar to the average normalized ORN responses to stimulus (1-hexanol at concentration 0.1 M) predicted by our model (bottom) Acknowledgements: This work was supported by HFSP, RGP0053/2015 and EPSRC, EP/J019690/1. ReferencesGalizia CG, Sachse S, Rappert A, Menzel R. The glomerular code for odor representation is species specific in the honeybee Apis mellifera. Nat Neurosci. 1999;2(5):473–8. Grémiaux A, Nowotny T, Martinez D, Lucas P, Rospars J-P. Modelling the signal delivered by a population of first-order neurons in a moth olfactory system. Brain Res. 2012;1434:123–35. Rospars J-P, Lansky P, Chaput M, Duchamp-Viret P. Competitive and noncompetitive odorant interactions in the early neural coding of odorant mixtures. J Neurosci. 2008;28(10):2659–66. Szyszka P, Gerkin RC, Galizia CG, Smith BH. High-speed odor transduction and pulse tracking by insect olfactory receptor neurons. Proc Natl Acad Sci USA. 2014;111(47):16925–30. Linster C, Sachse S, Galizia CG. Computational modeling suggests that response properties rather than spatial position determine connectivity between olfactory glomeruli. J Neurophysiol. 2005;93(6):3410–17. Ditzen M. Odor concentration and identity coding in the antennal lobe of the honeybee Apis mellifera (PhD thesis). Berlin: Freie Universität Berlin; 2005. P92 Stimulation-induced tuning of ongoing oscillations in spiking neural networks Christoph S. Herrmann1, Micah M. Murray2, Silvio Ionta2, Axel Hutt3, Jérémie Lefebvre4 1Research Center Neurosensory Science, Carl-von-Ossietzky University Oldenburg, Oldenburg, Germany; 2The Laboratory for Investigative Neurophysiology (The LINE), Department of Clinical Neurosciences and Department of Radiology, University Hospital Center and University of Lausanne, Lausanne 1011, Switzerland; 3Deutscher Wetterdienst, 63067 Offenbach, Germany; 4 Krembil Research Institute, University Health Network, Toronto, Ontario M5T 2S8, Canada Correspondence: Jérémie Lefebvre - jeremie.lefebvre@uhnresearch.com BMC Neuroscience 2016, 17(Suppl 1):P92 Rhythmic neural activity is believed to play a central role in neural computation. Oscillatory brain activity has been associated with myriad functions such as homeostasis, attention, and cognition as well as neurological and psychiatric disorders, including Parkinson’s disease, schizophrenia, and depression [1]. Numerous studies have shown that that non-invasive stimulation, such as repetitive transcranial magnetic stimulation (rTMS) and Transcranial Alternating Direct Current Stimulation (TACS), provide the means of modulating large-scale oscillatory brain dynamics by perturbing and/or entraining both resting state and task activity [2]. These stimulation-induced perturbations of neural oscillations have been shown to alter cognitive performance and perception, effects that are further known to depend on brain state prior and during stimulation [3]. Yet, the surge of interest in these approaches is compromised by the existence of complex interference patterns between exogenous and endogenous dynamics. To better understand oscillatory responses evoked during rhythmic stimulation, we simulated a spiking cortical network built of excitatory and inhibitory cells, expressing resting state alpha synchrony and subjected to pulsatile forcing at frequencies in the range of 1–100 Hz. Varying stimulation parameters—such as frequency and amplitude—we evaluated the influence of stimulation on the spectral properties of the network’ global neuroelectric output. The network was composed of recurrently connected Poisson neurons with propagation delays, linear adaptation, spatially profiled and sparse synaptic connections and noisy inputs. To model exogenous influences, we used continuous trains of phasic pulses and stimulated the network globally (all neurons identically), to mimic TMS-like signals. For every stimulation condition, we also measured the neurons mean firing rate, the mean network spike coherence and non-linearity metric. Multiple spectral patterns could be observed in the network’s responses, both in the power and frequency domains, indicating a plurality of responses to shifts in stimulation frequency and/or amplitude. Network responses to slower/weaker stimulation were expectedly found to be shaped by entrainment and resonance: resonance curves defining the amplitude of the system’s responses were revealed, alongside the characteristic Arnold tongues, where stimulus-locking can be achieved. The individual firing rates of the neurons and resulting spike coherence (assessing the degree of spiking synchronization) were both strongly tied to the stimulation forcing. In contrast, for stimulation frequencies higher than 50 Hz, a different mechanism was found to dominate the network dynamics: stimulation pulses shaped the system’s response via a non-linear acceleration (NLA) on ongoing oscillatory activity. The network peak frequency was gradually shifted, leading to a transition from the alpha to the beta band, and for forcing parameters that did not recruit neither resonance nor entrainment. Also, NLA led the network in a state of weak oscillatory power, where individual neurons were found in a state of intense, irregular spiking. By investigating closely the network non-linear interactions for each stimulation conditions, we found that high-frequency forcing induces synergetic and non-linear, large-scale effects [4]. Our results provide new computational perspectives about the response of synchronous spiking neural networks in which firing rates, spike coherence and emergent oscillatory activity can be exogenously modulated using dynamic inputs. Taken together, our results suggest that the action of forcing on oscillating neural systems must be regarded as strongly non-linear, and input features must be considered as control parameters. ReferencesWang X-J. Neurophysiological and computational principles of cortical rhythms in cognition. Physiol Rev. 2010;90:1195–1268. Thut G, et al. The functional importance of rhythmic activity in the brain. Curr Biol. 2012;22:658–63. Neuling T, et al. Orchestrating neuronal networks: sustained after-effects of transcranial alternating current stimulation depend upon brain states. Front Hum Neurosci. 2013;7:161. Lefebvre J, et al. Stimulus statistics shape oscillations in non-linear recurrent neural networks. J Neurosci. 2015;35(7):2895–903. P93 Decision-specific sequences of neural activity in balanced random networks driven by structured sensory input Philipp Weidel1, Renato Duarte1,4,5, Abigail Morrison1,2,3,4 1Institute of Advanced Simulation (IAS-6) & Institute of Neuroscience and Medicine (INM-6) & JARA BRAIN Institute I, Jülich Research Center, 52425 Jülich, Germany; 2Institute of Cognitive Neuroscience, Faculty of Psychology, Ruhr-University Bochum, 44801 Bochum, Germany; 3Simulation Laboratory Neuroscience – Bernstein Facility for Simulation and Database Technology, Institute for Advanced Simulation, Jülich Aachen Research Alliance, Jülich Research Center, Jülich, Germany; 4Bernstein Center Freiburg, Albert-Ludwig University of Freiburg, Freiburg im Breisgau, 79104, Germany; 5Institute for Adaptive and Neural Computation, School of Informatics, University of Edinburgh, EH8 9AB, UK Correspondence: Philipp Weidel - p.weidel@fz-juelich.de BMC Neuroscience 2016, 17(Suppl 1):P93 Perceptual decision-making is an intricate process implicating the coordinated activity of multiple brain areas [1, 2]. Recent experimental studies demonstrate the existence of a complex interplay between decision-related neural events and transient working memory processes [1], implemented by distributed circuits where specific sub-populations appear to be differentially involved in the evidence accumulation process and subsequent behavioral outcomes [2]. This results in observable divergences in choice-specific neuronal dynamics, unfolding as reproducible trajectories throughout the network’s state-space [1] and hinting at the dissipative nature of the underlying dynamical system, which executes cognitively relevant processing through transient trajectories. Despite this evidence, the majority of modeling studies addressing reward-modulated decision-making tend to simplify the formalization of environmental representations in the cortex as stable, attractor states corresponding to discrete environmental states [3]. Even models involving transient-based computations often simplify sensory stimuli to a discrete set of inputs transduced as stochastic point processes [4]. These simplifications potentially draw an incomplete picture of neural dynamics and therefore provide limited insights into the true nature of computation in neural circuits. To overcome this issue, we take one step towards realistic in silico experimental settings by using structured virtual environments to obtain rich sensory input to drive model neural systems using the ROS-MUSIC toolchain [5]. It allows us to simulate robotic agents in virtual 3D environments performing a realistic perceptual decision task, which can be directly equated to experimental data. The robotic simulation generates realistic and structured sensory data which is encoded to spiking neural activity using a nonlinear encoding process, as formalized in [6]. The encoded sensory data is then used as input to a balance recurrent neural circuit. In this study, we investigate the emergent dynamical features of neural activity when the agent is navigating a virtual T-maze. We observe decision-specific sequences of neural activity akin to experimental evidence [1], revealing possible processing strategies employed by the neural substrate. Furthermore, we investigate the role of different adaptation/plasticity mechanisms in shaping the system’s dynamics. In order to equate our results with those of other studies, we attempt to partition the network state-space into discrete activity clusters, which carry relevant information that could potentially be used to drive reinforcement learning algorithms. Acknowledgements: We acknowledge partial support the Helmholtz Alliance through the Initiative and Networking Fund of the Helmholtz Association and the Helmholtz Portfolio theme “Supercomputing and Modeling for the Human Brain”, EuroSPIN and the German Federal Ministry for Education and Research (BMBF Grant 01GQ1343). ReferencesHarvey CD, Coen P, Tank DW. Choice-specific sequences in parietal cortex during a virtual-navigation decision task. Nature. 2012;484(7392):62–8. Shadlen MN, Kiani R. Decision making as a window on cognition. Neuron. 2013;80(3):791–806. Jitsev J, Morrison A, Tittgemeyer M Learning from positive and negative rewards in a spiking neural network model of basal ganglia. In: The 2012 international joint conference on neural networks (IJCNN). IEEE; 2012. Duarte R, Morrison A. Dynamic stability of sequential stimulus representations in adapting neuronal networks. Front Comput Neurosci. 2014;8(124). Weidel P, Duarte R, Djurfeldt M, Morrison A. ROS-MUSIC toolchain (in preparation). Eliasmith C, Anderson CH. Neural engineering: computation, representation, and dynamics in neurobiological systems. MIT Press; 2004. P94 Modulation of tuning induced by abrupt reduction of SST cell activity Jung H. Lee1, Ramakrishnan Iyer1, Stefan Mihalas1 1Allen Institute for Brain Science, Seattle, WA 98109, USA Correspondence: Jung H. Lee - jungl@alleninstitute.org BMC Neuroscience 2016, 17(Suppl 1):P94 Inhibitory interneurons have been considered pivotal in orchestrating pyramidal neurons. Indeed, the optogenetic perturbation of inhibitory cell types confirmed its validity. Recent studies [1, 2] have found that the optogenetic stimulation of somatostatin positive (SST) interneurons, one of the three major inhibitory types, sharpens the tuning of visual neurons, but its effect was conspicuous only when the optogenetic activation of SST cells was turned off abruptly. Specifically, with 4-s presentation of visual stimuli, the 1-s activation of SST cells resulted in a sharper tuning, whereas 4-s activation did not induce significant sharpening [2], which leads to a question: “Why does the length of optogenetic stimulation render such a striking difference?” Lee et al. suggested that the 1-s activation sharpens the tuning curve due to the rebound activity of PV cells, and El-Boustani et al. suggested the reduction of co-activation between PV and Pyr cell activity; see Ref. [2] for the details. In our study, we investigate the potential mechanisms underlying the disparate effects between short and long activations of SST cells by using the firing rate equations that expresses the interactions among Pyr, SST and PV cells conveyed via cell-type specific connections reported by Pfeffer et al. [3]. Our model consists of five populations: two pyramidal populations (Pyr1, 2), two PV cell populations (PV1, 2) and SST cell population. We assume that Pyr1 and Pyr2 in close proximity respond to preferred and non-preferred stimuli, respectively. The two pyramidal populations excite the shared SST cell population which sends inhibition back to them. Since SST cells are known to be connected to distant presynaptic pyramidal cells via long-horizontal connections [4], the two SST populations in close proximity would receive (almost) identical inputs, making the two SST populations redundant. PV1 and PV2, which receive identical external background inputs, interact with Pyr1 and Pyr2, respectively. Pyr1 and Pyr2 are not directly connected, but they can indirectly interact with each other through SST cell population. Our model replicates the paradoxical finding that not 4-s activation of SST cells but 1-s activation leads to the sharper responses of V1 neurons. In our model, PV cells provide synchronized inhibition to pyramidal cells despite their distinctive receptive fields when SST cell activation is abruptly turned off. If SST cells are stimulated during the entire period of simulations (4 s), the induced synchronous inhibition from PV cells to pyramidal cells is not strong enough to induce sharper responses. We also found that this synchronous inhibition can be induced by the activation of VIP cells, raising the possibility that VIP cells regulate V1 neural responses with the proposed mechanism. Acknowledgements: We wish to thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their vision, encouragement and support. ReferencesWilson NR, Runyan CA, Wang FL, Sur M. Division and subtraction by distinct cortical inhibitory networks in vivo. Nature. 2012;488(7411):343–8. doi:10.1038/nature11347. Lee S-H, Kwan AC, Dan Y. Interneuron subtypes and orientation tuning. Nature. 2014;508(7494):E1–2. doi:10.1038/nature13128. Pfeffer CK, Xue M, He M, Huang ZJ, Scanziani M. Inhibition of inhibition in visual cortex: the logic of connections between molecularly distinct interneurons. Nat Neurosci. 2013;16(8):1068–76. doi:10.1038/nn.3446. Adesnik H, Bruns W, Taniguchi H, Huang ZJ, Scanziani M. A neural circuit for spatial summation in visual cortex. Nature. 2012;490(7419):226–31. doi:10.1038/nature11526. P95 The functional role of VIP cell activation during locomotion Jung H. Lee1, Ramakrishnan Iyer1, Christof Koch1, Stefan Mihalas1 1Allen Institute for Brain Science, Seattle, WA 98109, USA Correspondence: Jung H. Lee - jungl@alleninstitute.org BMC Neuroscience 2016, 17(Suppl 1):P95 Vasoactive intestinal polypeptide positive (VIP) inhibitory interneurons are commonly found in the superficial layers of cortices [1]. They are distinct from other major cortical inhibitory cell types in terms of connectivity and cellular mechanisms and exclusively inhibit somatostatin (SST) cells in visual cortex. This property is consistent with the interneuron-selective interneuron group, which has been proposed recently [2]. Indeed, bitufted cells in this group express VIP. Moreover, VIP cells have nicotinic receptors rarely found in SST and parvalbumin positive (PV) cells [3]. These recent studies lead to the hypothesis that VIP cells play unique functions in cortical areas, which can be supported with evidence. The optogenetic activation of cingulate cortex of mouse elicited a strong response in VIP cells of V1, suggesting the central roles of VIP cells in top-down gain control [4]. Also, VIP cells are nonspecifically depolarized when a mouse runs [5]. VIP cells disinhibit pyramidal cells by suppressing SST cell activity. That is, when VIP cells are activated, pyramidal cell activity increases due to reduction of inhibition from SST cells, which accounts for the gain modulation. However, the advantage of VIP cell activation induced by locomotion is not clear. We hypothesized that VIP cell activation leads to better perception of moving objects since all visual objects would appear to be in motion when a mouse runs. The strong surround suppression could prevent visual neurons from responding to those effective movements. In this sense, VIP cell activation may be beneficial to the mouse running, as the enhanced VIP cell activity reduces surround suppression. Here we use a computational model of V1 consisting of multiple cortical columns to address if VIP cell activation can enhance perception on moving objects. Our computational model is based on an earlier multiple column model [6], and we refined it by incorporating VIP, SST and PV cells into the superficial layers of the model. To build this refined model, we used two strategies. First, we inferred the time course of synaptic events and the number of connections from both experimental data [7] and parameters from the earlier model. Second, we identified the minimal set of intercolumnar connections necessary for reproducing lateral interactions observed in visual cortex [8]. To examine whether the enhanced VIP cells could enhance responses to moving objects, we simulated a single moving object by stimulating the columns sequentially in the model. Our simulation results support our hypothesis: the column responses during sequential stimulation increase as we increase inputs to VIP cells. Acknowledgements: We wish to thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their vision, encouragement and support. ReferencesRudy B, Fishell G, Lee S, Hjerling-Leffler J. Three groups of interneurons account for nearly 100 % of neocortical GABAergic neurons. Dev Neurobiol. 2011;71:45–61. Jiang X, Shen S, Cadwelll C, Berence P, Sinz F, Ecker AS, et al. Principles of connectivity among morphologically defined cell types in adult neocortex. Science. 2015;350:62–4. Kepecs A, Fishell G. Interneuron cell types are fit to function. Nature. 2014;505:318–26. Zhang S, Xu M, Kamigaki T, Hoang Do JP, Chang W-C, Jenvay S, et al. Long-range and local circuits for top-down modulation of visual cortex processing. Science. 2014;345:660–5. Fu Y, Tucciarone JM, Espinosa JS, Sheng N, Darcy DP, Nicoll RA, et al. A cortical circuit for gain control by behavioral state. Cell. 2014;156:1139–52. Wagatsuma N, Potjans TC, Diesmann M, Sakai K, Fukai T. Spatial and feature-based attention in a layered cortical microcircuit model. PLoS One. 2013;8:e80788. Pfeffer CK, Xue M, He M, Huang ZJ, Scanziani M. Inhibition of inhibition in visual cortex: the logic of connections between molecularly distinct interneurons. Nat Neurosci. 2013;16:1068–76. Xu S, Jiang W, Poo M-M, Dan Y. Activity recall in a visual cortical ensemble. Nat Neurosci. 2012, 15:449–455. P96 Stochastic inference with spiking neural networks Mihai A. Petrovici1,†, Luziwei Leng1,†, Oliver Breitwieser1,†, David Stöckel1,†, Ilja Bytschok1, Roman Martel1, Johannes Bill1, Johannes Schemmel1, Karlheinz Meier1 Kirchhoff-Institute for Physics, University of Heidelberg, Germany Correspondence: Mihai A. Petrovici - mpedro@kip.uni-heidelberg.de † Authors with equal contributions BMC Neuroscience 2016, 17(Suppl 1):P96 Brains are adept at creating an impressively accurate internal model of their surrounding based on incomplete and noisy sensory data. Understanding this inferential prowess is not only interesting for neuroscience, but may also inspire computational architectures and algorithms for solving hard inference problems. Here, we give an overview of our work on probabilistic inference with brain-inspired spiking networks, their advantages compared to classical neural networks and their implementation in neuromorphic hardware. In the neural sampling framework, we interpret spiking activity as sampling from distributions over binary random variables. By exploiting the dynamics of spiking neurons with conductance-based synapses, we have shown that their activation function can become symmetric in the high-conductance state, which in turn enables Glauber-like dynamics in ensembles of noise-driven LIF networks [1, 2]. This allows the straightforward construction of LIF networks that sample from previously defined probability distributions. When the parameters of the distribution are not well-defined, they need to be learned from data. Due to their analogy to classical neural networks such as Boltzmann machines, LIF networks are amenable to the same learning algorithms and can be shown to match the performance of their equally-sized abstract counterparts when trained on classical machine-learning datasets such as MNIST. However, spiking neural networks endowed with short-term plasticity can travel more efficiently through their associated state space, allowing them to simultaneously become good generative and discriminative models of learned data, which is notoriously difficult with conventional techniques such as Gibbs sampling. This finding points towards a distinct advantage of spike-based computation and communication, which is relevant in any scenario where spiking neural networks need to be able to escape local attractors. This computational advantage of spiking sampling networks can be further bolstered by emulation on an accelerated neuromorphic substrate. The core idea behind these devices is the direct emulation of biological neuronal dynamics in VLSI circuits. Such hardware can far surpass simulators running on conventional computing architectures both in terms of speed and power consumption, but with the caveat of having limited parameter precision, as well as other sources of disruptive noise [3, 4]. With some additional modifications, we have shown how LIF networks can become robust to certain types of parameter noise—both during training and during operation– thereby making them amenable to a neuromorphic implementation with an acceleration factor of 104 compared to biological real-time. An even more compelling argument for neuromorphic spike-based inference can be made when considering that learning (in particular, the simulation of synaptic plasticity) is by far the most time-consuming factor in simulations. In an effort to make expectation–maximization learning compatible with existing neuromorphic devices, we have developed a network model that can use double-exponential STDP with 4–6 bit weight resolution for learning and spike-based homeostasis for stabilization and robustness. This research was supported by EU grants #269921 (BrainScaleS), #237955 (FACETS-ITN), #604102 (Human Brain Project) and the Manfred Stärk Foundation. ReferencesPetrovici MA, Bill J, Bytschok I, Schemmel J, Meier K. Stochastic inference with deterministic spiking neurons. arXiv preprint arXiv:1311.3211, 2013. Petrovici MA, Bytschok I, Bill J, Schemmel J, Meier K. The high-conductance state enables neural sampling in networks of LIF neurons. BMC Neurosci. 20150;16(Suppl. 1):O2. Petrovici MA, Vogginger B, Müller P, Breitwieser O, Lundqvist M, Muller L, Ehrlich M, Destexhe A, Lansner A, Schüffny R, et al. Characterization and compensation of network-level anomalies in mixed-signal neuromorphic modeling platforms. PloS One. 2014;9(10):e108590. Pfeil T, Grübl A, Jeltsch S, Müller E, Müller P, Petrovici MA, Schmuker M, Brüderle D, Schemmel J, Meier K. Six networks on a universal neuromorphic computing substrate. Front Neurosci. 2013;7:11. ISSN 1662-453X. doi:10.3389/fnins.2013.00011. P97 Modeling orientation-selective electrical stimulation with retinal prostheses Timothy B. Esler1, Anthony N. Burkitt1, David B. Grayden1, Robert R. Kerr2, Bahman Tahayori3, Hamish Meffin4 1NeuroEngineering Laboratory, Electrical & Electronic Engineering, The University of Melbourne, Parkville VIC 3010, Australia; 2IBM Research, Melbourne, Australia; 3Monash Institute of Medical Engineering, Monash University, Melbourne, Australia; 4National Vision Research Institute, Melbourne, Australia Correspondence: Timothy B. Esler - tesler@student.unimelb.edu.au BMC Neuroscience 2016, 17(Suppl 1):P97 One present challenge in electrical stimulation for epiretinal prostheses is how to avoid stimulating axons of passage in the nerve fiber layer (NFL) that originate from distant regions of the ganglion cell layer (GCL). Co-stimulation of target retinal ganglion cells and overlying axons results in irregular visual percepts, which can significantly limit perceptual efficacy [1, 2]. This research explores how the characteristic distributions of fiber orientation in different retinal layers result in differences between the activation of the axon initial segment and axons of passage. Specifically, axons of passage of retinal ganglion cells are characterized by a narrow distribution of fiber orientations, dominated by the direction of passage towards the optic disk. In contrast, proximal axons in the GCL tend to have a wider distribution of orientations. A model of extracellular stimulation that captures the effects of neurite orientation has been developed using a modified version of the standard volume conductor model, known as the cellular composite model [3], embedded in a four layer model of the retina. The cellular composite model is used in this analysis as it addresses a number of limitations of conventional volume conductor models and more accurately captures the spatiotemporal properties of neural tissue. By generalizing the model to allow for analysis of fibers with arbitrary orientations, simulations have been conducted to investigate the interaction of neural tissue orientation, electrode placement, and stimulation pulse duration and amplitude. Through an exhaustive parameter search, a set of stimulation pulse durations, amplitudes and electrode positions are proposed to achieve selective activation of axon initial segments. Using appropriate multiple electrode configurations and higher frequency stimulation, preferential activation of the axon initial segment is shown to be possible for a range of realistic electrode-retina separation distances (Fig. 58).Fig. 58 A Simulation geometry showing the four modeled layers: insulator (glass), vitreous, NFL, and GCL. Distance from membrane threshold in mV for B parallel axons in a plane in the NFL and C perpendicular axon initial segments in a plane in the GCL, when stimulated with a 300 µs biphasic pulse with electrode-retina separation of 400 µm. Dotted contour marks the threshold level These results establish a quantitative relationship between the time-course of stimulation and physical properties of the tissue, such as fiber orientation. ReferencesFried SI, Lasker ACW, Desai NJ, Eddington DK. Axonal sodium-channel bands shape the response to electric stimulation in retinal ganglion cells. J Neurophysiol. 2009;101(4):1972–87. Rattay F, Resatz S. Effective electrode configuration for selective stimulation with inner eye prostheses. IEEE Trans Biomed Eng. 2004;51(9):1659–64. Meffin H, Tahayori B, Sergeev EN, Mareels IMY, Grayden DB, Burkitt AN. Modelling extracellular electrical stimulation: III. Derivation and interpretation of neural tissue equations. J Neural Eng. 2014;11(6):065004. P98 Ion channel noise can explain firing correlation in auditory nerves Bahar Moezzi1, Nicolangelo Iannella1,2, Mark D. McDonnell1 1Computational and Theoretical Neuroscience Laboratory, School of Information Technology and Mathematical Sciences, University of South Australia, Australia; 2School of Mathematical Sciences, University of Nottingham, Nottingham, UK Correspondence: Bahar Moezzi - bahar.moezzi@unisa.edu.au BMC Neuroscience 2016, 17(Suppl 1):P98 Neural spike trains are commonly characterized as a Poisson point process. However, the Poisson assumption is a poor model for spiking in auditory nerve fibers because it is known that interspike-intervals display positive correlation over long time scales and negative correlation over shorter time scales. It has been suggested that ion channel opening and closing might not be well described by Markov models. Instead, fractal ion channel gating could be used to take into account the involvement of proteins in the conformational changes of sub-states in the channel gating kinetics. Using a detailed biophysical model, we tested the hypothesis that fractal ion channel gating is responsible for short and long term correlations in the auditory nerve spike trains. We developed a biophysical model based on the well-known Meddis model of the peripheral auditory system [1]. We introduced biophysically realistic ion channel noise to an inner hair cell membrane potential model that includes (i) fractal fast potassium channels, (ii) deterministic slow potassium channels, and (iii) a stochastic Markov model for noisy calcium channels. We used Fano factor as a measure of firing correlation. We showed that the resulting simulated Fano factor time curves have all the common attributes of the Fano factor of experimentally recorded spike trains in the auditory nerve fibers, except the time scale of corelation. Our model thus replicates macro-scale stochastic spiking statistics in the auditory nerve fibers due to modeling stochasticity at the micro-scale of potassium and calcium ion channels. ReferenceMeddis R. Auditory-nerve first-spike latency and auditory absolute threshold: a computer model. J Acoust Soc Am. 2006;119(1):406–17. P99 Limits of temporal encoding of thalamocortical inputs in a neocortical microcircuit Max Nolte1, Michael W. Reimann1, Eilif Muller1, Henry Markram1 1Blue Brain Project, École Polytechnique fédérale de Lausanne (EPFL), Geneva, Switzerland Correspondence: Max Nolte - max.nolte@epfl.ch BMC Neuroscience 2016, 17(Suppl 1):P99 During naturalistic whisker motion, subsets of neurons in the same barreloid of the rat ventroposterior medial thalamus (VPM) respond synchronously with temporal precision to different kinetic features of whisker movement (spike-time coding) [1]. Multiple synchronously firing VPM cells can trigger temporally precise responses in the somatosensory cortex, such as those observed during full whisker deflection or active touch, but the minimum number of synchronously firing VPM cells needed to reliably drive the spiking of cortical cells is not known. In this study, we use the Blue Brain Project’s digital reconstruction of a somatosensory microcircuit of a juvenile rat [2] to characterize how many synchronously firing VPM cells are needed to reliably drive individual cells of different morphological types in the rat somatosensory cortex. We activate an increasing number of synchronously firing VPM fibers (with in vivo VPM spike trains from experiments published in [1]) in both simulations of single cells, and simulations of the whole reconstructed microcircuit with only a small number of active VPM fibers. We find that inhibitory neurons in layers 3 and 4 quickly approach maximum spike-timing reliability when receiving input from 10 to 15 synchronously firing VPM neurons. Excitatory neurons in layers 3 and 4 require substantially more synchronous VPM fibers, but less than excitatory neurons in layers 5 and 6 (see Fig. 59). With an average of eight synapses per connection, these numbers are significantly higher than what has been observed in a previous in silico study in the cat visual cortex [3]. In addition to the difference in animal and sensory system, we show that this decrease of reliability can be partly explained by a lower synaptic release probability in vivo than in vitro caused by a lower extracellular calcium concentration in vivo [4], which is taken into account in our simulations [2].Fig. 59 A Mean spike-timing reliability (similar correlation-based measure as in [3], but with firing rate adaption). The reliability of the VPM input is 0.55. B Mean probability of firing within 2–12 ms after the initial input VPM spike in each trial. C Mean ratio of spikes occurring within 2–12 ms after a VPM spike, out of all spikes. Mean of 30 (L3/4 excitatory), 50 (L5/6 exc.), 40 (L3/4 inhibitory) and 30 cells (L5/6 inh.) respectively Finally, we describe how the requirement for synchronous, redundant VPM inputs limits the maximum amount of asynchronous, temporally precise VPM activity (in subsets of synchronous VPM neurons) that can be reliably encoded in a neocortical microcircuit. Acknowledgements: This work was supported by funding from the ETH Domain for the Blue Brain Project (BBP). The BlueBrain IV IBM BlueGene/Q system is financed by ETH Board Funding to the Blue Brain Project and hosted at the Swiss National Supercomputing Center (CSCS). We thank M. Bale and R. Petersen for providing the VPM spike trains. ReferencesBale MR, Ince RAA, Santagata G, Petersen RS. Efficient population coding of naturalistic whisker motion in the ventro-posterior medial thalamus based on precise spike timing. Front Neural Circuits. 2015;50. Markram H, et al. Reconstruction and simulation of neocortical microcircuitry. Cell. 2015;163(2):456–92. Wang H, Spencer D, Fellous J, Sejnowski T. Synchrony of thalamocortical inputs maximizes cortical reliability. Science. 2010;328:106–9. Borst JGG. The low synaptic release probability in vivo. Trends Neurosci. 2010;33(6):259–66. P100 On the representation of arm reaching movements: a computational model Antonio Parziale1, Rosa Senatore, Angelo Marcelli1 Department of Information and Electrical Engineering, University of Salerno, 84084, Fisciano (SA), Italy Correspondence: Antonio Parziale - anparziale@unisa.it BMC Neuroscience 2016, 17(Suppl 1):P100 Experimental studies on the spinal cord (SC) have shown that SC is not a simple relay station for transmitting information to and from supraspinal centers but “it is a highly evolved and complex part of the CNS that has considerable computational ability” [1]. Limb movements are planned and initiated by the brain but they cannot be performed without a spinal cord and the intricate feedback systems that reside within it [2]. In the last years, computational models have been devised in order to explain the role of the spinal cord in the translation from motor intention to motor execution [3], in sensorimotor control and learning of movements [4], in investigating how the supraspinal centers can control the cord [5], for providing evidence that CNS can plan and control movements without a representation of complex bodily dynamics because the creation and coordination of dynamic muscle forces is entrusted to the spinal feedback mechanisms [6], for investigating how the central nervous system coordinates the activation of both α and γ motoneurons during movement and posture [7]. Here we propose a computational model of the local interneuron networks within SC to evaluate how spinal and supraspinal centers can interact for performing a movement. We model a one-degree of freedom system representing an arm learning and executing reaching movements. The model incorporates the key anatomical and physiological features of the neurons in SC, namely interneurons Ia, Ib and PN and Renshaw cells, and their interconnections [2]. The model envisages descending inputs coming from both rostral and caudal M1 motor cortex and cerebellum (through the rubro- and reticulo-spinal tracts), local inputs from both Golgi tendon organs and spindles, and its output is directed towards α motoneurons, which also receive descending inputs from the cortex and local inputs from spindles. The model envisages virtual muscle [8] for modeling musculoskeletal mechanics and proprioceptors. Our simulations show that the CNS may produce elbow flexion movements with different properties by adopting different strategies for the recruitment and the modulation of interneurons and motoneurons. One interesting results is that the speed-accuracy tradeoff predicted by the Fitts’ law [9] does not follow from the structure of the system, that is capable of performing fast and precise movements, but arises from the strategy adopted to produce faster movements, by starting from a pre-learned set of motor commands useful to reach the target position and by modifying only the activations of the PN and α neurons. Other simulations show that when a suddenly variation of the target position happens after the onset of a learned movement, the descending inputs from the cerebellum can be exploited for the online correction of the movement trajectory by regulating the activity of PN cells. This result agrees with the experimental studies suggesting that the CNS modulates interneurons networks to execute a visually guided online correction. ReferencesBurke RE. Spinal cord. Scholarpedia. 2008;3(4):1925. Pierrot-Deseilligny E, Burke DJ. The circuitry of the human spinal cord: neuroplasticity and corticospinal mechanisms. Cambridge: Cambridge University Press; 2012. Bullock D, Grossberg S. VITE and FLETE: Neural modules for trajectory formation and tension control. Volitional Action. 1989;253–97. Tsianos GA, Goodnes J, Loeb GE. Useful properties of spinal circuits for learning and performing planar reaches. J Neural Eng. 2014;11:1–21. Raphael G, Tsianos GA, Loeb GE. Spinal-like regulator facilitates control of a two-degree-of-freedom wrist. J Neurosci. 2010;30:9431–44. Buhrmann T, Di Paolo EA. Spinal circuits can accommodate interaction torques during multijoint limb movements. Front Comput Neurosci. 2014;8:1–18. Li S, Hao M, He X, Marquez JC, Niu CM, Lan N. Coordinated alpha and gamma control of muscles and spindles in movement and posture. Front Comput Neurosci. 2015;9:1–15. Cheng EJ, Brown IE, Loeb GE: Virtual Muscle: a computational approach to understanding the effects of muscles properties on motor control. J Neurosci Methods. 2000;101:117–30. Fitts PM. The information capacity of the human motor system in controlling the amplitude of movement. J Exp Psychol. 1954;47(6):381–91. P101 A computational model for investigating the role of cerebellum in acquisition and retention of motor behavior Rosa Senatore1,2, Antonio Parziale1, Angelo Marcelli1 1Department of Information and Electrical Engineering and Applied Mathematics, University of Salerno, Fisciano (SA), 81100, Italy; 2Laboratory of Neural Computation, Istituto Italiano di Tecnologia, Rovereto (TN), 38068, Italy Correspondence: Rosa Senatore - rsenatore@unisa.it BMC Neuroscience 2016, 17(Suppl 1):P101 Experimental studies on the cerebellum (CB) have provided a large body of knowledge about its anatomical and physiological features, the neural processes and the phenomena of synaptic plasticity occurring within both the cerebellar cortex and nuclei [1]. The emerging picture is that the CB plays a crucial role in the acquisition and/or retention of motor behaviors and is involved in several cognitive functions [2, 3], therefore several CB models and simulations of its neural processes have been proposed [3–5]. Here we investigated, through a modeling approach, the role of the CB in three different behaviors: vestibulo-ocular reflex (VOR) adaptation, motor learning, and eyeblink conditioning. Different cerebellar areas are involved in these functions: the control of the amplitude and timing of the VOR involves the vestibulocerebellum, learning novel limb movements involves the lateral cerebellar cortex and its connections to the dentate nucleus and acquisition of the eyeblink conditioned responses involves cerebellar cortex areas (lobule HVI) connected to the interposed nucleus[1, 3]. It is noteworthy that the CB is characterized by the remarkable regularity and geometrical structure of its circuits: cerebellar neurons are arranged in a highly regular manner as repeating units, the cerebellar microcomplexes [1]. Therefore the uniform structure of the CB and the contribution of different cerebellar areas to specific behaviors raise the possibility that different behaviors are based on a common ‘neural computation within the cerebellum’. We developed a model (using the Leabra framework in emergent neural simulation software [6]) that incorporates the key anatomical and physiological features of the cerebellar microcomplex, whose behavior was analyzed for investigating the neural processes occurring during the acquisition of novel motor behaviors, classically conditioned responses and VOR adaptation. Since the neural circuits involved in these behaviors present some differences, in terms of the input/output areas sending signals to or receiving signals from the CB, we developed three models, which share the same core network, made up of a set of cerebellar microcomplexes (comprising cerebellar cortex neurons and their connections to nuclear and olivary neurons), but which include different anatomical connections from/to different extra cerebellar regions: (a) “VOR model”, comprising the vestibulocerebellum (flocculus and vestibular nucleus) and its connections with the dorsal cap region of the inferior olive and oculomotor nuclei; (b) “Motor model”, comprising the lateral cerebellum (lateral cerebellar cortex and dentate nucleus) and its anatomical connections with the inferior olive, thalamus and motor cortex; (c) “Conditioning model”, comprising the lobule HVI of the cerebellar cortex and its connections to the interposed nucleus, and their external connections with the dorsal accessory olive, red nucleus and oculomotor nuclei. Our simulations suggest that the CB performs the same computational operation on whichever afferent information it receives, that the appearance of the ‘teaching’ signal conveyed by the climbing fibers could be the explanation for functional differentiation and that different types and sites of synaptic plasticity are involved in different behaviors. ReferencesGhez C, Thach WT. The cerebellum. In: Kandel ER, Schwartz JH, Jessell TM, editors. Principles of neural science. McGraw-Hill; 2000. p. 832–52. Koziol LF, Budding D, Andreasen N, D’Arrigo S, Bulgheroni S, Imamizu H, Ito M, Manto M, Marvel C, Parker K, et al. Consensus paper:the cerebellum’s role in movement and cognition. Cerebellum. 2014;13(1):151–77. Manto M, Bower JM, Conforto AB, Delgado-Garcia JM, da Guarda SN, Gerwig M, Habas C, Hagura N, Ivry RB, Marien P, et al. Consensus paper: roles of the cerebellum in motor control–the diversity of ideas on cerebellar involvement in movement. Cerebellum. 2012;11(2):457–87. Houk JC, Buckingham JT, Barto AG. Models of the cerebellum and motor learning. Behav Brain Sci. 1996;19(3):368–83. Medina JF, Mauk MD. Computer simulation of cerebellar information processing. Nat Neurosci. 2000;3(Suppl.):1205–11. Aisa B, Mingus B, O’Reilly R. The emergent neural modeling system. Neural Netw. 2008;21(8):1146–52. P102 The emergence of semantic categories from a large-scale brain network of semantic knowledge K. Skiker1, M. Maouene2 1LIST Laboratory, FST, Abdelmalek Essaadi’s University, Tangier, Morocco; 2Department of computer science, ENSAT, Abdelmalek Essaadi’s University, Tangier, Morocco Correspondence: K. Skiker - skiker.kaoutar85@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P102 In cognitive neuroscience, the issue of how semantic categories (e.g. animals, tools, fruits/vegetables) are organized in the brain is still debated (Caramazza and Mahon 2006). Some authors postulate that semantic categories are explicitly represented in specific brain areas developed through evolutionary pressure for rapid classification and categorization of animals, tools and foods (Caramazza and Shelton 1998). Other researches argue that semantic categories are not explicitly represented, instead emerge from distributed semantic knowledge (Martin 2007; Tyler and Moss 2001). However, a little is known about how semantic knowledge is structured within the brain for fast and efficient emergence of semantic categories. In this paper, we hypothesize that semantic knowledge is supported by a large-scale brain network that shows the properties of segregation and integration. To test this hypothesis, we first examine where semantic knowledge is nested in the brain; we present functional neuroimaging studies suggesting that semantic knowledge (e.g. visual, auditory, tactile; action, olfactory/gustatory) is grounded in modality specific association brain areas (e.g. visual association areas, auditory association areas, somatosensory association areas) (Barsalou 2008; Goldberg et al. 2006). Then, we derive the connectivity between brain areas where semantic knowledge is nested from Hagmann’s connectivity matrix (Hagmann et al. 2008) freely available. Finally, we examine the properties of the connectivity matrix using graph measures including clustering coefficient and characteristic path length. Our findings show that a large-scale brain network of features exhibit the small world property with high clustering coefficient (C = 0.48) and low path length (L = 2.49).These properties indicate a balance between segregation (high clustering) and integration (low path length) that are essential for the fast and efficient emergence of semantic categories from distributed semantic knowledge. ReferencesBarsalou, Lawrence W. Grounded cognition. Annu Rev Psychol. 2008;59: 617–45. Caramazza A, Mahon BZ. The organisation of conceptual knowledge in the brain: the future’s past and some future directions. Cogn Neuropsychol. 2006;23(1):13–38. Caramazza A, Shelton JR. Domain-specific knowledge systems in the brain the animate-inanimate distinction. J Cogn Neurosci. 1998;10(1):1–34. Goldberg RF, Perfetti CA, Schneider W. Perceptual knowledge retrieval activates sensory brain regions. J Neurosci. 2006;26(18):4917–21. Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, Wedeen VJ, Sporns O. Mapping the structural core of human cerebral cortex. PLoS Biol. 2008;6(7): e159. Martin A. The representation of object concepts in the brain. Annu Rev Psychol. 2007;58:25–45. Tyler LK, Moss HE. Towards a distributed account of conceptual knowledge. Trends Cogn Sci. 2001;5(6):244–52. P103 Multiscale modeling of M1 multitarget pharmacotherapy for dystonia Samuel A. Neymotin1,2, Salvador Dura-Bernal1, Alexandra Seidenstein1,3, Peter Lakatos4, Terence D Sanger5,6, and William W Lytton1,7 1Department Physiology & Pharmacology, SUNY Downstate, Brooklyn, NY 11203, USA; 2Department Neuroscience, Yale University School of Medicine, New Haven, CT, USA; 3Departmentt of Chemical and Biomedical Engineering, Tandon School of Engineering, NYU, Brooklyn, NY, USA; 4Nathan Kline Institute for Psychiatric Research, Orangeburg, NY, USA; 5Department Biomedical Engineering, University of Southern California, Los Angeles, CA, USA; 6Div Neurology, Child Neurology and Movement Disorders, Children’s Hospital Los Angeles, LA, CA, USA; 7Department Neurology, Kings County Hospital Center, Brooklyn, NY 11203, USA Correspondence: Samuel A. Neymotin - samn@neurosim.downstate.edu BMC Neuroscience 2016, 17(Suppl 1):P103 Dystonia is a movement disorder that produces involuntary sustained muscle contractions. Different types of dystonia likely involve primary or induced pathologies across multiple brain areas including basal ganglia, thalamus, cerebellum, and sensory and motor cortices. Due to lack of therapeutic alternatives, much current treatment involves paralyzing affected muscles directly with painful injections of botulinum toxin. Primary motor cortex (M1) represents a potential target for therapy. M1 pathological dynamics in some forms of dystonia include hyperexcitability and altered beta oscillations. In order to further develop understanding of motor cortex involvement in this disease and to look at potential drug cocktails (multitarget polypharmacy), we developed a multiscale model of M1 across spatial scales, ranging from molecular interactions, up to cellular and network levels. The model contains 1715 compartmental model neurons with multiple ion channels and intracellular molecular dynamics [1, 2]. Wiring and arrangements of cellular layers of the model was based on previously recorded electrophysiological data obtained from mouse M1 circuit mapping experiments. Simulations were run in the NEURON simulator and intracellular dynamics utilized the reaction–diffusion module [3]. The chemophysiological component of the simulation focused on calcium (Ca) handling, and Ca regulation of hyperpolarization-activated cyclic nucleotide-gated (HCN) channels. The Ca signaling was modeled in conjunction with intracellular cytosolic and endoplasmic reticulum (ER) volumes, inositol triphosphate (IP3) production via a metabotropic glutamate receptor signaling cascade, and ER IP3 and ryanodine receptors (RYR) which release ER Ca into the cytosol. The model reproduced the pathological dynamics providing hyperexcitability and synchronous beta oscillations across cortical layers. We applied independent random variations to multiple ion channel densities (multiple cell membrane channels: HCN, channels for Na, K, Ca; RYR, IP3 channels in ER), to identify pathological and physiological simulation sets. Experiments with these models demonstrated degeneracy, with multiple routes that produced the pathological syndrome. In most cases, there was no single parameter alteration which would induce the change from pathological to physiological dynamics. We used support vector machines to assess the high dimensional parameter space to provide overall direction for passage from an overall pathological to an overall physiological region of parameter space, enabling prediction of multitarget drug cocktails that would be likely to move the system from dystonic to physiological dynamics. Acknowledgements: Research supported by NIH grant R01 MH086638, NIH grant U01 EB017695, NIH grant R01 NS064046, NIH grant R01 DC012947. ReferencesNeymotin SA, McDougal RA, Bulanova AS, Zeki M, Lakatos P, Terman D, Hines ML, Lytton WW. Calcium regulation of HCN channels supports persistent activity in a multiscale model of neocortex. Neuroscience. 2016;316:344–66. Neymotin SA, McDougal RA, Sherif MA, Fall CP, Hines ML, Lytton WW. Neuronal calcium wave propagation varies with changes in endoplasmic reticulum parameters: a computer model. Neural Comput. 2015;27:898–924. McDougal RA, Hines ML, Lytton WW. Reaction–diffusion in the NEURON simulator. Front Neuroinform. 2013;7:28. P104 Effect of network size on computational capacity Salvador Dura-Bernal1, Rosemary J. Menzies2, Campbell McLauchlan2, Sacha J. van Albada3, David J. Kedziora2, Samuel Neymotin1, William W. Lytton1, Cliff C. Kerr2 1Department of Physiology & Pharmacology, SUNY Downstate Medical Center, Brooklyn, NY 11023, USA; 2Complex Systems Group, School of Physics, University of Sydney, Sydney, NSW 2006, Australia; 3Institute of Neuroscience and Medicine (INM-6), Jülich Research Centre and JARA, Jülich, Germany Correspondence: Cliff C. Kerr - cliff@thekerrlab.com BMC Neuroscience 2016, 17(Suppl 1):P104 There is exceptionally strong circumstantial evidence that organisms with larger nervous systems are capable of performing more complex computational tasks. Yet relatively few studies have investigated this effect directly, instead typically treating network size as a fixed property of a simulation while exploring the effects of other parameters. Recently, Diehl and Cook [1] found that network performance did increase modestly with network size; however, larger networks also required longer training times to achieve a given performance. In this work, we directly addresses the relationship between network size and computational capacity by using a biomimetic spiking network model of motor cortex to direct a virtual arm towards a target via reinforcement learning [2]. The reaching task was performed by a two-joint virtual arm controlled by four muscles (flexor and extensor muscles for shoulder and elbow joints). These muscles were controlled by a neural model that consisted of excitatory and inhibitory Izhikevich neurons in three cortical populations: a proprioceptive population, which received input from the current arm position; a motor population, which was used to drive the arm muscles; and a sensory population, which served as the link between the proprioceptive and motor populations. The model was trained to reach the target using exploratory movements coupled with reinforcement learning and spike-timing dependent plasticity (STDP). The model was implemented using NEURON. A major challenge in scaling network size is that not all properties of the network can be held constant. As shown by van Albada et al. [3], while first-order properties (such as average firing rate) can be maintained, there are limitations in preserving second- and higher-order statistical properties (such as noise correlations). Thus, we explored multiple different ways of scaling the connectivity of the network, including (a) preserving connection probability, scaling connection weight to be inversely proportional to model size, and increasing the variance of the external drive; and (b) reducing connection probability to preserve average node degree and leaving other parameters unchanged. In addition, we explored scaling each of the neuronal population groups versus only scaling the sensory (processing) population group. Large differences were observed in network dynamics and statistics based on different scaling choices. However, the relationship between network size and task performance was significant only for certain specific choices of model parameters. Overall, task performance is highly sensitive to the network’s metaparameters, such as STDP learning rates. We found that these must be optimized specifically for different network sizes; otherwise, differences in suitability of these parameters overwhelm the intrinsic advantages of larger networks. In conclusion, while network size does affect computational capacity, the relationship is strongly dependent on the manner in which the scaling is implemented. ReferencesDiehl PU, Cook M. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Front Comp Neurosci. 2015;9:99. Dura-Bernal S, Li K, Neymotin SA, Francis JT, Principe JC, Lytton WW. Restoring behavior via inverse neurocontroller in a lesioned cortical spiking model driving a virtual arm. Front Neurosci. 2016;10:28. Van Albada SJ, Helias M, Diesmann M. Scalability of asynchronous networks is limited by one-to-one mapping between effective connectivity and correlations. PLoS Comput Biol. 2015;11:e1004490. P105 NetPyNE: a Python package for NEURON to facilitate development and parallel simulation of biological neuronal networks Salvador Dura-Bernal1, Benjamin A. Suter2, Samuel A. Neymotin1, Cliff C. Kerr3, Adrian Quintana4, Padraig Gleeson4, Gordon M. G. Shepherd2, William W. Lytton1 1Department Physiology & Pharmacology, SUNY Downstate, Brooklyn, NY 11203, USA; 2Department Physiology, Northwestern University, Chicago, IL 60611, USA; 3Complex Systems Group, School of Physics, University of Sydney, Sydney, NSW 2006, Australia; 4Department of Neuroscience, Physiology & Pharmacology, University College London, London WC1E6BT, UK Correspondence: Salvador Dura-Bernal - salvadordura@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P105 NEURON is a widely used neuronal simulator, with over 1600 published models. It enables multiscale simulation ranging from the molecular to the network level. However, learning to use NEURON, especially running parallel simulations, requires much technical training. NetPyNE (Network development Python package for NEURON) greatly facilitates the development and parallel simulation of biological neuronal networks in NEURON, potentially bringing its benefits to a wider audience, including experimentalists. It is also intended for experienced modelers, providing powerful features to incorporate complex anatomical and physiological data into models. NetPyNE seamlessly converts a set of high-level specifications into a NEURON model. Specifications are provided in a simple, standardized, declarative format, based solely on Python’s lists and dictionaries. The user can define network populations and their properties, including cell type, number or density. For each cell type, the user can define morphology, biophysics and implementation, or choose to import these from existing files (HOC templates or Python classes). Cell models for each population can be easily changed, and several models can be combined to generate efficient hybrid networks, e.g. composed of Hodgkin–Huxley multicompartment cells and Izhikevich point neurons. NetPyNE provides an extremely flexible format to specify connectivity, with rules based on pre- and post-synaptic cell properties, such as cell type or location. Multiple connectivity functions are available, including all-to-all, probabilistic, convergent or divergent. Additionally, connectivity parameters (e.g. weight, probability or delay) can be specified as a function of pre/post-synaptic spatial properties. This enables implementation of complex biological patterns, such as delays or connection probabilities that depend on distance between cells, or weights that depend on the post-synaptic neuron’s cortical depth. The subcellular distribution of synapses along the dendrites can be specified, and is automatically adapted to the morphology of each model neuron. Learning mechanisms, including spike-timing dependent plasticity and reinforcement learning, can be readily incorporated. Using the high-level network specifications, NetPyNE instantiates the full model (all cells and connections) as a hierarchical Python structure including the NEURON objects necessary for simulation. Based on a set of simulation options (e.g. duration, integration step), NetPyNE runs the model in parallel using MPI, eliminating the burdensome task of manually distributing the workload and gathering data across computing nodes. Optionally NetPyNE plots output data, such as spike raster plots, LFP power spectra, connectivity matrix, or intrinsic time-varying variables (e.g. voltage) of any subset of cells. To facilitate data sharing, the package saves and loads the high-level specifications, instantiated network, and simulation results using common file formats (Pickle, Matlab, JSON or HDF5). NetPyNE can convert instantiated networks to and from NeuroML, a standard data format for exchanging models in computational neuroscience. NetPyNE has been used to develop a variety of multiscale models: primary motor cortex with cortical depth-dependent connectivity; the claustrum; and sensorimotor cortex that learns to control a virtual arm. The package is open source, easily installed, and includes comprehensive online documentation, a step-by-step tutorial and example networks (www.neurosimlab.org/netpyne). We believe this tool will strengthen the neuroscience community and encourage collaborations between experimentalists and modelers. Acknowledgements: Research supported by NIH grant U01 EB017695 and DARPA grant N66001-10-C-2008. P107 Inter-areal and inter-regional inhomogeneity in co-axial anisotropy of Cortical Point Spread in human visual areas Juhyoung Ryu1, Sang-Hun Lee1 1Brain and Cognitive Science, Seoul National University, Seoul 151-742, Republic of Korea Correspondence: Juhyoung Ryu - jh67753737@snu.ac.kr, visionsl@snu.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P107 A focal visual stimulus can evoke widespread neural activation far beyond the directly stimulated site, a phenomenon referred to as the “cortical point spread” (CPS). The lateral connections among neurons in the early visual cortex have been proposed as a likely anatomical conduit for the CPS, and recent functional studies on humans [1] and non-human primates [2] demonstrated that the CPS is spatially anisotropic, spread preferentially along with the axis of stimulus orientation, dubbed as ‘coaxial anisotropy.’ Although these two seminal studies documented the coaxial anisotropy robustly in two different species, there are several remaining questions to be further explored. First, previous human psychophysical studies reported substantial degrees of inhomogeneity in association field characteristics over the visual space (e.g., crowding effects), which implies the presence of corresponding inhomogeneity of coaxial anisotropy. Second, the animal study [2] examined the coaxial anisotropy only from V1 of two monkeys, and the human study [1] reported a substantial degree of individual differences and inter-areal differences. The current study is set out to address these two aspects of coaxially anisotropic CPS. We acquired time series of functional magnetic resonance imaging (fMRI) measurements in V1 while human individuals viewed a ring or wedge of Gabor patches that slowly drifted along the radial or tangential axis over a spatially extended (up to 8° in radius) region of retinotopic space (Fig. 60A). The orthogonal combination of two different drifting direction and stimulus orientation generated two interesting viewing conditions: coaxial and orthoaxial conditions (boxed and unboxed panels, respectively, in Fig. 60A). For individual gray matter units (2 mm iso volume voxels) in the early visual cortex (V1, V2, V3), we quantified the degree and sign of coaxial anisotropy by comparing the width of fMRI response profiles between the coaxial and orthoaxial conditions. In specific, we first estimated the width of CPS at the half of its maximum response respectively for two viewing conditions − coaxial condition (Wc) and orthoaxial condition (Wo), then computed coaxial anisotropy index by taking the singed contrast between these two width estimates: CAI = (Wc − Wo)/(Wc + Wo).Fig. 60 Stimuli and fMRI results. A The snapshot of traveling Gabors are shown for the four different conditions. The black arrows represent a moving direction of wedge or ring. B Significant (yellow, t test p < 0.001) coaxial anisotropy in all subjects. C Coaxial anisotropy across visual areas (V1, V2, V3) Results The results replicated those in the previous study [1]: in all of the subjects inspected, the width of CPS was significantly greater along the coaxial axis than along the orthoaxial axis (Student’s t test, p < 0.001), and the CAIs ranged from +0.05 to +0.15 (Fig. 60B). In addition, we found two interesting new findings: first, coaxial anisotropy tended to decrease along the processing hierarchy (V1 > V2 > V3; Fig. 60C); second, coaxial anisotropy tended to be more pronounced along the cardinal axes (horizontal meridians in particular) in retinotopic space. ReferencesPark SH, Cha K, Lee SH. Coaxial anisotropy of cortical point spread in human visual areas. J Neurosci. 2013; 33(3):1143–56a. Michel MM, Chen Y, Geisler WS, Seidemann E. An illusion predicted by V1 population activity implicates cortical topography in shape perception. Nat Neurosci. 2013; 16(10):1477–83. P108 Two bayesian quanta of uncertainty explain the temporal dynamics of cortical activity in the non-sensory areas during bistable perception Joonwon Lee1, Sang-Hun Lee1 1Department of Brain and Cognitive Sciences, Seoul National University, Seoul 151-742, Korea Correspondence: Joonwon Lee - jwl89@snu.ac.kr; visionsl@snu.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P108 Bistable perception—involuntary fluctuation over time in perceptual appearance despite unchanging physical stimulation on sensory organs—has been a popular tool for exploring neural loci substantiating transitions in perceptual awareness. To track observers’ ongoing percepts, experimenters make the observers actively report binary states of their perception. Unfortunately, observers’ engagement in perceptual tracking is likely to invite the brain activities that are not associated directly with perceptual transition per se, but rather reflect the cognitive processes ensuing from temporal sequences of perceptual transitions. We reasoned that, in a computational perspective, the latter, post-transition component of bistable perception is majorly driven by two separate quanta of uncertainty: unexpected (UU) and expected uncertainty (EU), which have been proposed to relate to neuromodulatory systems of norepinephrine and acetylcholine, respectively [1]. We found the Dynamic Belief Model (DBM) [2] particularly relevant to the concurrent measurement of those two quanta of uncertainty (i) because it is capable of updating the moment-to-moment, expected probability of binary events based on a recent history of those events (Fig. 61A) and (ii) because this probability directly estimates EU whereas the disparity between the expected probability and perceived outcome quantize UU.Fig. 61 Bayesian estimation to predict BOLD dynamics around switch. A Bayesian inference model of iteratively updated prior, input likelihood, and combined posterior. B (Upper) Uncertainty-driven BOLD estimated from Bayesian model locked to transition under different duration conditions. Time-series is built purely from real behavior history. (Lower) Average % BOLD signal of ACC region in 8 subjects (14 sessions) With the DBM in hand, we predicted the time courses of UU and EU (red curves in Fig. 61B), and explored the cortical loci substantiating those two kinds of uncertainty by acquiring fMRI measurements while human observers viewed a ‘structure-from-motion (SfM)’ display, in which ambiguous 2D motion of coherently moving dots gives perceptual alternations in 3D motion perception between bistable states, clockwise vs counterclockwise rotational motion. To compensate for the temporal resolution of fMRI activity, we slowed down the dynamcis of bistable perception using the intermittent stimulation technique, which allowed us to identify gray-matter units (voxels) whose variability in fMRI time course can be explained by UU or EU. As expected from previous studies, cortical activity increased substantially during the transition periods in many distributed brain regions. More importantly, the fMRI time series in these transition-locked regions were explained by the weighted linear sum of the time series of UU and EU quantity, some exhibiting greater weights for UU and others greater weights for EU. In additions, the time series of pupil size of the observers resembled the predicted time courses of UU, consistent with the previously reported tight linkage between UU and the LC-NE system. We conclude that the cortical activities previously claimed as being responsible for triggering perceptual transition are likely to reflect two post-transition cognitive quanta of uncertainty. ReferencesYu AJ, Dayan P. Uncertainty, neuromodulation, and attention. Neuron. 2005;46:681–92. Yu AJ, Cohen JD. Sequential effects: superstition or rational behavior? NIPS. 2009;21:1873–80. P109 Optimal and suboptimal integration of sensory and value information in perceptual decision making Hyang Jung Lee1, Sang-Hun Lee1 1Department of Brain and Cognitive Neuroscience, Seoul National University, Gwanak-gu, South Korea Correspondence: Hyang Jung Lee - hyangjung.lee@snu.ac.kr, visionsl@snu.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P109 Optimization of decision-making often requires the effective integration of sensory and value information, particularly when sensory inputs are ambiguous and the criterion for the successful decision changes stochastically over time (e.g., a hitter making ‘strike’ or ‘ball’ decisions by integrating visual information and an umpire’s calls). We adapted this ‘sensory-value integration’ situation to a laboratory, where 30 human subjects classified ring stimuli into ‘small’ or ‘large’ based on the perceived ring size and the trial-by-trial feedback (‘correct’ or ‘incorrect’) for judgment. The key manipulation was, unbeknownst to subjects, to induce a slight amount of bias, favoring either ‘small’ or ‘large’ choices, or staying ‘unbiased’. Inspired by previous animal studies (Corrado et al. 2005; Lau and Glimcher 2005; Busse et al. 2011), we developed a Linear-Nonlinear-Poisson model to describe the dynamics of how humans adapt their moment-to-moment perceptual decision to subtle, yet volatile environmental feedbacks. The integration process takes place at the Linear stage where a decision variable is formed by combining the sensory information in the current trial and the reward/choice information histories. This is followed by the nonlinear stage where softmax rule is applied to translate the decision variable into probability for the ensuing Poisson stage. Fitting the model to the data set of each individual allowed us to explore the individual differences in optimal sensory-value integration in our task. Our L–N–P model effectively depicted, and generated as well, the temporal dynamics of human subjects’ perceptual choices made in an environment with volatile and stochastic feedbacks. The correlation analysis identified a set of latent model parameters (e.g., reward kernel weight) that are tightly linked to the individual differences in ability to adapt their decision to abrupt changes in feedback. The ideal decision-maker analysis indicated that human subjects are generically suboptimal in forming an effective reward kernel for translating feedbacks in the past trials into action values in the upcoming trials. Acknowledgements: Supported by National Research Foundation of Korea, NRF-2013R1A2A2A03017022. ReferencesBusse L, Ayaz A, Dhruv, NT, Katzner S, Saleem, AB, Schölvinck ML, Carandini M. The detection of visual contrast in the behaving mouse. J Neurosci. 2011;31(31):11351–61. Lau B, Glimcher PW. Dynamic response-by-response models of matching behavior in rhesus monkeys. J Exp Anal Behav. 2005;84(3):555–79. Corrado GS, Sugrue LP, Seung, HS, Newsome, WT. Linear–nonlinear–Poisson models of primate choice dynamics. J Exp Anal Behav. 2005;84(3):581. P110 A Bayesian algorithm for phoneme Perception and its neural implementation Daeseob Lim1, Sang-Hun Lee1 1Department of Brain and Cognitive Sciences, Seoul National University, Seoul, 08826, South Korea Correspondence: Daeseob Lim - daeseob@snu.ac.kr, visionsl@snu.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P110 Human listeners seem effortless in recognizing a rapid stream of speech sounds uttered by their fellow speakers, thus being capable of readily participating in conversation. However, it remains poorly understood how the brain represents the basic processing unit of such fluent speech perception, phoneme. In a computational perspective, phoneme perception is a reverse engineering of speech production, where the goal is to infer from noisy acoustic signal which phonetic gesture is the one that was most probably intended by a speaker. This ‘probabilistic inferential’ nature of computations makes the Bayesian framework attractive. Here we developed a Bayesian algorithm that captures the two characteristic phenomena of phoneme perception, (1) sharp transitions in perception (categorization) and (2) enhanced discriminability (differentiation) at around phoneme boundaries, and then explored how this algorithm can be implemented in plastic human brains. Our model posits (i) that the brain has the probabilistic knowledge of frequencies of phonetic stimuli prior to forming the likelihood of phoneme stimuli based on noisy sensory signals (prior and likelihood beliefs), (ii) that the brain combines these two knowledges to form a posterior distribution of probability (posterior belief), and (iii) that the brain ‘optimally’ utilizes this single posterior belief to concurrently perform the categorization and discrimination tasks. The major latent variables of our interest were the shape of the prior probability distribution function (prior PDF) and the width of the likelihood PDF. The parameters for these two PDFs were estimated by fitting the model to the behavioral data in a pair of psychophysical experiments, where human subjects both categorized and discriminated the acoustic stimuli comprising the cyclic transition among three voiced stop consonant–vowel syllables, /ba/-/da/-/ga/, varying in the place of articulation (labial-alveolar-velar). The behaviorally constrained models revealed the prior PDF with the three modes whose peaks correspond to the three prototypical phoneme syllables and the posterior PDF with large variance. Having identified the prior and likelihood PDFs used by the optimal Bayesian listener, we explored plausible neural mechanisms for implementing those PDFs. Going beyond previous attempts, our proposal of Bayesian implementation offers a formal account for how the unequal frequency of acoustic stimuli, i.e., stimulus prior, is developmentally translated into an unequal distribution of sensory neurons via well-known canonical principles of neural plasticity (‘neural remapping’). Specifically we propose that sensory neurons, stimulus tuning preferences of which were equally distributed initially, iteratively shift their tuning curves toward an experienced stimulus (‘attractive shift’) as a function of their current responsitivity to that stimulus (Fig. 62). The population distribution of sensory tuning curves that were shaped by this remapping scenario, when plugged into typical probabilistic population coding schemes, reproduced qualitatively the human listeners’ performances in the both phoneme tasks. Our model exercise on tuning width also shed new light on how the optimal tuning width of sensory neurons (broad tuning in our case) can be constrained by the task requirements (categorical perception) and the stimulus environments (biased prior) imposed on a given sensory system (speech perception).Fig. 62 Interaction between tuning curve and prior in population tuning. A Bias map for combinations of tuning curve width and prior sharpness. Negative bias means that perception of a near-/ba/stimulus was biased toward/ba/, namely categorical perception. Three inset plots on ordinate and abscissa show the cases of the lowest/median/highest concentration parameters of tuning curve and prior peak, respectively. B Discrimination difference map. Negative number indicates that between-phoneme condition outperformed near-/ba/condition. C Tuning curves of population neurons that was marked as green squares in A, B. Location of tuning centers were marked as dots for 60 neurons, and tuning curves of 30 out of those 60 neurons were drawn below P111 Complexity of EEG signals is reduced during unconsciousness induced by ketamine and propofol Jisung Wang1, Heonsoo Lee1 1Physics department, Pohang University of Science and Technology, Pohang, South Korea Correspondence: Heonsoo Lee - beafool@postech.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P111 Identifying a universal feature of brain dynamics during anesthetic-induced unconsciousness has been an important work for both practical use of monitoring depth of general anesthesia and scientific knowledge about the nature of consciousness. However, it is difficult because anesthetics with different mechanism of action (MOA) induce distinct brain dynamics. From the perspective of complex system science, we claim that the dynamics generated by conscious brain is more complex compared to one from anesthetic-induced unconscious brain regardless of anesthetic types. To test the hypothesis, we used ketamine and propofol which fall into two distinct anesthetic groups [1]. Disorder and complexity of electroencephalogram (EEG) signals were analyzed before and after bolus injection of drugs. For the analysis, we employed Shannon entropy (SE) and fluctuation complexity (FC), which are information theory-based measures quantifying disorder and complexity, respectively [2]. The study shows that ketamine and propofol both reduced the complexity (p < 0.00001 for both) of EEG signals from the whole brain area (Fp1, F3, T3, P3) while each respectively increased (p = 0.000112) and decreased (p < 0.00001) disorder of the signal (Fig. 63). The finding supports our claim and suggests considering the EEG complexity as a common measure of consciousness.Fig. 63 A SE and FC values of Fp1 channel for three different states, which are wakeful, ketamine-induced, and propofol-induced states, are averaged over subjects (n = 29 for ketamine-induced, n = 20 for propofol-induced and n = 49 for wakeful states). Error bars represent standard errors. Wakeful state has the intermediate SE value between ones of two other states. For FC value, however, wakeful state has the highest one and both anesthetized states have smaller ones, forming a concave relationship between three states. B Each dot manifests averaged SE and FC values of one subject over 23 10 s-long epochs overlapping 5 s each other. Error bars here also indicate standard errors. Most wakeful states have higher FC values compared to ones of unconscious states and have intermediate SE values. Ketamine-induced states are mainly located in the lower right part when propofol-induced states are clustered at the lower left part of the area in SE–FC plot. C FC values of EEG signals from the whole brain area, covering pre-frontal, frontal, temporal, and parietal regions, significantly decreased during ketamine-induced loss of consciousness (p < 0.0001). D All FC values of the signals from different regions were also significantly reduced during propofol-induced unconscious state (p < 0.0001) ReferencesLee U, Ku S, Noh G, Baek S, Choi B, Mashour GA. Disruption of frontal–parietal communication by ketamine, propofol, and sevoflurane. J Am Soc Anesthesiol. 2013;118:1245–6. Bates JE, Shepard HK. Measuring complexity using information fluctuation. Phys Lett A. 1993;172:416–25. P112 Self-organized criticality of neural avalanche in a neural model on complex networks Nam Jung1, Le Anh Quang1, Seung Eun Maeng1, Tae Ho Lee1, Jae Woo Lee1 1Department of Physics, Inha University, Namgu, Incheon 22212, Korea Correspondence: Jae Woo Lee - jaewlee@inha.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P112 The concept of the self-organized criticality is applied to many natural and economical systems [1–4]. The distribution of neural avalanche in neural models obeys a power law with exponents of the mean-field theory. Neural avalanches in cultured neocortical network show self-organized criticality over long stable period with exponent −1.5 of the power law for the distribution of the neural avalanche [1]. We consider a modified integrate-and-fire model introduced by Levina, Herrmann and Geisel (LHG model) [2]. We extend the LHG model on the complex networks such as fully-connected network, random network, small-world network, and scale-free networks. In the LHG model the membrane potential of a neuron is accumulated from input potential and random external input. In a fully connected network we observed the power law with exponent −1.57 as shown in Fig. 64. The exponent of the power law depends on the network structure of the neural systems.Fig. 64 Distribution of avalanche size for LHG model on the fully-connected network. The distribution function of avalanche size shows the power law with exponent −1.57 Acknowledgements: This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and Future Planning (NRF-2014R1A2A1A11051982). ReferencesBeggs J, Plenz D. Neural avalanche in neocortical circuits. J Neurosci. 2003;23:11167–77. Levina A, Herrmann JM, Geisel T. Dynamical synapses causing self-organized criticality in neural networks. Nat Phys. 2007;3:857–60. Li X, Small M. Neuronal avalanches of a self-organized neural network with active-neuron-dominant structure. Chaos. 2012;22:023104. Liu H, Song Y, Xue F, Li X. Effects of bursting dynamic features on the generation of multi-clustered structure of neural network with symmetric spike-timing-dependent plasticity learning rule. Chaos. 2015;25:113108. P113 Dynamic alterations in connection topology of the hippocampal network during ictal-like epileptiform activity in an in vitro rat model Chang-hyun Park1,2, Sora Ahn3, Jangsup Moon1,2, Yun Seo Choi2, Juhee Kim2, Sang Beom Jun3,4, Seungjun Lee3, Hyang Woon Lee1,2 1Departments of Neurology, Ewha Womans University School of Medicine, Seoul, Korea; 2Department of Medical Science, Ewha Womans University School of Medicine, Seoul, Korea; 3Department of Electronics Engineering, Ewha Womans University College of Engineering, Seoul, Korea; 4Brain & Cognitive Sciences, Ewha Womans University College of Scranton, Seoul, Korea Correspondence: Chang-hyun Park - park.changhyun@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P113 The experimental approach using in vitro slices of the rat limbic system has been applied to identify mechanisms underlying epileptiform activity in ictal-like events [1]. We prepared combined slices of the rat hippocampus-entorhinal cortex and placed them in artificial corticospinal fluid that contained 4-aminopyridine (4AP). Field potential recordings were made with a microelectrode array composed of 6 × 10 microelectrodes with inter-electrode spacing of 500 μm (see Fig. 65A) when synchronous activity was induced by 4AP. Channels with high artifact rates were rejected, and the signal for each remaining channel was divided into 10 s windows without overlap. For each window, adjacency matrices, or binary networks, were estimated via inter-channel connections based on spectral coherence in different frequency bands, including delta, theta, alpha, beta, gamma, ripple, and fast ripple. Topological properties of the inter-channel networks were assessed by calculating global and local efficiency [2]. As ictal-like events were initiated, global efficiency started to decrease and local efficiency started to increase, and the changes were maintained during ictal-like epileptiform activity (see Fig. 65B). Such changes related to a shift in connection topology to a regularized pattern are in line with the findings for the whole brain in a rat model [3].Fig. 65 A Microelectrodes placed to cover both entorhinal cortex and hippocampus. B Temporal changes in global and local efficiency around the time of ictal-like epileptiform activity. The red vertical line indicates the initiation of ictal-like events and the time series in blue displays recorded field potentials Conclusions Although the initiation and propagation of epileptiform activity may not be fully appreciated due to the spatially isolated structure in the in vitro slice preparation, the pattern of ictal-like synchronous activity in the limbic system was related to changes in connection topology that may reflect a shift in brain states. Acknowledgements: This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT & Future Planning (2015R1C1A1A01052438 to C. Park and 2014R1A2A1A11052103 to H. W. Lee), and by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by the Ministry of Health & Welfare (HI14C1989 to H. W. Lee). ReferencesAvoli M, Barbarosle M, Lücke A, Nagao T, Lopantsev V, Köhling R: Synchronous GABA-mediated potentials and epileptiform discharges in the rat limbic system in vitro. J Neurosci. 1996;16(12):3912–24. Latora V, Marchiori M. Efficient behavior of small-world networks. Phys Rev Lett. 2001;87(19):198701. Otte WM, Dijkhuizen RM, van Meer MPA, van der Hel WS, Verlinde SAMW, van Nieuwenhuizen O, Viergever MA, Stam CJ, Braun KPJ. Characterization of functional and structural integrity in experimental focal epilepsy: reduced network efficiency coincides with white matter changes. PLoS One. 2012;7(7):e39078. P114 Computational model to replicate seizure suppression effect by electrical stimulation Sora Ahn1, Sumin Jo1, Eunji Jun1, Suin Yu1, Hyang Woon Lee2, Sang Beom Jun1, Seungjun Lee1 1Department of Electronics Engineering, Ewha Womans University, Seoul, 120-750, Korea; 2Department of Neurology, Ewha Womans University, Seoul, 120-750, Korea Correspondence: Seungjun Lee - slee@ewha.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P114 Deep brain stimulation (DBS) method for suppression of epileptic seizure is being developed mostly based on clinical experiences because the suppression mechanism by electrical stimulation is still unclear. As such, it is difficult to improve efficacy of the DBS method. The study of computational models allows to predict and analyze the effect of electrical stimulation by computer simulation such that it can help to determine optimum stimulation parameters to suppress seizure activity in various conditions. In this paper, we propose a hippocampal network model which portrays propagation characteristics of seizure-like events (SLEs) and suppression phenomena by electrical stimulation. The model is composed of four sub-networks representing EC, DG, CA3 and CA1 and well-known synaptic pathways between sub-networks. Each sub-network consists of excitatory and inhibitory neurons which are described by Izhikevich’s model [1]. Besides synaptic transmission [2], electrical field transmission [3] between neurons is also considered. Input gains of neurons are controlled by interaction strengths between sub-networks which are calculated by Granger causality analysis. We adopt the “potassium accumulation hypothesis” in order to replicate suppression effect by electrical stimulation [4, 5]. The effectiveness of the model is confirmed by comparing the simulation results with experimental data which were measured in rat hippocampal slice (horizontal, 400um) in bicuculline bath application. Local field potentials are recorded using micro-electrode array (MEA) and electrical stimulation (130 Hz, 500 µA, biphasic, 3–5 s) is applied manually in EC by an additional depth electrode when SLE is initiated. Following Fig. 66 shows time domain signals recorded in in vitro measurement and generated from the computer model, respectively. After stimulation, the SLE in EC is suppressed immediately, while SLEs in other areas still remained. The simulation results show similar waveforms with experimental data.Fig. 66 Recording data (A) and simulation results (B) of SLE suppression effect by electrical stimulation Acknowledgements: This work was supported by the National Research Foundation of Korea (No. 2014R1A2A1A11052763). ReferencesIzhikevich EM. Simple model of spiking neurons. IEEE Trans Neural Netw. 2003;14(6): 1569–72. Izhikevich EM, Gally JA, Edelman GM. Spike-timing dynamics of neuronal groups. Cereb Cortex. 2004;14(8):933–44. Fröhlich F, McCormick DA. Endogenous electric fields may guide neocortical network activity. Neuron. 2010;67(1):129–43. Fertziger AP, Ranck JB. Potassium accumulation in interstitial space during epileptiform seizures. Exp Neurol. 1970;26(3):571–85. Beurrier C, Bioulac B, Audin J, Hammond C. High-frequency stimulation produces a transient blockade of voltage-gated currents in subthalamic neurons. J Neurophysiol. 2001;85(4):1351–56. P115 Identifying excitatory and inhibitory synapses in neuronal networks from spike trains using sorted local transfer entropy Felix Goetze1,2, Pik-Yin Lai1 1Department of Physics, National Central University, Chung-Li, Taiwan, ROC; 2Taiwan International Graduate Program for Molecular Science and Technology, Institute for Atomic and Molecular Sciences, Academia Sinica, Taipei, Taiwan, ROC Correspondence: Felix Goetze - afgoetze@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P115 Transfer entropy [1] is the established method for quantifying the effective connectivity among neurons. It has been shown in simulations [2] that measuring it from simultaneously recorded spike trains of neurons can detect the underlying connections and therefore reconstruct a neuronal network from its observed dynamics. Being interpreted as the predicted information transfer, it quantifies the directed non-linear interactions between time series as a model-free method regardless of the underlying interaction type, which could be either inhibitory or excitatory. Making the distinction between excitatory and inhibitory synapses, however is important in order to understand the underlying principles of spatiotemporal patterns in functional networks. In our study we describe a method for the measuring of interaction types, based on the concept of local transfer entropies [3]. In contrast to the averaging across all configurations of variables of the source process and the target process as in the Transfer Entropy estimation, the local transfer entropy quantifies the effect of a specific configuration of variables on how they either inform or misinform on the future of the target process. For example observing the presynaptic neuron fire and then observing the postsynaptic neuron fire is informative for an excitatory connection, but misinformative for an inhibitory connection. On average, knowing the past of the source process of an inhibitory or excitatory connection are both predictive of the future of the target process, but local transfer entropies of specific variable configurations have opposite signed values for each interaction type respectively. Sorting the local entropies according to interaction type yields the quantity we call sorted local transfer entropy that aims to identify inhibitory and excitatory synapses from recorded spike trains. We validate this method with simulated spike trains from the Izhikevich model [4] of cortical neuronal networks, by following a previous paper [2]. The random network is noise-driven and consists of 800 excitatory and 200 inhibitory neurons. Synapses have random delays and the synaptic strengths evolved according to a spike-timing plasticity rule, before the recordings for the analysis are collected. Using Transfer Entropy and the new quantity Sorted Local Transfer Entropy, we reconstruct the networks and distinguish inhibitory from excitatory synapses. The use of two decision boundaries for classifying inhibitory and excitatory synapses separately improves the overall network reconstruction. ReferencesSchreiber T. Measuring information transfer. Phys Rev Lett. 2000;85:461. Ito S, Hansen ME, Heiland R, Lumsdaine A, Litke AM, Beggs JM, Zochowski M. Extending transfer entropy improves identification of effective connectivity in a spiking cortical network model. PLoS One. 2011;6:e27431 Lizier JT. Measuring the dynamics of information processing on a local scale in time and space. In: Directed information measures in neuroscience. Berlin: Springer; 2014. p. 161–93. Izhikevich EM. Polychronization: computation with spikes. Neural Comput. 2006;18(2):245–82. P116 Neural network model for obstacle avoidance based on neuromorphic computational model of boundary vector cell and head direction cell Seonghyun Kim1, Jeehyun Kwag1 Department of Brain and Cognitive Engineering, Korea University, Seoul, Korea Correspondence: Jeehyun Kwag - jkwag@korea.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P116 Developing robots that can perform autonomous exploration in unfamiliar environment is one of the challenges in robotics. Autonomously navigating robots are generally equipped with an obstacle avoidance (OA) system based on sensors such as Light Detection and Ranging (LIDAR) and camera to detect obstacles as well as complex algorithms to correct the noise of sensors [1]. Interestingly, rodent brain shows remarkable reliability to noise in OA [2] through using neurons specialized in processing spatial orientation and spatial boundary called head direction cell (HDC) [3] and boundary vector cell (BVC) [4], respectively. Therefore, building a bioinspired OA system with neural network that consists of neuromorphic HDC and BVC may help increase the efficiency of autonomous navigation. Hence, we built a neural network for OA consisting of all-to-all synaptic connections between six HDCs, four BVCs, and two motor neurons where HDC and BVC were constructed as multi-compartment Hodgkin–Huxley models using the NEURON based on full morphology and electrophysiological properties in vitro [5, 6]. Each HDC was modeled to spike at specific preferred directions separated by 60°, BVCs were modeled to spike at boundaries of cardinal directions and motor neurons were modeled to spike with Gaussian white noise as a background noise. We also built a virtual rat that navigated within a 1 m × 1 m environment whose trajectory was controlled by spikes of motor neurons receiving synaptic inputs from: (1) HDCs, (2) BVCs and (3) both HDCs and BVCs. Number of obstacle detection (detection number: DN) and the time spent during obstacle collision (collision time: CT) were analyzed to compare the efficiency of neural network for OA. We first verified that our neuromorphic HDC and BVC models could mimic the experimentally recorded electrophysiological properties in vitro and in vivo: HDCs reached maximum firing rate at each preferred direction, and BVCs increased their firing rate as the virtual rat approached boundaries. Using such neuromorphic HDC and BVC models, we investigated the roles of HDC and BVC in neural network for OA. Firstly, we performed the control simulation where virtual rat was controlled with neural network without neither HDC nor BVC and observed that DN was 40 and CT was 137 s. When HDC was added to the neural network, the result was similar to control simulation, (DN = 38 and CT = 139 s), indicating HDC alone cannot perform OA efficiently. When BVC alone was included in the neural network, DN substantially increased and CT decreased compared to control model (DN = 110 and CT = 73 s), indicating that OA efficiency increased. Finally, when both HDC and BVC were included in the neural network, the OA performance was most efficient (DN = 139 and CT = 39 s). These results suggest that our neural network model composed of neuromorphic HDC and BVC neurons can successfully perform OA even with background noise. Therefore, here we suggest the bioinspired neural network that consists of neuromorphic computational model of HDC and BVC could serve as a new approach to build an efficient OA system. Acknowledgements: This study was supported by the Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Science, ICT and Future Planning (NRF-2013R1A1A2053280). ReferencesZohaib M, Pasha M, Riaz R, Javaid N, Ilahi M, Khan R. Control strategies for mobile robot with obstacle avoidance. J Basic Appl Sci Res. 2013;3(4):1027–36. Vorhees CV, Williams MT. Assessing spatial learning and memory in rodents. ILAR J. 2014;55(2):310–32. Taube JS, Muller RU, Ranck JB Jr. Head-direction cells recorded from the postsubiculum in freely moving rats. I. Description and quantitative analysis. J Neurosci. 1990;10(2):420–35. Lever C, Burton S, Jeewajee A, O’Keefe J, Burgess N: Boundary vector cells in the subiculum of the hippocampal formation. J Neurosci. 2009;29(31):9771–77. Yoder RM, Taube JS. Projections to the anterodorsal thalamus and lateral mammillary nuclei arise from different cell populations within the postsubiculum: implications for the control of head direction cells. Hippocampus. 2011;21(10):1062–73. Menendez de la Prida L, Suarez F, Pozo MA. Electrophysiological and morphological diversity of neurons from the rat subicular complex in vitro. Hippocampus. 2003;13(6):728–44. P117 Dynamic gating of spike pattern propagation by Hebbian and anti-Hebbian spike timing-dependent plasticity in excitatory feedforward network model Hyun Jae Jang1, Jeehyun Kwag1 Department of Brain and Cognitive Engineering, Korea University, Seoul, Korea Correspondence: Jeehyun Kwag - jkwag@korea.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P117 Precise timings of spikes within in vivo spike train are believed to carry information critical for neural computation [1]. For such neural information to be effective, temporal patterns of spike train should be able to propagate across multiple neuronal layers in the feedforward network (FFN) of the brain without dissipation [2]. To support such reliable propagation of spike patterns, preferential and selective strengthening of synaptic pathways through which the spike patterns are routed may be necessary. Asymmetric Hebbian and anti-Hebbian spike timing-dependent plasticity (STDP), where synaptic strengths are strengthened or weakened depending on the precise relative timing and the order between pre- and postsynaptic spikes [3, 4], may serve as a good candidate for dynamically routing the propagation of spike patterns. Hence, we investigated the role of Hebbian and anti-Hebbian STDP in spike pattern propagation using a six-layered FFN model composed of 200 Hodgkin–Huxley excitatory neurons in each layer. Asymmetric and symmetric/Hebbian and anti-Hebbian STDP were modeled at excitatory synapses using exponential functions [5]. In vivo spike train obtained from public database (crcns.org) was used as an input spike pattern (TIN) in layer 1, which was simulated in a small subset of excitatory neurons in layer 1 of FFN model, while the rest were made to spontaneously spike with spike frequencies showing log-normal distribution to mimic in vivo background noise. The propagation of temporal spike pattern was quantified by analyzing the similarity ratio (SR) between TIN and output spike pattern in layer 6 (TOUT), which calculates how instantaneous inter-spike intervals of TIN and TOUT are similar. In FFN model without STDP, the spike pattern of TIN in layer 1 became dissipated in noise as it propagated across layers, and consequently failed to preserve its spike pattern to layer 6 with low SR (0.49). When asymmetric anti-Hebbian STDP was included in FFN model, TIN also failed to propagate to layer 6 with low SR (0.17). However, in the presence of asymmetric Hebbian STDP, TIN successfully propagated to the final layer with high SR (0.87), indicating that asymmetric Hebbian STDP preferentially enhanced TIN propagation in FFN model. Further analysis revealed that asymmetric Hebbian STDP selectively strengthened the synaptic weights of the synaptic pathways routing TIN while it weakened the synaptic weights of that routing noise, effectively serving as an open-gate for propagating TIN. In contrast, asymmetric anti-Hebbian STDP curve selectively weakened the synaptic weights of the synaptic pathways routing TIN, serving as a close-gate for propagating TIN. We also tested the effect of symmetric Hebbian STDP which induces only LTP or symmetric anti-Hebbian STDP which induces only LTD, and found that both types of symmetric STDP failed to propagate TIN with low SR (symmetric Hebbian = 0.23, symmetric anti-Hebbian = 0.14). Our results demonstrate that only asymmetric Hebbian STDP facilitates the reliable propagation of in vivo temporal pattern while asymmetric and symmetric anti-Hebbian STDP blocks temporal pattern propagation, suggesting that different types of STDP may dynamically gate the propagation of neural information. Acknowledgements: This study was supported by Human Frontier Science Program (RGY0073/2015) and the Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Science, ICT and Future Planning (NRF-2013R1A1A2053280). ReferencesMainen ZF, Sejnowski TJ. Reliability of spike timing in neocortical neurons. Science. 1995;268(5216):1503–6. Kumar A, Rotter S, Aertsen A. Spiking activity propagation in neuronal networks: reconciling different perspectives on neural coding. Nat Rev Neurosci. 2010;11(9):615–27. Bi GQ, Poo MM. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci. 1998;18(24):10464–72. Feldman DE. The spike-timing dependence of plasticity. Neuron. 2012;75(4):556–71. Song S, Miller KD, Abbott LF. Competitive Hebbian learning through spike-timing-dependent synaptic plasticity. Nat Neurosci. 2000;3(9):919–26. P118 Inferring characteristics of input correlations of cells exhibiting up-down state transitions in the rat striatum Marko Filipović1,2, Ramon Reig3, Ad Aertsen1,2, Gilad Silberberg4, Arvind Kumar1,5 1Bernstein Center Freiburg, Freiburg, Germany; 2Faculty of Biology, University of Freiburg, Freiburg, 79104, Germany; 3Instituto de Neurociencias de Alicante, University of Alicante, Alicante, Spain; 4Department of Neuroscience, Karolinska Institute, Stockholm, 17177, Sweden; 5Department of Computational Science and Technology, School of Computer Science and Communication, KTH Royal Institute of Technology, Stockholm, 10040, Sweden Correspondence: Marko Filipović - marko.filipovic@bcf.uni-freiburg BMC Neuroscience 2016, 17(Suppl 1):P118 Rat striatal projection neurons (SPNs) recorded under ketamine anesthesia exhibit slow oscillations, with transitions between depolarized and hyperpolarized membrane potential also referred to as up and down states, respectively. It is presumed that the activity during hyperpolarized down-states is determined by intracellular processes, whereas the large membrane voltage fluctuations during up-states are a product of increased synaptic input. Because local striatal activity during an up-state is weak, the statistics of the up-state fluctuations mainly reflect cortical feedforward input to the SPNs. To infer the statistics of the cortical input to SPNs we measured the statistics and spectrum of the membrane potential of SPNs in up and down states. The spectrum of the membrane potential reflects the filtering properties of the membrane and can be used to estimate the effective time constant (τeff) of the neuron. Our analysis showed that SPNs have significantly smaller τeff in the up-state than in the down-state, consistent with the assumption that the barrage of synaptic input causes an increase in membrane conductance during the up state. However, this observation is inconsistent with the idea that depolarization of SPNs should increase the membrane time constant because of the closing of some of the voltage dependent ion channels (e.g. the Kir) channels [1]. The mean (μup) and variance (σup) of the membrane potential during up states varied in a correlated manner. At the same time, for a given SPN, μup and σup of individual up-state membrane potentials were highly variable across different up states, indicating a corresponding variability in the cortical inputs. Using a point neuron model of an SPN, we show that the correlation and variability of the up-state mean and variance could be explained if we assume that SPNs receive correlated inputs. Across different SPNs, each recorded in a different animal, we observed a high variability in the correlation (ρ) between μup and σup. This variability could arise from the heterogeneity in the neuron morphology, intracellular properties, conductance state of the neurons, synaptic weights and the input rate and correlations. Using a point neuron model we tested the dependence of ρ on each of these properties. Our analysis showed that the variability of the correlation between μup and σup arises because of the diversity of synaptic weights and input correlations, and not because of intrinsic properties of SPNs. This suggests that neuronal heterogeneity could be obscured by the statistics of the synaptic inputs and synaptic weights. In summary, our analysis of up-down states allows us to make general inferences about characteristics of correlated synaptic input, such as strength of correlations and input firing rate, solely based on membrane potential recordings of SPNs exhibiting up and down states. Acknowledgements: This work was supported in parts by the Erasmus Mundus Joint Doctorate Programme EUROSPIN (MF), an ERC starting grant (GS), the Knut and Alice Wallenberg Academy Fellowship, the Karolinska Institutet Strategic Research program in Neuroscience (StratNeuro; GS, AK), and the Swedish Medical Research Council (GS). ReferenceNisenbaum ES, Wilson CJ. Potassium currents responsible for inward and outward rectification in rat neostriatal spiny projection neurons. J Neurosci. 1995, 15: 4449–63. P119 Graph properties of the functional connected brain under the influence of Alzheimer’s disease Claudia Bachmann1, Simone Buttler1, Heidi Jacobs2,3,4, Kim Dillen5, Gereon R Fink5,6, Juraj Kukolja5,6, Abigail Morrison1,7,8 1Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany; 2Faculty of Health, Medicine and Life Science, School for Mental Health and Neuroscience (MHeNS), Alzheimer Centre Limburg, Maastricht University Medical Centre, PO Box 616, 6200 MD Maastricht, The Netherlands; 3Department of Radiology & Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School, Boston, MA 02114, USA; 4Faculty of Psychology and Neuroscience, Department of Cognitive Neuroscience, Maastricht University, PO BOX 616, 6200 MD Maastricht, The Netherlands; 5Cognitive Neuroscience, Institute of Neuroscience and Medicine (INM-3), Jülich Research Centre, Jülich, Germany; 6Department of Neurology, University Hospital of Cologne, Cologne, Germany; 7Computational Neuroscience, Bernstein Center Freiburg, Freiburg, 79104, Germany; 8Institute of Cognitive Neuroscience, Faculty of Psychology, Ruhr-University Bochum, 44801 Bochum, Germany Correspondence: Claudia Bachmann - c.bachmann@fz-juelich.de BMC Neuroscience 2016, 17(Suppl 1):P119 Diagnosing Alzheimer’s disease (AD), especially in the early stage, is costly and burdensome for the patients, since it comprises a battery of psychological tests and an extraction of disease specific biomarkers from the cerebrospinal fluid. A cheaper and more convenient procedure would be a diagnosis based on images obtained through fMRI. Based on previous polymodal studies demonstrating disrupted inter- and intra-cortical connectivity in AD [1], we argue that the functional connectivity of the whole cortex might be a good predictor for the cause of the disease. In resting state fMRI, previous attempts to analyze graph properties of whole brain networks contradict each other [2]. In our opinion there are two general critical points in the methodology of these studies that are likely to contribute to the variability of the results. First, we criticize that the activities of the brain areas (graph nodes) that are used to calculate the functional connectivities (weights of the graph edges) are composed of functionally inhomogeneous signals, as individual brains are often mapped onto a standard atlas brain of known functional coherent areas [2, 3]. The second problem consists in converting the resulting weighted graphs into simple graphs, by setting weights above an arbitrary threshold wmin to 1, and those below it to 0 [2]. The drawback here is that there is no validation for an optimal threshold, and information that might be relevant in AD may be lost. In this work we address the first problem by applying an activity-driven, region-growing clustering algorithm derived from image processing [4]. In order to guarantee functionally homogeneous clusters, the threshold for inclusion of a voxel in a region is regulated by a heterogeneity criterion [3]. Applying this algorithm, we end up with undirected weighted graphs with varying numbers of nodes for three sets of data: healthy elderly controls, mild cognitive impairment and Alzheimer’s disease. Targeting the second problem, we analyze the dependence of graph theoretic measures (shortest path length, in- and out-degree distribution, clustering coefficient, modularity and minimal spanning tree [5]) on wmin. Finally, we investigate the distribution of these measures for each data set to determine candidates for a predictive measure. Acknowledgements: We acknowledge partial support by the Helmholtz Alliance through the Initiative and Networking Fund of the Helmholtz Association and the Helmholtz Portfolio theme “Supercomputing and Modeling for the Human Brain”. ReferencesBokde AL, Ewers M, Hampel H.: Assessing neuronal networks: understanding Alzheimer’s disease. Prog Neurobiol. 2009;89:125–33. Tijms BM, Wink AM, de Haan W, van der Flier WM, Stam CJ, Scheltens P, Barkhof F. Neurobiol: Alzheimer’s disease: connecting findings from graph theoretical studies of brain networks. Aging. 2013;34: 2023–36. Marrelec G, Fransson P. Assessing the influence of different ROI selection strategies on functional connectivity analyses of fMRI data acquired during steady-state conditions. PLoS One. 2011;6(4):e14788. Lu Y, Jiang T, Zang Y. Region growing method for the analysis of functional MRI data. Neuroimage. 2003;20(1):455–65. Wang J, Zuo X, Dai Z, Xia M, Zhao Z, Zhao X, Jia J, Han Y, He Y. Disrupted functional brain connectome in individuals at risk for Alzheimer’s disease. Biol Psychiatry. 2013;73(5):472–81. P120 Learning sparse representations in the olfactory bulb Daniel Kepple1, Hamza Giaffar1, Dima Rinberg2, Steven Shea1, Alex Koulakov1 1Cold Spring Harbor Laboratory, Cold Spring Harbor, NY 11724, USA; 2NYU Neuroscience Institute, New York, NY 10016, USA Correspondence: Daniel Kepple - akula@cshl.edu BMC Neuroscience 2016, 17(Suppl 1):P120 In mouse olfaction, olfactory receptor responses are aggregated in spherical structures of the main olfactory bulb (MOB) called ‘glomeruli.’ These signals are then received by mitral cells (MC) and communicated to the cortex. MC responses are modified by local inhibitory interneurons called granule cells GC, which vastly outnumbered MCs (50fold–100fold). In our previous work [1], we proposed a model in which GCs inhibit MC to form a sparse, incomplete representation (SIR) of odors (Fig. 67). Here, we reason that since sparse representations are efficient, sparseness may increase with learning. We extend the SIR model to allow network synaptic weights to be adjusted, increasing representational sparseness and increasing stimulus discriminability. We derive learning rules for dendrodendritic connectivity between GCs and MCs and also for centrifugal cortico-granule synapses. We computationally test these learning rules and make several predictions of GC and MC plasticity. Specifically, we predict that a minority of GCs outcompete the rest of the population to generate a negative image of a learned odor. Additionally, we predict that participation of the GC network will confer combination selectivity and the ability to discriminate overlapping input patterns. Finally, we experimentally validate these predictions for the dynamics of GCs during locus coeruleus-induced MOB plasticity.Fig. 67 Sparse incomplete representations (SIR). In our previously formulated model of the main olfactory bulb network [1], MCs receive inputs from receptor neurons in the glomeruli (black circles) and interact with GCs through dendrodendritic synapses. GCs build representations of MC glomerular inputs (red arrows). The representations are contained in the inhibitory inputs returned by the GCs to the MCs (blue arrows). Because GCs inhibit each other through second-order inhibitory interactions, only a few GCs respond to an odorant (full blue circles with a dendrite shown). The vast majority of GCs do not change their firing rate in response to an odorant (empty circles). Thus, the responses of GCs are sparse. Because some MCs manage to retain the responses to odorants, the representation by GCs is called incomplete. According to this model, MCs transmit to higher areas the errors in the GC representation ReferenceKoulakov AA, Rinberg D. Sparse incomplete representations: a potential role of olfactory granule cells. Neuron. 2011;72(1):124–36 P121 Functional classification of homologous basal-ganglia networks Jyotika Bahuguna1,2,3, Tom Tetzlaff1, Abigail Morrison1,2, Arvind Kumar2,3, Jeanette Hellgren Kotaleski3 1Institute of Neuroscience and Medicine (INM-6), Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany; 2Computational Neuroscience, Bernstein Center Freiburg, Freiburg, 79104, Germany; 3Computational Brain Science, Department of Computational Science and Technology, School of Computer Science and Communication, KTH, Royal Institute of Technology, Stockholm, Sweden Correspondence: Jyotika Bahuguna - j.bahuguna@fz-juelich.de BMC Neuroscience 2016, 17(Suppl 1):P121 The basal ganglia (BG) are a set of nuclei that play an important role in motor and cognitive functions. Indeed many brain diseases such as Parkinson’s disease (PD) can be attributed to dysfunction of one or more BG nuclei. The classical model of basal ganglia has been regularly updated with discoveries of new sub-populations within a nucleus or new projections from existing nuclei in recent years. It is unclear how these new insights on the structure of the BG network foster our understanding of its function. The effective connectivities among these recently identified BG sub-populations are only partially known. In the framework of a simple firing-rate model subjected to a genetic algorithm, we identified effective BG connectivities which are consistent with experimentally established firing-rate and phase relationships in Subthalamic Nucleus (STN) and two GPe subpopulations (arkypallidal [GPe-TA] and prototypical [GPe -TI]) in both healthy and PD states [1]. This is in extension to an earlier model that identified effective connectivities for the STN-TA-TI-sub circuit [2]. As expected, we found that multiple parameter combinations can fit the data [1]. We re-classified these homologous networks that reproduced the healthy and PD state, on the basis of two dynamical features: suppression of GPi activity and susceptibility of the BG network to oscillate in the presence of cortical input. These features were chosen because task execution requires GPi suppression while oscillations in the STN-GPe subnetwork are characteristic of PD. We found that most putative pathological networks showed insufficient suppression of GPi activity and high susceptibility to oscillations whereas most putative healthy networks showed sufficient suppression of GPi activity and low susceptibility to oscillations. This is consistent with experimental data that shows that lack of GPi suppression [3] or oscillations [4, 5] is correlated with Parkinsonian symptoms such as stymied movement and tremor. A small fraction of networks, however, in both cases show deficiency in only one of the features. This could indicate the configurations of healthy networks that might be more pathology prone and in contrast configurations of pathological networks that might be easier to push into a healthy state. Further analysis of estimated BG connectivity revealed that transitions between the putative PD and healthy networks were possible by modifying the strength of the relevant projections. Most of the transitions involved changes in corticostriatal, striatopallidal and pallidopallidal projections. Finally, the variance observed in the functional classification of putative pathological and healthy networks might hint at the variance observed in manifestation of Parkinson’s disease (PD). Acknowledgements: Klinische Forschergruppe (KFO219, TP12) of the Deutsche Forschungsgemeinschaft; Helmholtz Association, EuroSPIN and Erasmus Mundus Joint Doctorate Programme. ReferencesMallet A, Pogosyan A, Márton LF, Bolam JP, Brown P, Magill PJ. Parkinsonian beta oscillations in the external globus pallidus and their relationship with subthalamic nucleus activity. J Neurosci. 2008;28(52):14245–58. Nevado-Holgado AJ, Mallet N, Magill PJ, Bogacz R. Effective connectivity of the subthalamic nucleus-globus pallidus network during Parkinsonian oscillations. J. Physiol. 2014;592(7):1429–55. Boraud T, Bezard E, Bioulac B, Gross CE. Ratio of inhibited-to-activated pallidal neurons decreases dramatically during passive limb movement in the MPTP-treated monkey. J Physiol. 2000;83(3):1760–63. Chen CC, Litvak V, Gilbertson T, Kühn A, Lu CS, Lee ST, Tsai CH, Tisch S, Limousin P, Hariz M, Brown P. Excessive synchronization of basal ganglia neurons at 20 Hz slows movement in Parkinson’s disease. Exp Neurol. 2007;205(1):214–21. Moran A, Bergman H, Israel Z, Bar-Gad I. Subthalamic nucleus functional organization revealed by parkinsonian neuronal oscillations and synchrony. Brain. 2008;131(Pt-12):3395–409. P122 Short term memory based on multistability Tim Kunze1,2, Andre Peterson3, Thomas Knösche1 1Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany; 2Institute of Biomedical Engineering and Informatics, Ilmenau University of Technology, Ilmenau, Germany; 3Department of Medicine, University of Melbourne, Melbourne, Australia Correspondence: Tim Kunze - tkunze@cbs.mpg.de BMC Neuroscience 2016, 17(Suppl 1):P122 Neural circuits can be formally described by modeling the collective behavior of relatively homogeneous neural populations, so-called neural masses [1]. Some neural mass models consider the minimum set of one excitatory and one inhibitory subpopulation each. In contrast, three-population models make a distinction between excitatory pyramidal cells (PC), projecting to distant areas, and excitatory interneurons (EIN), providing local feedback. We investigate a three-population neural mass model, driven by input to the EIN, with respect to its input/output behavior. We find that such a circuit exhibits, for sufficiently salient inputs, a memory effect based on multi-stability. Furthermore, we test the hypothesis that this mechanism essentially depends on the separation between input and output neurons and is thus not captured in the simpler two-population model. We use a neural mass model [2], where a pyramidal cell subpopulation receives negative feedback from an inhibitory interneuron subpopulation and positive feedback either directly through self-connections or indirectly via a secondary excitatory subpopulation of interneurons. The respective feedback topology of interest (including which subpopulation is targeted by external input) is controlled by a single parameter. We systematically applied transient sensory inputs, modeled by pulses of various magnitude and duration, as external inputs to the EIN and monitored the behavior of the PC. Depending on the duration and intensity of the applied stimuli (see Fig. 68A), the output either transiently follows the input (i) or it jumps to a more depolarized state, where it remains oscillating with a higher mean membrane potential even after the stimulus has ceased (ii) and where further input does not effect the output any more (iii). This state can be terminated by an impulse to the inhibitory interneurons (iv). The accessibility of this memory effect depends on the saliency of the stimulus in terms of duration and intensity (see Fig. 68B) and disappears in case of direct feedback in a structurally similar two population model.Fig. 68 A Response of pyramidal cells to transient input to excitatory interneurons shows different modes. B Depending on intensity and duration of the stimulus The identified short-term memory mechanism would be important for temporal integration in cortical processing, potentially applicable in predictive coding schemes. The distinction between the input receiving excitatory subpopulation and the output sending excitatory subpopulation appears to be crucial for the described mechanism, which is further modulated by inhibitory feedback. The further examination of the ratio between excitation and inhibition, governing this mechanism, thus represents an important step to elucidate how the topology between excitatory and inhibitory neural populations affects emerging dynamics on a mesoscopic scale with potential effects on brain states and higher-order brain functionality. ReferencesFreeman WJ. Mass action in the nervous system. New York: Academic Press; 1975. Spiegler A, Kiebel SJ, Atay FM, Knösche TR. Bifurcation analysis of neural mass models: Impact of extrinsic inputs and dendritic time constants. NeuroImage. 2010;52:1041–58. P123 A physiologically plausible, computationally efficient model and simulation software for mammalian motor units Minjung Kim1, Hojeong Kim1 1Division of IoT and Robotics Convergence Research, DGIST, Daegu, 42988, Korea Correspondence: Hojeong Kim - hojeong.kim03@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P123 Background A spinal motoneuron contacts a bunch of muscle fibers forming a motor unit that underlies all mammalian movements. The essential role of the motor unit is the transduction of synaptic inputs from descending and reflex pathways into muscle force. Since the input–output properties of both motoneurons and muscle fibers are non-linear, it has been difficult to make predictions on how changes in synaptic inputs to motoneuron, cellular properties of the motoneuron and muscle fibers and muscle length may affect motor output [1]. Methods To tackle this fundamental issue in the field of motor neuroscience, we developed a physiologically plausible but computationally efficient model of the motor unit and a software package that allows for virtual experiments on the input–output properties of the motor unit over a full range of physiological inputs and biophysical parameters. Results The computational model of motor unit was first built in this study coupling the motoneuron model and the muscle unit model with a simplified axon model. The motoneuron model was developed using the recently reported two-compartment modeling approach [2]. The key feature of the new reduced motoneuron model is that all cable parameters of the reduced model are analytically determined based on the system properties such as input resistance, membrane time constant and electrical coupling properties between the soma and the dendrites, which are all empirically measurable from real motoneurons. For the muscle unit, the recently developed muscle modeling approach was employed that consists of three sub-modules representing [3]: (1) the transformation of the spike signals from motoneurons into the dynamics of calcium concentration in the sarcoplasm, (2) the conversion of the calcium concentration to the muscle activation level, and (3) the transformation of the muscle activation level into the muscle force using Hill-type muscle mechanics. The new muscle model was constructed in this study to reflect all experimentally identified dependencies of muscle activation dynamics on muscle length and movement over a full range of stimulation frequencies in cat soleus muscles. Then, to enhance the usability and extendibility the software package for simulating and analyzing the developed motor unit model was designed and implemented based on the object-oriented programing paradigm and open source Python language along with graphic user interfaces (GUI). The software package developed in this study provides a GUI-based simulation environment in which a single motoneuron, muscle unit, and motor unit can be individually simulated and analyzed in a wide range of experimental conditions reflecting biological realisms. Conclusions Our model of the motor unit and user-friendly simulation software may provide not only a computational framework to gain systemic insights into motor control by the central nervous system in a cellular perspective but also a basis on which to build biologically realistic large-scale neuro-musculo-skeletal models. Acknowledgements: This work was supported by the DGIST R&D Program of the Ministry of Science, ICT and Future Planning of Korea (15-RS-02 and 16-RS-02). ReferencesHeckman CJ, Enoka RM: Motor unit. Compr Physiol. 2012;2(4):2629–82. Kim H, Jones KE, Heckman CJ. Asymmetry in signal propagation between the soma and dendrites plays a key role in determining dendritic excitability in motoneurons. PLoS One. 2014;9(8):e95454. Kim H, Sandercock TG, Heckman CJ. An action potential-driven model of soleus muscle activation dynamics for locomotor-like movements. J Neural Eng. 2015;12(4). P125 Decoding laser-induced somatosensory information from EEG Ji Sung Park1, Ji Won Yeon, Sung-Phil Kim1 1Department of Human Factors Engineering, Ulsan National Institute of Science and Technology, Ulsan 689-798, South Korea Correspondence: Sung-Phil Kim - spkim@unist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P125 Recently, our research group has proposed a new way of providing a non-nociceptive tactile sensation with laser [1]. In this study, we aimed to investigate laser-induced somatosensory information represented in cortical activity using the human EEG. The EEG data were acquired using the V-Amp amplifier (Brain Products GmbH, Gilching, Germany) with 16 wet electrodes that were placed on the scalp following the international 10–20 system. Twenty one subjects participated in the study (7 female and mean age of 22.4 years). During the experiment, a mechanical stimulus, a laser stimulus and a heat stimulus were given in a random order to subjects sixty times per stimulus. Subjects described the feeling of laser stimulation as non-painful sensation, painful sensation and no sensation. As described in the previous study, 56.3, 12.3 and 31.4 % of the subjects reported laser stimulation as non-painful, painful and no sensation, respectively [1]. To examine similarity of cortical activity in response to different stimuli, we employed a decoding analysis of the EEG data. In the decoding analysis, we used the linear discriminant analysis (LDA) method to classify the beta (21–28 Hz) event-related desynchronization/synchronization (ERD/S) patterns of EEG into one of the two classes representing every pair of stimuli (a total of six pairs from four stimuli) [2]. Classification error indicated how similar beta ERD/S patterns were between two stimuli: a larger error reflected more difficulty in discriminating patterns and consequently a greater similarity between patterns. The beta ERD/S patterns were estimated using the short time Fourier transform. Baseline correction was implemented using the 0.5 s period before stimulus onset. For each pair of stimuli, one-way ANOVA was used to select four channels that exhibited the most differences in beta ERD/S patterns between classes and classification accuracy was assessed by the leave-four-out cross validation [3] (see Fig. 69 for the classification error between every stimulus pair). The classification results showed that to the beta ERD/S pattern induced by mechanical stimulation, the pattern by non-painful laser stimulation was most similar. Also, the results indicated closeness of cortical activities between non-painful and painful laser stimulations as well as painful laser and thermal stimulations (see Fig. 69). These results suggest that laser might induce similar beta responses whether it evoked painful or non-painful feelings but non-painful laser might share presumably non-nociceptive somatosensory information with mechanical stimulation whereas painful laser shared presumably nociceptive somatosensory information with thermal stimulation. We expect that further information theoretical analyses may reveal more details about somatosensory information encoded in cortical rhythms induced by laser.Fig. 69 Each node represents stimulation type and each edge means classification error rate. Length of edge shows similarity between a pair of stimulation ReferencesJun J-H, Park J-R, Kim S-P, Bae YM, Park J-Y, Kim H-S, Choi S, Jung SJ, Park SH, Yeom D-I. Laser-induced thermoelastic effects can evoke tactile sensations. Sci Rep. 2015;5. Pfurtscheller G, Lopes Da Silva FH. Event-related EEG/MEG synchronization and desynchronization: basic principles. Clin Neurophysiol. 1999;110(11):1842–57. Celisse A, Robin S. Nonparametric density estimation by exact leave-p-out cross-validation. Comput Stat Data Anal. 2008;52(5):2350–68. P126 Phase synchronization of alpha activity for EEG-based personal authentication Jae-Hwan Kang1, Chungho Lee1, Sung-Phil Kim1 1Department of Human and Systems Engineering, Ulsan National Institute of Science and Technology, Ulsan Correspondence: Sung-Phil Kim - spkim@unist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P126 There has been a growing interest in the EEG-based biometric system as an alternative approach to personal authentication (PA). In this study, we focused on the potentialities of functional connectivity, especially phase synchronization in the alpha rhythm represented by the phase locking value (PLV) as a novel EEG signature for PA. We analyzed an EEG dataset of 39 trials from 7 subjects who participated in the 5–7 sessions repeatedly on different days. In the sessions, a total of 16 EEG signals were acquired by a portable EEG device, when the subjects were in a resting state with their eyes closed for 2 min. The characteristics of alpha phase synchronization were estimated by the following procedures. (1) The alpha rhythm was extracted from the EEG signals using a band-pass filter of 8–13 Hz. (2) We randomly selected 20 2-s time segments of the alpha rhythm and calculated the mean phase coherence [1] between channels within each time segment. From all possible pairs of 16 EEG channels, a total of 120 mean alpha phase coherence values were extracted. 3) From these mean alpha phase coherence values, we calculated a criteria index (CI) of each of them where the CI calculated the ratio of an inter-subject variability to an intra-subject variability, which was developed to discriminate critical EEG features for PA in our previous study [2]. 4) Using the mean alpha phase coherence and its CI values, we constructed the association matrix of phase coherence and extracted the 12 top-ranked CI connections (Fig. 70). The topographical result showed that there were apparently two functional connectivity networks of the alpha rhythm in the brain for PA. The first network was distributed over the anterior regions including the pre-frontal, frontal and left central regions. The second one was located in the posterior region covering the occipital region. It should be noted that two regions have been well known as main sources of the alpha rhythm from many studies on EEG alpha rhythms. Our results suggest an important role of the alpha rhythm in the EEG-based biometrics system.Fig. 70 Overall characteristics of alpha phase synchronization for PA. A The upper triangle of association matrix indicates the PLV calculated by grand mean phase coherence. The lower triangle of association matrix indicates the its CI values in pairs. B Topographical connections with the 12 top-ranked CI from the lower triangle of a Acknowledgements: This work was supported by Institute for Information and communications Technology Promotion (IITP) grant funded by the Korea government (MSIP) (R0190-15-2054, Development of Personal Identification Technology based on Biomedical Signals to Avoid Identity Theft). ReferencesMormann F, Lehnertz K, David P, Elger CE. Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy patients. Phys D. 2000;144:358–69. Kang J-H, Lee C, Kim S-P. EEG Feature Selection and the use of Lyapunov exponents for eeg-based biometrics. In: IEEE international conference on biomedical and health informatics, Las Vegas, NV, USA; 2016. p. 1–4. P129 Investigating phase-lags in sEEG data using spatially distributed time delays in a large-scale brain network model Andreas Spiegler1, Spase Petkoski1,2, Matias J. Palva3, Viktor K. Jirsa1 1INSERM UMR 1106 Institut de Neurosciences des Systèmes - Aix-Marseille Université, Marseille, France; 2Aix-Marseille Université, CNRS, ISM UMR 7287, 13288, Marseille, France; 3Neuroscience Center, University of Helsinki, Helsinki 00014, Finland Correspondence: Andreas Spiegler - Andreas.Spiegler@univ-amu.fr BMC Neuroscience 2016, 17(Suppl 1):P129 On a large scale, the brain appears as a network composed of white matter tracts connecting brain areas within and between cerebral hemispheres. The finite transmission speed delays the interaction of areas via these pathways. The delays are on the same scale as the brain oscillates, that is, 10–250 ms [1], and have been suggested to play a role in the functional organization. One potential key mechanism is synchronization [2], which could explain the phase-lags of brain signals. In humans, stereotactical EEG (sEEG) revealed frequency-specific inter-areal synchronization often associated with nearly zero phase-lag, or with variable phase-lags between ±π. With the advance of non-invasive imaging techniques, large-scale modeling of the entire brain has become feasible using realistic connectivity and time delays [3], Fig. 71A. From structural and diffusion MRI, we obtained human connectomes composed of the strength and length of connections among 68 cortical areas. We approximated the bimodal tract length distribution (Fig. 71B) by Dirac deltas (Fig. 71C). Intra-hemispheric connections are in the 1st, and inter-hemispheric ones are in the 2nd mode. The delay of a connection was determined from its length divided by the speed of 5 m/s. The Kuramoto phase oscillator described the activity in each area. The phase difference of areas was analyzed and compared with the map of inter-areal phase lags obtained from resting state sEEG of epileptic patients.Fig. 71 A Model with local (left) and long-range connections (right). B, C Averaged tract lengths and weights from 4 connectomes B Joint distribution, and C histogram of weighted lengths for intra- and inter-hemispheric links. D Sketch of the spatial delay structure. E Phase-lag distributions (top) and phase-lags between areas (bottom rows) The model of fixed oscillators (e.g., f = 20 Hz) switched from global incoherence to alternating in- and anti-phase coherence with increasing coupling strength. Increasing the natural frequencies for constant coupling resulted in alternating switching from in- to anti-phase coherence, but also to incoherence. Intra-hemispheric links were in-phase (phase-lags ~0), and inter-hemispheric links were either in- or anti-phase (±π), see clusters in Fig. 71E. Links among areas of low in-strength (sum of all the weights for that node) showed flatly distributed phase-lags. For f = 20 Hz, we found the phase-lags in the sEEG in the regime of in- and anti-phase coherence in the model, Fig. 71E. We demonstrated that it is not simply the connectivity strength that matters in oscillatory large-scale brain networks, but time delays are of equal importance. The spatial structure in the time delays is reflected in the clustering of phase-lags. The model captured the statistics of the phase-lags as observed in the experimental data. The phase-lag structure of links at f = 20 Hz is explained in the model by a spatial organization of in- and anti-phase coherence. ReferencesBuzsáki G, Draguhn A. Neuronal oscillations in cortical networks. Science. 2004;304(5679):1926–29. Varela F, Lachaux J, Rodriguez E, Martinerie J. The brainweb: phase synchronization and large-scale integration. Nat Rev Neurosci. 2001;2(4): 229–39. Deco G, Jirsa V, McIntosh AR, Sporns O, Kötter R. Key role of coupling, delay, and noise in resting brain fluctuations. Proc Natl Acad Sci USA. 2009;106(25):10302–07. P130 Epileptic seizures in the unfolding of a codimension-3 singularity Maria L. Saggio1, Silvan F. Siep1, Andreas Spiegler1, William C. Stacey2, Christophe Bernard1, Viktor K. Jirsa1 1INSERM UMR 1106 Institut de Neurosciences des Systèmes - Aix-Marseille Université, Marseille, France; 2Department of Neurology, 2Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI 48109, USA Correspondence: Maria L. Saggio - marisa.saggio@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P130 Seizures can arise under a variety of conditions. Despite this fact, there are invariant features resulting in a characteristic electrophysiological signature. Investigations of these universal properties lead to a classification of a planar description of point-cycle fast-slow bursters [1] and a taxonomy of seizures [2]. The phenomenological model, the epileptor [2], is able to reproduce the main features of the predominant class of human seizure (~80 % of all cases), according to data from epileptic patients. We aim at generalizing this model to include other bursting classes of the taxonomy. We extended the work by [3] on bursters, living in the unfolding of high codimension singularities, and systematically investigated the unfolding of the codimension-3 degenerate Takens–Bogdanov bifurcation (focus, elliptic, saddle and cusp cases) [4]. The biological relevance of this codimension-3 bifurcation has been highlighted by other authors, and several classes appearing in the context of neuronal spiking have been identified in its unfolding (e.g., [5]). However, a systematic search in this unfolding for all the planar point-cycle bursting classes predicted by [1] was still missing. The existence of 16 bursters of the type slow-wave (self-oscillating slow subsystem), and 16 of type hysteresis-loop (slow-subsystem oscillating thanks to feedback from the fast one) was predicted. We could find all slow-wave bursters in the unfolding together with seven of the hysteresis-loop ones. With regard to these hysteresis-loop bursters, we propose a model able to reproduce each of them depending only on the initial and final points of the path in the unfolding’s parameter space. This model is based on the known normal form of the codimension-three bifurcation [4], therefore we can readily describe the role of all its variables and how the tuning of its parameters affects the models activity. We found that the codimension-three model incorporates not only the repertoire (80 % of seizure) of the model proposed by [2] but also the classes that account for the remaining 20 % of seizures. Moreover, based on an ultra-slow modulation of the bursting path (see also [6]) in the model, possible transitions between bursting classes and, more importantly, transitions to regimes (in the parameter space) where bursting behavior is not possible at all could be predicted. These predictions could be tested using data from epileptic patients for whom different types of seizures coexist. Overall, the main points of the present work are threefold: (i) a model description comprising the complete set of slow-wave bursters and seven (out of 16) hysteresis-loop bursters predicted by Izhikevich [1], (ii) a generalization of the model proposed by [2] to include the missing seizure types found in human data and to make prediction about their robustness, (iii) a framework to investigate the coexistence of different seizure types in the same patient and the transitions between them. The possibility of describing different seizure types with a unique model, thus with a unique set of variables and parameters, will facilitate the search for physiological correlates and treatments. ReferencesIzhikevich EM. Neural excitability, spiking and bursting. IJBC. 2000;10(6):1171–266. Jirsa VK, Stacey WC, Quilichini PP, Ivanov AI, Bernard C. On the nature of seizure dynamics. Brain. 2014;137(8):2210–30. Golubitsky M, Josic K, Kaper TJ. An unfolding theory approach to bursting in fast-slow systems. In: Krauskopf B, Broer HW, Vegter G, editors. Global analysis of dynamical systems; 2001. p. 227–308. Dumortier F,Roussarie R, Sotomayor J, Żaładek H. Bifurcations of planar vector fields—nilpotent singularities and abelian integrals. Berlin: Springer; 1991. Osinga HM, Sherman A, Tsaneva-Atanasova K. Cross-currents between biology and mathematics: the codi-mension of pseudo-plateau bursting. DCDS-A. 2012;32(8):2853–77. Franci A, Drion G, Sepulchre R. Modeling the modulation of neuronal bursting: a singularity theory approach. SIAM J Appl Dyn Syst. 2014;13(2):798–829. P131 Incremental dimensional exploratory reasoning under multi-dimensional environment Oh-hyeon Choung1, Yong Jeong1 1Department of Bio and Brain Engineering, KAIST, Daejeon, 34141, South Korea Correspondence: Oh-hyeon Choung - iohyeonki@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P131 In our daily life, we encounter thousands of complex problems, which are not ‘one dimensional’. However, multi-dimensional problems were known to suffer from “curse of dimensionality” [1]. Therefore, the researches of reward learning and goal-directed behavior were mostly focused on single dimensional environment for a decade [2]. Even a few researches on multi-dimensional tasks was emphasizing that human representation learning is done by reducing the dimensionality, but not focusing on multiple compositional reasoning under multi-dimensional environment [3]. Here, the multi-dimensional decision task was conducted (Fig. 72A, B) and the framework of Reinforcement Learning (RL) was used for analysis. We investigated that the reasoning under multi-dimensional environment is processed in incremental order, rather than one-shot learning. Also, the exploration of the best strategy occurs depends more on internal value, that is exploring under low value and exploiting under high value (softmax decision rule) rather than random exploration (randomized ɛ-greedy algorithm). Functional MRI were taken on each subject, while conducting the behavioral task. Brain regions of the incremental learning and the value sensitive explorative behavior will be verified.Fig. 72 Multidimensional decision making task design and model comparison. A Multidimensional decision making task schematics. B Systemic structure of the task. C The result of model comparison, proposed model has significantly high accuracy on prediction. D, E The models’ prediction accuracy of proposed model (D) and naïve model (E) We demonstrated that incremental learning rule can explain the multidimensional reasoning process better than other models (Fig. 72C–E). This result indicate that people deal with the complicated multi-dimensional problem, we solve them by adding dimensional information one by one. ReferencesSutton RS, Barto AG. Reinforcement learning: An introduction. MIT Press; 1998. Niv Y, Daniel R, Geana A, Gershman SJ, Leong YC, Radulescu A, Wilson RC. Reinforcement learning in multidimensional environments relies on attention mechanisms. J Neurosci. 2015;8145–57. Lee SW, Shimojo S, O’Doherty JP. Neural computations underlying arbitration between model-based and model-free learning. Neuron. 2014;687–99. P132 A low-cost model of eye movements and memory in personal visual cognition Yong-il Lee1,2, Jaeseung Jeong1,2 1Department of Bio and Brain Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea; 2Program of Brain and Cognitive Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea Correspondence: Jaeseung Jeong - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P132 Eye movements are the most useful and clearest signal of our body to understand cognitive process and memory mechanism. Because it is convenient to measure the stream signal and quantify. Other sensory inputs, like auditory, gustatory, olfactory, and kinesthetic stimuli, are hard to estimate, but visual stimuli are easy by the eye-tracker [1]. And people are primarily visually oriented. Every day, people get over 80 percent of information from their own eye. Until now, eye movements data has been measured and analyzed by very expensive eye-trackers mostly. The major companies’ devices price, including SMI, Tobii and EyeLink, are at least 10,000$ with their analysis tool SW. Of course, there have been many substitution trials in open SW and open HW area [2]. But their performance is certainly lower than major brands. Also, their data representation doesn’t have standard. Therefore, open SWs are difficult to apply for other utility services and scientific researchers have hesitated to use them to analyze result data to understand complex cognitive process. A few research results, which is investigating the relation between eye movement and cognitive model based on open eye-tracker platform, have reported by this time. However, eye-tracker is not only for science, and other areas need the usability of eye movement, for instance, UI/UX, healthcare, driving, game, learning consulting, TV viewer rating, market research and so on [3]. If there is a more general and low-cost eye-tracker which is confirmed cognitive model, above areas would be effective and we could do better decision making. This research implements a low-cost eye-tracker using a front camera (webcam) and a pin camera (Fig. 73, If the pc or laptop has laptop has a front camera, it doesn’t need more pin camera). The implementation includes the auto detection and classifying of useful memory based on eye movement of visual information on the device’s display. To do this function, the camera measures the saccade variation spectrum, as the X–Y axis acceleration, and categorize individual pattern while the user is taking train session. It is developed using OpenCV library and C#. XLabs Inc., already has made the gaze/head tracker using front camera without cognitive pattern analysis [4]. In the future, we will try this function on the mobile devices, which are cellular phone, tablet pc, and game interface. These devices have more sensors, like GPS, illumination, and activity accelerator. Combination of sensors input would make more precise prediction for memory cognition.Fig. 73 Low-cost eye movement tracker using front cameras on the each devices ReferencesWedel M, Pieters R: Eye tracking for visual marketing. Now Publishers Inc; 2008. Dalmaijer ES, Mathôt S, Van der Stigchel S. PyGaze: an open-source, cross-platform toolbox for minimal-effort programming of eyetracking experiments. Behav Res Methods. 2014;46(4):913–21. Pannasch S, Helmert JR, Velichkovsky BM. Eye tracking and usability research: an introduction to the special issue. J Eye Mov Res. 2008;2(4):1–4. https://xlabsgaze.com/about/. P133 Complex network analysis of structural connectome of autism spectrum disorder patients Su Hyun Kim1,2, Mir Jeong1, Jaeseung Jeong1,2 1Department of Bio and Brain Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, Korea; 2Program of Brain and Cognitive Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, Korea Correspondence: Jaeseung Jeong - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P133 Background Human connectome which is the map of full connection of neuronal network in the human brain exhibits the characteristics of complex network [1]. The human connectome is known to be wired in a way that neurons efficiently transmit and communicate information. Statistical measures of complex network describe the topological features of a certain network and this enables researchers to compare effectiveness of information processing within a network. Autism spectrum disorder (ASD) subjects exhibit repetitive behaviors, impaired social communication skills, and sensory problems. Those symptoms of neurodevelopmental disorder is doubted to be originated from genetic causes [2]. Also, recent investigations find that ASD is a ‘connection problem’. But still exact cause of ASD is unknown. The aim of the research to be conducted is to reveal the genetic cause of (ASD) by the statistical analysis of network measures of ASD patients and normal groups’ structural connectome data using diffusion tensor imaging (DTI). Methods The method to be used in the research is to compare the network measure values of various subject groups’ connectome and relate the difference to the genetic mutations in common. DTI data describes the neuronal connection of the brain regions in the mesoscale level. Structural connectome that is constructed from DTI information. Expected result is that there are genotypic changes of genes which affect development of neuronal connection in ASD subjects. This finding will shade a new light on the investigation of ASD diagnosis and treatment. ReferencesBullmore E, Sporns O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nat Rev Neurosci. 2009;10(3):186–98. Freitag CM, Staal W, Klauck SM, Duketis E, Waltes R. Genetics of autistic disorders: review and clinical implications. Eur Child Adolesc Psychiatry. 2010;19(3):169–78. P134 Cognitive motives and the neural correlates underlying human social information transmission, gossip Jeungmin Lee1,2, Jaehyung Kwon1, Jerald D. Kralik1,2, Jaeseung Jeong1,2 1Department of Bio and Brain Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, Republic of Korea; 2Program of Brain Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, Republic of Korea Correspondence: Jaeseung Jeong - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P134 Gossip is a specific example of human conversation containing social factors and has been considered as malicious, useless idle-talk by general population. Several researchers have suggested the role of gossip as social police that control the members of social groups to behave cooperative rather than selfish. However, there is not enough data to explain the actual cognitive motives that drive people to spread gossip. Throughout this study, human gossiping behavior is defined as transmission of social information about an absent third-party (i.e. the target of the gossip). In order to define the types of gossip, various scenarios containing social information are divided into 48 different categories by the third-party identity, valence and contents. Big five personality inventory, prosocial personality battery, cultural orientation scale and moral foundations questionnaire were used to measure personal traits that may influence the gossiping behavior of individuals. We found out that people, regardless to the scores of their personal traits, tend to spread gossip about in-group and celebrities more than out-group members. We also found out that positive gossip about in-group members is spread with significantly higher rates than in-group negative gossip, whereas the spread pattern was the opposite when the gossip is about celebrities. With such findings, we conducted fMRI study using in-group and out-group gossip with positive and negative valence. Increased activity in various brain regions including medial frontal gyrus, dorsolateral prefrontal cortex and precuneus was found when participants made decisions whether or not to spread gossip. With the obtained data, we tried to construct a computational model that may be used for classification of spread gossip. P135 EEG hyperscanning detects neural oscillation for the social interaction during the economic decision-making Jaehwan Jahng1,2, Dong-Uk Hwang3, Jaeseung Jeong1,2 1Department of Bio and Brain Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea; 2Program of Brain and Cognitive Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea; 3Division of Computational Mathematics, National Institute for Mathematical Sciences (NIMS), Daejeon, 34047, South Korea Correspondence: Jaehwan Jahng - jahngjh.627@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P135 Social interaction is an important feature of the economic exchange. However, little is known about the varying neural mechanism during the economic decision-making depending on the different degrees of the social interaction. In this study, we used an iterated version of Prisoner’s dilemma game (PDG) with an EEG hyperscanning to investigate how the presence of face-to-face interaction modulates the social interactions and in turn the aspects of an economic decision-making. Participants played the game either face-to-face (FF) or face-blocked (FB). On the behavioral level, face-to-face interaction led both participants to choose cooperative strategies more often. On the neural level, FF groups showed significantly different alpha power during the first 0.5 s after seeing each outcome compared with FB groups in right temporo-parietal region. By computing the phase locking value (PLV), we measured the brain synchrony and found that the inter-brain phase synchronies across right temporo-parietal area were significantly associated with both the group differences and strategical differences of both players (Fig. 74). These results suggest that inter-brain alpha synchronies across right temporo-parietal area might serve as an implicit neural marker for both the social interaction level and intention to either cooperate or defect. Moreover, our results warrant the future hyperscanning studies on the social interactions of autism spectrum disorder (ASD) patients as all neural substrates revealed are known to be deeply associated with their social traits.Fig. 74 A Brain synchrony analyses. Intra-brain and inter-brain phase synchronies in alpha band [0.5, 1] s. Links between electrodes means that the phase activities there are synchronized. All synchronies here were higher in FF groups than in FB groups (gray line). Blue line denotes the synchronies that were higher in CC epochs compared with DD epochs of FF groups (CC > DD) whereas red line denotes the synchronies that were higher in DD epochs compared with CC epochs of FF groups (CC < DD). Intra-brain synchronies are drawn in both brains and only one of each pair of inter-brain synchronies are drawn. Significant level was at p < 0.05, Bonferroni corrected. B Magnitudes of phase synchronies that showed significant strategical differences. These correspond to the links depicted as blue and red line in A. * p < 0.05; ** p < 0.01, Bonferroni corrected Acknowledgements: This research was supported by the CHUNG Moon Soul Research Center for Bio Information and Bio Electronics (CMSC) in KAIST and a Korea Science and Engineering Foundation (KOSEF) grant funded by the Korean government (No. 2006-2005399). The funders had no role in study design, data collection and analysis, the decision to publish, or the preparation of the manuscript. P136 Detecting purchase decision based on hyperfrontality of the EEG Jae-Hyung Kwon1,2, Sang-Min Park1,2, Jaeseung Jeong1,2 1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea; 2Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea Correspondence: Jae-Hyung Kwon - jh2393@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P136 Understanding and predicting purchase decision process is one of the fundamental issues in economics, marketing, decision sciences, yet an easily accessible means to monitoring purchase decisions has not been developed yet [1–3]. Using event-related functional fMRI, such purchase behavior with a shopping task was investigated [4] but it has limit on potential practical uses due to the cost and portability of the MRI. The electroencephalogram (EEG) has been suggested to have many advantages for applications in marketing due to its relatively low cost, portability, and high temporal resolution. The aim of the current study was to determine the possibility of the EEG as a tool for detecting and predicting purchase decision in potential consumers. Twenty-three participants were recruited to record their EEGs as they saw the pictures of products followed by the products’ prices and made the choice of whether to buy them or not. We estimated the power spectra and approximate entropy (ApEn), an information-theoretic measure to quantify the complexity [5], of their EEGs and compared them for purchase and non-purchase trials. The support vector machine (SVM) method was to predict their purchase decisions. We found that the relative spectral powers and ApEn values of the EEG significantly differed between purchase and non-purchase trials, in particular frontal regions. SVM could distinguish and predict purchase and non-purchase decisions based on the spectral powers and ApEn values of the EEGs in frontal regions prior to the decision moment with a high accuracy (>87 %). This finding suggests that relatively inexpensive, portable EEG recording technique has great potential as a neural predictor of purchase behavior in neuromarketing and neuroeconomics. ReferencesLee N, Broderick AJ, Chamberlain L. What is “neuromarketing”? A discussion and agenda for future research. Int J Psychophysiol. 2007;63:199–204. Mirja Hubert PK. A current overview of consumer neuroscience. J Consum Behav. 2008;7:272–92. Ariely D, Berns GS. Neuromarketing: the hope and hype of neuroimaging in business. Nat Rev Neurosci. 2010;11:284–92. Knutson B, Rick S, Wimmer GE, Prelec D, Loewenstein G. Neural predictors of purchases. Neuron. 2007;53:147–56. Gu F, Meng XIN, Shen E, Cai Z. Can we measure consciousness with EEG complexities? Int J Bifurc Chaos. 2003;13:733–42. P137 Vulnerability-based critical neurons, synapses, and pathways in the Caenorhabditis elegans connectome Seongkyun Kim1, Hyoungkyu Kim1, Jerald D. Kralik1, Jaeseung Jeong1 Department of Bio and Brain Engineering, Program of Brain and Cognitive Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea Correspondence: Jerald D. Kralik - jerald.kralik@raphe.kaist.ac.kr, Jerald D. Kralik - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P137 Determining the fundamental architectural design of complex nervous systems will lead to significant medical and technological advances. Yet it remains unclear how nervous systems evolved highly efficient networks with near optimal sharing of pathways that yet produce multiple distinct behaviors to reach the organism’s goals. To determine this, we investigated the vulnerability of the nematode roundworm Caenorhabditis elegans connectome [1] by attacking each of 279 individual neurons and 6393 chemical synapses and 890 electrical junctions in the connectome, and quantifying the lethality of the network in terms of global information processing using graph-theoretic measures: i.e., examining vulnerability with respect to clustering (C), efficiency (E), and betweenness (B). The vulnerability analyses, VC, VE, VB, identified 12 critical neurons and 29 critical synapses that are the most important components for establishing fundamental network properties. These critical elements were found to be control elements—i.e., those with the most influence over multiple underlying pathways. In addition, we found that the critical synapses formed into circuit-level control units, suggesting fractal-like control in the connectome. More specifically, three main critical pathways emerged from the results (Fig. 75A, B).Fig. 75 Three critical pathways emerged from the results. A For V B they were: (1) AVA-based; (2) PVP-based; and (3) RMD → OLL. B Two of these pathways were again implicated for V E: the AVA-based and the PVP-based pathways Conclusions The critical pathways that emerged from our computational analysis provide evidence for (a) the importance of backward locomotor control, avoidance behavior, and social feeding to the organism; (b) the potential roles of specific neurons whose functions have been unclear; and (c) both parallel and serial design elements in the connectome—i.e., specific evidence for a mixed architectural design. This design structure may be fundamental to nervous systems, providing necessary building blocks for the evolution of higher intelligence. Acknowledgements:: Supported by the National Research Foundation of Korea (NRF-2013R1A1A2011570). ReferenceAltun Z, Hall D. Worm Atlas; 2002. http://www.wormatlas.org. P138 Motif analysis reveals functionally asymmetrical neurons in C. elegans Pyeong Soo Kim1, Seongkyun Kim1, Hyoungkyu Kim1, Jaeseung Jeong1 1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon 305-701, South Korea Correspondence: Jaeseung Jeong - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P138 Majority of animal species have bilaterally symmetrical nervous system. Symmetric and asymmetric features among their morphological symmetric nervous system have been interesting issue for long time. The simplest bilaterally symmetrical organism is nematode called Caenorhabditis elegans. Previously, symmetry for C. elegans has only been thoroughly studied in morphological and functional manner [1]. According to previous observation, there are 92 bilaterally symmetrical neuronal pairs and remaining 95 neurons are mostly located on the axis of symmetry. Functionally there are only 2 neuronal pairs that show asymmetrical gene expression among 92 pairs of symmetrical neurons. We examined the symmetry of C. elegans nervous network which has not been studied. Total of 279 neurons and 2990 links in C. elegans were used. Neurons were classified into bilaterally symmetrical neurons, unlateral neurons, and unilateral neurons. According to the neuronal positions, we could define the symmetry of each individual link and expand that definition to define the symmetry of motif [2]. After defining symmetry of nervous network, we suggest a novel approach to classify asymmetric neurons of C. elegans nervous system by examining asymmetric network topology for every node. We defined 5 explicit locally topological parameters for a neuron; (1) the degree is defined as the number of asymmetric links attached to the neuron, (2) the motif is defined as distribution of the numbers of asymmetric motifs for a neuron, (3) the degree ratio is defined as ratio of asymmetric links over totally attached links to the neuron including both of symmetric links and asymmetric links, (4) the motif ratio is distribution of the rates for asymmetric motifs over total motifs including both of symmetric and asymmetric motifs, and (5) the relative distance is defined by the difference of asymmetric motif fingerprint of bilaterally symmetrical neurons. Thresholds were defined using mean and standard deviation (SD) values of asymmetries to find statistically asymmetric components. Neurons with asymmetry value over the threshold were considered as asymmetric neurons (asymmetric neurons > SD from the mean values). We checked our asymmetric neurons with ASE and AWC neurons that are only known to show bilaterally asymmetrical gene expression. As a result, our study suggested that (4) ratio of asymmetric motif and (5) relative distance measures successfully classified ASE and AWC as asymmetric neurons. Except for ASE and AWC neurons, BDU, PLM, and PVW neurons are classified asymmetric in both measures. These results could be interpreted that BDU neurons, PLM neurons, and ALN neurons might possess asymmetric features that have not been discovered. ReferencesOliver H, Johnston RJ, Chang S. Left–right asymmetry in the nervous system: the Caenorhabditis elegans model. Nat Rev Neurosci. 2002;3(8):629–40. Sporns O, Kötter R. Motifs in brain networks. PLoS Biol. 2004;2(11):e369. P139 Computational approach to preference-based serial decision dynamics: do temporal discounting and working memory affect it? Sangsup Yoon1,2, Jaehyung Kwon1,2, Sewoong Lim1,2, Jaeseung Jeong1,2 1Department of Bio and Brain Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea; 2Program of Brain and Cognitive Engineering, College of Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, South Korea Correspondence: Jaeseung Jeong - jsjeong@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P139 When we face the multiple options (such as different taste of chocolates in a box) and choose one by one sequentially, there is no reason to prefer particular order of choice than any other possible choice strategies since all the items will eventually be consumed by ourselves. Recent studies have, however, revealed that there are distinct patterns of choice strategy in this preference-based serial decision tasks. Interestingly, there are two opposite choice patterns (favorite-first and favorite-last) in human subjects [1], while non-human animals (rhesus monkeys) only chose their favorite options at first [2]. Although several hypotheses for underlying neural mechanisms have been suggested to explain about how these distinct choice strategies appeared, they are not directly tested yet. The goal of the current study was to examine whether temporal discounting and working memory affect choice strategy of serial decision-making and if so, to examine how they influence it. To measure the choice strategy, we used the modified version of ‘the sushi problem’ task [1], which use the pictures of opposite sex as a reward instead of the sushi [3]. We also measured the temporal discounting parameters and working memory performance by using the same set of opposite-sex pictures. The whole pictures were rated by each subject twice before the main experiment, the average rating score were subsequently used to divide the whole picture set into four groups based on the difference of subject specific preference. In ‘sushi’ task, subjects were asked to choose among four options (squares contain different number of stars; 1 ~ 4), each of which represents following short presentation (1.5 s) of pictures right after their choice. Each trial ends when subject choose all of four options, so they couldn’t skip or miss any options. The temporal discounting parameter was measured by the subject choice between sooner-small reward and later-larger reward, in this case, the magnitude of reward was the number of stars which indicate the subject-specific attractiveness of each pictures, the delay of reward was relatively shorter (1–30 s) than typical temporal discounting task since our task offered actual outcome (see the picture) of each choices [4]. The picture version of n-back task was used to measure the working memory performance. Consistent with previous studies, we observed distinct patterns of choice order in ‘sushi’ task, favorite-last strategy was most dominant (58 %) and favorite-first was second (31 %). We also found the relationship of both temporal discounting and working memory with choice strategies. The favorite-last group showed significantly lower rate of temporal discounting and higher performance of working memory than favorite-first group. The effects of working memory and temporal discounting parameters on choice strategy were examined by logistic regression analysis, which revealed how the propensity to discount future events and the memory effect about recent events predicted the pattern of serial choice. We also constructed simple computational models using support vector machine and naïve Bayes classifier to predict their decision patterns based on working memory performance and temporal discounting parameters. We showed that these computational models successfully predicted preference-based decision patterns. ReferencesJaeseung J, Younmin O, Miriam C, Jerald DK. Preference-based serial decision dynamics: your first sushi reveals your eating order at the sushi table. Plos One. 2014;9(5):e96653 Kanghoon J, Jerald DK. Get it while it’s hot: a peak-first bias in self-generated choice order in rhesus macaques. Plos One. 2013;8(12):e83814. Itzhak A, Nancy E, Dan A, Christopher FC, Ethan O, Hans CB. Beautiful faces have variable reward value: fMRI and behavioral evidence. Neuron. 2001;32:537–51. Benjamin YH, Purak CP, Robert OD, Michael LP. Economic principles motivating social attention in humans. Proc R Soc B. 2007;274:1751–56. P141 Social stress induced neural network reconfiguration affects decision making and learning in zebrafish Choongseok Park1, Thomas Miller2, Katie Clements2, Sungwoo Ahn3, Eoon Hye Ji4, Fadi A. Issa2 1Department of Mathematics, North Carolina A&T State University, Greensboro, NC, 27411, USA; 2Department of Biology, East Carolina University, Greenville, NC, 27858, USA; 3Department of Mathematics, East Carolina University, Greenville, NC, 27858, USA; 4David Geffen School of Medicine, UCLA, Los Angeles, CA, 90095, USA Correspondence: Choongseok Park - issaf14@ecu.edu, Fadi A. Issa - cpark@ncat.edu BMC Neuroscience 2016, 17(Suppl 1):P141 In many social species, behavioral mechanisms of how social hierarchies formed and maintained have been studied extensively [1]. However, the neural bases underlying behavioral decisions and dynamics of neural circuits that permit animals to adapt to changes in social rank are poorly understood. In this study we focused on two social stress induced behaviors in zebrafish [the Mauthner cell (M-cell) mediated startle escape response and swimming behavior] to investigate how social regulation affects intrinsic cellular and network properties that result in the behavioral differences between dominant (DOMs) and subordinate (SUBs) animals. We utilized a non-invasive technique that allowed us to monitor the activation pattern of the two neural circuits in freely behaving animals. High behavioral responsiveness and a low stimulus threshold for the initiation of escape in M-cell were observed in SUBs while DOMs showed the quicker habituation to repeated auditory stimulation compared to SUBs. We also observed that on average SUBs generated significantly less number of swim bursts compared to DOMs. These results suggest that social status induced stress can modify the startle plasticity as well as the local swimming circuit. The change in M-cell’s excitability due to the change in the presynaptic inhibitory drive may be responsible for the lowered threshold. On the other hand, the local neural circuits and their intrinsic modulatory components (motor neurons and interneurons) may be configured differently according to social status to produce status-dependent swim patterns [2]. To test these ideas, we developed a biologically-based mathematical model whose network architecture is based on recent experimental data [3]. The model is able to reproduce several hallmarks of social status induced behavioral differences that were experimentally observed between DOMs and SUBs, as well as some inherent activity patterns. Changing some intrinsic synaptic and network parameters was sufficient to obtain the transition between DOMs and SUBs activity patterns while maintaining the network architecture. Recent experiments show that the startle plasticity in M-cell can be modulated by endocannabinoids, 2-AG [3]. We chose the availability of 2-AG in M-cell as one of main parameters for the simulation, whose dynamics is governed by the intracellular calcium level in M-cell. Model simulation shows that high behavioral responsiveness in SUBs results from the increased excitability in M-cell, which can be interpreted as the reduced inhibitory input to M-cell. To reproduce less swimming activity in SUBs, the hallmark of social status induced behavioral difference observed in our experiments, we chose another intrinsic parameter, the availability of 2-AG in inhibitory interneurons to represent 2-AG modulated local network property. Model simulation shows that less swimming activity in SUBs is produced by the increased inhibitory input to the swimming neural circuit via the 2-AG driven elevated interneuron activity. Acknowledgements: This work was partially supported by Simons Foundation Collaboration Grants for Mathematicians (#317566) to CP and East Carolina University, Department of Biology fund to FAI. ReferencesBergman TJ, Beehner JC, Cheney DL, Seyfarth RM. Hierarchical classification by rank and kinship in baboons. Science. 2003;302(5648):1234–36. Issa FA, Drummond J, Cattaert D, Edwards DH. Neural circuit reconfiguration by social status. J Neurosci. 2012;32(16):5638–45. Song J, Ampatzis K, Ausborn J, Manira AE. A hardwired circuit supplemented with endocannabinoids encodes behavioral choice in zebrafish. Curr Biol. 2015;25:2610–20. P142 Descriptive, generative, and hybrid approaches for neural connectivity inference from neural activity data JeongHun Baek1, Shigeyuki Oba1, Junichiro Yoshimoto2,3, Kenji Doya2, Shin Ishii1 1Graduate School of Informatics, Kyoto University, Yoshidahonmachi 36-1, Sakyo, Kyoto, Japan; 2Neural Computation Unit, Okinawa Institute of Science and Technology Graduate University, 1919-1 Tancha, Onna-son, Kunigami-gun, Okinawa, Japan; 3Graduate School of Information Science, Nara Institute of Science and Technology,8916-5 Takayama, Ikoma, Nara, Japan Correspondence: JeongHun Baek - ku21fang@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P142 Identification of the connectivity between neurons is important not only for elucidating neural bases for various functions but also for reconstructing the dynamics emerged in the connectivity. This thesis considers efficient methods for estimating synaptic connections from neural activity data. There are two kinds of approaches to neural connectivity inference: analytic one based on descriptive statistics and reproductive one based on statistical generative models. Analyses based on descriptive statistics, such as Pearson correlation, can identify neural connectivity based on activity data, with low computational cost. It, however, cannot reproduce the dynamic behaviors of the underlying connectivity, and hence, it is not suited for simulating the identified network. Reproductive approach based on statistical generative models, such as generalized linear model, can naturally simulate the dynamic behaviors of the identified network, once we determine the network parameters from activity data. Contrary to this advantage, the computational cost of reproductive approach is often much heavier than analytic one. To utilize the preferable characters of the two approaches, in this study, we propose a hybrid approach of using a descriptive statistic for prescreening of existing connections and then performing generative model inference for dynamic model construction. We applied the hybrid approach to artificially generated spike data of various network sizes. Results and conclusions Figure 76A shows the accuracy of functional connectivity analysis, in terms of ROC-AUC of binary classification, presence/absence of connectivity, where we see the hybrid approach performed slightly worse than the GFAM10. Note that the hybrid approach performed almost same with GFAM10 when the number of neurons was 2000. Figure 76B shows the computation time of the two methods, where we see that GFAM10 took five times as much time as the hybrid approach. Our hybrid approach successfully reduced computational time, into about one-fifth of that of the sole reproductive approach based on the GFAM, while maintaining the estimation accuracy of the response functions within the identified functional connectivity.Fig. 76 A Comparison of the prescreening accuracy in terms of ROC-AUC value (higher is better accuracy). B Comparison of the computation time. ‘GFAM10 [1]’, regarded as the original method, denotes generative functional additive model which is extended version of generalized linear model. ‘Correlation-GFAM10’ denotes a hybrid approach which performs the Pearson correlation for prescreening and then performs GFAM10 ReferenceSong D, Wang H, Tu CY, Marmarelis VZ, Hampson RE, Deadwyler SA, Berger TW. Identification of sparse neural functional connectivity using penalized likelihood estimation and basis functions. J Comput Neurosci. 2013;35(3):335–57. P145 Divergent-convergent synaptic connectivities accelerate coding in multilayered sensory systems Thiago S. Mosqueiro1, Martin F. Strube-Bloss2, Brian Smith3, Ramon Huerta1 1University of California San Diego, La Jolla CA, USA; 2Biocenter University of Würzburg, Würzburg, Germany; 3School of Life Sciences, Arizona State University, Tempe, AZ, USA Correspondence: Brian Smith - brian.h.smith@asu.edu BMC Neuroscience 2016, 17(Suppl 1):P145 A central dogma in perception postulates that a minimal number of higher-order neurons provide the coding basis required for decision making and survival [1]. However, sensory information must travel through several neural layers before converging onto a smaller number of neurons in a premotor decision layer [2]. This multi-layered processing and convergence induces a time lag between peripheral input and adaptive behavior, which is inconsistent with the need for reaction speed. We propose that the divergent–convergent organization often occurring in multilayered neuropils enhances processing speed. Insect olfactory processing is a good model for investigating perceptual timing [3], where effective classification in the 4th layer ‘anticipates’ classification in input layers by 50 ms (Fig. 77A, B) [4]. Here we show that this anticipation emerges from divergent-convergent connectivity and the relative sizes of the layers, which rapidly amplifies subtle input signals and improves precision (Fig. 77C). We reproduced experimental results of peak classification in MBONs anticipating PNs by 50 ms on average (Fig. 77D). This becomes more pronounced as the KC layer grows, although increased noise is also observed. For an oversized KC layer, thus, this anticipation becomes lower and the signal is eventually destroyed by the emphasized noise. Interestingly, the key feature to this anticipation is indeed the ratio between KCs and PNs, showing that larger brains may balance these populations to achieve jointly higher pattern recognition capabilities and fasts discrimination times. We have analyzed fast coding properties of fan-out/fan-in structures that are ubiquitous in the brain. We developed a model to reproduce experimental data and analyze the optimal reaction times of the network model, finding a balance between fast information transmission and high accuracy in pattern recognition. Our contribution improves understanding of the role of divergent convergent feedforward networks on the stability of fast and accurate decision-making.Fig. 77 Early discrimination of stimulus in the MBs. A Recordings of PNs and MBONs activities from untrained honey bees to odor stimulation. At t = 0 s (green bar), an odor stimulation is presented. B Connectionist blueprint of the MBs, emphasizing synapses and population size. Note the divergence present between PN and KC layers, followed by convergence onto MBONs. C We reproduced in silico the early response (blue bar) of MBONs in the vertical lobe with respect to the PNs (orange bar) using spiking neuron networks. D Time differences in our simulations for each experiment repetition. Difference in response time between is on average 50 ms (one tail Mann–Whitney test, p < 0.025) ReferencesBarlow HB. Single units and sensation: a neuron doctrine for perceptual psychology? Perception. 2009;38:371–94. Shepherd GM. The synaptic organization of the brain. Oxford: Oxford Press; 2003. Mosqueiro TS, Huerta R. Computational models to understand decision making and pattern recognition in the insect brain. Curr Opin Insect Sci. 2014;6:80–5. Strube-Bloss MF, Herrera-Valdez MA, Smith BH. Ensemble response in mushroom body output neurons of the honey bee outpaces spatiotemporal odor processing two synapses earlier in the antennal lobe. PLoS One. 2012;7:e50322. P146 Swinging networks Michal Hadrava1,2,3, Jaroslav Hlinka2,3 1Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague, Prague, 166 27, Czech Republic; 2Department of Nonlinear Dynamics and Complex Systems, Institute of Computer Science, The Czech Academy of Sciences, Prague, 182 07, Czech Republic; 3National Institute of Mental Health, Klecany, 250 67, Czech Republic Correspondence: Michal Hadrava - hadrava@cs.cas.cz BMC Neuroscience 2016, 17(Suppl 1):P146 Nature is a powerful illusionist who, unfortunately for life sciences, hates revealing her secrets. One of her most rewarding tricks involves interconnecting a bunch of non-oscillatory neurons in such a way that they collectively behave like an oscillator [1]. Contemporary neuroscience strives to decipher this magic and does so not out of mere curiosity: the trick can go wrong, causing myriads of neurons to march to the deadly rhythm of epileptic seizure. It is of vital importance to determine which connectivity patterns promote and suppress epileptiform activity if surgery is to be effective when nothing else is [2]. On a lighter note, oscillatory dynamics explains many aspects of musical experience [3–5]. A question comes up again and again in ethnomusicological discourse as to whether these aspects are learned or not. In the context of oscillatory dynamics on networks, we might ask whether a single connectivity leads to the emergence of dynamics relevant to each musical culture or a different (learned) connectivity is at work each time. In conclusion, establishing a link between connectivity and oscillatory dynamics on networks seems to be an important problem with repercussions in such diverse fields as epileptology and ethnomusicology. The mainstream approach to the problem can be characterized as follows: first, choose a dynamical model of single unit—e.g. neuron, synapse, or population thereof. Next, connect units of the selected type(s) in a network. Finally, study the effect of connectivity parameters on the global dynamics analytically, computationally, or using a combination of both. The major drawback of any analysis performed in this way is that the validity of its results is put in doubt whenever that of the single unit model is. Needless to say, none of the ever-growing variety of models has gained a wide acceptance yet. The mainstream approach could be dubbed the “object-oriented” one. The alternative approach, advocated by category theorists and adopted by us, could be called the “relational” one: instead of analysing a particular dynamical system, one investigates a whole class of dynamical systems on a particular manifold characterized only by its relations to classes of dynamical systems on different manifolds. This latter approach is epitomized by a recently introduced algebraic structure [6] which relates global network dynamics to its connectivity. We are currently trying to prove the existence of global periodic solutions in selected classes of simple networks with a given structure using this new theory. Acknowledgements: This work was supported by the Grant Agency of the Czech Technical University in Prague, grant No. SGS14/192/OHK3/3T/13, the Czech Science Foundation Project No. P303-14-02634S, and the Czech Health Research Council Project No. NV15-29835A. ReferencesBuzsáki G. Rhythms of the brain. New York: Oxford University Press; 2006. Dixit AB, Banerjee J, Tripathi M, Chandra PS. Presurgical epileptogenic network analysis: a way to enhance epilepsy surgery outcome. Neurol India. 2015;63(5):743–50. Cartwright JHE, Gonzalez DL, Piro O. Pitch perception: a dynamical-systems perspective. Proc Natl Acad Sci USA. 2001;98(9):4855–59. Large EW: A dynamical systems approach to musical tonality. In: Huys R, Jirsa VK, editors. Studies in computational intelligence: nonlinear dynamics in human behavior, vol 328. Berlin: Springe; 2011. p. 193–211. Large EW, Snyder JS. Pulse and meter as neural resonance. In: DallaBella S, Kraus N, Overy K, Pantev C, Snyder JS, Tervaniemi M, Tillmann B, Schlaug G, editors. Neurosciences and music III: disorders and plasticity, vol 1169; 2009. p. 46–57. Lerman E, Spivak DI. An algebra of open continuous time dynamical systems and networks. arXiv:1602.01017v1 [math.DS]. P147 Inferring dynamically relevant motifs from oscillatory stimuli: challenges, pitfalls, and solutions Hannah Bos1, Moritz Helias1,2 1Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, 52425 Jülich, Germany; 2Department of Physics, Faculty 1, RWTH Aachen University, 52074 Aachen, Germany Correspondence: Hannah Bos - h.bos@fz-juelich.de BMC Neuroscience 2016, 17(Suppl 1):P147 Applications of oscillatory stimuli in optogenetical studies have been used to gather evidence that γ oscillations are generated by the interaction of inter-neurons (also termed the inter-neuron γ or ING mechanism) [1, 2]. We elaborate the pitfalls of inferring the origin of the oscillation from absolute (response spectra) as well as relative (power ratios) changes in spectra of neural activity induced by oscillatory input. We consider minimalistic models that isolate the difficulties and limitations arising in the interpretation of response spectra. The described effects generalize to more realistic models. This is demonstrated in simulations of a multi-laminar model of V1 composed of leaky-integrate-and-fire (LIF) model neurons [3], where the ground truth regarding the sub-circuits generating the oscillations is known [4]. In this structured model these effects combine and yield misleading results. By extending mean-field theoretical descriptions of population dynamics [5] by oscillatory input, we can close the loop to the condensed models. We identify three main complications: First, the input can modify the excitability of the population in a linear or non-linear fashion, yielding significantly different changes in the spectra. Second, depending on the properties of the system, the input to the populations is potentially low pass filtered before it enters the system. Since this low pass filter is reflected in the response spectra, without revealing information regarding the internal dynamics of the network, we propose a stimulation protocol counteracting this effect by emphasizing high frequencies. Third, in general, the stimulation of a single population excites a mixture of dynamical modes. One frequency is generated by one dynamical mode that can be mapped to its anatomical origin [4]. Since the observable response is composed of an inseparable mixture of modes, the mode generating the oscillation cannot easily be isolated. Hence reconstructing the underlying connectivity as well as identifying the role of the stimulated population in the generation of the rhythm is not straightforward. Instead, the stimulus vector needs to reflect the structure of the circuit generating the oscillation in order to allow insights into the dynamically relevant components of the system. These problems can be regarded as a sub-set of challenges that need to be faced when interpreting the results of circuits composed of more complex units. The proposed solutions may be used to construct new experimental stimulation protocols. Acknowledgements: We acknowledge funding by the Helmholtz Association: portfolio theme SMHB and Young Investigator’s Group VH-NG-1028, and 604102 (Human Brain Project). All network simulations were carried out with NEST (http://www.nest-initiative.org). ReferencesCardin JA, Carlé M, Meletis K, Knoblich U, Zhang F, Deisseroth K, Tsai L-H, Moore CI. Driving fast-spiking cells induces gamma rhythm and controls sensory responses. Nature. 2009;459:663–7. Buzsáki G, Wang XJ. Mechanisms of gamma oscillations. Annu Rev Neurosci. 2012;35:203–25. Potjans TC, Diesmann M. The cell-type specific cortical microcircuit: relating structure and activity in a full-scale spiking network model. Cereb Cortex. 2014;24:785–806. Bos H, Diesmann M, Helias M. Identifying anatomical origins of coexisting oscillations in the cortical microcircuit. 2015, arXiv preprint arXiv:1510.00642. Brunel N. Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. J Comput Neurosci. 2000;8:183–208. P148 Spatiotemporal mapping of brain network dynamics during cognitive tasks using magnetoencephalography and deep learning Charles M. Welzig1, Zachary J. Harper1,2 1Departments of Neurology and Physiology, Medical College of Wisconsin, Milwaukee, WI 53226, USA; 2College of Engineering and Applied Science, University of Wisconsin-Milwaukee, Milwaukee, WI 53211, USA Correspondence: Charles M. Welzig - welzig@mcw.edu BMC Neuroscience 2016, 17(Suppl 1):P148 Magnetoencephalography (MEG) offers the high spatiotemporal resolution necessary to capture dynamic mesoscale cortical activation features [1] for spatiotemporal mapping of brain networks. In order to adapt such data for effective pathological or cognitive state classifiers, novel techniques are required to extract complex connectivity dynamics that vary in duration and latency. We have developed an advanced deep-learning system to explore such network dynamics through parcellated connectomes in individual subjects. The machine learning system produces transparent classifiers that can define spatiotemporal characteristics of state-specific connectivity, model neurophysiological pathology and expand understanding of connectivity dynamics. Our implementation uses source localized activation patterns extracted from event related epochs to classify cognitive states corresponding to working memory tasks. Subject-wise MEG data is mapped to segmented morphology from magnetic resonance imaging for source localization, then preprocessed to optimize neural network performance. The following steps minimize dimensionality and accommodate the deep-learning system’s input requirements. First, spatiotemporal data are encoded using a wavelet transformation that extracts oscillatory data in the theta, alpha and beta/gamma bands per parcel (Fig. 78A). Next, synchronicity between parcels is calculated to populate 2D connectivity matrices respective to the frequency bands. These matrices are normalized and combined into frame images where theta, alpha and beta/gamma synchronicity is encoded as blue, green and red intensity values respectively. These pixel grids are smoothed and expanded using gridded, cubic interpolation (Fig. 78B). The deep learning system consists of a recursive neural network utilizing a long-short term memory (LSTM) architecture [2] that preserves temporal input characteristics (Fig. 78C). LSTM presents dynamically changing oscillatory patterns to the deep learning system by integrating features of a specified range of contiguous frames relative to each training frame. As this classification system allows for visualization of activation at each layer, we are able to identify specific patterns that mediate the classification process (Fig. 78D) [3].Fig. 78 Visualisations of processing and deep learning stages. A Wavelet decomposition across bands of interest. B Progression of image-encoded oscillatory synchronization in BA10. C Deep learning network improvement across training epochs. D One trained network layer displaying parcel dynamics that mediate classification Our classification methods have demonstrated significantly low error rate of 0.236 ± 0.425 (mean ± SD) in binary working memory state classification after 3 min of GPU-accelerated training. Additionally, weight patterns at specific layers within the deep learning network highlighted relevant parcel interactions with significant effect on functional connectivity dynamics within classified cognitive states. This project represents an advancement in preserving critical spatiotemporal information required to classify complex cognitive states that characterize dynamically changing oscillatory and synchronous functional activity patterns across the connectome. ReferencesLarson-Prior LJ, Oostenveld R, Della Penna S, Michalareas G, Prior F, Babajani-Feremi A, Schoffelen JM et al. Adding dynamics to the human connectome project with MEG. NeuroImage. 2013;80:190–201. Donahue J, Hendricks LA, Guadarrama S, Rohrbach M, Venugopalan S, Saenko K, Darrell T. Long-term recurrent convolutional networks for visual recognition and description; 2014. arXiv:14114389. Plis SM, Hjelm D, Salakhutdinov R, Allen EA, Bockholt HJ, Long JD, Johnson HJ, Paulsen J, Turner JA, Calhoun VD. Deep learning for neuroimaging: a validation study. Front Neurosci. 2014;8. P149 Multiscale complexity analysis for the segmentation of MRI images Won Sup Kim1, In-Seob Shin1, Hyeon-Man Baek2, Seung Kee Han1 1Department of Physics, Chungbuk National University, Cheongju, Chungbuk 28644, Republic of Korea; 2Korea Basic Science Institute, Cheongju, Chungbuk 28119, Republic of Korea Correspondence: Seung Kee Han - skhan@chungbuk.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P149 Segmentation of human brain images into regions with homogeneous intensity or texture is very crucial for the diagnosis of various brain diseases. However, the presence of noises or artifacts remains as one of the biggest obstacles for the successful segmentation. Here we propose a novel method of segmentation based on the multiscale complexity analysis. The idea is to characterize the complexity in visual images by the multiscale profile representing the scale dependence of compositional complexity. Our claim is that the multiscale profile of human brain images combining scale dependent information on intensity and texture information could be effectively utilized for the segmentation of human brain images. We have applied the multiscale complexity analysis for the segmentation of two dimensional MRI images. Our method consists of three steps. (I) An MRI image is partitioned into homogeneous regions utilizing the information bottleneck method. (II) Multiscale complexity profiles of individual pixels are computed from the partitioned image of the MRI. (III) Feature vectors combining both intensity and texture information are extracted for the segmentation. For the segmentation, the feature vectors of individual pixels are clustered using a simple K-mean clustering algorithm. Using the simulated MRI images provided the BrainWeb database [1], the performance of the segmentation was tested. The performance shown in Fig. 79 indicates that the multiscale complexity analysis is very robust against noise. Details will be presented during the meeting.Fig. 79 A An MRI image from the BrainWeb [1] and the result of segmentation into five clusters. B An MRI image with 7 % noise added and the result of segmentation into five clusters ReferenceCocosco C, Kollokian V, Kwan RS, Evan A. BrainWeb: on line interface to a 3D MRI simulated brain database. NeuroImage. 1997;5:S425. P150 A neuro-computational model of emotional attention René Richter1, Julien Vitay1, Frederick Beuth1, Fred H. Hamker1,2 1Department of Computer Science, Chemnitz University of Technology, Chemnitz, Germany; 2Bernstein Center for Computational Neuroscience, Charité University Medicine, Berlin, Germany Correspondence: René Richter - rene.richter@cs.tu-chemnitz.de BMC Neuroscience 2016, 17(Suppl 1):P150 Emotional stimuli attract attention so the brain can focus its processing resources on them. The questions that arise is how these stimuli acquire their emotional value and how they can influence attentional processes. Evidence suggests that this association might be learned through conditioning in the amygdala, more specifically the basal lateral amygdala (BLA). Furthermore, feedback connections from the BLA to the visual cortex seem to enhance the activation of neural representations which is a possible top-down attention mechanism of the emergent attention hypothesis. While neuro-computational models of attention mechanisms attract increasing interest due to their importance for the focused processing of information in the brain, the possible emotional feedback from the amygdala is to date largely unexplored. Therefore, we propose a rate-coded, biological realistic neuro-computational model constructed of 3 smaller functional models. First, we combined a model of the visual processing pathway for object recognition [1] that includes the retina, the lateral geniculate nucleus, the visual areas V1, V2 and V4 as well as the frontal eye field with an amygdala model for the associative conditioning of a visual stimulus with a bodily reaction representing a particular emotional state. Second, in order to provide the model with realistic temporal learning properties, a reward-timing model [2] simulating the afferent system to the dopaminergic area VTA has been integrated to temporally adjust the learning process through dopamine-mediated modulation of plasticity. This timing model includes a number of brain areas, most prominently the ventral tegmental area, the nucleus accumbens, the lateral hypothalamus, the ventral medial prefrontal cortex and the amygdala. In order to enable emotional attention, 2 simulation phases were implemented: (1) a conditioning phase to learn the association between an important stimulus and the body reaction, and (2) an attention phase where the representation of the visual stimulus activates the BLA which then sends back a feedback to enhanced this specific stimulus. Afterwards, the enhanced representation in V4 suppresses the competing ones and allows the frontal eye field to initiate a saccade in its direction. As a result of the biologically based connectivity and the realistic learning process, the model outcomes are coherent with several experimental findings and increase our understanding of the brain network’s interaction. In the future, the model could furthermore be used for facial analysis and the process of learning the importance of specific facial features for emotional expressions. ReferencesBeuth F, Hamker FH. A mechanistic cortical microcircuit of attention for amplification, normalization and suppression. Vis Res. 2015;116:241–57. Vitay J, Hamker FH. Timing and expectation of reward: a neuro-computational model of the afferents to the ventral tegmental area. Front Neurorobot. 2014;8:1–25. P151 Multi-site delayed feedback stimulation in parkinsonian networks Kelly Toppin1, Yixin Guo1 1Department of Mathematics, Drexel University, Philadelphia, PA 19104, USA Correspondence: Yixin Guo - yixin@math.drexel.edu BMC Neuroscience 2016, 17(Suppl 1):P151 The conventional deep brain stimulation (DBS), as a surgical procedure to alleviate debilitating and disrupting symptoms of Parkinson’s disease (PD), has several drawbacks. Multi-site delayed feedback stimulation (MDFS) has been proposed as a feasible alternative to overcome the drawbacks of the conventional DBS [2, 3]. We first build two types of large scale biophysical networks to explore the effectiveness of MDFS. The persistent parkinsonian network has strongly synchronized bursting clusters with elevated firing rates present in subthalmic nucleus (STN), internal and external segments of globus pallidus (GPi and GPe) neurons. However, the brain of a PD patient may not be in a constant strong synchronized and clustered state, and short desynchronized events may present when the brain is in between high synchronization [1]. We build an intermittent parkinsonian network that can transit between synchronized and desynchronized dynamics. Using both parkinsonian networks, we compute the TC error index, the fraction of miss responses and excessive responses when a TC neuron relays multiple excitatory inputs, in five different stimulation settings: MDFS from STN to STN, from STN to GPe, from GPi to STN, from GPi to GPe and from GPi to GPi, shown by the dashed-arrows labeled 1–5 in Fig. 80A. Each “to” population is stimulated by the signal based on the LFP calculated at the “from” population. Our results of lower TC relay errors with the five different stimulations in Fig. 80B show that MDFS improves the fidelity of the TC relay neurons’ communication and responses to the input motor signal in both persistent and intermittent parkinsonian networks. We also find that MDFS with STN or GPe as a stimulation target is more effective in reducing TC relay errors.Fig. 80 A The network model. Plus symbol indicates excitatory connection, and minus symbol indicates inhibitory connection. LFP is computed from GPi and STN populations separately. These two LFP signals are used as the source of the five MDFS stimulations (dashed-arrows): STN-to-STN, STN-to-GPi, GPi-to-STN, GPi-to-GPe and GPi-to-GPi, shown by arrow labeled 1–5. B Error index values for 80 different model TC neurons in an intermittent network. Comparison of MDFS among different stimulation targets using either STN or GPi LFP signal. Whisker plots show mean (red line), 25–75 percentile range (blue box), 95 % confidence interval (black lines) and outliers (red plus signs) Acknowledgements: The study was supported by NSF grant DMS-1226180 awarded to Yixin Guo. ReferencesAhn S, Rubchinsky L. Short desynchronization episodes prevail in synchronous dynamics of human brain rhythms. Chaos. 2013;23:013138. Guo Y, Rubin JE. Multi-site stimulation of subthalamic nucleus diminishes thalamocortical relay errors in a biophysical network model. Neural Netw. 2011;24(6):602-16. Hauptmann C, Omel’Chenko O, Popovych V, Maistrenko Y, Tass PA. Control of spatially patterned synchrony with multisite delayed feedback. Phys Rev E. 2007;76(6):066209. P152 Bistability in Hodgkin–Huxley-type equations Tatiana Kameneva1, Hamish Meffin2, Anthony N. Burkitt1, David B Grayden1,3 1NeuroEngineering Laboratory, Department of Electrical & Electronic Engineering, University of Melbourne, Parkville, VIC 3010, Australia; 2National Vision Research Institute, Australian College of Optometry, Carlton, VIC 3053, Australia; 3Centre for Neural Engineering, University of Melbourne, Parkville, VIC 3010, Australia Correspondence: René Richter - tkam@unimelb.edu.au BMC Neuroscience 2016, 17(Suppl 1):P152 Background Purkinje cells have two states of the resting membrane potential: a hyperpolarized quiescent state (down state) and a depolarized spiking state (up state) [1]. This bistability has been observed in in vitro and in vivo recordings, in anesthetized animals, and in slices. It has been proposed that bistability in Purkinje cells play a key role in the short-term processing and storage of sensory-motor information. Methods To investigate bistability of the neuronal resting state, we use computer simulations in neuron. We simulate single compartment neurons and use the Hodgkin–Huxley-type formalism to study how initial conditions and a combination of ionic channels affect neuronal response. We systematically apply intracellular current pulse stimulation to set the membrane potential to different levels and observe the neuronal dynamics after the stimulation is released. Results We show that the neural response after release of the pulse stimulation depends on the amplitude of the current pulses. For some stimulation levels, the cells return to the level of the activity prior to stimulation, while for other levels, the neuronal dynamics are different to prior activity levels for a long time post stimulation. We show that different initial conditions lead to different neuronal dynamics even when all other parameters in the Hodgkin–Huxley-type model are set the same (Fig. 81). We explore the region of attraction for two stable states and find that they differ for different parameters of the model, in particular that different ionic channel combinations do not change our qualitative results.Fig. 81 Response of a modeled neuron for different initial conditions, V(0). Sodium, three types of potassium, calcium, and hyperpolarisation-activated currents are included in the model Conclusions This work demonstrates a potential method to explore the mechanisms underlying bistability in Purkinje cells. In particular, the proposed methodology allows the exploration of the circumstances under which Purkinje cells transit from the down state to the up state and return. This work implies that results obtained using the Hodgkin–Huxley formalism should be carefully considered since the choice of initial conditions may significantly affect the final outcome. Acknowledgements: This research was supported by the Australian Research Council (ARC). TK acknowledge support through ARC Discovery Early Career Researcher Award (DE120102210). ReferenceLoewenstein Y, Mahon S, Chadderton P, Kitamura K, Sompolinsky H, Yarom Y, Hausser. Bistability of cerebella Purkinje cells modulated by sensory stimulation. Nat Neurosci. 2005;8(2):202–11. P153 Phase changes in postsynaptic spiking due to synaptic connectivity and short term plasticity: mathematical analysis of frequency dependency Mark D. McDonnell1, Bruce P. Graham2 1Computational and Theoretical Neuroscience Laboratory, School of Information Technology and Mathematical Sciences, University of South Australia, Mawson Lakes, SA, 5095, Australia; 2Computing Science and Mathematics, School of Natural Sciences, University of Stirling, Stirling, FK9 4LA, UK Correspondence: Mark D. McDonnell - mark.mcdonnell@unisa.edu.au BMC Neuroscience 2016, 17(Suppl 1):P153 We examined how short-term synaptic depression due to vesicle depletion [1] interacts with the configuration of the synaptic pathways onto an output neuron. Using both simulations and mathematical analysis, we found significant frequency-dependent phase-shifts of the spiking response of a neuron driven by independent frequency-modulated Poisson input signals. The synaptic inputs to the neuron are assumed to consist of a fixed number of release sites that are divided between active zones, with each active zone being the presynaptic axonal target of a single input neuron (Fig. 82A). For the same number of release sites, at one extreme the output neuron receives input from a large number of neurons through independent active zones, each containing a single release site, similar to cortical cells. At the other extreme (similar to a Calyx of Held in the auditory brainstem), the neuron is driven by a single input neuron through a giant synapse containing a single active zone with a very large number of release sites.Fig. 82 The phase of post-synaptic firing in response to frequency-modulated inhomogenous Poisson pre-synaptic spike trains depends on the configuration of input synapses. A The connectivity involves M independent pre-synaptic neurons each with NM synaptic release sites at a post-synaptic neuron. Vesicles are released probabilistically when activated by a pre-synaptic action-potential, if one is available at that site. The post-synaptic neuron depolarizes and produces action potentials after arrival of neurotransmitter according to standard models. B The phase lead of output spiking relative to periodic modulation at frequency f, for M active-zones is both frequency-dependent and configuration-dependent. The figure is from simulations but we also derived the same result mathematically Using standard stochastic models of short term depression due to vesicle depletion [2], and post-synaptic current dynamics, we found strong phase dependencies for input modulation rates up to 5 Hz. The phase shift also depends strongly on the configuration (Fig. 82B). However, the phase shift otherwise remains invariant for a wide range of post-synaptic conditions, such as for Hodgkin–Huxley or leaky integrate-and-fire models, and whether or not the dynamics of post synaptic currents included rise-times, or longer or shorter decay times. Acknowledgements: M. D. McDonnell was supported by an Australian Research Fellowship from the Australian Research Council (project DP1093425). B. P. Graham’s contribution was supported by the BBSRC project grant BB/K01854X/1. ReferencesAbbott LF, Regehr WG. Synaptic computation. Nature. 2004;431:796–803. McDonnell MD, Mohan A, Stricker C. Mathematical analysis and algorithms for efficiently and accurately implementing stochastic simulations of short-term synaptic depression and facilitation. Front Comput Neurosci. 2013;7:58. P154 Quantifying resilience patterns in brain networks: the importance of directionality Penelope J. Kale1, Leonardo L. Gollo1 1Systems Neuroscience Group, QIMR Berghofer Medical Research Institute, Brisbane, QLD, 4006, Australia Correspondence: Penelope J. Kale - Penelope.Kale@qimr.edu.au BMC Neuroscience 2016, 17(Suppl 1):P154 Defining how interactions take place, directionality is major feature of network connections. Brain networks are intrinsically directed because of the nature of chemical synapses, which comprise most of the neuronal connections. The specific fingerprint of the interactions between cortical regions and neurons thereof are crucial to the neuronal dynamics. The neuronal ability to synchronize is extremely sensitive to the presence of reciprocal connections in neuronal motifs and circuits [1]. The type of synchronization (or the phase relation between phase locked neurons and cortical regions) also depends on the relation between the synaptic strengths between these regions [2, 3]. Moreover, whole brain network dynamics is also shaped by reciprocal connections, which stabilizes the network dynamics and reduce transitions between metastable states [4]. However, due to limitations in current brain imaging techniques, the directionality of edges between structurally connected regions of the human brain cannot be confirmed. Additionally, despite the demonstrated importance of synaptic direction, its effect over main network features is not yet elucidated. Comparing several directed brain networks from different species (macaque, cat, mouse, and C. elegans) and with variable node size (parcellation), we estimate the error that is made in characterizing and identifying brains as complex network when undirected networks are assumed. We use different approaches to turn directed networks undirected: (i) remove unidirectional links; (ii) add reciprocal links; (iii) add one reciprocal for each removed unidirectional link thus keeping the same network density. We find that directionality plays a major role in shaping the brain networks. All regions are affected, including hub nodes, which have large degree and enhanced importance in information integration for cognitive functions [5]. We compute and rank graph theoretical measures and determine their resilience with respect to the loss of directionality of the network. Overall, our results suggest that the characterization of connectomes can be compromised in the absence of data regarding the directionality of brain networks. ReferencesGollo LL, Mirasso C, Sporns O, Breakspear M. Mechanisms of zero-lag synchronization in cortical motifs. PLoS Comput Biol. 2014;10(4):e1003548. Matias FS, Carelli PV, Mirasso CR, Copelli M. Anticipated synchronization in a biologically plausible model of neuronal motifs. Phys Rev E. 2011;84(2):021922. Matias FS, Gollo LL, Carelli PV, Bressler SL, Copelli M, Mirasso CR. Modeling positive Granger causality and negative phase lag between cortical areas. NeuroImage. 2014;99:411–8. Gollo LL, Zalesky A, Hutchison RM, van den Heuvel M, Breakspear M. Dwelling quietly in the rich club: brain network determinants of slow cortical fluctuations. Philos Trans R Soc Lond B Biol Sci. 2015;370(1668):20140165. van den Heuvel MP, Sporns O. Network hubs in the human brain. Trends Cogn Sci. 2013;17(12):683–96. P155 Dynamics of rate-model networks with separate excitatory and inhibitory populations Merav Stern1, L.F. Abbott2 1Faculty of Medicine, Technion, Haifa, Israel; 2Department of Neuroscience and Department of Physiology and Cellular Biophysics, Columbia University, New York, NY, USA Correspondence: Merav Stern - merav.stern@mail.huji.ac.il BMC Neuroscience 2016, 17(Suppl 1):P155 Randomly connected networks of rate-model neurons have a rich dynamics [1], a feature that has been exploited to model a variety of phenomena [2]. These model networks typically do not distinguish between excitatory and inhibitory neuron classes. Doing this requires constraining the network connectivity matrix to have columns with exclusively positive entries, representing input from excitatory neurons, and with negative entries, representing input from inhibitory neurons. The eigenvalue spectra of random matrices satisfying this constraint have a number of interesting properties [3, 4]. Here we study the dynamics of rate-model networks that result from using such connectivity matrices. We find that neural activity is correlated across all neurons, including both excitatory and inhibitory subpopulations. This correlation depends on the difference between the mean strengths of the excitation and inhibition connections and it increases as this difference is increased. For very large values of this difference, the network reaches a stable fixed point, otherwise it is chaotic. Chaos arises from the residual activity deviating from the correlated mean network activity and it acts to reduce these correlations. The magnitude of the residual chaotic activity is determined by the variances of the synaptic strengths within the excitatory and inhibitory populations. In summary, unlike models with a single mixed excitatory/inhibitory population, in which the activity between pairs of neurons is uncorrelated for every value of synaptic gain, networks with distinct excitatory and inhibitory subpopulations exhibit strongly correlated activity across the entire network reminiscent of the up/down states seen in neural recordings [5]. ReferencesSompolinsky H, Crissanti A, Sommers HJ. Chaos in random neural networks. Cerebral Phys Rev Lett. 1988;61:259–62. Sussillo D. Neural circuits as computational dynamical systems. Curr Opin Neurobiol. 2014;25:156–63. Rajan K, Abbott LF. Eigenvalue spectra of random matrices for neural networks. Phys Rev Lett. 2006;97:188104. Tao T. Outliers in the spectrum of IID matrices with bounded rank perturbations. arXive: 1012.4818v6. Steriade M, Nunez A, Amzica F. A novel slow (<1 Hz) oscillation of neo-cortical neurons in vivo: depolarizing and hyperpolarizing components. J. Neurosci. 1993;13:3252–65. P156 A model for multi-stable dynamics in action recognition modulated by integration of silhouette and shading cues Leonid A. Fedorov1,2, Martin A Giese1,2 1Section for Computational Sensomotorics, Department of Cognitive Neurology, CIN&HIH, Tübingen, Germany; 2GTC, International Max Planck Research School, University of Tübingen, Tübingen, Germany Correspondence: Leonid A. Fedorov - leonid.fedorov@uni-tuebingen.de BMC Neuroscience 2016, 17(Suppl 1):P156 The visual perception of body motion can show interesting multi-stability. For example, a walking body silhouette (bottom inset Fig. 83A) is seen alternately as walking in two different directions [1]. For stimuli with minimal texture information, such as shading, this multi-stability disappears. Existing neural models for body motion perception [2–4] do not reproduce perceptual switching. Extending the model [2], we developed a neurodynamic model that accounts for this multi-stability (Fig. 83A). The core of the model is a two-dimensional neural field that consists of recurrently coupled neurons with selectivity for instantaneous body postures (‘snapshots’). The dimensions of the field encode the keyframe number θ and the view of the walker ϕ. The lateral connectivity of the field stabilizes two competing traveling pulse solutions that encode the perceived temporally changing action patterns (walking in the directions ±45°). The input activity of the field is generated by two visual pathways that recognize body postures from gray-level input movies. One pathway (‘silhouette pathway’) was adapted from [2] and recognizes shapes, mainly based on the contrast edges between the moving figure and the background. The second pathway is specialized for the analysis of luminance gradients inside the moving figure. Both pathways are hierarchical (deep) architectures, built from detectors that reproduce known properties of cortical neurons. Higher levels of the hierarchies extract more complex features with higher degree of position/scale invariance. The field activity is read out by two Motion Pattern (MP) neurons, which encode the two possible perceived walking directions. Testing the model with an unshaded silhouette stimulus, it produces randomly switching percepts that alternate between the walking directions (±45°) (Fig. 83B, C). Addition of shading cues disambiguates the percept and removes the bistability (Fig. 83D). The developed architecture accounts for the disambiguation by shape-from shading.Fig. 83 A Model architecture with 2D neural field that receives input from two hierarchical path-ways. B Response traces of MP neurons for silhouette stimulus without shading during a 200 s simulation. C Corresponding average response times of the output neurons. D Response times for shaded stimulus Acknowledgements: Supported by EC Fp7-PEOPLE-2011-ITN PITN-GA-011-290011 (ABC), FP7-ICT-2013-FET-F/604102 (HBP), FP7-ICT-2013-10/611909 (Koroibot), BMBF, FKZ: 01GQ1002A, DFG GI 305/4-1 + KA 1258/15-1. ReferencesVangeneugden J, et al. Activity in areas MT+ and EBA, but not pSTS, allows prediction of perceptual states during ambiguous biological motion. Soc Neurosci Meet. 2012;127(04). Giese MA, Poggio T. Neural mechanisms for the recognition of biological movements and action. Nat Rev Neurosci. 2003;4:179–92. Lange J, Lappe M. A model of biological motion perception from configural form cues. J Neurosci. 2006;26:2894–906. Jhuang H, et al. A biologically inspired system for action recognition. In: ICCV 2007. p. 1–8. P157 Spiking model for the interaction between action recognition and action execution Mohammad Hovaidi Ardestani1,2, Martin Giese1 1Section Computational Sensomotorics, CIN & HIH, Department of Cognitive Neurology, Tübingen, 72076, Germany; 2IMPRS for Cognitive and Systems Neuroscience, University Clinic Tübingen, Tübingen, 72076, Germany Correspondence: Mohammad Hovaidi Ardestani - Mohammad.Hovaidi-Ardestani@uni-tuebingen.de BMC Neuroscience 2016, 17(Suppl 1):P157 Action perception and the control of action execution are intrinsically linked in the human brain. Experiments show that the concurrent motor execution influences the visual perception of actions and biological motion (e.g. [1]). This interaction likely is mediated by action-selective neurons in the STS, premotor and parietal cortex. We have developed a model based on biophysically realistic spiking neurons that accounts for the observed interactions between action perception and motor planning. The model is based on two dynamic representation levels (Fig. 84A), one modeling a representation of perceived action patters (vision field), and one representing associated motor programs (motor field). Both levels are modeled by recurrent spiking networks that approximate neural fields, where each field consists of 30 coupled neural ensembles, each consisting of 80 excitatory and 20 inhibitory adaptive Exponential Integrate-and-Fire (aEIF) neurons [2]. Within each field asymmetric recurrent connections between the ensembles stabilize a traveling pulse solution, which is stimulus-driven in the visual field and autonomously propagating in the motor field after initiation by a go-signal. Both fields are coupled by interaction kernels that results in mutual excitation between the fields of the traveling pulse propagate synchronously and in mutual inhibition otherwise. We used the model to reproduce the result of a psychophysical experiment that tested the detection of point-light stimuli in noise during concurrent motor execution [1]. The point-light patterns showed arm movements of the observer, which were synchronized with varying time delays with the executed movements. Compared to a baseline without concurrent motor execution, the detectability of the visual stimulus was higher for very small time delays between the visual stimulus and the executed arm movement, and it was lower when the observed movement was strongly delayed (>300 ms) against the executed motor patterns (Fig. 84B). The same pattern arises from the detectability of the visual stimulus as predicted from our model, where we assumed that the level of neural activity (compared to a noise level) provides a measure for the detectability of the stimulus (Fig. 84C). The proposed model, which is derived by simplification from physiologically-inspired neural models for action execution and motor planning, reproduces correctly the modulation of visual detection by the synchrony of the stimulus with executed motor behavior. Present work extends the model by a full visual pathway and an effector model, allowing for the simulation of a broader spectrum of experimental results.Fig. 84 A Model architecture consisting of two coupled neural fields, implemented with biophysically realistic neurons. B Psychophysical results from [1] showing the dependence of the detectability of visual point-light stimuli in dependence of the delay between a visually observed and the concurrently executed action. C Simulated detectability derived from the model for the same experimental conditions Acknowledgements: Supported by EC FP7-ICT-2013-FET-F/604102 (HBP), Fp7-PEOPLE-2011-ITN PITN-GA-011-290011 (ABC), FP7-ICT-2013-10/611909 (Koroibot), BMBF, FKZ: 01GQ1002A, DFG GI 305/4-1 + KA 1258/15-1. ReferencesChristensen A, Ilg W, Giese MA. Spatiotemporal tuning of the facilitation of biological motion perception by concurrent motor execution. J Neurosci. 2011;31:3493–9. Brette R, Gerstner W. Adaptive exponential integrate-and-fire model as an effective description of neuronal activity. J Neurophysiol. 2005;94:3637–42. P158 Surprise-modulated belief update: how to learn within changing environments? Mohammad Javad Faraji1, Kerstin Preuschoff2, Wulfram Gerstner1 1School of Life Sciences, Brain Mind Institute and School of Computer and Communication Sciences, Ecole Polytechnique Federal de Lausanne (EPFL), CH-1015 Lausanne, Switzerland; 2Geneva Finance Research Institute (GFRI) and Swiss Center for Affective Sciences (CISA), University of Geneva, CH-1211 Geneva, Switzerland Correspondence: Mohammad Javad Faraji - mohammadjavad.faraji@epfl.ch BMC Neuroscience 2016, 17(Suppl 1):P158 Surprise is informative because it drives attention [1] and modifies learning [2]. Correlates of surprise have been observed at different stages of neural processing, and found to be relevant for learning and memory formation [3]. Although surprise is ubiquitous, there is neither a widely accepted theory that quantitatively links surprise to observed behavior, such as the startle response, nor an agreement on how surprise should influence learning speed or other parameters in iterative statistical learning algorithms. Building on and going beyond earlier surprise measures [4–6], we propose a novel information theoretic measure for calculating surprise in a Bayesian framework so as to capture uncertainty of the world as well as imperfections of the subjective model of the world, two important aspects of surprise. The principle of future surprise minimization leads to a learning rule that can be interpreted as a surprise modulated belief update suitable for learning within changing environments. Importantly, we do not need an assumption on how quickly the world changes. We apply our surprise-modulated learning rule to an exploration task in a maze-like environment. Our results are consistent with the behavioral finding that surprising events induce humans and animals to learn faster and to adapt more quickly to changing environments. Information content [5] captures the inherent unexpectedness of a piece of data for a given set of models (uncertainty of the world), whereas Bayesian surprise [4] measures the change in belief caused by a new data point (observer dependent). These are two complementary approaches for calculating surprise. In our approach both aspects are combined with a third aspect: if we are uncertain about what to expect, receiving a low-probability data sample is less surprising than in a situation when we are almost certain about the world. A surprise minimizing learning (SMiLe) rule is derived by solving a constrained optimization problem defined as follows: the objective is to maximally reduce surprise when facing the same data again in the not so far future, under the constraint that the posterior belief (after the update step) is not too different from the prior. The resulting SMiLe rule balances the influence of newly acquired data with prior knowledge where the balance depends on surprise. In case of a fundamental change in the world signaled by surprising samples, data acquired before the change is downgraded as less informative about the current state of the world. A simultaneous increase of the influence of newly acquired data on learning leads to a fast adaptation of the model to an environmental change. While in a stationary environment our algorithm approaches the known Bayesian update rule, it also allows the model to react to changes in the environment. In summary, surprising data increases the uncertainty we have about our current model of the world and gives a bigger influence of newly acquired data on belief update. The interaction between surprise and uncertainty is important for modeling the behavior of humans and animals in changing environments. The surprise signal could be broadly transmitted in the brain by a neuromodulator with widespread axonal ramifications (e.g., norepinephrine (NE) released from locus coeruleus (LC) neurons) and influence synaptic plasticity rules. Acknowledgements: This research was supported by the European Research Council (Grant Agreement No. 268 689). ReferencesItti L, Baldi P. Bayesian surprise attracts human attention. Vis Res. 2009;49(10):1295–1306. Schultz W, Dickinson A. Neuronal coding of prediction errors. Annu Rev Neurosci. 2000;23(1):473–500. Ranganath C, Rainer G. Neural mechanisms for detecting and remembering novel events. Nat Rev Neurosci. 2003;4(3):193–202. Baldi P, Itti L. Of bits and wows: a bayesian theory of surprise with applications to attention. Neural Netw. 2010;23(5):649–66. Shannon CE. A mathematical theory of communication. ACM SIGMOBILE Mobile Comput Commun Rev. 2001;5(1):3–55. Palm G. Novelty, information and surprise. Berlin: Springer; 2012. P159 A fast, stochastic and adaptive model of auditory nerve responses to cochlear implant stimulation Margriet J. van Gendt1, Jeroen J. Briaire1, Randy K. Kalkman1, Johan H. M. Frijns1,2 1ENT-Department, Leiden University Medical Centre, Leiden, 2300 RC, the Netherlands; 2Leiden Institute for Brain and Cognition, Leiden, 2300 RC, the Netherlands Correspondence: Margriet J. van Gendt - m.j.van_gendt@lumc.nl BMC Neuroscience 2016, 17(Suppl 1):P7 Cochlear implants (CI) rehabilitate hearing impairment through direct electrical stimulation of the auditory nerve. In many modern CIs sound is coded through the continuous interleaved sampling (CIS) strategy. Although many different sound-coding strategies have been introduced in the last decade, no major advances have been made since the introduction of the CIS strategy [1]. New stimulation strategies are commonly investigated by means of psychophysical experiments and clinical trials, which is time-consuming for both patient and researcher. Alternatively, strategies can be evaluated using computational models. In this study a computationally efficient model that accurately predicts auditory nerve responses to CI pulse train input is developed. The model includes the 3D volume conduction and active nerve model developed in the Leiden University Medical Center [2], and is extended with stochasticity, adaptation and accommodation. This complete model includes spatial as well as temporal characteristics of both the cochlea and the auditory nerve. The stochastic and adaptive auditory nerve model is used to investigate full-nerve responses to amplitude modulated long duration stimulation. Understanding responses to amplitude modulation is important because current speech coding strategies are based on the principle of speech information distribution through amplitude modulation of the input pulse trains. The model is validated by comparison to experimentally measured single fiber action potential (SFAP) responses to pulse trains published in literature [3–6]. The effects of different pulse-train parameters such as pulse rate, pulse amplitude and amplitude modulation are investigated. The neural spike patterns produced in response to CI stimulation are very similar to spike patterns obtained with single fiber action potential measurements in animal experiments in response to CI stimulation. Besides predicting single fiber responses to constant amplitude pulse trains, the model also very well predicts single fiber responses to amplitude modulated pulse trains. Response alterations seen over the duration of the stimulus are similar to those seen in animal experiments. Modeled effects of stimulus amplitude, pulse rate and amplitude modulation is similar to the effects seen in animal experiments. Adaptation is found to be an important factor in modeling nerve outcomes to amplitude modulated pulse trains and their spatial effects. The model is shown to accurately predict spike timings in response to long duration pulse trains as observed in animal experiments. The model can be used to predict full auditory nerve responses to electrical pulse trains, and thus to different sound coding strategies. The next step will be to apply this model to evaluate complete auditory nerve responses to different sound coding strategies. Acknowledgements: This study was financially supported by Advanced Bionics Corporation. ReferencesZeng FG, Rebscher S, Harrison WV. Cochlear implants: system design, integration and evaluation. IEEE Rev Biomed Eng. 2008;115–42. Kalkman RK, Briaire JJ, Dekker DMT, Frijns JHM. Place pitch versus electrode location in a realistic computational model of the implanted human cochlea. Hear Res. 2014;315:10–24. Miller CA, Hu N, Zhang F, Robinson BK, Abbas PJ. Changes across time in the temporal responses of auditory nerve fibers stimulated by electric pulse trains. J Assoc Res Otolaryngol. 2008;9:122–37. Litvak L, Delgutte B, Eddington D. Auditory nerve fiber responses to electric stimulation: modulated and unmodulated pulse trains. J Acoust Soc Am. 2001;110:368. Zhang F, Miller CA, Robinson BK, Abbas PJ, Hu N. Changes across time in spike rate and spike amplitude of auditory nerve fibers stimulated by electric pulse trains. JARO J Assoc Res Otolaryngol. 2007;8:356–72. Hu N, Miller CA, Abbas PJ, Robinson BK, Woo J. Changes in auditory nerve responses across the duration of sinusoidally amplitude-modulated electric pulse-train stimuli. J Assoc Res Otolaryngol. 2010;11:641–56. P160 Quantitative comparison of graph theoretical measures of simulated and empirical functional brain networks Won Hee Lee1, Sophia Frangou1 Department of Psychiatry, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA Correspondence: Won Hee Lee - wonhee.lee@mssm.edu BMC Neuroscience 2016, 17(Suppl 1):P160 Graph theoretical approaches to resting-state fMRI have been widely used to quantitatively characterize functional network organization in the resting brain, but mechanistic explanations for how resting-state brain works are still lacking. Whole-brain computational models have shown promise in enriching our understanding of mechanisms contributing to the formation and dissolution of resting-state functional patterns [1]. It is therefore important to determine the degree to which computational models reproduce the topological features of empirical functional brain networks. Here, we focused on the performance of the Kuramoto model [2] as it is considered most representative model of coupled phase oscillators and is widely used in the literature. Empirical and simulated functional networks were defined based on 66 brain anatomical regions (nodes). Simulated resting-state functional connectivity (FC) was generated using the Kuramoto model constrained by empirical structural connectivity. The simulated FC matrix was tuned to best fit empirical FC matrix. In order to improve stability and reliability, we simulated 10 runs of fMRI BOLD time series (obtained from 320 s simulations, discarding 20 s initial transients) with varying random initial conditions, and generated the best-fit simulated FC matrix for each run. We applied graph theoretical approaches to optimally simulated FC and empirical FC data to characterize key topological features of brain networks [3]. Finally, we quantified and compared the difference, in terms of relative error, in graph theoretical measures between the simulated and empirical functional networks. Figure 85 shows the quantitative difference in graph theoretical measures between the empirical FC and the simulated FC over the entire (1–100 %) and selected range of connection densities (37–50 %). The averaged relative differences were found to be 2–77 % over the entire range of connection densities as well as 0.1–22 % over a range of 37–50 % connection densities. We found that simulated functional data can be used with confidence to model graph measures of global and local efficiency, characteristic path length, eigenvector centrality, and resilience to targeted attack and random failure. Our results also highlight the critical dependence of the solutions obtained in simulated data on the specified connection density.Fig. 85 Relative error (RE) in percentage between graph theoretical measures of simulated FC versus empirical FC for the entire (1–100 %) and selected range of connection densities (37–50 %). Bars and error bars correspond respectively to the averages and standard deviations across the ten RE values. E glob global efficiency, E loc local efficiency, CC clustering coefficient, L characteristic path length, EC eigenvector centrality, PC participation coefficient, SW: small-worldness, Rtc and Rtg represent resilience to targeted attack in the size of largest connected component and global efficiency, respectively, Rrc and Rrg represent resilience to random failure in the size of largest connected component and global efficiency, respectively This study demonstrates the value of computational models in assessing whole-brain network connectivity, and provides a method for the quantitative evaluation and external validation of graph theory metrics derived from simulated data that can be used to inform future study designs. ReferencesDeco G, Jirsa VK, McIntosh AR: Emerging concepts for the dynamical organization of resting-state activity in the brain. Nat Rev Neurosci. 2011;12(1):43–56. Kuramoto Y. Chemical oscillations, waves, and turbulence. Berlin: Springer; 1984. Rubinov M, Sporns O. Complex network measures of brain connectivity: uses and interpretations. Neuroimage. 2010l52:1059–69. P161 Determining discriminative properties of fMRI signals in schizophrenia using highly comparative time-series analysis Ben D. Fulcher1, Patricia H. P. Tran1, Alex Fornito1 1Monash Institute of Cognitive and Clinical Neurosciences, Monash University, Clayton, Vic 3168, Australia Correspondence: Ben D. Fulcher - ben.fulcher@monash.edu BMC Neuroscience 2016, 17(Suppl 1):P161 Analysis of fMRI data typically focuses on inter-regional functional connectivity, measured as pairwise correlations, or through multivariate decompositions (e.g., ICA). Relatively little attention is given to the univariate time-series properties of BOLD signals within a specific brain region, despite a broad scientific literature on time-series analysis (including power spectral techniques, information theoretic methods, model fitting, nonlinear time-series analysis, and fractal scaling). Here we undertake the largest systematic comparison of over 7000 such measures of temporal structure to identify the temporal features of individual BOLD signals, and their locations in the brain, that are most discriminative of people with schizophrenia. MRI data were obtained from the open COBRE database [4] for 72 people with schizophrenia (SCZ) and 74 healthy controls (CON). For each subject, we extracted 7779 temporal features from the BOLD time series recorded in each of 264 brain regions using the publicly available highly comparative time-series analysis framework, hctsa (http://benfulcher.github.io/hctsa/) [1]. Spatial analysis ROIs that were most discriminative of SCZ versus CON were identified by training a separate linear support vector machine (SVM) classifier for each ROI, across all features, using tenfold cross validation. We identified 23 ROIs with a classification accuracy exceeding chance levels (p < 0.05, FDR-corrected) with some individual ROI accuracies reaching 69.5 %. These discriminative brain regions were mostly located in the frontal and parietal cortices. Temporal analysis The most discriminative temporal features were deduced using t-tests in each ROI, and then averaging across all ROIs. P values were computed using permutation tests with 1000 shuffles. We identified over 100 time-series features of the BOLD signal with statistically significant separability between SCZ and CON (p < 0.05). These features were mostly measures of time series ‘predictability’, including autocorrelation, local prediction error (using exponential smoothing, Gaussian Processes, and AR models), the SD1 measure from the heart rate variability literature, and low frequency power. This emergent class of discriminative properties of BOLD dynamics is consistent with the use of the ALFF metric in existing work using fMRI data [3]. We present the first systematic comparison of thousands of interdisciplinary time-series analysis measures to fMRI data and use machine learning to uncover characteristic BOLD signatures of schizophrenia, in both space and time. In a completely data-driven manner, we identify informative brain regions and time-series analysis techniques that best discriminate people with schizophrenia from healthy controls, using just the properties of BOLD signals in individual ROIs. The framework presented here represents a general and powerful data-driven means of identifying discriminative time-series features from neuroscience data. ReferencesFulcher BD, Little MA, Jones NS. Highly comparative time-series analysis: the empirical structure of time series and their methods. J R Soc Interface. 2013;10:20130048. Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, Church JA, Vogel AC, Laumann TO, Miezin FM, Schlaggar BL, Petersen SE. Functional network organization of the human brain. Neuron. 2011;72:665–78. Yu-Feng Z, Yong H, Chao-Zhe Z, Qing-Jiu C, Man-Qiu S, Meng L, Li-Xia T, Tian-Zi J, Yu-Feng W. Altered baseline brain activity in children with ADHD revealed by resting-state functional MRI. Brain Dev. 2007;29:83–91. COBRE database. http://fcon_1000.projects.nitrc.org/indi/retro/cobre.html. P162 Emergence of narrowband LFP oscillations from completely asynchronous activity during seizures and high-frequency oscillations Stephen V. Gliske1, William C. Stacey1,2, Eugene Lim3, Katherine A. Holman4, Christian G. Fink3,5 1Department of Neurology, University of Michigan, Ann Arbor, MI 48104, USA; 2Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI 48104, USA; 3Department of Physics, Ohio Wesleyan University, Delaware, OH 43015, USA; 4Department of Physics, Towson University, Towson, MD 21252, USA; 5Neuroscience Program, Ohio Wesleyan University, Delaware, OH 43015, USA Correspondence: Christian G. Fink - cgfink@owu.edu BMC Neuroscience 2016, 17(Suppl 1):P162 Recent experimental studies have demonstrated the emergence of narrowband local field potential oscillations during epileptic seizures in which the underlying neural activity appears to be completely asynchronous [1]. We derive a mathematical model explaining how this counterintuitive phenomenon may occur, showing that a population of independent, completely asynchronous neurons may produce narrowband oscillations if each neuron fires quasi-periodically. This quasi-periodicity can occur through cells with similar frequency–current (f–I) curves receiving a similar, high amount of uncorrelated synaptic noise. Thus, this source of oscillatory behavior is distinct from the usual cases (pacemaker cells entraining a network, or oscillations being an inherent property of the network structure), as it requires no oscillatory drive nor any specific network or cellular properties other than cells that repetitively fire with continual stimulus. We deduce bounds on the degree of variability in neural spike-timing which will permit the emergence of such oscillations, both for action potential- and postsynaptic potential-dominated LFPs. (See Fig. 86 for example voltage traces and energy spectra resulting from asynchronous neural activity, demonstrating how our model naturally explains why PSPs tend to dominate the LFP at low frequency, while APs dominate at high frequency.) These results suggest that even an uncoupled network may generate collective rhythms, implying that the breakdown of inhibition and high synaptic input often observed during epileptic seizures may generate narrowband oscillations. We propose that this mechanism may explain why so many disparate epileptic pathologies can produce similar high frequency oscillations [2].Fig. 86 Normalized energy spectra and voltage traces resulting from asynchronous neural activity. A, B Results of superimposed, asynchronous action potential waveforms for quasi-periodic frequencies of 100 Hz (A) and 200 Hz (B). C, D Results of superimposed, asynchronous postsynaptic potential waveforms for quasi-periodic frequencies of 100 Hz (A) and 200 Hz (B). Gray dashed lines represent energy spectra that would result from Poisson process spike trains convolved with AP/PSP waveforms ReferencesTruccolo W, et al. Neuronal ensemble synchrony during human focal seizures. J Neurosci. 2014;34:9927–44. Engel Jr J, Bragin A, Staba R, Modi I. High-frequency oscillations: what is normal and what is not? Epilepsia. 2009;50:598–604. P163 Neuronal diversity in structure and function: cross-validation of anatomical and physiological classification of retinal ganglion cells in the mouse Jinseop S. Kim1,2, Shang Mu2, Kevin L Briggman3, H. Sebastian Seung2,4, and the EyeWirers5 1Department of Structure and Function of Neural Networks, Korea Brain Research Institute, Daegu 41068, Republic of Korea; 2Princeton Neuroscience Institute, Princeton University, Princeton, NJ 08544, USA; 3Circuit Dynamics and Connectivity Unit, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, MD 20824, USA; 4Computer Science Department, Princeton University, Princeton, NJ 08544, USA; 5http://eyewire.org Correspondence: Jinseop S. Kim - jinseop.s.kim@kbri.re.kr BMC Neuroscience 2016, 17(Suppl 1):P163 The neural computation of visual perception begins in the retina. The retinal neural circuits receive inputs from the photoreceptors, spread out along interneurons, and converge to retinal ganglion cells (RGCs). The axons of RGCs are the only output of the retina and carry all the visual information from the retina to the rest of the brain. Each type of RGCs is thought to be associated with one microcircuit and to process distinct visual information. Therefore, classifying the types is an important step towards understanding the neural computation in the retina and retina’s role in vision [1, 2]. We anatomically classified roughly 400 RGCs based mainly on dendritic stratification profiles [3]. The RGC dendritic arbors were reconstructed from serial electron microscope (EM) images of a (0.3 mm)2 slice of the inner plexiform layer of the mouse retina [4]. The reconstruction was carried out on EyeWire, a web-based EM reconstruction pipeline that combines artificial intelligence of deep learning and human intelligence of a community of ‘citizen neuroscientists’ [5]. This is the first time EM reconstruction was done on a large enough area to potentially sample and identify all RGC types. For cross-validation of the anatomical classification, we compared it with the visual responses of the same cells recorded by calcium imaging performed before EM preparation. The comparison confirmed that our classification recovered all well-known ganglion cell types including on–off direction selective ganglion cells (DSGCs), sustained/transient On DSGCs, asymmetric Off DSGC types, sustained/transient and On/Off alpha cells, and local edge detectors. We also found orientation selective or direction selective responses in some cell types that were not previously well-characterized or were previously unknown. In all, our classification includes over 40 types of RGCs. ReferencesSanes JR, Masland RH. The types of retinal ganglion cells: current status and implications for neuronal classification. Annu Rev Neurosci. 2015;38:221–46. Baden T, Berens P, Franke K, Rosón MR, Bethge M, Euler T. The functional diversity of retinal ganglion cells in the mouse. Nature. 2016;529:345–50. Sümbül U, Song S, McCulloch K, Becker M, Lin B, Sanes JR, Masland RH, Seung HS. A genetic and computational approach to structurally classify neuronal types. Nat Commun. 2014;5:3512. Briggman KL, Helmstaedter M, Denk W. Wiring specificity in the direction-selectivity circuit of the retina. Nature. 2011;471:183–88. Kim JS, Greene MJ, Zlateski A, Lee K, Richardson M, Turaga SC, Purcaro M, Balkam M, Robinson A, Behabadi BF, et al. Space–time wiring specificity supports direction selectivity in the retina. Nature. 2014;509:331–6. P164 Analysis and modelling of transient firing rate changes in area MT in response to rapid stimulus feature changes Detlef Wegener1, Lisa Bohnenkamp1,2, Udo A. Ernst2 1Brain Research Institute, University of Bremen, 28334 Bremen, Germany; 2Institute for Neurophysics, University of Bremen, 28334 Bremen, Germany Correspondence: Detlef Wegener - wegener@brain.uni-bremen.de BMC Neuroscience 2016, 17(Suppl 1):P164 Neurons in area MT of the primate visual system are strongly tuned to the direction and speed of moving stimuli, and they exhibit pronounced transients in their firing rates after changes in visual stimulation. These transients increase the sensitivity of neurons and they are closely correlated to behavioral performance. For example, arbitrary instantaneous speed changes are associated with transients of different sign and amplitude, which closely correlate with the sign and magnitude of the preceding stimulus change and with behavioral performance [1, 2]. Interestingly, the transients’ size cannot be directly referred from the neuron’s underlying speed tuning, and is significantly more pronounced if the base speed before the change is far from the neuron’s preferred speed. Understanding the neural dynamics shaping these responses, and their effects on information transmission of arbitrary time-varying signals, is key to understanding how the visual system copes with dynamic scenes. We here present a dynamical model for MT neurons that reproduces detailed characteristics of experimentally observed transients (Fig. 87). The model takes the single cell’s kinetics and its speed tuning into account. Based on divisive inhibition of excitation, it is capable to reproduce and explain the specific transients of single neurons. Single direction column are made up of one excitatory and one inhibitory population, with the inhibitory population providing divisive inhibition onto the excitatory population. By combining multiple direction columns to one hypercolumn, the model consists of N × 2 populations, with the excitatory populations receiving different input depending on their tuning parameters and the stimulus, and the inhibitory populations receiving an input averaged over the neighboring columns’ input by a Gaussian kernel plus a fixed offset. Using an optimization procedure, the model reliably reproduces MT cell responses to arbitrary accelerations and decelerations of a moving stimulus, starting from both low and high base speeds, reproducing recently unexplained experimental data. If the inhibitory time constant is a multiple of the excitatory time constant, the model is analytically tractable for a piecewise constant input current: The analytical solution allows quantifying the transients’ magnitude as a function of general neuron parameters such as response gain and time constants, providing precise predictions for population responses to brief events of arbitrary contrast.Fig. 87 A, B Fits to motion onset responses to estimate each neuron’s kinetics. C Experimentally estimated MT transients to positive and negative speed changes of various magnitude. D Transient response amplitudes as derived from the model. E, F Relation between transient and sustained MT responses as a function of speed change magnitude as estimated experimentally (E) and by the model (F) ReferencesGalashan FO, Saßen HC, Kreiter AK, Wegener D. Monkey area MT latencies to speed changes depend on attention and correlate with behavioral reaction times. Neuron. 2013;78(4):740–50. Traschütz A, Kreiter AK, Wegener D: Transient activity in monkey area MT represents speed changes ind is correlated with human behavioral performance. J Neurophysiol 2015, 113(3):890-903. P165 Step-wise model fitting accounting for high-resolution spatial measurements: construction of a layer V pyramidal cell model with reduced morphology Tuomo Mäki-Marttunen1, Geir Halnes2, Anna Devor3,4, Christoph Metzner5, Anders M. Dale3,4, Ole A. Andreassen1, Gaute T. Einevoll2,6 1NORMENT, Institute of Clinical Medicine, University of Oslo, Norway; 2Department of Mathematical Sciences and Technology, Norwegian University of Life Sciences, Ås, Norway; 3Department of Neurosciences, University of California San Diego, La Jolla, CA, USA; 4Department of Radiology, University of California San Diego, La Jolla, CA, USA; 5Biocomputation Research Group, University of Hertfordshire, Hatfield, UK; 6Department of Physics, University of Oslo, Norway Correspondence: Tuomo Mäki-Marttunen - tuomomm@uio.no BMC Neuroscience 2016, 17(Suppl 1):P165 Novel imaging methods such as intracellular Ca2+ imaging and voltage-sensitive dye measurements provide ever finer spatiotemporal data about single-neuron activity. The challenge for model fitting methods is to incorporate these data in order to describe the neuron behavior in a manner that faithfully preserves the signal propagation and membrane potential dynamics across the neuronal dendrites. A difficulty in this task is the evidently large number of different ion channels residing along the dendritic and perisomatic locations: Unless extra care is taken, the role of specific species of ion channel could be under- or overestimated at the expense of another type of ion channel. In this work, we propose an automatic step-wise model fitting procedure as a solution to this challenge. Our approach resembles that of [1], but our objective functions are designed to account for correct membrane potentials not only at soma but also along the dendrites. In addition, we replace the need for spatial occlusion of parts of dendrite (“pinching”) [2] in the experimental setup by a cumulative use of ion channel blockers. We apply this procedure to construct a reduced-morphology version of the layer V pyramidal cell model of [3]. We simulated the cumulative blocking of ion channels by setting the corresponding ion channel conductances to zero in the full model, and measured the membrane potential (and Ca2+ concentrations when needed) along the soma and dendrites at each step. We then fitted the maximal conductances in the model with reduced morphology in four steps, starting with passive parameters (1st step), continuing with Ih current conductances (2nd step), Ca2+ dynamics and related conductances (3rd step), and ending with ion channel conductances that are in charge of the spiking behavior (4th step). We show that our model with reduced morphology correctly reproduces important aspects of the membrane potential dynamics across the neuron, both in the control condition (see Fig. 88), and under the effect of the abovementioned ion channel blockers. In the final step of our study, we present and apply a method for reducing the number of synaptic contacts (from 1000s to a few 100s) yet maintaining the spatio-temporal activation pattern of the neuron. The obtained network model is cost-efficient in terms of both simulation time and memory requirements. Our model is publicly accessible in ModelDB, accession number 187474, as NEURON and NeuroML-2 descriptions (https://senselab.med.yale.edu/ModelDB/showModel.cshtml?model=187474).Fig. 88 Comparison of model with reduced (red) morphology to the model with full (blue) morphology. The y-axis shows the membrane potential at soma (solid) and apical dendrite (dashed) as a response to a somatic 200-ms DC pulse ReferencesBahl A, Stemmler MB, Herz AV, Roth A. Automated optimization of a reduced layer 5 pyramidal cell model based on experimental data. J Neurosci Methods. 2012;210(1):22–34. Bekkers JM, Häusser M. Targeted dendrotomy reveals active and passive contributions of the dendritic tree to synaptic integration and neuronal output. Proc Natl Acad Sci. 2007;104(27):11447–52. Hay E, Hill S, Schürmann F, Markram H, Segev I. Models of neocortical layer 5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS Comput Biol. 2011;7:e1002107. P166 Contributions of schizophrenia-associated genes to neuron firing and cardiac pacemaking: a polygenic modeling approach Tuomo Mäki-Marttunen1, Glenn T. Lines2, Andy Edwards2, Aslak Tveito2, Anders M. Dale3, Gaute T. Einevoll4, Ole A. Andreassen1 1NORMENT, Institute of Clinical Medicine, University of Oslo, Norway; 2Simula Research Laboratory and Center for Cardiological Innovation, Oslo, Norway; 3Multimodal Imaging Laboratory, UC San Diego, La Jolla, CA, USA; 4Department of Mathematical Sciences and Technology, Norwegian University of Life Sciences, Ås, Norway Correspondence: Tuomo Mäki-Marttunen - tuomomm@uio.no BMC Neuroscience 2016, 17(Suppl 1):P166 A recent genome-wide association study (GWAS) of schizophrenia (SCZ) has identified more than a hundred genetic loci exceeding genome-wide significance, confirming the polygenic nature of the disorder [1]. The loci implicate genes that encode numerous ion channel subtypes and calcium transporters, and are major contributors not only to the function of brain cells, but also to the functioning of organs outside the central nervous system, such as heart. Meta-studies have reported a 2.5-fold–threefold increase in mortality rates in schizophrenic patients, and majority of these excess deaths are natural and mostly due to cardiovascular disease [2]. In agreement with this observation, GWASs of cardiac phenotypes, such as electrocardiographic (ECG) measures, highlight a set of genes that overlaps with the one discovered in GWASs of SCZ. Nevertheless, both the genetic and mechanistic connections between cardiac and neural phenotypes in SCZ patients remain poorly understood. In this work, we use computational modeling to study the contribution of SCZ-associated genes to cardiac and neuronal excitability. We focus our analyses on two central, well-studied cell types, namely, layer V pyramidal cells (L5PCs) in the cortex and sinoatrial node cells (SANCs) in the myocardium. The apical tuft of an L5PC serves as an integration hub for non-local synaptic inputs, and is considered a biological substrate for cortical associations providing high-level “context” for low-level (e.g., sensory) inputs that arrive to the perisomatic compartment. Therefore, the ability of L5PC to integrate the apical and perisomatic inputs has been proposed as one of the mechanisms that could be impaired in hallucinating patients. The SANCs, in turn, have a key role in controlling the heart rate as the primary pacemakers of the mammalian heart. Both of these cell types are well described in terms of biophysical modeling, and are therefore a suitable target for a detailed computational studies incorporating genetic effects. We apply two recent multicompartmental L5PC models and two recent SANC models to argue for the generality of our findings. We show that small changes in the parameters governing the voltage-dependence and time constants of activation and inactivation of different ion channels caused observable effects in both L5PC and SANC function. In the case of Ca2+ channel gene variants, these changes typically had opposite effects on cell excitability in L5PCs compared to SANCs (higher L5PC firing frequency ↔ lower SANC pacemaking frequency), while in the case of Na+ or HCN channel variants, the effects were mostly similar (higher L5PC firing frequency ↔ higher SANC pacemaking frequency). Furthermore, many of the studied variants showed an impact on signal propagation in a chain of coupled SANCs. Our results may help explain some of the cardiac comorbidity in schizophrenia, and may facilitate generation of effective antipsychotic medications with less arrythmia side-effects. ReferencesRipke S, Sanders AR, Kendler KS, Levinson DF, Sklar P, Holmans PA, Lin DY, Duan J, Ophoff RA, Andreassen OA et al.: Genome-wide association study identifies five new schizophrenia loci. Nat Gen. 2011;43:969–76. Laursen TM, Munk-Olsen T, Vestergaard M. Life expectancy and cardiovascular mortality in persons with schizophrenia. Curr Opin Psychiatry. 2012;25(2):83–8. Hay E, Hill S, Schürmann F, Markram H, Segev I: models of neocortical layer 5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS Comput Biol. 2011;7:e1002107. Almog M, Korngreen A. A quantitative description of dendritic conductances and its application to dendritic excitation in layer 5 pyramidal neurons. J Neurosci. 2014;34(1):182–96. Kharche S, Yu J, Lei M, Zhang H. A mathematical model of action potentials of mouse sinoatrial node cells with molecular bases. Am J Physiol Heart Circ Physiol. 2011;301(3):H945–63. Severi S, Fantini M, Charawi LA, DiFrancesco D. An updated computational model of rabbit sinoatrial action potential to investigate the mechanisms of heart rate modulation. J Physiol. 2012;590(18):4483–99. P167 Local field potentials in a 4 × 4 mm2 multi-layered network model Espen Hagen1, Johanna Senk1, Sacha J van Albada1, Markus Diesmann1,2,3 1Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, 52425, Germany; 2Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, Aachen, 52074, Germany; 3Department of Physics, Faculty 1, RWTH Aachen University, Aachen, 52074, Germany Correspondence: Espen Hagen - e.hagen@fz-juelich.de BMC Neuroscience 2016, 17(Suppl 1):P167 The local field potential (LFP), the low-frequency part of extracellular potentials in neural tissue, is routinely recorded as a measure of population activity. LFPs reflect correlated activity of both local and remote neurons and depend on the anatomy and electrophysiology of neurons near the recording location. While forward models have shed light on various aspects of LFPs, e.g., their spatial reach [1], such models often ignore network interactions. Large-scale network models commonly use point neurons for tractability (see, e.g., [2]). However, predicting the LFP signal from such models is not straightforward, as point neurons do not generate extracellular potentials. In [3] we provided methods to compute extracellular potentials from point-neuron networks incorporating the biophysical principles of LFP generation using multicompartment neurons. This hybrid scheme uses spike times of point neurons as spatially dependent synaptic input with layer specificity of connections from anatomical data. The methods were demonstrated using a laterally homogeneous, layered point-neuron network representing 1 mm2 of early sensory cortex at full cell and synapse density [4]. Preserving biological cell and connection densities is critical: networks may not be strongly downscaled without affecting correlations [5], and diluted LFP-generating populations fail to preserve the effect of correlations on the LFP [3]. Even small network correlations dominate in the compound LFP spectrum due to the different scaling of average single-cell LFP spectra and average pairwise coherence of single-cell LFP. Here, we extend this work to a network covering 4 × 4 mm2 (Fig. 89A) accounting for connection probabilities falling off with lateral distance. Even for low pairwise spike-train correlations (Fig. 89B), the model accounts for highly correlated LFPs across lateral distance (Fig. 89C) as observed experimentally. Further we show that such features strongly depend on network state.Fig. 89 A Instantaneous spiking and LFP in a 4-layer network model covering 4 × 4 mm2 at realistic cell and synapse density with distance-dependent connectivity. B Pairwise correlations between spike trains of exc. (E) and inh. (I) layer 5 neurons as function of distance (red: E–E, blue: I–I, black: E–I). C Distance-dependent LFP correlation computed for a 10 × 10 electrode grid in layer 5 (0.4 mm between contacts) Acknowledgements: EU FP7 grant 604102 (HBP); Helmholtz Portfolio Supercomputing and Modeling for the Human Brain (SMHB). ReferencesLindén H, Tetzlaff T, Potjans TC, Pettersen KH, Gruen S, Diesmann M, Einevoll GT. Modeling the spatial each of the LFP. Neuron. 2011;72:859–72. Bos H, Diesmann M, Helias M. Identifying anatomical origins of coexisting oscillations in the cortical microcircuit. arXiv:1510.00642 [q-bio.NC] 2016. Hagen E, Dahmen D, Stavrinou ML, Lindén H, Tetzlaff T, van Albada SJ, Grün S, Diesmann M, Einevoll GT. Hybrid scheme for modeling local field potentials from point-neuron networks. arXiv:1511.01681 [q-bio.NC] 2016. http://inm-6.github.io/hybridLFPy. Potjans TC, Diesmann M. The cell-type specific cortical microcircuit: relating structure and activity in a full-scale spiking network model. Cereb Cortex. 2014;24:785–806. Van Albada SJ, Helias M, Diesmann M. Scalability of asynchronous networks is limited by one-to-one mapping between effective connectivity and correlations. PLoS Comput Biol. 2015;11(9):e1004490. P168 A spiking network model explains multi-scale properties of cortical dynamics Maximilian Schmidt1, Rembrandt Bakker1,2, Kelly Shen3, Gleb Bezgin4, Claus-Christian Hilgetag5,6, Markus Diesmann1,7,8, Sacha Jennifer van Albada1 1Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6), and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany; 2Donders Institute for Brain, Cognition and Behavior, Radboud University, Nijmegen, Netherlands; 3Rotman Research Institute, Baycrest, Toronto, Ontario M6A 2E1, Canada; 4McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University, Montreal, Canada; 5Department of Computational Neuroscience, University Medical Center Eppendorf, Hamburg, Germany; 6Department of Health Sciences, Boston University, Boston, MA, USA; 7Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, Aachen, Germany; 8Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany Correspondence: Maximilian Schmidt - max.schmidt@fz-juelich.de BMC Neuroscience 2016, 17(Suppl 1):P168 Neural networks in visual cortex are structured into areas, layers, and neuronal populations with specific connectivity at each level. Cortical dynamics can similarly be characterized on different scales, from single-cell spiking statistics to the structured patterns of interactions between areas. A challenge of computational neuroscience is to investigate the relation of the structure of cortex to its dynamics. Network models are promising tools, but for technical and methodological reasons, they have been restricted to detailed models of one or two areas or large-scale models that reduce the internal structure of areas to a small number of differential equations. We here present a multi-scale spiking network model of all vision-related areas of macaque cortex that represents each area by a full-scale microcircuit with area-specific architecture based on a model of early sensory cortex [1]. The layer- and population-resolved network connectivity integrates axonal tracing data from the CoCoMac database with recent quantitative tracing data, and is systematically refined using dynamical constraints [2]. Gaps in the data are bridged by exploiting regularities of cortical structure such as the exponential decay of connection densities with inter-areal distance and a fit of laminar patterns versus logarithmized ratios of neuron densities. Simulations reveal a stable asynchronous irregular ground state with heterogeneous activity across areas, layers, and populations. In the presence of large-scale interactions, the model reproduces longer intrinsic time scales in higher compared to early visual areas, similar to experimental findings [3]. Activity propagates preferentially in the feedback direction, mimicking experimental results associated with visual imagery [4]. Cortico-cortical interaction patterns agree well with fMRI resting-state functional connectivity [5]. The model bridges the gap between local and large-scale accounts of cortex, and clarifies how the detailed connectivity of cortex shapes its dynamics on multiple scales. Acknowledgements: VSR computation time Grant JINB33, Helmholtz Portfolio SMHB, EU Grant 269921 (BrainScaleS), EU Grant 604102 (Human Brain Project, HBP), SFB936/A1, Z1 and TRR 169/A2. ReferencesPotjans TC, Diesmann M. The cell-type specific cortical microcircuit: relating structure and activity in a full-scale spiking network model. Cereb Cortex. 2014;24:785–806. Schuecker J, Schmidt M, van Albada SJ, Diesmann M, Helias M. Fundamental activity constraints lead to specific interpretations of the connectome. arXiv preprint 2015, arXiv:1509.03162. Murray JD, Bernacchia A, Freedman DJ, Romo R, Wallis JD, Cai X, Padoa-Schioppa C, Pasternak T, Seo H, Lee D, Wang X-J. A hierarchy of intrinsic timescales across primate cortex. Nat Neurosci. 2014;17:1661–3. Dentico D, Cheung BL, Chang J-Y, Guokas J, Boly M, Tononi G, Van Veen B. Reversal of cortical information flow during visual imagery as compared to visual perception. Neuroimage. 2014;100:237–243. Shen K, Bezgin G, Hutchison RM, Gati JS, Menon RS, Everling S, McIntosh AR. Information processing architecture of functionally defined clusters in the macaque cortex. J Neurosci. 2012;32:17465–76. P169 Using joint weight-delay spike-timing dependent plasticity to find polychronous neuronal groups Haoqi Sun1,2,3,5, Olga Sourina2,5, Guang-Bin Huang3,5, Felix Klanner4,5, Cornelia Denk5 1Energy Research Institute @ NTU (ERI@N), Interdisciplinary Graduate School, Nanyang Technological University, Singapore 639798; 2Fraunhofer IDM @ NTU, Nanyang Technological University, Singapore 639798; 3School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798; 4School of Computer Engineering, Nanyang Technological University, Singapore 639798; 5Future Mobility Research Lab, A Joint Initiative of BMW Group & NTU, Nanyang Technological University, Singapore 639798 Correspondence: Haoqi Sun - hsun004@e.ntu.edu.sg BMC Neuroscience 2016, 17(Suppl 1):P169 It is known that polychronous neuronal groups (PNGs), i.e. neuron groups having reproducible time-locked but not synchronous firing patterns, can function as representative entities [1]. They have huge capacity by sharing neurons. They compete between each other to represent sensory inputs. Therefore, PNG is considered as one of the potential, yet elusively difficult to analyze, hypothetical mechanisms of memory in the brain. In computational models, the difficulties of finding PNGs mainly come from (1) low percentage of spikes from PNGs (about 4 % [1]) when driven by random inputs; and (2) combination explosion to enumerate all possible PNGs for template-matching (possible PNGs triggered by 3 neurons in a 1000-neuron network is 3C1000 = 1.66 × 108). Here we aim at solving the second difficulty without template-matching by connecting PNG readout neurons with joint weight-delay spike-timing dependent plasticity (joint STDP) to the network. The joint STDP consists of (1) weight STDP with the conventional exponential learning window; and (2) (axonal) delay STDP with learning window of shape te−t/τ, scaled by weight-related gains. The joint STDP strengthens and pulls together spikes arriving before postsynaptic firing, on the other hand weakens and postpones spikes after postsynaptic firing. In this way, we can recover the PNG by looking at (1) the strengthened synapses, which tells which neurons belong to the PNG; and (2) the delays of the strengthened synapses, which are complementary to the spike timing inside the PNG, because the presynaptic spike arrival times for the readout neuron (=spike timing + delay) are pulled close to each other. In the experiment, we repeatedly fed structured inputs to a sparsely connected network of 800 excitatory and 200 inhibitory neurons. There were 150 readout neurons connected to the network with lateral inhibition between them. After 405 s of simulation, we used the incoming weights and delays of the readout neurons to find PNGs (see Fig. 90). It turned out that the readout neurons can learn the subsets of the persistently activated PNGs. The readout neurons do not rely on template-matching. Instead, they become differentiated members of the PNG to indicate the actual activation of its subsets.Fig. 90 A The spike raster plot showing 0.6 s of simulation. The vertical axis shows neuron index. Neurons from index 1 to 100 receive structured inputs. Colored spikes refer to PNGs founded by the same colored spikes of readout neurons (above the dash line), where the letter-marked ones are shown in other panels. B A recovered PNG with the predicted spike timing (receptive field) and the actual spikes. The numbers are neuron indices. C The same PNG in B but activated at another time. D, E Another PNG ReferenceIzhikevich EM. Polychronization: computation with spikes. Neural Comput. 2006;18(2):245–82. P170 Tensor decomposition reveals RSNs in simulated resting state fMRI Katharina Glomb1, Adrián Ponce-Alvarez1, Matthieu Gilson1, Petra Ritter2,3,4,5, Gustavo Deco1,6 1Center for Brain and Cognition, Universitat Pompeu Fabra, 08018 Barcelona, Spain; 2Minerva Research Group Brain Modes, Max Planck Institute for Human Cognitive and Brain Sciences, 04103 Leipzig, Germany; 3Department of Neurology, Charité - University Medicine, 10117 Berlin, Germany; 4Bernstein Focus State Dependencies of Learning & Bernstein Center for Computational Neuroscience, 10115 Berlin, Germany; 5Berlin School of Mind and Brain & Mind and Brain Institute, Humboldt University, 10117 Berlin, Germany; 6Catalan Institution for Advanced Studies (ICREA), Universitat Barcelona, 08010 Barcelona, Spain Correspondence: Won Hee Lee - katharina.glomb@upf.edu BMC Neuroscience 2016, 17(Suppl 1):P170 The subject of this study is the temporal dynamics of functional connectivity (FC) in human resting state (RS) as measured with BOLD fMRI. In spite of rising interest in the topic [1], it remains unclear whether observed FC is stationary or if state switching is present, nor is it clear what constitutes these putative states. Modelling is an invaluable tool for answering these questions: here we combine a dynamic mean field model of the cortex with data analysis in order to determine whether and to what extent spatio-temporal FC patterns found in empirical data can be mimicked by a stationary model as described in [2]. To this end, we cast our data into tensor form by computing time-dependent FC inside of sliding windows (dynamic FC, dFC), comparing three methods to compute dFC (two correlation based, and mutual information). We employ canonical polyadic decomposition (also known as parallel factor analysis) with or without non-negativity constraint to decompose the tensors, which allows us to simultaneously consider the temporal and spatial dimensions [3]. First, we decompose such tensors obtained from empirical data of 24 subjects [4] and cluster resulting spatial features (i.e., communities) in order to obtain a small number of templates. These templates are used in a second step to compare to simulated data that is processed in the same way. We find that even on a very low level of spatial resolution (66 cortical regions), and using only the 2 % biggest dFC values in terms of region pairs and time windows, we succeed in extracting communities that generalize across subjects and can be found in the simulated data. Furthermore, we show that using model-based effective connectivity to inform the model [5] leads to more realistic and stable communities than diffusion weighted MRI-based structural connectivity alone. The method shown here is widely applicable to compare patient groups, data obtained from different tasks as well as mental states, and opens the door to understanding the differences between the temporal dynamics of these conditions. Acknowledgements: KG is funded by the Marie Curie Initial Training Network INDIREA, grant agreement no ITN-2013-606901. APA was supported by SEMAINE ERA-Net NEURON Project. MG acknowledges funding from FP7 FET ICT Flagship Human Brain Project (604102). GD was supported by the ERC Advanced Human Brain Project (n. 604102) and the Plan Estatal de Fomento de la investigación Científica y Técnica de Excelencia (PSI2013-42091-P). ReferencesHutchison RM, Womelsdorf T, Allen EA, Bandettini PA, Calhoun VD, Corbetta M, Della Penna S, Duyn JH, Glover GH, Gonzalez-Castillo J, Handwerker DA, Keilholz S, Kiviniemi V, Leopold DA, de Pasquale F, Sporns O, Walter M, Chang C. Dynamic functional connectivity: promise, issues, and interpretations. NeuroImage. 2013;80:360–78. Deco G, Ponce-Alvarez A, Hagmann P, Romani G, Mantini D, Corbetta M. How local excitation–inhibition ratio impacts the whole brain dynamics. J Neurosci. 2014;34(23):7886–98. Cichocki A. Tensor decompositions: a new concept in brain data analysis? arxiv Prepr. 2013, arXiv1305.0395, 507–17. Schirner M, Rothmeier S, Jirsa VK, McIntosh AR, Ritter P. An automated pipeline for constructing personalized virtual brains from multimodal neuroimaging data. NeuroImage. 2015;117:343–57. Gilson M, Moreno-Bote R, Ponce-Alvarez A, Ritter P, Deco G. Estimation of directed effective connectivity from fMRI functional connectivity hints at asymmetries in cortical connectome. PLoS Comput Biol. 2016. P171 Getting in the groove: testing a new model-based method for comparing task-evoked versus resting-state activity in fMRI data on music listening Matthieu Gilson1,†, Maria A. G. Witek2,†, Eric F. Clarke3, Mads Hansen4, Mikkel Wallentin5, Gustavo Deco1, Morten L Kringelbach2,5,6, Peter Vuust2,5 1Center for Brain Cognition, Universitat Pompeu Fabra, Barcelona, Spain; 2Center for Music in the Brain, Aarhus University & Royal Academy of Music, Aarhus/Aalborg, Denmark; 3Faculty of Music, University of Oxford, UK; 4Department of Psychology and Behavioural Sciences, Aarhus University, Denmark; 5Center of Functionally Integrative Neuroscience, Aarhus University, Denmark; 6Department of Psychiatry, University of Oxford, UK Correspondence: Matthieu Gilson - matthieu.gilson@upf.edu BMC Neuroscience 2016, 17(Suppl 1):P171 † Equal contribution. Much of present neuroimaging studies have used fMRI to simply measure the activity in brain regions and computing the functional connectivity (FC) between regions and behaviour. This has provided important insights into the task-evoked activity compared to rest and thus the flow of information between functional regions (e.g., sensory, multimodal integration, memory). Yet, this does not capture all of the complex spatiotemporal patterns of brain activity and in particular not been the underlying effective connectivity. The present study provides evidence for application of a novel method of determining the functional roadmap of effective connectivity (EC), which measures the strengths of dynamic cortical interactions. We use a whole-brain dynamical model that combines fMRI data with anatomical information obtained using diffusion-tensor imaging (DTI) [1]. Our recently developed method [2] provides estimates for the EC as well as the local excitability and stimulus load in a study of groove-based music. The brain is divided into 90 areas and the input noise is shaped by the EC to generate the FC. This model allows us to explore the role of the network parameters in shaping FC: after constraining the model to reproduce resting-state activity, we examine the effect of an arbitrary change in individual inputs and EC strengths on FC. Our method focuses on spatio-temporal FC, meaning covariances of BOLD signals with possible time shifts. The estimated EC and inputs are taken as fingerprints of the brain dynamics. We analyze fMRI data of participants listening to 15 rhythms with three levels of syncopation: Low, Medium and High (five drum-breaks in each level) [3]. In accordance with other studies, we find behaviourally that the Medium level—with more “groove”—elicits the most pleasure and wanting to move. We tune the model to reproduce the FC recorded for each syncopation level, as well as rest. We analyze significant changes for each groove condition as compared to rest. FC for Medium syncopation exhibits a faster shuffling between successive brain patterns of activity, linked to more metastability and corresponding maybe to the subjective experience of higher pleasure. In addition, our model gives a detailed functional neuroanatomy of dynamical changes in the brain networks. Interestingly, Medium syncopation induces changes in excitability in the basal ganglia, such as the pallidum and the caudate nucleus, which may be related to the increased desire for moving. Interestingly, significant changes are also observed in regions of the orbitofrontal and anterior cingulate cortices, which have been strongly implicated in the pleasure network [4]. Overall, our new method has for the first time allowed us to uncover the network and the corresponding effective connectivity of a highly pleasurable state of groove; possibly even revealing the brain topography of eudaimonia, the sense of well-being. Acknowledgements: MG and GD were supported by the FP7 FET ICT Flagship Human Brain Project (604102). GD was supported by the ERC Advanced Grant: DYSTRUCTURE (n. 295129), by the Spanish Research Project SAF2010-16085. MLK was supported by the ERC Consolidator Grant: CAREGIVING (n. 615539). MLK, MW and PV were supported by the Center for Music in the Brain, funded by the Danish National Research Foundation (DNRF117). ReferencesDeco G, Jirsa V, McIntosh A. Emerging concepts for the dynamical organization of resting-state activity in the brain. Nat Rev Neurosci. 2011;12:43–56. Gilson M, Moreno-Bote R, Ponce-Alvarez A, Ritter P, Deco G. Estimation of directed effective connectivity from fmri functional connectivity hints at asymmetries in cortical connectome. PLoS Comput Biol. 2016 (in press). Witek MAG, Clarke E, Wallentin M, Kringelbach ML, Vuust P. Syncopation, body-movement and pleasure in Groove music. PLoS One. 2014;9:e94446. Berridge KC, Kringelbach ML. Pleasure systems in the brain. Neuron. 2015;86:646–64. P172 Stochastic engine for pathway simulation (STEPS) on massively parallel processors Guido Klingbeil1, Erik De Schutter1 1Computational Neuroscience Unit, Okinawa Institute of Science and Technology, 1919-1 Tancha, Onna-son, Kunigami-gun, Okinawa 904-0495, Japan Correspondence: Guido Klingbeil - guido.klingbeil@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P172 STEPS is a stochastic reaction–diffusion simulator. Its emphasis is on accurately simulating signaling pathways in realistic morphologies [1]. It is becoming apparent that larger computational models are demanded to either capture more such morphologies or to simulate more complex systems. As an example, the dendrite calcium burst model presented by Anwar et al. [2] requires approximately 285,000 sub-volumes with 15 diffusing molecular species and 20 reactions per sub-volume. It required several weeks to compute a simulation of 500 ms. Thus it is desirable to reduce the computational burden. Accelerators such as graphics processing units (GPU) offer unprecedented computing performance and are now common amongst the fastest super computers [3]. This project enables STEPS to benefit from the computational power of GPUs. GPUs are massively parallel co-processors aggregating thousands of simplified processing cores onto a single chip. They share many characteristics with vector computers and a key challenge is that the processing cores are not independent. Similar to vector computers an operation is applied to a group of data elements rather than to the individual data element. Furthermore, the programmer has to mitigate the memory hierarchy of GPUs. While memory with a high access latency is, in general, abundant, fast memory space shared between threads is small which may limit the size of the reaction system one is able to simulate. Previous research has shown that we can exploit the computational power of GPUs to accelerate spatially homogenous stochastic simulations by two orders of magnitude while avoiding the limitation imposed to the size of the reaction system to be simulated by the small fast memory space [4]. STEPS implements a spatial version of Gillespie’s stochastic simulation algorithm (SSA) computing reaction–diffusion systems on a mesh of tetrahedral sub-volumes [1, 5]. Currently a parallelised multi-processor version of STEPS is under development. Operator splitting techniques allow to separate the reaction of molecules within a sub-volume from the diffusion of molecules between them. This prevents computationally costly rollbacks in case of molecules diffusing between sub-volumes handled by different processors. We develop a layered hybrid software architecture using both, the classic central processing unit as well as GPUs, integrated into STEPS applying GPU acceleration at the sub-volume level and integrating them into a coherent spatial simulation using operator splitting. Our architecture will be a plug-in solution to STEPS not requiring any changes to the interfaces towards the user or other software systems of STEPS itself. ReferencesHepburn I, et al.: STEPS: efficient simulation of stochastic reaction–diffusion models in realistic morphologies. BMC Syst Biol. 2012;6:36. Anwar H, et al.: Stochastic calcium mechanisms cause dendritic calcium spike variability. J Neurosci. 2013;33(40):15848–67. TOP500 Supercomputer Site [http://www.top500.org]. Klingbeil, et al. Stochastic simulation of chemical reactions with cooperating threads on GPUs (in preparation). Gillespie DT. Exact stochastic simulation of coupled chemical reactions. J Phys Chem. 1977;81(25):2340–61. P173 Toolkit support for complex parallel spatial stochastic reaction–diffusion simulation in STEPS Weiliang Chen1, Erik De Schutter1 Computational Neuroscience Unit, Okinawa Institute of Science and Technology, Okinawa 904-0411, Japan Correspondence: Weiliang Chen - w.chen@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P173 The studies of large neuronal pathway models with complex morphologies, such as our previous work on the stochastic effects of calcium dynamics in Purkinje cells [1], present a great challenge to currently available spatial stochastic reaction–diffusion simulators, for example, STEPS [2], as the model scales and complexities quickly surpass the capability any serial simulator can achieve. One possible solution for this challenge is parallelization. At CNS2015 we reported a parallel implementation of STEPS which demonstrated great speedup when simulating a reduced calcium burst model with a tetrahedral cylinder [3], but it is clear that to explore the full potential of our implementation, a larger scale simulation with more complex geometry is required. Here we extend our work by simulating the reduced calcium burst model with a reconstructed Purkinje dendrite tree branch mesh. Comparing to previous simulations with regular cylinder meshes, the simulation with dendrite tree mesh requires several new support routines from the simulator. First of all, the simple axis based partitioning approach used in the cylinder simulations is no longer a good partitioning solution due to the complex tree structure of the mesh. A sophisticated mesh partitioning and validation solution is therefore necessary for the new simulation. Second, the new simulation demands good regional annotation and data collection support as calcium concentration, the main focus of the simulation, varies both spatially and temporally. The parallel environment further increases the difficulty of such support as the simulation is distributed over a massive number of processors, and each annotated region may not be completely simulated within a single processor. Furthermore, it is necessary to minimize the user interface difference between serial and parallel STEPS solvers and extend the STEPS visualization toolkit to facilitate comparison with results from previous work. In this poster we demonstrate the general procedure of converting a serial STEPS simulation to its parallel counterpart, using the reduced calcium burst model with complex tree mesh as example, and showcase new supporting toolkits developed for the procedure. We believe that the presented procedure and toolkits will be helpful to STEPS users in their future research. ReferencesAnwar H, Hepburn I, Nedelescu H, Chen W, De Schutter E. Stochastic calcium mechanisms cause dendritic calcium spike variability. J Neurosci. 2013;33(40):15488–867. Hepburn I, Chen W, Wils S, De Schutter E. STEPS: efficient simulation of stochastic reaction–diffusion models in realistic geometries. BMC Syst Biol. 2012;6:36. Chen W, Hepburn I, De Schutter, E. Implementation of parallel spatial stochastic reaction–diffusion simulation in STEPS. BMC Neurosci. 2015;16(Suppl 1):P54. P174 Modeling the generation and propagation of Purkinje cell dendritic spikes caused by parallel fiber synaptic input Yunliang Zang1, Erik De Schutter1 Computational Neuroscience Unit, Okinawa Institute of Science and Technology Graduate University, Onna-son, Okinawa, Japan Correspondence: Yunliang Zang - yunliang.zang@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P174 The dendrite of Purkinje cell (PC) has been shown to express different types of voltage gated ion channels. After strong parallel fiber (PF) stimulus, calcium currents can cause dendritic spikes to occur in the spiny dendrite [1]. Different with climbing fiber caused calcium signals that propagate throughout the dendritic tree, PF caused dendritic spikes are local. The elevated calcium concentration due to the local dendritic spike may trigger local synaptic plasticity, possibly playing a significant role in information processing. However, until now, how these dendritic spikes originate and propagate is not well understood. In this work, we built a new PC dendrite model, which can generate local dendritic calcium spikes. The generated spike by model shows similar properties to experimental observations [1], including spike threshold, amplitude and latency. We identify the role of P type Ca2+ current, A type K+ current, high threshold K+ current (Kv3), calcium activated K+ current and axial current on the depolarization and repolarization of the spike. In the model, the required threshold synaptic input to trigger local dendritic spikes decreases with distance from soma, which facilitates the occurrence of spikes in the spiny dendrite by PF synaptic input. This model can also successfully replicate the failure of propagation of PF caused dendritic spikes at the parent branch point. By analyzing the spatial spread of the dendritic spikes and EPSP signal to soma, we identify the relative contribution of active currents and impedance mismatch on the signal decay. Because dendritic spikes can robustly propagate over an entire branchlet in the direction away from the soma, dendritic branchlets may be the basic organization unit for integrating synaptic input [2, 3]. ReferencesRancz EA, Hausser M: Dendritic calcium spikes are tunable triggers of cannabinoid release and short-term synaptic plasticity in cerebellar Purkinje neurons. J Neurosci. 2006;26(20):5428–37. De Schutter E, Bower JM. Simulated responses of cerebellar Purkinje cells are independent of the dendritic location of granule cell synaptic inputs. PNAS. 1994;91(11):4736–40. Branco T, Hausser M. The single dendritic branch as a fundamental functional unit in the nervous system. Curr Opin Neurobiol. 2010;20(4):494–502. P175 Dendritic morphology determines how dendrites are organized into functional subunits Sungho Hong1, Akira Takashima1, Erik De Schutter1 1Computational Neuroscience Unit, Okinawa Institute of Science and Technology Graduate University, Onna-son, Okinawa 904-0495, Japan Correspondence: Sungho Hong - shhong@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P175 Studies have established that dendrites are not simple cables that deliver synaptic inputs to a spike initiation zone in a neuron but can also perform active transformation, which is termed “dendritic computation” [1]. In particular, it has been claimed that individual dendritic branch should function as a local computational subunit [2] and therefore single neurons (especially pyramidal neurons) can act like two-layer neural networks [3]. Evidence supporting these hypotheses is largely based on existence of active membrane mechanisms in dendrites that give rise to their rich computational capabilities (e.g. [4]) and independent operations [5]. However, dendritic morphology is also known to play a significant role: For example, spike backpropagation is effectively prevented in the cerebellar Purkinje cells mostly due to morphology, even when artificial active mechanisms supporting propagation are embedded in simulations [6]. Nevertheless, to our best knowledge, there has been only few studies that quantified how the real morphological structure can control the functional properties of dendrites by forming subunits. Here we address this question by combining a data-driven statistical analysis and computational modeling approach: First, we simulated central neurons of diverse morphological types with the passive membranes where localized inputs were injected. Response patterns in the dendritic membrane were collected as “features” corresponding to the input sites. Then, our dimensionality reduction/clustering procedure grouped them into clusters, which we call “subunits”. We found that those subunits usually consist of a few nearby branches in many neuron types, containing 2.12 ± 0.13 dendritic terminals per subunit (mean ± SEM), whereas they consist of one or more branchlets in the cerebellar Purkinje cells (12.9 ± 0.82 terminals). We also found that the subunits are comparable with other functional properties such as sublinear summation of multiple synaptic inputs and spreading of a dendritic spike. Conclusions The morphological branching pattern of a neuronal dendritic tree determines how dendrites are organized into functional subunits. This implies that principles governing synaptic integration and active events, such as dendritic spiking, can widely vary depending on the morphological type of the neuron. ReferencesLondon M, Häusser M. Dendritic computation. Annu Rev Neurosci. 2005;28:503–32. Branco T, Häusser M. The single dendritic branch as a fundamental functional unit in the nervous system. Curr Opin Neurobiol. 2010;20:494–502. Poirazi P, Brannon T, Mel BW. Pyramidal neuron as two-layer neural network. Neuron. 2003;37:989–99. Branco T, Clark BA, Häusser M. Dendritic discrimination of temporal input sequences in cortical neurons. Science. 2010;329:1671–5. Behabadi BF, Mel BW. Mechanisms underlying subunit independence in pyramidal neuron dendrites. Proc Natl Acad Sci USA. 2014;111:498–503. Vetter P, Roth A, Häusser M. Propagation of action potentials in dendrites depends on dendritic morphology. J Neurophysiol. 2001;85:926–37. P176 A model of Ca2+/calmodulin-dependent protein kinase II activity in long term depression at Purkinje cells Criseida Zamora1, Andrew R. Gallimore1, Erik De Schutter1 Computational Neuroscience Unit, Okinawa Institute of Science and Technology Graduate University, Okinawa 904-0895, Japan Correspondence: Won Hee Lee - criseida.chimal@oist.jp BMC Neuroscience 2016, 17(Suppl 1):P176 Cerebellar long-term depression (LTD) is a form of synaptic plasticity involved in motor learning. It is characterized as a robust and persistent decrease in the synaptic transmission between parallel fibers (PF) and Purkinje cells (PC), which is expressed as a reduction in the number of synaptic AMPA receptors (AMPAR). LTD signaling network includes a PKC-ERK-cPLA2 positive feedback loop and mechanism of AMPAR trafficking. Previous studies suggest that Ca2+/calmodulin-dependent protein kinase II (CaMKII) is required for the LTD induction [1]. However, the molecular mechanism of how CaMKII contributes to LTD is not fully understood. Noise in the signaling networks plays an important role in cellular processes. LTD models including the CaMKII pathway have been developed [2], but they have not included the intrinsic stochasticity of molecular interactions. Our lab recently developed a stochastic model of the LTD signaling network including a PKC-ERK-cPLA2 feedback loop and AMPAR trafficking [3]. In this work, we have extended the model by adding the molecular network regulating CaMKII activity, which is known to influence LTD. This new model was solved stochastically by STEPS (STochastic Engine for Pathway Simulation) to simulate the influence of noise in the LTD signaling network [4]. Some of the most important new components of this network include phosphatase 2A (PP2A), phosphodiesterase 1 (PDE1), cGMP/protein kinase G (PKG) and nitric oxide (NO) pathway. Through stochastic modeling we showed that the requirement of CaMKII activity for LTD induction is controlled by its indirect inhibition of PP2A activity, with PP2A markedly suppressing the activation of LTD when CaMKII activity is decreased. The impairment of LTD could be rescued by the additional PDE1 reduction when CaMKII is reduced. In addition, the cGMP/PKG pathway supports LTD through its activation by NO. These results are congruent with previous studies of CaMKII activity [2] and make our stochastic model a potential tool to study the effects of CaMKII, phosphatases and phosphodiesterases in LTD molecular network. ReferencesHansel C, de Jeu M, Belmeguenai A, Houtman SH, Buitendijk GH, Andreev D, De Zeeuw CI, Elgersma Y. αCaMKII is essential for cerebellar LTD and motor learning. Neuron. 2006;51:835–43. Kawaguchi SY, Hirano T. Gating of long-term depression by Ca2+/calmodulin-dependent protein kinase II through enhanced cGMP signalling in cerebellar Purkinje cells. J Physiol. 2013;591(7):1707–30. Antunes G, De Schutter E. A stochastic signaling network mediates the probabilistic induction of cerebellar long-term depression. J Neurosci. 2012;32(27):9288–300. Hepburn I, Chen W, Wils S, De Schutter E. STEPS: efficient simulation of stochastic reaction–diffusion models in realistic morphologies. BMC Syst Biol. 2012;6:36. P177 Reward-modulated learning of population-encoded vectors for insect-like navigation in embodied agents Dennis Goldschmidt1, Poramate Manoonpong2, Sakyasingha Dasgupta3,4 1Champalimaud Neuroscience Programme, Champalimaud Center for the Unknown, Lisbon, Portugal; 2Center of Biorobotics, Mærsk Mc-Kinney Møller Institute, University of Southern Denmark, Odense, Denmark; 3Riken Brain Science Institute, 2-1 Hirosawa, Wako, Saitama, Japan; 4IBM, IBM Research - Tokyo, Tokyo, 103-8510, Japan Correspondence: Dennis Goldschmidt - dennis.goldschmidt@neuro.fchampalimaud.org BMC Neuroscience 2016, 17(Suppl 1):P177 Many insects exhibit robust and efficient visual-based navigation in complex environments [1]. Specifically, behavioral studies on ants and bees showed that they are guided by orientation vectors based on a process called path integration. This process allows them to estimate their current location by integrating cues from odometry and a sun-based compass. While it is mainly applied to return back to the nest, it also guides learning of so-called vector memories for subsequent foraging [2, 3]. Vector memories can be anchored globally to the nest or locally to landmarks. Recent neurophysiological studies revealed that the central complex, an insect neuropil, contains neural representations of compass [4] and odometric cues [5]. However, it is still unclear, how these representations are involved in path integration and vector memories, and how they produce goal-directed navigation. Computational modeling has been powerful in testing hypotheses about the underlying neural substrates and their generated behavior, and to predict further experimental data. Previous models [6, 7] sufficiently produced insect-like vector navigation, but they neglected biologically plausible explanations about underlying neural mechanisms that could generate this behavior. We present here a novel computational model of neural mechanisms in closed-loop control for vector navigation in embodied agents. It consists of a path integration mechanism, reward-modulated learning of global and local vectors, random search, and action selection. The path integration mechanism computes a vectorial representation of the agent’s current location. The vector is encoded in the activity pattern of circular arrays, where the angle is population-coded and the distance is rate-coded. We apply a reward-modulated learning rule for global and local vector memories, which associates the local food reward with the path integration state. A motor output is computed based on the combination of vector memories and random exploration. We show that the modeled neural mechanisms enable robust homing and localization in a simulated agent, even in the presence of external sensory noise. The proposed learning rules produce goal-directed navigation and route formation under realistic conditions. This provides an explanation for, how view-based navigational strategies are guided by path integration. As such, the model is the first to link behavioral observations to their possible underlying neural substrates in insect vector navigation. Acknowledgements: We thank Florentin Wörgötter at the Department of Computational Neuroscience in Göttingen, where most of this work was conducted. SD acknowledges funding from the RIKEN Brain Science Institute. ReferencesWehner R. Desert ant navigation: how miniature brains solve complex tasks. J Comp Physiol A. 2003;189(8):579–88. Collett M, Collett TS, Bisch S, Wehner R. Local and global vectors in desert ant navigation. Nature. 1998;394(6690):269–72. Collett TS, Collett M. Route-segment odometry and its interactions with global path-integration. J Comp Physiol A. 2015;201(6):617–30. Seelig JD, Jayaraman V. Neural dynamics for landmark orientation and angular path integration. Nature. 2015;521(7551):186–91. Martin JP, Guo P, Mu L, Harley CM, Ritzmann RE. Central-complex control of movement in the freely walking cockroach. Curr Biol. 2015;25(21):2795–803. Cruse H, Wehner R. No need for a cognitive map: decentralized memory for insect navigation. PLoS Comput Biol. 2011;7(3):e1002009. Kubie JL, Fenton AA. Heading‐vector navigation based on head‐direction cells and path integration. Hippocampus. 2009;19(5):456–79. P178 Data-driven neural models part II: connectivity patterns of human seizures Philippa J. Karoly1, Dean R. Freestone1,2, Daniel Soundry2, Levin Kuhlmann3, Liam Paninski2, Mark Cook1 1Department of Medicine, The University of Melbourne, Parkville VIC, 3010, Australia; 2Department of Statistics, Columbia University, New York, NY, USA; 3Swinburne University of Technology, Hawthorn, VIC 3122, Australia Correspondence: Philippa J. Karoly - pkaroly@student.unimelb.edu.au BMC Neuroscience 2016, 17(Suppl 1):P178 Here we present a model-based estimation framework for electrocorticography (ECoG) data that provides insight into mechanisms of seizures; and can be used as a clinical tool to monitor and design new treatment strategies on a patient-specific basis. Seizures are brief periods of abnormal, hypersynchronous neural firing that spreads across multiple cortical regions. People with epilepsy experience recurrent seizures, which are often untreatable and of unknown cause. The data-driven estimation framework, shown in Fig. 91, describes dynamic neural connectivity patterns during patient seizures. Data were obtained from a clinical trial for an implantable seizure warning device [1], which captured thousands of seizures. We estimated mean membrane potentials and connectivity strengths between excitatory, inhibitory and pyramidal populations using a non-linear, assumed density filter for the neural mass equations [2].Fig. 91 Example estimation of a seizure recording. A Sixteen channel electrocortiography (ECoG) of seizure (red lines indicate start and end points). B The ECoG channels are modelled as cortical regions, each with three coupled populations. C–G Estimation results of coupling strength (proportional to color) between neural populations for 16 cortical regions (vertical axis), over the time span of the seizure (horizontal axis) Estimated parameters provide insights into the mechanisms of the seizure, which are not apparent from ECoG alone. For instance, Panel E shows the seizure is preceded by a focal decrease in inhibition compared to the surrounding channels, with widespread disinhibition during the seizure. Joint state and parameter estimation was repeated for every seizure, and the results showed consistent, stereotypical effective connectivity patterns that differed between short (<20 s) and long seizures. This is an important finding, as understanding the regulatory factors implicated in stopping seizures can guide new pharmaceutical treatments and electrical counter-stimulation strategies. The successful application of the neural mass model to study epileptic seizures supports the use of data-driven estimation for the clinical management of epilepsy. ReferencesCook MJ, et al. Prediction of seizure likelihood with a long-term, implanted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study. Lancet Neurol. 2013;12(6):563–71. Freestone et al.: Data-driven neural models part I CNS 2016 abstract submission. P179 Data-driven neural models part I: state and parameter estimation Dean R. Freestone1, Philippa J. Karoly1,2, Daniel Soundry3, Levin Kuhlmann4, Mark Cook1 1Department of Medicine, The University of Melbourne, Parkville VIC, 3010, Australia; 2Department of Electrical and Electronic Engineering, The University of Melbourne, Parkville VIC, 3010, Australia; 3Department of Statistics, Columbia University, New York, NY, USA; 4Swinburne University of Technology, Hawthorn, VIC 3122, Australia Correspondence: Dean R. Freestone - deanrf@unimelb.edu.au BMC Neuroscience 2016, 17(Suppl 1):P179 This work describes a novel algorithm for inferring neural activity and effective cortical connectivity from neuroimaging data. The ability to infer cortical network structure from data is an important step towards understanding and treating neurological disorders, such as epilepsy. However, statistical measures for correlation in neuroimaging data are ambiguous and bear little or no relation to physiology. On the other hand, estimating physiologically realistic connectivity is highly challenging due to the complex, non-linear dynamics of the brain. The algorithm we present overcomes this challenge by providing an exact solution to non-linear inversion for a class of biologically inspired neural network models. The presented algorithm performs joint state and parameter estimation for a class of neural model that represents interacting cortical regions as coupled nodes (shown in Fig. 92). The states of the model represent mean cortical activity [population membrane potentials, v(t)], and the parameters are the effective connectivity (synaptic gain kernels, αi,e). The output voltage, vn(t), represents the electrophysiology recording, which is inverted using a novel formulation of the Kalman filter equations for neural models [1]. The novelty of this method is the derivation of an exact solution to the integral over the distribution of hidden model states conditioned on previous data.Fig. 92 Data-driven model estimation. A The basic unit of a neural model is described by the mean membrane potential, vn(t), of a neural ensemble and synaptic inputs. B, C Pre-synaptic firing rates are convolved with the excitatory/inhibitory kernel to generate membrane potential fluctuations. D The resulting membrane potential is converted to an output firing rate via a sigmoidal transform. E Electrical recording of a seizure. F Estimated gain parameters during seizure We provide results showing that the new algorithm demonstrates higher estimation accuracy and greater computational tractability than existing inference methods for neural models. We also show example estimation results from an electrical recording of a human seizure (shown in Fig. 92). This new method for data-driven inference represents an important contribution to online diagnostic applications, in particular for the treatment of epilepsy [2]. ReferencesFreestone DR, et al. Estimation of effective connectivity via data-driven neural modeling. Front Neurosci. 2014;8:383 Karoly, et al. Data-driven neural models part II: connectivity patterns of human seizures. CNS 2016 abstract. P180 Spectral and spatial information processing in human auditory streaming Jaejin Lee1, Yonatan I. Fishman2, Yale E. Cohen1 1Department of Otorhinolaryngology – Head and Neck Surgery, University of Pennsylvania, Philadelphia, PA 19104, USA; 2Department of Neurology, Albert Einstein College of Medicine, Bronx, NY 10461, USA Correspondence: Jaejin Lee - jaejin@mail.med.upenn.edu BMC Neuroscience 2016, 17(Suppl 1):P180 The purpose of auditory system is to transform acoustic stimuli from the external environment to sound perception. To achieve this goal, the auditory system needs to analyze a mixture of stimuli that originate from independent sources and distinguish individual sound sources in the auditory scene. It is believed that the auditory system groups and segregates auditory stimuli based on their regularities, but the neural basis of how regularities relate to sound perception is not well known. The ventral pathway in the brain is involved in auditory perception whereas the dorsal pathway is involved in spatial processing and audiometer processing. We are interested in how the spatial information is represented in the ventral pathway during perceptual auditory streaming tasks that use spatial information. We first developed a novel task based on [1] in which human listeners can segregate streams using spectral or spatial information and detect the deviant tone. An array of 13 free field speakers with different spatial distributions were used to play the stimuli. The frequency difference between streams and the spatial separation were varied to explore how the spectral and spatial information interplay in the auditory streaming task. We also manipulated other acoustic features of the stimuli to understand how different acoustic cues can affect the auditory streaming performance. We found that the ability to segregate the streams is vastly improved when there is spatial information available in addition to spectral information. Also, we further analyzed the behavioral data to get psychophysical kernels and fit the data to variants of sequential sampling models related to the drift diffusion model(DDM) [2] to quantify the effects of sequence coherence on the decision making process. ReferencesSussman E, Steinschneider M. Attention effects on auditory scene analysis in children. Neuropsychologia. 2009;47:771–85. Liu A, Tsunada J, Gold J, Cohen Y. Temporal integration of auditory information is invariant to temporal grouping cues. eNeuro. 2015;2(2):e0077-14. P181 A tuning curve for the global effects of local perturbations in neural activity: mapping the systems-level susceptibility of the brain Leonardo L. Gollo1, James A. Roberts1, Luca Cocchi1 1Systems Neuroscience Group, QIMR Berghofer Medical Research Institute, Herston, QLD 4006, Australia Correspondence: Leonardo L. Gollo - leonardo.l.gollo@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P182 The activity of the human brain in a state of rest exhibits a defined pattern of functional connectivity, and a small set of functional networks, which comprise regions that are highly correlated and are mostly distinctive from one another [1]. However, despite recent efforts [2, 3], the effects of local perturbations into endogenous whole brain dynamics are not yet clearly understood [4]. To gain insights into the global effects of a focal perturbation, we simulate the human brain dynamics using a weighted high-resolution connectome of 513 cortical regions [5]. The cortical dynamics is modelled by a canonical oscillatory model, introducing heterogeneous dynamics between cortical regions as a function of the anatomical nodal strength (sum of weights). Such heterogeneity leads to a hierarchy of time scales of cortical regions recapitulating the known anatomical hierarchy, with peripheral regions having fast time scales and core regions with slow time scales [6]. Results showed that nodal diversity is not just a crucial element to improve the model’s performance [7, 8], but also to reproduce the experimental data of variations in functional connectivity following local inhibitory transcranial magnetic stimulation (TMS). We find a large variation in the overall effect of functional connectivity following local stimulation. Specifically, the inhibition of hub nodes causes increased anti-correlated activity, whereas inhibition of peripheral nodes caused increased correlated activity with the rest of the brain. The intensities of the variations in functional connectivity with respect to baseline were also highly variable and stronger for intermediary nodes that were not hubs or peripheral regions. Moreover, depending on the weights of the cortical regions, changes in functional connectivity form a tuning curve (Fig. 93). Overall, our findings suggest a key role of local temporal dynamics to explain the widespread effects of focal perturbations in neural activity.Fig. 93 Changes in functional connectivity with respect to baseline after inhibitory stimulation as a function of cortical weight of the structural connectivity matrix. Red line: mean uniform bins curve smoothed; dashed line: mean weight ReferencesPower JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, Church JA, Vogel AC, Laumann TO, Miezin FM, Schlaggar BL, Petersen SE. Functional network organization of the human brain. Neuron. 2011;72:665–78. Cocchi L, Sale MV, Lord A, Zalesky A, Breakspear M, Mattingley JB. Dissociable effects of local inhibitory and excitatory theta-burst stimulation on large-scale brain dynamics. J Neurophysiol. 2015;113:3375–85. Kunze, T, Hunold A, Haueisen J, Jirsa V, Spiegler A. Transcranial direct current stimulation changes resting state functional connectivity: a large-scale brain network modeling study. NeuroImage. doi:10.1016/j.neuroimage.2016.02.015. Sale MV, Mattingley JB, Zalesky A, Cocchi L. Imaging human brain networks to improve the clinical efficacy of non-invasive brain stimulation. Neurosci Biobehav Rev. 2015;57:187–98. Roberts JA, Perry A, Lord AR, Roberts G, Mitchell PB, Smith RE, Calamante F, Breakspear M. The contribution of geometry to the human connectome. NeuroImage. 2016;124:379–93. Gollo LL, Zalesky A, Hutchison RM, van den Heuvel M, Breakspear M. Dwelling quietly in the rich club: Brain network determinants of slow cortical fluctuations. Philos Trans R Soc B Biol Sci. 2015;370:20140165. Mejias JF, Longtin A. Optimal heterogeneity for coding in spiking neural networks. Phys Rev Lett. 2012;108:228102. Gollo LL, Copelli M, Roberts JA. Diversity improves performance in excitable networks. arXiv preprint arXiv:1507.05249. 2015 Jul 19. P182 Diverse homeostatic responses to visual deprivation mediated by neural ensembles Yann Sweeney1, Claudia Clopath1 1Department of Bioengineering, Imperial College London, UK Correspondence: Yann Sweeney - y.sweeney@imperial.ac.uk BMC Neuroscience 2016, 17(Suppl 1):P182 Visual deprivation paradigms provide crucial insight into the homeostatic response in visual cortex. We explore how neurons within functional ensembles may exhibit correlated homeostatic responses to visual deprivation, and how the source of common inputs to these ensembles determine the extent of their homeostatic recovery. We hypothesise that common inputs from non-visual stimuli are responsible for driving recovery from visual deprivation. We simulate development during spontaneous and evoked activity in a recurrent network model of visual cortex in which Hebbian and homeostatic synaptic plasticity is implemented. This leads to the emergence of highly interconnected ensembles of neurons driven by either common visual or common non-visual inputs. When we then deprive the developed network of visual input, the homeostatic response is a strengthening of activity within ensembles which share common non-visual inputs. A broad reduction in inhibition across the network is also observed. Interestingly, the magnitude of the homeostatic response depends on the size of these ensembles, with larger ensembles more likely to fully recover from visual deprivation. Our results demonstrate the importance of investigating functional plasticity of ensembles triggered by sensory deprivation paradigms. Acknowledgements: This research was supported by the Engineering and Physical Sciences Research Council (EPSRC), the Leverhulme Trust and Google Faculty Award. P183 Opto-EEG: a novel method for investigating functional connectome in mouse brain based on optogenetics and high density electroencephalography Soohyun Lee1,3, Woo-Sung Jung1,2, Jee Hyun Choi3 1Department of Physics, POSTECH, Pohang, 37673, South Korea; 2Department of Industrial and Management Engineering, POSTECH, Pohang, 37673, South Korea; 3Center for Neuroscience, KIST, Seoul, 02792, South Korea Correspondence: Jee Hyun Choi - jeechoi@kist.re.kr BMC Neuroscience 2016, 17(Suppl 1):P183 Connectome, comprehensive structural description of the network of elements and connections forming the brain [1], is fundamental for understanding the brain functions. Recent advances in optical imaging techniques allow us to be feasible to structural connectivity. But differently from structural connectivity, the functional connectivity is altered by condition such as brain states, input types and pathological conditions. To construct functional connectome, the techniques to map individual functional circuit and control specific neuronal activity have been needed. However, the current functional brain mapping techniques have limitations to obtain the map of the functionally correlated brain activity in freely moving mouse model. Here, we introduce novel functional brain mapping technique for mouse model by high density electroencephalography [2] under optogenetic stimulus, which we referred as opto-EEG. Opto-EEG tool enables us to investigate the functionally connected neuronal circuit with high spatial and temporal resolution. We stimulated ventral posterioromedial thalamic nucleus (VPM) with various frequencies for verifying different frequency dependency of functional connectome. Stimulation of VPM induced sequential activations of ipsilateral somatosensory cortex (S1) followed by ipsilateral motor cortex (M1), contralateral M1 and contralateral S1. The power based analysis result showed information flow between S1 and M1 was maximized under beta frequency stimulus. On the other hand, latency-based result showed minimized interhemispheric transfer latency under gamma band (Fig. 94). This example indicates that opto-EEG makes it possible to be used to characterize the functional connectivity under temporally precise control of specific neuronal circuits, provide new insights into brain exploration capabilities of functional connectome, and be applied to discover neuromodulation method for treatment of disease or pathologies.Fig. 94 Propagation patterns of optical stimulation at each frequency in thalamocotical circuit. Beta frequency stimulus propagated S1–M1 strongly, but gamma frequency case, contralateral propagation is dominant. Blue bars indicate optical stimulus in left VPM ReferencesSporns O, Tononi G, Kotter R. The human connectome: A structural description of the human brain. PLoS Comput Biol. 2005;1(4):e42. Lee M, Kim D, Shin HS, Sung HG, Choi JH. High-density EEG recordings of the freely moving mice using polyimide-based microelectrode. J Vis Exp. 2011;(47). P184 Biphasic responses of frontal gamma network to repetitive sleep deprivation during REM sleep Bowon Kim1,2, Youngsoo Kim3, Eunjin Hwang1, Jee Hyun Choi1,2 1Center for neuroscience, Korea Institute of Science and Technology, Seoul, South Korea; 2Department of Neuroscience, University of Science and Technology, Daejon, South Korea; 3Department of Psychiatry, VA Boston Healthcare System & Harvard Medical School, Brockton, MA, USA Correspondence: Jee Hyun Choi - jeechoi@kist.re.kr BMC Neuroscience 2016, 17(Suppl 1):P184 Prefrontal cortex has been known to be less activated [1] and decoupled from the other cortical area in REM sleep [2]. In our previous study of chronic sleep deprivation (SD) in mice model, we observed that the 5 successive days of SD (SD 1–5, 18 h sleep deprivation in each day) induced a monotonic increase of the prefrontal gamma oscillation (30–40 Hz) in REM sleep as the sleep pressure increased. However, the functional role of this increased gamma oscillation was not answered. Here, we investigated the functionality of the increased prefrontal gamma in sleep deprived nights by calculating the connectivity between the prefrontal cortex and the other cortical regions [3]. Phase synchrony index (PSI) was employed to minimize the volume conduction in high density EEG microarray. In the first day of the sleep deprivation (SD 1), we observed statistically significant increases in gamma connectivity within the bilateral prefrontal regions and between prefrontal and ipsilateral somatosensory cortex. However, as the sleep deprivation continued, an opposite response of prefrontal-somatosensory gamma connectivity was observed in a way that the PSI between these two areas become insignificant in SD 3 and statistically significantly decreased in the SD 5, which remained even after the 3rd day of recovery after the sleep deprivation (R 3). The area of decreased gamma connectivity became broader as well. On the other hand, the intracortical connectivity within prefrontal connectivity remained elevated throughout the sleep deprivation and recovery days. This result implies that the increased prefrontal gamma oscillation due to the homeostatic response of REM sleep does not participate in the information transfer from prefrontal to the other cortical area (Fig. 95).Fig. 95 The pairs with statistically significantly increased (red) or decreased (blue) PSI of gamma oscillations (Student t test, p < 0.05). Only the pairs from the prefrontal cortex were depicted here. SD and R stand for sleep deprivation and recovery days, respectively Acknowledgements: This research was supported by the Korean National Research Council of Science and Technology (No. CRC-15-04-KIST). ReferencesMaquet P, et al. Functional neuroanatomy of human rapid-eye-movement sleep and dreaming. Nature. 1996;383(6596):163–6. Castro S, et al. Coherent neocortical 40‐Hz oscillations are not present during REM sleep. Eur J Neurosci. 2013;37(8):1330–9. Hwang E, McNally JM, Choi JH. Reduction in cortical gamma synchrony during depolarized state of slow wave activity in mice. Front Syst Neurosci. 2013;7. P185 Brain-state correlate and cortical connectivity for frontal gamma oscillations in top-down fashion assessed by auditory steady-state response Younginha Jung1,2, Eunjin Hwang1, Yoon-Kyu Song2, Jee Hyun Choi1,3 1Center for Neuroscience, Korea Institute of Science and Technology, Seoul 02792, Korea; 2Program in Nano Science and Technology, Seoul National University, Seoul 08826, Korea; 3Department of Neuroscience, University of Science and Technology, Daejon 34113, Korea Correspondence: Jee Hyun Choi - jeechoi@kist.re.kr BMC Neuroscience 2016, 17(Suppl 1):P185 Cortical gamma rhythm, particularly in the frequency range of 30–50 Hz, has received intensive attention as neural correlates of cognitive process [1]. On the other hand, diminished cognitive flexibility, one of the typical symptoms in psychiatric disorders, is closely associated with disturbances in neural oscillations, specifically gamma band [2]. To quantify gamma-band oscillation, auditory steady-state response (ASSR) evoked by repetitive auditory stimulus given at a rate of 40 Hz has been used as a prominent approach which reflects neural efficiency for maintaining gamma oscillation [3]. Despite its diagnostic advantages, there is less discussion whether ASSRs are modulated by endogenous top-down effect. The present research attempts to investigate top-down influences on ASSR by analyzing in vivo mouse data. Experimental data in this study were obtained from 38-channel mouse epidural electroencephalogram during auditory steady-state stimulus. Interestingly, there were two distinctive topographic maps of EEG spectral power and the notable difference between topographies was the presence or absence of frontal responses. By comparing topographic results, we hypothesized that frontal ASSRs reflect top-down functioning. The analytic approaches taken in this work are based on brain-state alteration and regional connectivity. The first research question in the data analysis is that frontal ASSRs switch states of arousal via top-down control. Video-based behavior analysis was adapted to classify arousal states into wakefulness and drowsiness and the proportion of arousal behavior in two topographies were determined. In addition, comparison of delta spectral power for topographic patterns could explain frontal engaged sleep state modulation. The second study question is about early stages of auditory ascending pathway in each topographic pattern. Magnitude and latency of auditory evoked potentials and gamma spectral power were analyzed in inferior colliculus and primary auditory cortex. The third question in data analysis is functional connectivity among cortical regions and, in detail, phase-locking value and directed phase lag index were calculated in frontotemporal and inter-frontal coupling. Overall, the current results show that frontal lobe contributes substantially to ASSR and imply that it is important to consider the frontal involvement in auditory steady-state signal processing. Together, these methodologies could provide important insights to clinical research by demonstrating top-down modulation. Investigating gamma oscillatory activity across cortical regions potentially provides deeper understanding for dysfunction in neurological disorders and furthermore gives clues to determine neural circuit disruption. ReferencesFries P, Reynolds JH, Rorie AE, Desimone R: Modulation of oscillatory neuronal synchronization by selective visual attention. Science. 2001;291(5508):1560–3. Kwon JS, O’Donnell BF, Wallenstein GV, Greene RW, Hirayasu Y, Nestor PG, Hasselmo ME, Potts GF, Shenton ME, McCarley RW. Gamma frequency-range abnormalities to auditory stimulation in schizophrenia. Arch Gen Psychiatry. 1999;56(11):1001–5. Picton TW, John MS, Dimitrijevic A, Purcell D. Human auditory steady-state responses. Int J Audiol. 2003;42(4):177–219. P186 Neural field model of localized orientation selective activation in V1 James Rankin1, Frédéric Chavane2 1Center for Neural Science, New York University, 4 Washington Place, 10003 New York, NY, USA; 2Institut de Neuroscienes de la Timone (INT), CNRS & Aix-Marseille University, 27 Boulevard Jean Moulin, 13005 Marseille, France Correspondence: James Rankin - james.rankin@nyu.edu BMC Neuroscience 2016, 17(Suppl 1):P186 Voltage imaging experiments in primary visual cortex [1] have shown that local, oriented visual stimuli elicit stable orientation-selective activation within the stimulus retinotopic footprint. The cortical activation dynamically extends far beyond the retinotopic footprint, but the peripheral spread stays non-selective—a surprising finding given a number of studies showing the orientation specificity of long-range connections, e.g. [2]. We study the dynamics of these input-driven localized states in a planar neural field model building on an earlier theoretical study using radially symmetric inputs [3]. Here we use a new anatomically-motivated connectivity profile and extend the model to multiple sub-populations encoding orientation. For canonical choices of connectivity profile (such as a radial difference of Gaussians), localized orientation selectivity arises. However, unlike the experimental observations, the selective activation is unstable during transient dynamics. In the new connectivity profile defined in our study, the range of excitatory and inhibitory connections and the orientation selectivity of those connections are controlled with separate parameters. We demonstrate how peaks in the number of excitatory connections at each hyper-column distance [4] are crucial in stabilizing the transient, local orientation selective activation. If these peaks in excitation are non-realistically exaggerated, we demonstrate that spurious selectivity (not matching preference map) could arise in the peripheral spread. Furthermore, although orientation selectivity of connections increases accuracy of the selective activation within the retinotopic footprint, it can also lead to orientation selective activation in the periphery. Our parameter exploration shows that with a balance in the sharpness of peaks in long-range excitatory connections and the selectivity of these connections, we can capture the correct localized selective activation, the non-selective peripheral spread and the stable transient dynamics. Conclusions Typical choices of connectivity profile in planar models of cortex fail to produce important aspects of the observed cortical spread of activation. We developed a more realistic connectivity profile inspired by anatomical data that, used in conjunction with our planar multiple sub-population model, captures all key spatial and temporal aspects of the cortical spread of activation. For the first time, our study shows that the unexpected experimental findings of [1] can be accounted for with a realistic balance between the sharpness of peaks in long-range excitation and orientation selectivity of connections. ReferencesChavane F, Sharon D, Jancke D, Marre O, Frégnac Y, Grinvald A. Lateral spread of orientation selectivity in V1 is controlled by intracortical cooperativity. Front Syst Neurosci. 2011;5:4 Bosking WH, Zhang Y, Schofield B, Fitzpatrick D. Orientation selectivity and the arrangement of horizontal connections in tree shrew striate cortex. J Neurosci. 1997;17:2112–27. Rankin J, Avitabile D, Baladron J, Faye G, Lloyd DJ. Continuation of localized coherent structures in nonlocal neural field equations. SIAM J Sci Comput. 2014;36:B70–93. Buzás P, Eysel U, Adorján P, Kisvárday Z. Axonal topography of cortical basket cells in relation to orientation, direction, and ocular dominance maps. J Comp Neurol. 2001;437:259–85. P187 An oscillatory network model of Head direction and Grid cells using locomotor inputs Karthik Soman1, Vignesh Muralidharan 1, V. Srinivasa Chakravarthy1 1Department of Biotechnology, Indian Institute of Technology Madras, Chennai, Tamil Nadu, India Correspondence: V. Srinivasa Chakravarthy - schakra@iitm.ac.in BMC Neuroscience 2016, 17(Suppl 1):P187 The model (Fig. 96A) takes proprioceptive inputs coming from the joint angles of the two limbs. Locomotor rhythms are modeled as two sinusoidal oscillators whose amplitudes are modulated by the curvature and the speed of simulated animal. These inputs are gated using a leaky integrate and fire (LIF) neuron that spikes at a fixed frequency so that the curvature and speed information from respective limb oscillations are extracted out and given to two oscillatory neural networks separately. The oscillatory neural networks are modeled as Kuramoto networks in which the phase is modulated by the curvature of the path traversed by the simulated animal. Synchrony between the two clusters of oscillators is quantified in terms of phase coherence and phase difference. The synchrony parameters are further used to train a one dimensional self organizing map (SOM), whose neurons display head direction-tuned responses. Each HD response is transformed to a cosine response which is further given to path integration (PI) layer. PI layer is again a network of Kuramoto oscillators whose phase is integrated as a function of HD responses. PCA is done on the PI values. The top few principal components (PC) corresponding to the largest Eigen values are rearranged in increasing order and taken as the weight connections from the PI layer to an outer 1-D layer of neurons. While remapping the neural response of the third and fourth PCs, square grid fields were observed while the fifth and sixth PCs gaveFig. 96 A The model architecture. B Hexagonal firing field of a single neuron in the outer layer of the model while remaping its response on the visual space We present a model of head direction (HD) and grid cells formed purely from idiothetic (locomotor) inputs. Grid cells are a class of spatial cells located in the medial Entorhinal Cortex which is assumed to perform path integration and characterized by its unique hexagonal firing fields [1]. Empirically it is proven that HD cells, another class of spatial cells which encode the heading direction of an animal, form the major input to the grid cell. Existing computational models of grid cells make artificial assumptions like existence of HD cells with a phase differences that are integral multiples of 60° [2]. The aim of the study is to model grid cell firing without imposing these special constraints. Hexagonal grid fields (Fig. 96B). Further analysis showed that PCs were sinusoidal vectors. Investigation of the correlation values between the adjacent rows of the covariance matrix of PI pointed out its similarity to circulant matrices. This was in corroboration with the circulant matrix theorem which states that a circulant matrix of any size gives rise to sinusoidal Eigen vectors. Hence this is a generalized model which provides a theoretical basis for the formation of both hexagonal and non hexagonal grid fields and possibly other spatial cells which are actually projections of the PI values onto sinusoidal orthonormal basis vectors. ReferencesHafting T, Fyhn M, Molden S, Moser M-B, Moser EI. Microstructure of a spatial map in the entorhinal cortex. Nature. 2005;436(7052):801–6. Burgess N, Barry C, O’Keefe J. An oscillatory interference model of grid cell firing. Hippocampus. 2007;17(9):801–12. P188 A computational model of hippocampus inspired by the functional architecture of basal ganglia Karthik Soman1,*, Vignesh Muralidharan1,*, V. Srinivasa Chakravarthy1 1Department of Biotechnology, Indian Institute of Technology Madras, Chennai, Tamil Nadu- 600036, India Correspondence: V. Srinivasa Chakravarthy - schakra@iitm.ac.in BMC Neuroscience 2016, 17(Suppl 1):P188 * Both authors have equal contribution. We present a networkmodel of hippocampus (HC) inspired by the functional architecture of the basal ganglia (BG). The model describes the role of hippocampus in spatial navigation and is cast in reinforcement learning (RL) framework (Fig. 97A). There is a corpus of literature which states that hippocampus is a key player in spatial learning because of the enriched sensory information that arrives at the portals of HC, the entorhinal cortex (EC), from the sensory cortical areas [1]. The model simulates the Morris water maze task wherein a virtual agent navigates inside a circular pool to find an invisible platform using the spatial context from the environment.Fig. 97 The model architecture (A) used to simulate the water maze task indicating the notion of a direct and an indirect pathway. The value function (B) developed after training the agent for 10 trials, the value peaks near to the platform. The escape latency (C) through trials shows that the agent has learnt the task with increased hippocampal dependence in the earlier stages and cortical dependence in the later stages of learning. The spectrogram (D) of the activity of CA3 as the function of time shows desynchronization while active exploration of the maze (0–45 s) and synchronized activity upon reaching the platform (45–65 s) In order to model the ability of HC to learn spatial context, we simulate a circular pool surrounded by six distinguishable poles of equal heights. As the agent/animal navigates, the size of retinal image of each pole varies with distance between the agent and the pole. Reward is given to the agent when it reaches the platform. The abstract form of the visual input is given to EC which has afferent and efferent projections from ventral tegmental area (VTA), one of the dopamine centers in the mid brain. Temporal difference (TD) error generated from VTA is used to update the synaptic weights for value computation of the sensory input in EC. Additionally an action vector defining the direction of the agent’s next step also forms a feedback input to the EC. We describe the functional anatomy of HC in terms of two pathways: a direct pathway between EC and CA1 and an indirect pathway between EC and CA1 via dentate gyrus (DG), and CA3. A quantity known as Value difference, that represents afferent dopamine signals in EC, is thought to control switching between these pathways. Desynchronized activity generated by the DG–CA3 loop in the indirect pathway aids the agent to explore the space. Direct pathway facilitates the agent’s navigation. The difference in the responses from these two pathways is computed in CA1 and is relayed to subiculum (Sbc) which computes the direction of the next step. Output of Sbc is communicated to higher motor areas (MC), modeled as a 1-D Continuous attractor neural network, via deeper layers of EC. MC also receives direct inputs from sensory areas so that the output of MC is the weighted sum of responses from the sensory cortical areas and HC respectively. MC response is used to update the next step. Cortico-cortical pathway (CCP) connections are updated using the TD error as well as the velocity generated from HC as a target. As the value function matures (Fig. 97B), contribution from HC declines; the CCP connections gain the upper hand and the agent reaches the platform faster (Fig. 97C). Thus, after training, the CCP can drive navigation without the involvement of HC. Analysis of CA3 activity shows desynchronization during active exploration and synchronizationupon reaching the platform (Fig. 97D). This resonates with experimental results suggesting that low-amplitude theta waves correspond to desynchronized activity during exploration, whereas the sharp waves during non-exploratory states correspond to synchronized activity [2]. ReferencesSukumar D, Rengaswamy M, Chakravarthy VS. Modeling the contributions of Basal ganglia and Hippocampus to spatial navigation using reinforcement learning. PloS One. 2012;7(10):e47467. Buzsáki G. Theta oscillations in the hippocampus. Neuron. 2002;33(3):325–40. P189 A computational architecture to model the microanatomy of the striatum and its functional properties Sabyasachi Shivkumar1, Vignesh Muralidharan1, V. Srinivasa Chakravarthy1 1Department of Biotechnology, Indian Institute of Technology Madras, Chennai, Tamil Nadu, India-600036 Correspondence: V. Srinivasa Chakravarthy - schakra@iitm.ac.in BMC Neuroscience 2016, 17(Suppl 1):P189 We propose a computational model of the functional architecture of the striatum. Anatomical and physiological evidence suggests that the microstructure of the striatum maps the sensory-motor information from the cortex in complex patterns. The dorsal striatum can be differentiated into centre surround regions called striosomes and matrisomes [1]. In the proposed striatum model, striosomes map the state space and the matrisomes map the action space. The model consists of a hierarchical two-level self organizing map (SOM), wherein the higher level SOM is trained on the state values and a sub-SOM layer containing multiple smaller SOMs are trained on action values. Neurons of ‘action SOMs’ are activated by neurons of ‘state SOMs’ The scheme of mapping of state space and action space onto the proposed architecture is given in Fig. 98A where the red area represents the striosomes and green area represents the matrisomes. We have also shown previously that such feature representation in a SOM layer can be used to develop value functions for sensory state spaces [2]. Thus to compute the state and action values, the activities of the respective SOMs were mapped to individual neurons by state and action weight vectors respectively. These weights were trained by the temporal difference error which represents the dopamine signals from the Substantia Nigra pars compacta (SNc) based on the reward from the environment. Action selection was performed by using the action values with exploration.Fig. 98 A Centre surround mapping in striosomes and matrisomes. The striosomes highlighted by red, map the states and the matrisomes highlighted by green, map the actions. B Value function map in the multiple context setting where the reward is present at the top left corner and the bottom right corner. C Switching of the modules based on the environmental contexts. The reward changes every 1000 episodes and the corresponding change in module with episode is shown The model was further extended to reflect striatal modularity, which could also be exploited to solve modular RL tasks with varying contexts [3]. This is done by using multi-SOMs, where multiple SOMs compete with each other to represent the input space. Biologically, this competition between different local striatal maps can be thought to be carried out by striatal interneurons. Using the above described striatal model as a single module, multiple modules were created. The higher level SOMs in these modules generate a responsibility signal, which represent the ability of the module to best represent that context. This was used to select the modules, a selection process which is probably carried out by the tonically active neurons (TANs). To validate this overall architecture, we tested this on the gridworld problem with a 10 × 10 grid and 4 actions. The reward is present at the corner of the grid in the first case and in subsequent case of modular RL framework, the reward is placed at one of two opposite corners. The value function map is built between the two modules and their switching at regular intervals is given in Fig. 98B, C respectively. ReferencesGraybiel A, Flaherty A, Gimenez-Amaya J-M. Striosomes and matrisomes. In: The basal ganglia III edn. Berlin: Springer; 1991. p. 3–12. Krishnan R, Ratnadurai S, Subramanian D, Chakravarthy VS, Rengaswamy M. Modeling the role of basal ganglia in saccade generation: Is the indirect pathway the explorer? Neural Networks. 2011;24(8):801–13. Amemori K-I, Gibb LG, Graybiel AM. Shifting responsibly: the importance of striatal modularity to reinforcement learning in uncertain environments; 2011. P190 A scalable cortico-basal ganglia model to understand the neural dynamics of targeted reaching Vignesh Muralidharan1, Alekhya Mandali1, B. Pragathi Priyadharsini1, Hima Mehta1, V. Srinivasa Chakravarthy1 1Department of Biotechnology, Indian Institute of Technology Madras, Chennai, Tamil Nadu-600036, India Correspondence: V. Srinivasa Chakravarthy - schakra@iitm.ac.in BMC Neuroscience 2016, 17(Suppl 1):P190 We present a scalable network model of the basal ganglia (BG) to highlight its role in performing simple reaching movements. The model consists of the following components: a 2-joint arm model (AM), a layer of motor-neurons in the spinal cord (MN), the proprioceptive cortex (PC), the motor cortex (MC), the prefrontal cortex (PFC) and the BG (Fig. 99A). The arm model has two joints each consisting of an agonist and an antagonist muscle pair innervated by a pair of motor neurons; the muscles in turn control the position of the arm in 2D space. The PC receives information about the muscle length and tension, thought to be originating from muscle spindles and Golgi tendon organs of the muscle. The MC then uses the sensory map information from the PC to develop a motor map of the arm. The MC activity is also modulated by the BG which uses reward information to make the arm learn to reach the target. The MC then sends these signals to respective muscles of the arm via the motor neurons (MN) to perform the movement. Since the existence of maps has been well established in the cortex, the sensory map of the PC, and the map from PC to MC were modelled using the self-organizing map (SOM) algorithm [1]. The motor command is thought to arise from the PFC, which specifies the goal to be reached. The MC therefore combines inputs from three sources: the PC, the prefrontal cortex (PFC), and the BG (from GPi via the thalamus). To enable this summation dynamically, MC was modelled as a continuous attractor neural network (CANN), wherein stable activity in CANN space corresponds to an equilibrium position of the arm in the workspace.Fig. 99 The model architecture (A) with the different modules aiding in reaching movements. The comparison of controls and PD’s approach to a target (B) and the appearance of PD symtoms including tremor and rigidity as a function of distance to the target Training of the model proceeds as follows. A target is chosen by activating corresponding neurons in the PFC. The arm makes exploratory movements driven by the Indirect Pathway of BG and gets rewarded when it reaches the target. Now BG uses this reward information and the corresponding arm position to transform it into a value profile over the arm’s workspace such that the trained value peaks at the target positions. As the training of model proceeds, the arm reaches the goal position faster and faster as BG stochastically climbs over the trained value function [2]. Furthermore, the connections from PFC and MC are also trained on successful reach, so that the motor command can directly activate the motor cortex thereby producing rapid movement avoiding the slow search conducted by the BG. The model exhibits all stages of motor learning i.e., slow movements dominated by the BG during early stages and cortically driven fast movements at later stages. The simulation results show PD symptoms like tremor which could be attributed to synchronized oscillations in STN-GPe (Fig. 99B). ReferencesChen Y, Reggia JA. Alignment of coexisting cortical maps in a motor control model. Neural Comput. 1996;8(4):731–55. Magdoom K, Subramanian D, Chakravarthy VS, Ravindran B, Amari S-I, Meenakshisundaram N. Modeling basal ganglia for understanding parkinsonian reaching movements. Neural Comput. 2011;23(2):477–516. P191 Emergence of radial orientation selectivity from synaptic plasticity Catherine E. Davey1, David B. Grayden1,2, Anthony N. Burkitt1 1Department of Electrical and Electronic Engineering, University of Melbourne, Victoria, 3010, Australia; 2Centre for Neural Engineering, University of Melbourne, Victoria, 3010, Australia Correspondence: Catherine E. Davey - cedavey@unimelb.edu.au BMC Neuroscience 2016, 17(Suppl 1):P191 The ability to learn and recall are primary functions of the brain. Synaptic plasticity is one of the key mechanisms by which we learn and adapt to our environment, and describes the process by which neuronal connection strengths are modified in response to environmental inputs [1]. There has been significant research effort invested into identifying the general principles of plasticity in neural networks, in order to garner insight into the learning process. The ability for animals to see and hear prior to birth is evidence of learning before exposure to ongoing external sensory signals. Consequently, cortical structure can be created, to some extent, in the absence of structured input. In a three-paper series, Linsker outlined a process by which cortical learning may occur prior to birth [2–4]. Linsker’s model identified particular spatial distributions of synaptic connectivity that are sufficient to induce the development of circularly symmetric cells in a system driven only by noisy input [2]. Furthermore, Linsker [3] revealed that orientation selective cells may develop by the sixth layer of processing. However, the resulting preferred orientation was a random function of stochastic weight initialisations [5]. Radial selectivity describes a tendency for cells to have the preferred orientation biased towards a central point, and has been observed in several cortical structures, including the visual cortex [6] and the auditory cortex [7]. In this study we reveal the mechanism by which radial orientation selectivity can emerge from synaptic plasticity in the absence of structured input. Linsker’s model assumed that cells within a laminar had an identical distribution of synaptic connection densities. This assumption is modified in this study to allow synaptic connection densities to change as a function of a cell’s radial distance to the centre of the laminar. The proposed network provides for spatially larger receptive fields as cells become progressively distal in the laminar, which is in keeping with electrophysiological and anatomical results. We show, both analytically and computationally, that this slightly modified network prompts the evolution of orientation selective cells with a predictable radial preference, in the third layer of neural processing. Importantly, this proposal maintains Linsker’s intent for a minimal set of model assumptions, ensuring that the resulting structure is robust to details and parameter values of the model used, and that general principles of plasticity are established. Consequently, our results are applicable to cortical learning generally. The mechanisms developed in this study could play a central role in the development of radial orientation selectivity in the visual cortex. Acknowledgements: This research was supported under Australian Research Council’s Discovery Projects funding scheme (Project Number DP140102947). ReferencesHughes JR. Post-tetanic potentiation. Phys Rev. 1958;38(1):91–113. Linsker R. From basic network principles to neural architecture: emergence of spatial-opponent cells. Proc Natl Acad Sci USA. 1986;83:7508–12. Linsker R. From basic network principles to neural architecture: emergence of orientation-selective cells. Proc Natl Acad Sci USA. 1986;83:8390–4. Linsker R. From basic network principles to neural architecture: emergence of orientation columns. Proc Natl Acad Sci USA. 1986;83:8779–83. Domany E, Hemmen JL van, Schulten K: Models of neural networks III. Berlin: Springer; 2012. Schall JD, Vitek DJ, Leventhal AG. Retinal constraints on orientation specificity in cat visual cortex. J Neurosci. 1986;6(3):823–36. Wang Y, Brzozowska-Prechtl A, Karten HJ. Laminar and columnar auditory cortex in avian brain. Proc Natl Acad Sci USA. 2010;107:12676–81. P192 How do hidden units shape effective connections between neurons? Braden A. W. Brinkman1,2, Tyler Kekona1, Fred Rieke2,3, Eric Shea-Brown1,2,4, Michael Buice4 1Department of Applied Mathematics, University of Washington, Seattle, WA 98195, USA; 2Department of Physiology and Biophysics, University of Washington, Seattle, WA 98195, USA; 3Howard Hughes Medical Institute, University of Washington, Seattle, WA 98195, USA; 4Allen Institute for Brain Science, Seattle, WA, 98109, USA Correspondence: Braden A. W. Brinkman - bradenb@uw.edu BMC Neuroscience 2016, 17(Suppl 1):P192 A major challenge in neuroscience is understanding how “hidden units”—neurons or other influences not observed in an experiment—influence the behavior of the observed neurons. Much work has been done on the inferring network interactions from data [1, 2], but it remains unknown how hidden neurons shape the network interactions inferred. Using techniques from non-equilibrium statistical physics, we have developed a theoretical framework to predict how effective connections in subsampled networks depend on the true connections in the full network. Beyond calculating effective connections, this approach can be systematically expanded to study how hidden units generate effective noise in subsampled networks. As an example, we apply this framework to a network of three spiking neurons described by a generalized linear model (GLM) with rates driven by a neuron’s own filtered spiking activity and those from which it receives input. By approximating the subsampled network as a GLM with effective spike-filters corrupted by Gaussian noise, we can analytically calculate how hidden units transform the filters (Fig. 100) and give rise to correlations in the effective noise (not shown). Based on our 3-neuron results, we conjecture that for general networks within this framework the filter between neurons i and j is modified by corrections from every path that neuron i can send a signal to neuron j through hidden units.Fig. 100 A Self-history filters (diagonal) and directed coupling filters between neurons (off-diagonal) in the full 3-neuron network. Neuron 1 is excitatory and its couplings to the other neurons are strictly positive. Neurons 2 and 3 are inhibitory and make strictly negative couplings to other neurons. There is no coupling from neuron 2–3. B Effective self-history filters (diagonals) and coupling filters (off-diagonals) when neuron 3 is hidden. The bottom row is unaltered because neuron 2 makes no coupling to neuron 3. The filters in the top row are changed due to the influence of signals neuron 1 sends to itself through neuron 3 and to neuron 2 through neuron 3. Although neuron 1’s true self-history filter and coupling from neuron 2 are negative the effective filters change sign. C The effective self-history filter of neuron 1 when both neurons 2 and 3 are hidden. Times and filter amplitudes are given in arbitrary units (a.u.) Acknowledgements: Support provided by the Sackler Scholar Program in Integrative Biophysics (BAWB), CRCNS grant DMS-1208027 (ESB, FR), NSF-DMS-1056125 (ESB), NIH grant EY11850 (FR), HHMI (FR). ESB and MB thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their vision, encouragement and support. ReferencesPillow JP, Shlens J, Paninski L, Sher A, Litke AM, Chichilnisky EJ, Simoncelli EP. Spatio-temporal correlations and visual signalling in a complete neuronal population. Nature. 2008;454:995–9. Pillow JW, Latham P. Neural characterization in partially observed populations of spiking neurons. Adv Neural Inf Process Syst. 2007;3256:1–8. P193 Characterization of neural firing in the presence of astrocyte-synapse signaling Maurizio De Pittà1,2, Hugues Berry2,3, Nicolas Brunel1,3 1Department of Neurobiology, University of Chicago, Chicago, IL 60637, USA; 2Project-Team BEAGLE, INRIA Rhône-Alpes, Villeurbanne, F-69603, France; 3Department of Statistics, University of Chicago, Chicago, IL 60637, USA Correspondence: Maurizio De Pittà - maurizio.depitta@gmail.com BMC Neuroscience 2016, 17(Suppl 1):P194 We study analytically the dynamics of neural activity in the presence of synaptic inputs modulated by astrocyte-released neurotransmitters (i.e. the so-called “gliotransmitters”). We start with the simple scenario of gliotransmitter-mediated modulation of synaptic release probability at N excitatory synapses impinging on a single postsynaptic neuron as well as on the same astrocyte domain. In this scenario, release from pre-synaptic terminals leads to activation of the astrocyte, that in turn modulates synaptic release through gliotransmitter release. In the limit of N → ∞ synapses, we derive equations relating gliotransmitter release to the instantaneous presynaptic rate, identify conditions for co-existence of multiple states of synaptic release, and study their stability. In the bistable regime, long-lasting potentiation of synaptic release by gliotransmission accounts for the emergence of persistent postsynaptic firing. Analysis of the coefficient of variation (CV) of the ensuing interspike interval distribution reveals increased firing variability following stimulation and in the presence of gliotransmission, in close analogy with increased CV values experimentally observed during the delay period in working-memory related tasks. We then extend our analysis to the scenario of a balanced neural network coupled with a network of astrocytes, and demonstrate the existence of an analogous mechanism for persistent neural firing by mean field theory. Taken together, our analysis suggests a novel astrocyte-based mechanism for persistent activity, and provides experimentally testable hypotheses on the possible involvement of astrocytes in cognitive tasks related to working memory. P194 Metastability of spatiotemporal patterns in a large-scale network model of brain dynamics James A. Roberts1, Leonardo L Gollo1, Michael Breakspear1 1Systems Neuroscience Group, QIMR Berghofer Medical Research Institute, Brisbane, QLD 4006, Australia Correspondence: James A. Roberts - james.roberts@qimrberghofer.edu.au BMC Neuroscience 2016, 17(Suppl 1):P194 Advances in mapping the human connectome have yielded increasingly-detailed descriptions of large-scale brain networks, prompting growing interest in the dynamics that emerge from this structural connectivity. Moreover, there is a desire to move beyond simple static functional connectivity measures to better describe and understand the more complex repertoire of brain dynamics, which unfolds on multiple time scales. Here, we analyze the dynamics that emerge from a neural mass model [1, 2] with network connectivity derived from densely-seeded probabilistic tractography on human diffusion imaging data [3]. We find a rich array of three-dimensional wave patterns, including traveling waves, spiral waves, sources, and sinks (Fig. 101). These patterns are metastable, with the dynamics cycling between several relatively long-lived states. Varying the overall coupling strength and coupling delay reveals a complex parameter space, with other emergent patterns such as cycling between strongly-correlated clusters and multistability between different regimes (as distinct from metastability within a single regime). These dynamics accord with empirical data from multiple imaging modalities, including observations of electrical waves in cortical tissue [4] and the presence of sequential spatiotemporal patterns in resting state MEG data [5]. By characterizing the dynamic states and time scales in our simulated data, we demonstrate the richness of dynamics that emerge from the human connectome. This work lays a platform for detailed analyses of large-scale functional neuroimaging data and their mechanistic underpinnings.Fig. 101 Large-scale wave patterns for strong coupling, showing four time snapshots for a traveling wave (top), a spiral wave (middle), and a sink pattern (bottom). Warmer colors denote higher amplitudes ReferencesBreakspear M, Terry JR, Friston KJ: Modulation of excitatory synaptic coupling facilitates synchronization and complex dynamics in a biophysical model of neuronal dynamics. Network. 2003;14:703–32. Gollo LL, Zalesky A, Hutchison RM, van den Heuvel M, Breakspear M. Dwelling quietly in the rich club: Brain network determinants of slow cortical fluctuations. Philos Trans R Soc Lond B Biol Sci. 2015;370:20140165. Roberts JA, Perry A, Lord AR, Roberts G, Mitchell PB, Smith RE, Calamante F, Breakspear M. The contribution of geometry to the human connectome. NeuroImage. 2016;124:379–93. Townsend RG, Solomon SS, Chen SC, Pietersen AN, Martin PR, Solomon SG, Gong P. Emergence of complex wave patterns in primate cerebral cortex. J Neurosci. 2015;35:4657–62. Baker AP, Brookes MJ, Rezek IA, Smith SM, Behrens T, Smith PJ, Woolrich M. Fast transient networks in spontaneous human brain activity. eLife. 2014;3:e01867. P195 Comparison of three methods to quantify detection and discrimination capacity estimated from neural population recordings Gary Marsat1, Jordan Drew1, Phillip D. Chapman1, Kevin C. Daly1, Samual P. Bradley1 1Department of Biology, West Virginia University, Morgantown, WV 26506, USA Correspondence: Gary Marsat - gary.marsat@mail.wvu.edu BMC Neuroscience 2016, 17(Suppl 1):P195 The responses of sensory neurons carry information about the presence and the identity of relevant external events [1]. The pattern of activity in populations of such neurons must be decoded by post synaptic networks for the information to contribute to the elaboration of behavioral responses. For two stimuli to be discriminated, or a stimulus discriminated from background (i.e. detected), the patterns of activity it elicits in the encoding neural population must be different enough that target decoders are activated differentially [2]. In this research we used recording from electrosensory neurons in Gymnotid fish [3] and recordings from the antennal lobe of moth [4] to compare three methods that can be used to quantify how accurately the neural responses can support detection and discrimination tasks. The first two methods, namely Euclidian distances [5] and spike metrics distances [3], have traditionally been used by neurophysiologist to characterize the information carried by spike trains and to compare the responses to different stimuli. The third method we explored relies on the clustering neural network provided in Matlab toolboxes that uses unsupervised learning. Neural networks of this type are widely used by engineers to perform practical tasks but are rarely used by neuroscientist to study actual neural systems. The clustering tool relies on a quantification similar in many ways to Euclidian distances or spike distance metrics. However, since it learns to weight the inputs to allow optimal clustering, the weight patterns of networks that cluster accurately can reveal features that actual neural networks ought to have (including synaptic facilitation and depression or amount of convergence of inputs). We show that, both for the olfactory system of moth and the electrosensory system of fish, the neural network decoder outperforms the other two decoding analyses by weighting more heavily information rich inputs and more weakly noisy ones. We argue that this tool can be advantageously used to quantify neural coding, can make testable prediction regarding the characteristics of the decoding network, and, most importantly, can be easily used and implemented by researchers who have little training in neural modeling. Acknowledgements: This work was supported by NSF Grant IOS-1557846 to G.M. ReferencesOllerenshaw DR, Zheng HJV, Millard DC, Wang Q, Stanley GB. The adaptive trade-off between detection and discrimination in cortical representations and behavior. Neuron. 2014;81(5):1152–64. Clemens J, Ronacher B. Feature extraction and integration underlying perceptual decision making during courtship behavior. J Neurosci. 2013;33(29):12136–45. Marsat G, Maler L. Neural heterogeneity and efficient population codes for communication signals. J Neurophysiol. 2010;104(5):2543–55. Staudacher EM, Huetteroth W, Schachtner J, Daly KC. A 4-dimensional representation of antennal lobe output based on an ensemble of characterized projection neurons. J Neurosci Methods. 2009;180(2):208–23. Daly KC, Bradley S, Chapman PD, Staudacher EM, Tiede R, Schachtner J. Space takes time: concentration dependent output codes from primary olfactory networks rapidly provide additional information at defined discrimination thresholds. Front Cell Neurosci. 2016;9. P196 Quantifying the constraints for independent evoked and spontaneous NMDA receptor mediated synaptic transmission at individual synapses Sat Byul Seo1, Jianzhong Su1, Ege T. Kavalali2, Justin Blackwell1 1Department of Mathematics, University of Texas at Arlington, Arlington, TX 76019, USA; 2Department of Neuroscience, University of Texas Southwestern Medical Center, Dallas, TX 75390, USA Correspondence: Sat Byul Seo - satbyul.seo@mavs.uta.edu BMC Neuroscience 2016, 17(Suppl 1):P196 Presynaptic terminals release neurotransmitters either in response to action potentials or spontaneously independent of presynaptic activity. In the case of glutamate, released neurotransmitters activate N-methyl-D-asparate (NMDA) receptors within a single postsynaptic site and give rise to miniature postsynaptic currents. In this study, we used a mathematical model to simulate spontaneous and evoked neurotransmission processes resulting from glutamate release within a synapse and evaluate the quantitative constraints that determine their degree of overlap independent signaling mediated by spontaneous and evoked release events. First we simulated isotropic diffusion of 4000 glutamates molecules release from a point source. We then simulated release of the glutamate molecules through a vesicle by addition of two compartments that one modeled the vesicle and the other represented the fusion pore. After we obtains the glutamate concentration from the standard heat equation then determine the opening probability of individual receptor using a state model (3C2O). Those two problems in MATLAB are solved. If we assume a fivefold–tenfold ratio as a good indicator for independent currents, then we cannot assure independency with the structure for medium and small synapses in our current hypothesis. Figure 102 shows that small synapse (200 nm × 200 nm) might not have independent signaling when glutamate release instantaneously because evoked and spontaneous receptors are not far away from each other and thus not far from the release site in either evoked or spontaneous releases, the ratio of open probability is close to 1. The open probability is consistent up to 90 nm far from the release site as in Fig. 102. However for small synapses, as glutamate release through 10 and 2 nm vesicle fusion pore, the open probability ratio decreases more drastically and become close to zero, and in 2 nm pore, the ratio achieves 10-fold reduction at 90 nm distance, giving plausibility for independent signaling.Fig. 102 In small synapses (200 nm × 200 nm), Ratios of maximum NMDA receptor opening probabilities as functions of receptor distance for different release speed (slow, 2 nm fusion pore—triangle, regular, 10 nm fusion pore—asterisk, and instantaneous—circle) of glutamate vesicle release. The open probabilities were calculated by the kinetics equation, when glutamates are released above the center location Conclusion From the results we conclude that peak open value is most sensitive to the distance from the receptor to the release site. Glutamate release speed or fusion pore size is relevant but to a lesser degree. The calculation was first performed for a large synapse of 0.36 µm2 (with R6 near the center for evoked neurotransmission, and R16 for spontaneous neurotransmission) which established in theory that two non-overlapping domains that give rise to independent signaling in large synapses [1]. Then calculations in medium size 0.16 µm2 and small size 0.04 µm2 push the biophysical envelope for independent currents, as the degree of independence decreases when the size of synapse gets smaller, or the distances from evoked and spontaneous receptors to the release site are closer together. ReferenceAtasoy D, Ertunc M, Moulder KL, et al. Spontaneous and evoked glutamate release activates two populations of NMDA receptors with limited overlap. J Neurosci. 2008;28:10151–166. P199 Gamma oscillation via adaptive exponential integrate-and-fire neurons LieJune Shiau1, Laure Buhry2, Kanishka Basnayake3 1Department of Mathematics, University of Houston, Clear Lake, Houston, TX, 77059, USA; 2Department of Computational Neurosciences, University of Lorraine, Nancy, 54600, France; 3Computational Neurosciences Laboratory, Ecole Polytechnique Federale de Lausanne, CH-1015, Switzerland Correspondence: LieJune Shiau - shiau@uhcl.edu BMC Neuroscience 2016, 17(Suppl 1):P199 Coherent oscillation of neuronal spiking in the brain is known related to cognitive functions, including perception, attention, and memory. It is therefore important to determine the properties of neurons and network architectures in emerging the coherent activities that influence the network collective behaviors. It is known that, in local cortical circuits, the probability for any pair of pyramidal cells to be connected is low and about 0.1–0.2 [1]. Wang and Buzsaki [2] numerically demonstrate that, in a heterogeneous and inhibitory network with sparse and random connection, the minimal connection required per neuron to observe coherence oscillation is approximately 60 with Hodgkin–Huxley (H–H) type interneurons in certain parameter regime. More importantly, this minimal number is relatively independent to the network size. In contrast, Golomb and Hanel [3] theoretically show that, identical inhibitory neurons in a sparse and random network, the minimal connection required per neuron to exhibit coherence oscillation is about 360 with integrate-and-fire (IF) neurons. The minimal connection required in either study depends on the intrinsic and synaptic properties of the neurons. It is shown that hippocampal CA1 neurons make an average of about 60 contacts to other neurons within a spatial span of approximately 500 μm [4]. Hence, to study rhythmic oscillation in hippocampal networks, H–H type neurons, instead of IF neurons, could produce Gamma rhythm through sparsely connected network in a population of interneurons. These findings are believed to address the importance of detailed physiological properties of single neurons in determining collective network behaviors. We adopt an increasingly popular two-dimensional adaptive exponential integrate-and-fire (aEIF) model [5] which is equipped with a subthreshold adaptation coupling the voltage and a slow current, and a spike-triggered adaptation regulated through each spike. To demonstrate that aEIF neurons can provide adequate networks in inducing Gamma frequency as in hippocampus effectively, we establish the minimal synaptic contacts required in sparse and random networks of aEIF neurons to exhibit coherent oscillations, and the impacts of neuronal and synaptic properties have on the minimal value. The aEIF neuron provides more physiological neuronal details than IF neuron, but much less than the H–H neurons. Intuitively, it may be anticipated that the minimal synaptic contacts of aEIF required in such networks lies somewhere between that of the IF and H–H neurons. We demonstrate that the minimal synaptic contacts required in such networks of aEIF neurons to exhibit Gamma rhythm is also surprisingly low with an approximation of 60 in certain parameter regime. More specifically, the minimal connection required per neuron for the onset of network synchrony is not a faction of the total number of network neurons. It either remains constant or only depends weakly on the total network neurons. This study indicates that the inclusion of subthreshold and spike-triggered adaptations provides aEIF neuron with features to compensate for the lack of physiological details, as supposed to its H–H neuron counterpart, in studying Gamma rhythm in the brain. Our result is very encouraging in building neural network studies through a simple two-dimensional model. ReferencesSong S, Sjostrom P, Reigl M, Nelson S, Chklovskii D. Highly nonrandom features of synaptic connectivity in local cortical circuits. PLoS Biol. 2005;3(3):0507–19. Wang X, Buzsaki G. Gamma oscillation by synaptic inhibition in a hippocampal interneuronal network model. J Neurophysiol. 1996;20:6402–13. Golomb D, Hansel D. The number of synaptic inputs and the synchrony of large sparse neuronal networks. Neural Comput. 1999;12(5):1095–1139. Sik A, Penttonen M, Ylinen A, Buzsaki G. Hippocampal CA1 interneurons: an in vivo intracellular labeling study. J Neurosci. 1999;24:49–65. Brette R, Wulfram G. Adaptive exponential integrate-and-fire model as an effective description of neuronal activity. J Neurophysiol. 2005;94:3637–42. P200 Visual face representations during memory retrieval compared to perception Sue-Hyun Lee1,2, Brandon A. Levy3, Chris I. Baker3 1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon 34141, Republic of Korea; 2 Program of Brain and Cognitive Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon 34141, Republic of Korea; 3Laboratory of Brain and Cognition, National Institute of Mental Health, National Institutes of Health, Bethesda, MD 20892, USA Correspondence: Sue-Hyun Lee - suelee@kaist.ac.kr BMC Neuroscience 2016, 17(Suppl 1):P200 In our daily life, we can easily discriminate and recognize familiar faces. Much evidence suggests that the fusiform face area (FFA) and the occipital face area (OFA) are involved in face processing [1–3]. However, it remains unclear how individual face information is represented in the visual cortex during retrieval compared to perception. To address this question, we performed an event-related functional magnetic resonance imaging (fMRI) experiment, comprising separate perception, learning and retrieval sessions. During the perception session, which took place inside the scanner, participants were presented with fixed pairings of six auditory cues (pseudowords) with six face images, and six auditory cues with six shoe images. During the learning session, which took place on a separate day outside the scanner, participants were trained to memorize the pseudoword-image associations for about 1 h. Finally, 1 day after the learning session, participants were scanned again and instructed to retrieve each image in response to auditory presentation of the paired pseudoword cue. To test the veracity of the retrieved visual information, participants were asked to perform forced-choice tests after the retrieval scan session, in which they heard one of the pseudoword cues and chose the paired category or image. Every participant showed near perfect performance in the forced-choice test. We focused on the patterns of response in face-selective cortical areas. Using multivoxel pattern analyses, we found that FFA showed more discriminable patterns of response to individual faces during retrieval compared to those elicited during perception. In contrast object-selective areas, which respond well to images of shoes, did not show any significant difference between perception and retrieval for individual shoe images. To determine whether the increased discrimination reflected a difference between perceived and retrieved face information and not an effect of learning, we conducted a similar fMRI experiment in which the second session was also perception and not retrieval. Importantly, there was no difference in face discrimination between the first and second perception sessions in FFA. Taken together, these results suggest that retrieval of face information generates more discriminative neural responses for individual faces than that evoked by perception of the very same faces. Acknowledgements: This work was supported by the US National Institutes of Health Intramural Research Program of the National Institute of Mental Health, and a NARSAD Young Investigator Grant from the Brain & Behavior Research Foundation. ReferencesKanwisher N, McDermott J, Chun MM. The fusiform face area: a module in human extrastriate cortex specialized for face perception. J Neurosci. 1997;17:4302–11. Tarr MJ, Gauthier I. FFA: a flexible fusiform area for subordinate-level visual processing automatized by expertise. Nat Neurosci. 2000;3:764–9. Kanwisher N, Yovel G. The fusiform face area: a cortical region specialized for the perception of faces. Philos Trans R Soc Lond B Biol Sci. 2006;361:2109–28. P201 Top-down modulation of sequential activity within packets modeled using avalanche dynamics Timothée Leleu1, Kazuyuki Aihara1 1Institute of Industrial Science, the University of Tokyo, Tokyo, Japan Correspondence: Timothée Leleu - timothee@sat.t.u-tokyo.ac.jp BMC Neuroscience 2016, 17(Suppl 1):P201 Recent experiments show that short activity packets are triggered by external stimuli or internal spontaneous events during which the temporal order of spikes is only partially stereotypical [1]. Moreover, it has been suggested that the timing of neurons during these packets depends on top-down modulatory inputs that “gate” the sensory information and represents either the replay of previously stored patterns or information about ongoing external stimuli [1]. Finally, it has been observed that spontaneous activity consists in the superposition of multiple overlapping packets [1]. We propose a simple model of cortical neural networks that reproduces these experimental observations [1] and an analytical description of the top-down modulation of packets using avalanche dynamics [2]. The proposed theory allows predicting the average size of packets using the synaptic weight matrix and vice versa. The model describes the neural activity within a cortical area that receives top-down modulatory and bottom-up sensory inputs from higher-order areas and thalamic projections, respectively. The activity of excitatory neurons is simulated using the leaky integrate-and-fire model. Excitatory synaptic connections are modified on shorter and longer time-scales by short-term depression and spike-timing dependent plasticity, respectively. Sequential patterns are stored within the recurrent connections of the middle area by repeated presentation of the external stimuli. Figure 103A1–C1, A2–C2 show that only the first and second stored sequences are replayed when the first and second top-down input is active, respectively, although the bottom-up inputs trigger the starting neurons of both sequential patterns. When there is no top-down input, the spontaneous activity is composed of the time-compressed superposition of both sequential patterns (see Fig. 103A4–C4).Fig. 103 A1–A4 Network structure. Neurons of the first and second stored pattern are represented by colors ranging from blue to green and yellow to red, respectively. Effective synaptic connections can be calculated and are shown by colored segments. B1–B4 Cross-correlograms (CCG) of single neuron activity with the summed activity of other neurons (see [1]). C1–C4 Center of mass of CCGs, noted μCC Acknowledgements: This research was supported by ImPACT Program of Council for Science, Technology and Innovation (Cabinet Office, Government of Japan). ReferencesLuczak A, Bartho P, Harris KD. Gating of sensory input by spontaneous cortical activity. J Neurosci. 2013;33(4):1684–95. Leleu T, Aihara K. Unambiguous reconstruction of network structure using avalanche dynamics. Phys Rev E. 2015;91:022804. Q28 An auto-encoder network realizes sparse features under the influence of desynchronized vascular dynamics Ryan T. Philips1, Karishma Chhabria1, V.Srinivasa Chakravarthy1 1Department of Biotechnology, Indian Institute of Technology, Madras, Chennai 600036, India Correspondence: V. Srinivasa Chakravarthy - schakra@iitm.ac.in BMC Neuroscience 2016, 17(Suppl 1):Q28 Please note that this abstract was presented at the previous year’s 24th Annual Computational Neuroscience Meeting: CNS-2015. Cerebral vascular dynamics are generally thought to be controlled by neural activity in a unidirectional fashion. However, both computational modeling and experimental evidence points to the feedback effects of vascular activity on neural dynamics [1, 2]. Vascular feedback in the form of glucose and oxygen controls neuronal ATP, which in turn can control the threshold of neural firing. We present a computational model of a neuro-vascular system in which a network of ‘vascular units’ supply ‘energy’ to a neural network (NN), which reduces neural firing threshold. The vascular network (VN) is modeled by a network of oscillators as in [3]. Neuronal pools fed by the complex dynamics of VN are turned ON and OFF randomly. We show that such a feedback mechanism results in sparse weight matrix, thereby enhancing the performance of an auto-encoder NN. In the proposed NN model, the hidden layer is coupled to a vascular network in a one-to-one fashion (Fig. 104A) and is trained using back-propagation. The cross-entropy (ce) error measure is used to update the energy demand parameter (Md) which is fed back to the VN. Md in turn governs the state of the vascular units, which determines if the neuron should be turned ON/OFF. This paradigm of randomly turning neural units ON/OFF is adapted from [4]. High Md results in an increase in ce as the neuronal dropout level is too low; similarly low Md results in an increase in ce due to high dropout. The time scale of vascular dynamics (Md) is much longer than that of neural dynamics (ce), reflecting physiology. The network was trained on two datasets: overlapping bar patterns and MNIST data. The settled weight matrices corresponding to both the synchronized (Fig. 104B) and desynchronized vascular dynamics (Fig. 104C) are shown in the Fig. 104B1, B2, C1, C2, respectively. Our earlier modeling study highlighted the link between desynchronized vascular dynamics and efficient energy delivery in skeletal muscle [3]. We now show that desynchronized vascular dynamics leads to efficient training in an auto-encoder NN.Fig. 104 A Auto-encoder NN coupled to the VN. B, C Depict desynchronized (ɛ = 1) and synchronized (ɛ = 0) states of VN respectively. The corresponding output weight patterns learnt by the auto-encoder, driven by the VN, trained on bar pattern data (B1, C1) and MNIST data (B2, C2) ReferencesChander BS, Chakravarthy VS. A computational model of neuro-glio-vascular loop interactions. PloS One. 2012;7(11):e48802. Moore CI, Cao R: The hemo-neural hypothesis: on the role of blood flow in information processing. J Neurophysiol. 2008;99(5):2035. Pradhan RK, Chakravarthy V, Prabhakar A. Effect of chaotic vasomotion in skeletal muscle on tissue oxygenation. Microvasc Res. 2007;74(1):51–64. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res. 2014;15(1):1929–58.

PMC005xxxxxx/PMC5001228.txt

==== Front Front Behav NeurosciFront Behav NeurosciFront. Behav. Neurosci.Frontiers in Behavioral Neuroscience1662-5153Frontiers Media S.A. 10.3389/fnbeh.2016.00161NeuroscienceOriginal ResearchVentral Midbrain NMDA Receptor Blockade: From Enhanced Reward and Dopamine Inactivation Hernandez Giovanni 1Cossette Marie-Pierre 2Shizgal Peter 2Rompré Pierre-Paul 12*1Département de Neurosciences, Université de MontréalMontréal, QC, Canada2FRQ-S Research Group in Behavioral Neurobiology, Department of Psychology, Concordia UniversityMontréal, QC, CanadaEdited by: John D. Salamone, University of Connecticut, USA Reviewed by: Alicia Izquierdo, University of California, Los Angeles, USA; Akiko Shimamoto, Meharry Medical College, USA *Correspondence: Pierre-Paul Rompré pierre-paul.rompre@umontreal.ca26 8 2016 2016 10 16125 5 2016 08 8 2016 Copyright © 2016 Hernandez, Cossette, Shizgal and Rompré.2016Hernandez, Cossette, Shizgal and RompréThis is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution and reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.Glutamate stimulates ventral midbrain (VM) N-Methyl-D-Aspartate receptors (NMDAR) to initiate dopamine (DA) burst firing activity, a mode of discharge associated with enhanced DA release and reward. Blockade of VM NMDAR, however, enhances brain stimulation reward (BSR), the results can be explained by a reduction in the inhibitory drive on DA neurons that is also under the control of glutamate. In this study, we used fast-scan cyclic voltammetry (FSCV) in anesthetized animals to determine whether this enhancement is associated with a change in phasic DA release in the nucleus accumbens. Rats were implanted with a stimulation electrode in the dorsal-raphe (DR) and bilateral cannulae above the VM and trained to self-administer trains of electrical stimulation. The curve-shift method was used to evaluate the effect of a single dose (0.825 nmol/0.5 μl/side) of the NMDAR antagonist, (2R,4S)-4-(3-Phosphopropyl)-2-piperidinecarboxylic acid (PPPA), on reward. These animals were then anesthetized and DA release was measured during delivery of electrical stimulation before and after VM microinjection of the vehicle followed by PPPA. As expected, phasic DA release and operant responding depended similarly on the frequency of rewarding electrical stimulation. As anticipated, PPPA produced a significant reward enhancement. Unexpectedly, PPPA produced a decrease in the magnitude of DA transients at all tested frequencies. To test whether this decrease resulted from excessive activation of DA neurons, we injected apomorphine 20 min after PPPA microinjection. At a dose (100 μg s.c.) sufficient to reduce DA firing under control conditions, apomorphine restored electrical stimulation-induced DA transients. These findings show that combined electrical stimulation and VM NMDARs blockade induce DA inactivation, an effect that indirectly demonstrates that VM NMDARs blockade enhances reward by potentiating stimulation-induced excitation in the mesoaccumbens DA pathway. dopamineglutamateNMDArewardin vivo voltammetryNatural Sciences and Engineering Research Council of Canada10.13039/501100000038RGPIN-2015-05018RGPIN-308-11 ==== Body Introduction Glutamate, the major excitatory neurotransmitter in the brain, plays a major role in behavioral, cognitive and motivational functions. Within the ventral midbrain (VM), glutamatergic afferents potently modulate reward-relevant circuitry by controlling dopamine (DA) excitability via two opposing mechanisms. Through its action on GABAergic afferents and GABAergic interneurons, glutamate maintains a strong inhibitory drive on DA neurons (Grace et al., 2007) so that a majority of them are unresponsive to excitatory inputs (Grace and Bunney, 1984). In contrast, through its direct action on DA neurons, glutamate switches DA neural activity from a slow, irregular, firing pattern to a phasic burst-firing mode that is associated with enhanced DA release (Grace and Bunney, 1984; Charara et al., 1996; Geisler et al., 2007; Omelchenko et al., 2009). Moreover, this mode of neural activity is associated with the acquisition of appetitive and aversive tasks (Zweifel et al., 2009). It has also been proposed that DA burst firing encodes reward prediction errors (Montague et al., 1996) and conveys motivationally relevant signals to anterior forebrain regions that control executive functions (Overton and Clark, 1997). Excitation and inhibition of DA neurons may entail multiple glutamate receptor subtypes with different sensitivity to agonists and antagonists and the expression on different inputs to VM DA neurons. Empirical evidence for such an arrangement comes from experiments in which systemic or local injection of either N-methyl-D-aspartate receptor (NMDAR) agonists or antagonists increases DA burst firing (French et al., 1993), terminal DA release (Westerink et al., 1996; Mathé et al., 1998; Kretschmer, 1999), and forward locomotion (Kretschmer, 1999; Cornish et al., 2001). Furthermore, VM microinjections of short inhibitory RNA that reduces the number of NMDAR on VM neurons results in an attenuation of reward induced by dorsal-raphe (DR) electrical stimulation, while VM microinjection of (2R,4S)-4-(3-Phosphopropyl)-2-piperidinecarboxylic acid (PPPA) and 2-carboxypiperazin-4-propyl-1-phosphonic acid ((R)-CPP), NMDAR antagonists that display a high affinity for receptor composed of the GluN2A subunits, produces the opposite effect (Bergeron and Rompré, 2013; Hernandez et al., 2015). The most likely mechanism for the reward-attenuating effect is a reduction in the NMDARs that control DA burst firing, whereas a likely mechanism for the reward-enhancing effect is a blockade of a different subtype of NMDARs that maintain the inhibitory drive on DA neurons. Given the large body of evidence supporting a role of VM mesoaccumbens DA neurons in reward (Wise and Rompre, 1989; Lak et al., 2014; Eshel et al., 2016) and the more recent evidence that activation of VM glutamatergic inputs from the DR induces reward and enhances extracellular DA in the nucleus accumbens (Qi et al., 2014), we used fast-scan cyclic voltammetry (FSCV) to determine whether the enhancement of DR reward by VM NMDAR blockade is associated with a change in phasic DA release in the nucleus accumbens shell (NAcS). Materials and Methods Subjects and Surgery Sixteen (16) male Long-Evans rats (Charles River, St-Constant, QC, Canada) weighing between 350–400 g at the time of the surgery were used. Rats were individually housed in a temperature- and humidity- controlled room with a 12-h light-dark cycle (lights on at 06:00 h) and ad libitum access to food and water. After a minimum 7-day period of acclimation to the housing environment rats were anesthetized with isoflurane and stereotaxically implanted according to Paxinos and Watson (2007) coordinates with 26-gauge guided cannulae (HRS Scientific, Montreal, Canada) aimed bilaterally at the VM (−5.5 mm AP, ±3.2 mm ML at a 18° angle, −6.5 mm DV from the skull surface) and a stainless steel monopolar electrode aimed at the DR (−7.6 mm AP, 0 mm ML, −6.6 mm DV from the skull surface). Detailed surgical procedures can be found in Bergeron and Rompré (2013). Of these 16 rats, nine rats were trained to nose-poke to deliver electrical pulses to the DR; whereas the other seven were used only in the electrochemistry experiment. For the behaviorally trained rats, one failed to self-stimulate and in three rats we were unable to measure DA. For the non-behaviorally trained rats we were unable to measure DA in one rat. For an schematic of the experimental sequence see Figure 1. All procedures were approved by the Animal Care and Use Committee of the Université de Montréal and Concordia University in accordance with the guidelines of the Canadian Council on Animal Care. Figure 1 Experimental sequence for behaviorally trained and naïve subjects. Drugs PPPA (Tocris, Ellisville, MI, USA) was dissolved in sterile 0.9% saline and stored frozen in 40–50 μl aliquots. Drug solutions were thawed just before testing and used only once. PPPA was injected into the VM at a dose of 0.825 nmol/0.5 μl/side. Urethane (ethyl carbamate; Sigma, St. Louis, MO, USA) was dissolved in sterile 0.9% saline; it was injected intraperitoneally (i.p.) at a dose of 1.5 g/kg. Apomorphine (Sigma, St. Louis, MO, USA) was dissolved in sterile 0.9% saline; it was injected subcutaneously (s.c.) at a total dosage of 100 μg /kg. Drug doses are expressed as salts. Self-Stimulation Training Eight rats were shaped to nose poke, under a fixed ratio 1 (FR1), for a 0.4-s train of cathodal, rectangular, constant-current pulses, 0.1 ms in duration, delivered at a frequency of 98 Hz. Once the rat nose poked consistently at current intensities between 125 and 400 μA, a rate vs. pulse-frequency curve was obtained by varying the stimulation frequency across trials over a range that drove the number of rewards earned from maximal to minimal levels. A detailed shaping and training procedure can be found in Hernandez et al. (2015). At least three behavioral tests were carried out before the FSCV experiments. A first saline test was performed to habituate the animals to the microinjection procedure. The detailed bilateral injections procedure can be found in Hernandez et al. (2015). Immediately after the microinjection, rats were allowed to self-stimulate for an hour. Results from this test were not included in the analysis. Baseline data were collected 1 week after this test. The rate vs. pulse-frequency data were fitted using the following sigmoid function y = Min+(Max−Min)1+10(x50−x)*p where Min is the lower asymptote, Max is the upper asymptote, x50 is the position parameter denoting the frequency at which the slope of the curve is maximal, and p determines the steepness of the sigmoid curve. The resulting fit was used to derive an index of reward threshold, defined as the pulse-frequency sustaining a half-maximal rate of responding (M50). Self-stimulation behavior was considered stable when the M50 values varied less than 0.1 log unit for three consecutive days. Once stable performance was obtained, we evaluated, on separate days, the effect of bilateral VM microinjections of a single dose (0.825 nmol/0.5 μl/side) of PPPA and an equal volume of saline (counterbalance order). FSCV After at least 4 days after completion of the last behavioral test, rats were anesthetized, and phasic DA release was measured during delivery of DR stimulation, both before and after bilateral VM microinjection of saline and PPPA. Rats were anesthetized with urethane and placed in the stereotaxic apparatus. Holes were drilled for the placement of the carbon-fiber electrode, sintered Ag/AgCl reference electrode (In Vivo Metrics, Healdsburg, CA, USA) and the anode electrode. For naïve rats (not behaviorally tested; n = 7), a DR stimulation electrode and VM cannulae were also implanted using the stereotaxic coordinates mentioned above. The carbon fiber electrode was built by encasing a carbon fiber (Thorne, Amoco Corporation, Greenville, SC, USA) in a single barrel borosilicate glass capillary (ID = 0.40 mm OD = 0.60 mm; A-M System Carlsborg, WA, USA). The seal between the carbon fiber and the glass was produced by heating the glass capillary with a pipette puller (PUL-1, WPI, Sarasota, FL, USA). A wire covered with silver paint (GC Electronics, Rockford, IL, USA) was inserted in the capillary to make contact with the carbon fiber and secured with shrink tubing coated with epoxy. The carbon fiber electrodes had an exposed tip length of 150–200 μm exposed tip length and 7 μm diameter. The carbon-fiber electrode (working electrode) was aimed at the NAcS (+1.7 mm AP, +1 mm ML, −7.0 mm DV). FSCV was computer- controlled as described previously (Heien et al., 2003). In brief, an 8.5 ms triangular input waveform (initial ramp, −0.4 to 1.3 V, 400 V/s; Heien et al., 2003) was applied to the working electrode at 10 Hz. The potential was held at −0.4 V between each scan to promote cation absorption at the surface of the FSCV electrode. A computer using software written in LABVIEW (National Instruments) and a multifunction data-acquisition board (PCI-6052E, National Instruments) controlled the waveform parameters and digitalized the recorded data. A PCI-6711E (National Instruments) board was used to synchronize the waveform acquisition, data collection and stimulation delivery. Background-subtracted cyclic voltammograms were obtained by digitally subtracting voltammograms collected during stimulation from those collected during baseline recording. Electrical stimulation was triggered after a synchronization signal was sent to the external input of a Master-8 pulse generator (A.M.P.I. Jerusalem, Israel). Voltages generated by the Master-8 were converted to constant currents via a stimulus isolation unit (AM-2200, AM-Systems, Carlsborg, WA, USA). Electrical stimulation was delivered 5 s after the start of each recording, and each was delivered in the 91.5 ms inter-waveform interval so that it did not interfere with the voltammetry scans. Once the carbon fiber electrode was in place, electrical stimulation (40 cathodal rectangular, constant-current pulses, 0.1 ms in duration) was delivered at a frequency of 98 Hz at the current intensity used during the behavioral tests or at an initial current of 400 μA.) through the DR electrode while DA transients were monitored; if a DA signal was not observed, the working electrode was lowered by 0.1 mm, and the electrical stimulation was repeated. This sequence was reiterated until stimulation-induced DA transients were detected. For naïve rats, the current intensity was adjusted so that the magnitude of DA signal measured across the different stimulation parameters was similar to that of the behaviorally trained rats. The DA signal was recorded in response to a descending set of pulse parameters similar to the ones used during the behavioral test. Each set of stimulation parameters was repeated three times with a 60-s inter-stimulation interval. This inter-stimulation interval produces a stable amplitude of DA transients (Cossette et al., 2016). Following this initial sweep, stimulation was delivered using three sets of parameters, corresponding to those that had produced maximal, half-maximal, and minimal responding in the behavioral test. The first sweep was used as the baseline, and then the vehicle was injected into the VM, followed by PPPA; The DA release was monitored across the three sets of stimulation parameters. In some rats, apomorphine or its vehicle was injected systemically (s.c.) after PPPA, and DA release was monitored again. At the end of the FSCV recordings, electrodes were post-calibrated by placing the electrode into a flow injection system (Upchurch Scientific, Oak Harbor, WA, USA), in which known concentrations of DA 100, 200, and 500 nM dissolved in artificial cerebrospinal fluid (aCSF: 145 mM Na+, 2.7 mM K+, 1.22 mM Ca2+, 1.0 mM Mg2+, 150 mMCl−, 0.2 mM ascorbate, 2 mM Na2HPO4, pH 7.4 ± 0.05) were used to related current values to concentration values. The average concentration of DA transients observed in the NacS after electrical stimulation of the DR at maximum frequency was 151.38 (SEM = 19.58) nM a value in the range of previously reported studies (Cheer et al., 2007; Park et al., 2010; Oleson et al., 2012; Saddoris et al., 2015). Histology After the completion of the experiment the location of the cannulae, stimulating and recording electrodes were obtained via deposit of iron ions after 1 mA of anodal current was delivered for 60 s. Animals were deeply anesthetized with urethane (2 g/kg, i.p.) and the current was delivered through the stimulating electrode, through the injection cannulae that were inserted into the guides, or through a stimulating electrode that was lowered to the site at which the voltammetric recordings were obtained. The animals were then perfused intracardially with 0.9% sodium chloride, followed by a formalin-Prussian Blue solution (10% formalin, 3% potassium ferricyanide, 3% potassium ferrocyanide and 0.5% trichloroacetic acid) that forms a blue reaction product with the iron particles. Brains were removed and fixed with 10% formalin solution for at least 7 days. Coronal sections of 40-μm thickness was cut with a cryostat to confirm placements of the FSCV electrodes in the NAcS, the stimulating electrode in the DR and cannulae in the VM. Data Analysis Matlab (Natick, MA, USA) was used to fit curves to the nose-poke and DA-release data. The phasic DA release was normalized to the maximum current obtained during baseline recordings. When standardized, nose-poke responses were normalized to the maximum number of responses. Comparisons between effectiveness indices for the behavioral and neurochemical data were made with paired or unpaired Student’s t-tests or one-way repeated-measures ANOVAs, followed by Tukey’s honestly significant difference post hoc tests. If the sphericity assumption was violated, the Greenhouse and Geisser correction was used. The correlation between the average change in M50 and the average change in the DA phasic release following PPPA injection was computed. Initial voltammetric data analysis was performed using LabVIEW written software and was low-pass filtered at 2 KHz. DA was chemically identified by its characteristic background-subtracted cyclic voltammogram; oxidation peak occurs at ~0.65 V and the reduction peak at ~−0.2 V, and principal component regression (PCR) was used as previously described to extract the DA component from the raw voltammetric data (Heien et al., 2005). The DA signal used for analysis was time-locked to the electrical stimulation trains. The quality of the signature was comparable at all pulse frequencies. All analyses were performed and graphics were prepared in Origin v9 (Northampton, MA, USA). Results Histology Histological analysis revealed that the tips of the FSCV recording electrodes were within the shell of the nucleus accumbens (Figures 2A,A′). The injection sites were located within the ventral part of the VM (Figures 2B,B′), a region that contains neurons activated by rewarding electrical stimulation (Wise and Rompre, 1989; Marcangione and Rompré, 2008). Finally, stimulation sites were located within the postero-medial mesencephalon, within the ventral central gray, between the anterior-posterior regions (Figures 2C,C′). Figure 2 Location of the tips of the carbon fibers (fast-scan cyclic voltammetry, FSCV; A,A′), injection sites (B,B′), and stimulating electrodes (C,C′) for each animal included in the study. Left panel shows the animals that were behaviorally trained. Right panel shows naïve animals. Sweep Dopamine Transients Figure 3A shows the obtained DA phasic transiens across the different stimulation parameters for those rats that received previous behavioral training (circles) and those that were naïve (squares). The DR electrical stimulation induced DA transients in the NAcS and the magnitude of the transients increased systematically as a function of the pulse frequency. Normalized phasic DA transients overlap, and there is no statistical difference in the steepness of the curves [t(11) = 1.09; p > 0.05], the pulse frequency that produced half-maximal release [t(11) = 1.66; p > 0.05], the lower asymptote [t(11) = 0.57; p > 0.05], or the upper asymptote [t(11) = 0.89; p > 0.05] between trained and naïve subjects. Figure 3B shows, in one subject, how self-stimulation performance and stimulation-induced DA transients vary as a function of pulse frequency. Although the slopes, in all the behaviorally tested subjects, between fitted curves relating stimulation frequency to the behavior and DA release differ [t(5) = 3.79; p < 0.05]; their M50 values do not [t(5) = 1.78; p > 0.05]. Figure 3 Dopamine (DA) release induced by dorsal raphe (DR) stimulation, as a function of pulse frequency. (A) The DA release profile obtained from behaviorally trained and naïve subjects is very similar. (B) The relation between self-stimulation performance and stimulation-induced DA release in one representative subject. Although the slopes of the two curves differ, their midpoints fall at similar positions along the pulse-frequency axis. PPPA Enhanced Brain Stimulation Reward, Yet it Decreased Dopamine Transients Figure 4A shows for a representative subject the behavioral effects of intra-VM injections of PPPA (0.825 nmol/0.5 μl/side). This drug produced a leftward and upward shift of the curve that relates the nose-poke rate as a function of pulse frequency: less stimulation was necessary to obtain a given level of performance and more responses were emitted when contrasted against the vehicle, at some stimulation frequencies. Figures 4B,C show respectively the average changes in M50 values and in maximal response expressed as a percentage of baseline, for all the behaviorally trained subjects in which DA transients were recorded. In contrast to the vehicle, PPPA produced a significant 25.3% (SEM = 6.3) reduction in M50 value [t(5) = 3.43; p < 0.05] and a significant 39.7% (SEM = 11.7) increase in nose-poke responses [t(5) = 3.27; p < 0.05]. The observed potentiation of DR reward is very similar to what has been previously observed and described (Bergeron and Rompré, 2013; Ducrot et al., 2013; Hernandez et al., 2015). Figure 4 Effects of intra ventral midbrain (VM) (2R,4S)-4-(3-Phosphopropyl)-2-piperidinecarboxylic acid (PPPA) injection on behavior. (A) In one representative subject intra VM injection of PPPA shifted the response-rate vs. pulse-frequency curve leftward and increased its upper asymptote. (B) The bar graph shows a significant reduction in M50 value in all the behaviorally trained subjects in which DA release was successfully measure. The reduction in M50 value suggests an increase in the effectiveness of the stimulation to elicit nose-poke behavior. (C) The bar graph shows a significant increase in the maximum response rate elicited by intra VM injection of PPPA. Under urethane anesthesia (Figure 5A) VM vehicle microinjection produced no discernable change in DA transients evoked by DR stimulation, whereas PPPA injection produced a significant decrease in the magnitude of the DA transients at all tested frequencies Maximum [F(1.002,5.02) = 47.94; p < 0.05]; M50[F(1.09,5.46) = 12.48; p = <0.05]; Minimum [F(1.41,7.07) = 15.65; p < 0.05]. It is noteworthy that the decrease in DA is negatively correlated with the magnitude of reward enhancement (rxy = −0.821); 59% of the variance observed in the magnitude attenuation of DA transients can be explained by the observed magnitude of reward enhancement (r2adj. = 0.593; p < 0.05, Figure 5B). Figure 5 Effects of intra VM PPPA injection on DA release. (A) Under urethane anesthesia intra VM PPPA injection decreased stimulation-elicited DA release in comparison to measures obtained before and after vehicle injection; this effect was seen at all three frequencies tested. (B) A negative correlation between the drug-induced decrease in stimulation-evoked DA release and the change in M50 was observed. In the behaviorally trained animals, the enhancement in reward effectiveness, produced by VM injection of PPPA covariates with the reduction in DA release; so that the larger the enhancement the greater the DA release reduction. We injected apomorphine to test whether the negative correlation between the drug-induced changes in M50 and DA transient magnitude resulted from depolarization inactivation (DI) in DA neurons due to the strong excitatory drive produced by the DR stimulation. At low doses, apomorphine is known to reduce DA firing by stimulating DA autoreceptors. This action hyperpolarizes DA neurons and increases input resistance, restoring DA firing and excitability when neurons are in a state of DI (Grace and Bunney, 1985). We injected a dose of apomorphine (100 μg s.c., Figure 6) that when injected alone produced a long-lasting and significant reduction of DA transients at the three frequencies tested Maximum [F(1.65,4.96) = 9.91; p < 0.05]; M50[F(1.89,5.69) = 17.87; p < 0.05]; Minimum [F(1.41,7.07) = 8.04; p = <0.05]. Figure 6 Subcutaneous injection of apomorphine (100 μg) produced a significant and long lasting reduction of DA oxidation at the three frequencies tested. Filled symbols represent a significant reduction of DA oxidation when contrasted against the pre-injection values, time 0. Figure 7 documents, for different selected subjects, the effects of apomorphine on stimulation-induced DA transients. The false-color plots represent background-subtracted redox currents as a function of voltage and time. Above each false-color plot is the time course of the DA-oxidation current at the potential corresponding to peak oxidation (the ordinal value corresponding to the center of the red blob in the leftmost false-color plot). DA transient concentration is directly proportional to this oxidation current. Superimposed on the false-color plots, in the upper right quadrant, is a background-subtracted voltammogram. The form of the voltammograms matches the voltammetric signature of DA; oxidation peak occurs at ~0.65 V and the reduction peak at −0.2 V. The time-course plots and voltammograms are horizontal and vertical sections, respectively, passing through the peak DA-oxidation current in the false-color plot. The dashed line denotes the onset of the electrical stimulation train. Figure 7 Examples of DA release in the nucleus accumbens shell (NAcS) in response to electrical stimulation of the DR at M50 parameters. The false-color plots show redox currents as function of applied voltage and time. Current at the peak oxidation potential of DA is shown above the false-color plot as function of time. The insets show the cyclic voltammogram. Dashed lines denote the onset of the electrical stimulation train. (A) DA release for M50 stimulation at baseline, 10 min after intra-VM saline; 10 and 20 min after s.c. saline. The DA peak is quite stable over time and across conditions. (B) Intra-VM injection of PPPA alone produces a significant decrease in DA release. (C) A similar decrease in DA phasic release is observed after s.c. injection of apomorphine. (D) Administration of apomorphine 20 min after intra-VM injection of PPPA partially restored the magnitude of DA transient. Figure 7A (top row) shows that systemic or intra-VM injection of saline had a negligible effect on the stimulation-induced DA transient. The rightmost two panels in Figures 7B,C demonstrate that both intra-VM PPPA (Figure 7B) and systemic apomorphine (Figure 7C) decreased the stimulation-induced DA transient. Systemic administration of apomorphine 10 min following intra-VM injection of PPPA partially restored the DA transient (Figure 7D, 4th column). Figure 8 quantifies the partial restoration of the DA transients by systemic administration of apomorphine (as illustrated by the difference between the peak DA oxidation currents in the bottom right panel of Figure 7 and the panel to its immediate left). The increase observed at each frequency was significant when contrasted against the last transient recorded after PPPA injection (tMax(5) = 4.34, p = <0.05; tM50(5) = 7.17, p < 0.05; tMin(5) = 3.58, p < 0.05). Figure 8 Subcutaneous injection of apomorphine 20 min after intra-VM injection of PPPA partially restores stimulation-induced DA release at the maximal stimulation frequency (A), M50 stimulation frequency (B), and at the lowest stimulation frequency (C). Discussion DA plays a critical role in reward and reward-seeking. Direct optical stimulation of DA neurons produces a conditioned place-preference (Tsai et al., 2009) and rodents will perform an operant response to optically activate VM DA neurons (Witten et al., 2011; Kim et al., 2012). Rewarding electrical stimulation of different brain areas increases DA levels in the Nac and elicits DA phasic release (Hernandez et al., 2006; Owesson-White et al., 2008; Hernandez and Shizgal, 2009; Cossette et al., 2016). Also, drugs that boost DA release enhance brain stimulation reward (BSR); whereas the opposite behavioral effect is obtained with drugs that decrease DA availability (Wise and Rompre, 1989). Our results show that electrical stimulation of the DR, at frequencies that elicit self-stimulation behavior, produces DA phasic release in the NacS. The magnitude of DA transients increase as a function of pulse frequency in an orderly manner and eventually levels off, as it has been observed previously (Fiorino et al., 1993; Yavich and Tanila, 2007). In our anesthetized preparation, DA release correlates with the behavioral performance observed during self-stimulation, which suggests that information about the reward intensity is reflected in the DA phasic release, at least over the range of stimulation parameters tested in the present study. To evaluate this hypothesis rigorously, it will be necessary to determine whether equiprefered stimulation parameters would produce similar phasic DA output (Moisan and Rompre, 1998). In freely moving animals, preferential blockade of VM NMDA GluN2A receptors with the PPPA enhances the reward-seeking produced by electrical stimulation of the DR (Bergeron and Rompré, 2013; Ducrot et al., 2013; Hernandez et al., 2015). This enhancement implicates GluN2A receptors in inhibition of VM DA neurons. Blockade of these NMDA receptors will lead to a disinhibition of DA neurons, enhanced phasic firing and DA release in terminal areas. Unexpectedly, electrically induced DA transients were lost when rewarding stimulation was combined with VM microinjection of PPPA. One possible explanation for this unexpected result is the product of an increase in net excitation that drives the DA neurons into a state of DI (White and Wang, 1983; Grace and Bunney, 1986). To test this hypothesis, we determined whether apomorphine can restore electrically induced DA transients following microinjection of PPPA. An in vitro electrophysiological study has shown that under normal conditions, apomorphine hyperpolarizes DA neurons (Grace and Bunney, 1985). However, in a state of DI, apomorphine-induced hyperpolarization restores the responsiveness of DA neurons to excitatory input (Grace and Bunney, 1986) by allowing the slow sodium gates to reopen. Consistent with this hypothesis, apomorphine partially restored the magnitude of electrically induced DA transients in animals that had received a prior VM microinjection of PPPA. We speculate that DA release was not totally restored because apomorphine produces multiple effects, some of which have opposing influences on the excitability of DA neurons. In addition to hyperpolarization due to stimulation of DA autoreceptors, apomorphine activates nerve-terminal autoreceptors, an effect that reduces extracellular DA release (Grace and Bunney, 1985). Accordingly, we found that apomorphine reduced the magnitude of electrically induced DA transients following VM microinjection of the vehicle. The occurrence of DI in our preparation is most likely the product of coordinated disinhibition, via blockade of NMDAR, and glutamatergic excitation, via activation of AMPA receptors (Ducrot et al., 2013; Qi et al., 2014). Similar synergistic effects leading to DI had been previously reported with the DA antagonist, pimozide, and the opiate agonist, morphine (Rompre and Wise, 1989; Henry et al., 1992). A VM microinjection of morphine that enhanced BSR in naïve animals produced a complete cessation of responding in animals previously injected with pimozide. In this latter condition, operant responding was reinstated by VM microinjection of a dose of the GABA agonist, muscimol, that inhibited reward under a control condition (no other drug treatment). The present findings not only support the hypothesis that PPPA and reward synergize to enhance DA excitation, but they also suggest that different NMDAR receptor subtypes are involved in modulation of mesoaccumbens DA impulse flow. NMDARs are heterodimers composed of two GluN1 subunits with GluN2 and/or GluN3 subunits. Previous pharmacological and SiRNA data suggest that PPPA-sensitive NMDARs are most likely located on VM afferents to DA neurons, are composed of GluN2A subunits, and are devoid of GluN2B (Bergeron and Rompré, 2013; Hernandez et al., 2015). Although previous results show that the DR reward signal is transmitted to VM neurons through AMPA receptors, a role for NMDAR cannot be excluded. NMDAR activation is essential for induction of DA burst firing (Zweifel et al., 2009) and a reduction in VM GluN1, the subunit common to all NMDARs, produces a significant attenuation of DR reward (Hernandez et al., 2015); it is thus most likely that the NMDAR involved induction of DA burst firing is composed of different subunits than the NMDAR that controls the inhibitory drive. The present results show that among its many roles, glutamate mediates a strong inhibitory drive on DA neurons and that DA-related reward signals can be strongly enhanced by reducing this inhibitory drive through blockade of VM NMDARs. This provides additional evidence that glutamate modulates DA neural activity in multiple ways and thus plays a key, albeit complex, role in reward signaling. Author Contributions P-PR and GH designed the study. GH and M-PC carried out the experiments, GH analyzed the data. P-PR, GH, M-PC and PS contributed to interpretation of the results and writing of the article. Funding and Disclosure This article was supported by Natural Sciences and Engineering Research Council of Canada (NSERC) grants to P-PR (#119057, RGPIN-2015-05018) and to PS (RGPIN-308-11); NSERC postdoctoral Fellowship to GH, a grant from the “Fonds de recherche du Québec—Santé” to the “Groupe de Recherche en Neurobiologie Comportementale”/Center for Studies in Behavioral Neurobiology, and support to PS from the Concordia University Research Chairs program. Conflict of Interest Statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The authors thank Wayne Brake, for hosting part of the behavioral experiments in his laboratory and David Munro and Claude Bouchard for technical assistance. ==== Refs References Bergeron S. Rompré P.-P. (2013 ). Blockade of ventral midbrain NMDA receptors enhances brain stimulation reward: a preferential role for GluN2A subunits . Eur. Neuropsychopharmacol. 23 , 1623 –1635 . 10.1016/j.euroneuro.2012.12.005 23352316 Charara A. Smith Y. Parent A. (1996 ). Glutamatergic inputs from the pedunculopontine nucleus to midbrain dopaminergic neurons in primates: phaseolus vulgaris-leucoagglutinin anterograde labeling combined with postembedding glutamate and GABA immunohistochemistry . J. Comp. Neurol. 364 , 254 –266 . 10.1002/(SICI)1096-9861(19960108)364:2<254::AID-CNE5>3.0.CO;2-4 8788248 Cheer J. F. Aragona B. J. Heien M. L. Seipel A. T. Carelli R. M. Wightman R. M. (2007 ). Coordinated accumbal dopamine release and neural activity drive goal-directed behavior . Neuron 54 , 237 –244 . 10.1016/j.neuron.2007.03.021 17442245 Cornish J. L. Nakamura M. Kalivas P. W. (2001 ). Dopamine-independent locomotion following blockade of N-methyl-D-aspartate receptors in the ventral tegmental area . J. Pharmacol. Exp. Ther. 298 , 226 –233 . 11408546 Cossette M.-P. Conover K. Shizgal P. (2016 ). The neural substrates for the rewarding and dopamine-releasing effects of medial forebrain bundle stimulation have partially discrepant frequency responses . Behav. Brain Res. 297 , 345 –358 . 10.1016/j.bbr.2015.10.029 26477378 Ducrot C. Fortier E. Bouchard C. Rompré P.-P. (2013 ). Opposite modulation of brain stimulation reward by NMDA and AMPA receptors in the ventral tegmental area . Front. Syst. Neurosci. 7 :57 . 10.3389/fnsys.2013.00057 24106463 Eshel N. Tian J. Bukwich M. Uchida N. (2016 ). Dopamine neurons share common response function for reward prediction error . Nat. Neurosci. 19 , 479 –486 . 10.1038/nn.4239 26854803 Fiorino D. F. Coury A. Fibiger H. C. Phillips A. G. (1993 ). Electrical stimulation of reward sites in the ventral tegmental area increases dopamine transmission in the nucleus accumbens of the rat . Behav. Brain Res. 55 , 131 –141 . 10.1016/0166-4328(93)90109-4 7689319 French E. D. Mura A. Wang T. (1993 ). MK-801, phencyclidine (PCP) and PCP-like drugs increase burst firing in rat A10 dopamine neurons: comparison to competitive NMDA antagonists . Synapse 13 , 108 –116 . 10.1002/syn.890130203 8446919 Geisler S. Derst C. Veh R. W. Zahm D. S. (2007 ). Glutamatergic afferents of the ventral tegmental area in the rat . J. Neurosci. 27 , 5730 –5743 . 10.1523/JNEUROSCI.0012-07.2007 17522317 Grace A. A. Bunney B. S. (1984 ). The control of firing pattern in nigral dopamine neurons: burst firing . J. Neurosci. 4 , 2877 –2890 . 6150071 Grace A. A. Bunney B. S. (1985 ). Low doses of apomorphine elicit two opposing influences on dopamine cell electrophysiology . Brain Res. 333 , 285 –298 . 10.1016/0006-8993(85)91582-3 3995296 Grace A. A. Bunney B. S. (1986 ). Induction of depolarization block in midbrain dopamine neurons by repeated administration of haloperidol: analysis using in vivo intracellular recording . J. Pharmacol. Exp. Ther. 238 , 1092 –1100 . 3746660 Grace A. A. Floresco S. B. Goto Y. Lodge D. J. (2007 ). Regulation of firing of dopaminergic neurons and control of goal-directed behaviors . Trends Neurosci. 30 , 220 –227 . 10.1016/j.tins.2007.03.003 17400299 Heien M. L. A. V. Khan A. S. Ariansen J. L. Cheer J. F. Phillips P. E. M. Wassum K. M. . (2005 ). Real-time measurement of dopamine fluctuations after cocaine in the brain of behaving rats . Proc. Natl. Acad. Sci. U S A 102 , 10023 –10028 . 10.1073/pnas.0504657102 16006505 Heien M. L. A. V. Phillips P. E. M. Stuber G. D. Seipel A. T. Wightman R. M. (2003 ). Overoxidation of carbon-fiber microelectrodes enhances dopamine adsorption and increases sensitivity . Analyst 128 , 1413 –1419 . 10.1039/B307024G 14737224 Henry D. J. Wise R. A. Rompre P. P. White F. J. (1992 ). Acute depolarization inactivation of A10 dopamine neurons: interactions of morphine with dopamine antagonists . Brain Res. 596 , 231 –237 . 10.1016/0006-8993(92)91552-p 1467986 Hernandez G. Hamdani S. Rajabi H. Conover K. Stewart J. Arvanitogiannis A. . (2006 ). Prolonged rewarding stimulation of the rat medial forebrain bundle: neurochemical and behavioral consequences . Behav. Neurosci. 120 , 888 –904 . 10.1037/0735-7044.120.4.888 16893295 Hernandez G. Khodami-Pour A. Lévesque D. Rompré P.-P. (2015 ). Reduction in ventral midbrain NMDA receptors reveals two opposite modulatory roles for glutamate on reward . Neuropsychopharmacology 40 , 1682 –1691 . 10.1038/npp.2015.14 25578795 Hernandez G. Shizgal P. (2009 ). Dynamic changes in dopamine tone during self-stimulation of the ventral tegmental area in rats . Behav. Brain Res. 198 , 91 –97 . 10.1016/j.bbr.2008.10.017 18996152 Kim K. M. Baratta M. V. Yang A. Lee D. Boyden E. S. Fiorillo C. D. (2012 ). Optogenetic mimicry of the transient activation of dopamine neurons by natural reward is sufficient for operant reinforcement . PLoS One 7 :e33612 . 10.1371/journal.pone.0033612 22506004 Kretschmer B. D. (1999 ). Modulation of the mesolimbic dopamine system by glutamate: role of NMDA receptors . J. Neurochem. 73 , 839 –848 . 10.1046/j.1471-4159.1999.0730839.x 10428083 Lak A. Stauffer W. R. Schultz W. (2014 ). Dopamine prediction error responses integrate subjective value from different reward dimensions . Proc. Natl. Acad. Sci. U S A 111 , 2343 –2348 . 10.1073/pnas.1321596111 24453218 Marcangione C. Rompré P.-P. (2008 ). Topographical Fos induction within the ventral midbrain and projection sites following self-stimulation of the posterior mesencephalon . Neuroscience 154 , 1227 –1241 . 10.1016/j.neuroscience.2008.05.014 18556137 Mathé J. M. Nomikos G. G. Schilström B. Svensson T. H. (1998 ). Non-NMDA excitatory amino acid receptors in the ventral tegmental area mediate systemic dizocilpine (MK-801) induced hyperlocomotion and dopamine release in the nucleus accumbens . J. Neurosci. Res. 51 , 583 –592 . 10.1002/(SICI)1097-4547(19980301)51:5<583::AID-JNR5>3.0.CO;2-B 9512002 Moisan J. Rompre P. P. (1998 ). Electrophysiological evidence that a subset of midbrain dopamine neurons integrate the reward signal induced by electrical stimulation of the posterior mesencephalon . Brain Res. 786 , 143 –152 . 10.1016/s0006-8993(97)01457-1 9554987 Montague P. R. Dayan P. Sejnowski T. J. (1996 ). A framework for mesencephalic dopamine systems based on predictive Hebbian learning . J. Neurosci. 16 , 1936 –1947 . 8774460 Oleson E. B. Gentry R. N. Chioma V. C. Cheer J. F. (2012 ). Subsecond dopamine release in the nucleus accumbens predicts conditioned punishment and its successful avoidance . J. Neurosci. 32 , 14804 –14808 . 10.1523/JNEUROSCI.3087-12.2012 23077064 Omelchenko N. Bell R. Sesack S. R. (2009 ). Lateral habenula projections to dopamine and GABA neurons in the rat ventral tegmental area . Eur. J. Neurosci. 30 , 1239 –1250 . 10.1111/j.1460-9568.2009.06924.x 19788571 Overton P. G. Clark D. (1997 ). Burst firing in midbrain dopaminergic neurons . Brain Res. Rev. 25 , 312 –334 . 10.1016/S0165-0173(97)00039-8 9495561 Owesson-White C. A. Cheer J. F. Beyene M. Carelli R. M. Wightman R. M. (2008 ). Dynamic changes in accumbens dopamine correlate with learning during intracranial self-stimulation . Proc. Natl. Acad. Sci. U S A 105 , 11957 –11962 . 10.1073/pnas.0803896105 18689678 Park J. Aragona B. J. Kile B. M. Carelli R. M. Wightman R. M. (2010 ). In vivo voltammetric monitoring of catecholamine release in subterritories of the nucleus accumbens shell . Neuroscience 169 , 132 –142 . 10.1016/j.neuroscience.2010.04.076 20451589 Paxinos G. Watson C. (2007 ). The Rat Brain in Stereotaxic Coordinates Sixth Edition. (Vol. 170), San Diego, CA : Elsevier Academic Press , 547 –612 . Qi J. Zhang S. Wang H.-L. Wang H. de Jesus Aceves Buendia J. Hoffman A. F. . (2014 ). A glutamatergic reward input from the dorsal raphe to ventral tegmental area dopamine neurons . Nat. Commun. 5 :5390 . 10.1038/ncomms6390 25388237 Rompre P.-P. Wise R. A. (1989 ). Behavioral evidence for midbrain dopamine depolarization inactivation . Brain Res. 477 , 152 –156 . 10.1016/0006-8993(89)91402-9 2702480 Saddoris M. P. Cacciapaglia F. Wightman R. M. Carelli R. M. (2015 ). Differential dopamine release dynamics in the nucleus accumbens core and shell reveal complementary signals for error prediction and incentive motivation . J. Neurosci. 35 , 11572 –11582 . 10.1523/JNEUROSCI.2344-15.2015 26290234 Tsai H.-C. Zhang F. Adamantidis A. Stuber G. D. Bonci A. de Lecea L. . (2009 ). Phasic firing in dopaminergic neurons is sufficient for behavioral conditioning . Science 324 , 1080 –1084 . 10.1126/science.1168878 19389999 Westerink B. H. Kwint H. F. deVries J. B. (1996 ). The pharmacology of mesolimbic dopamine neurons: a dual-probe microdialysis study in the ventral tegmental area and nucleus accumbens of the rat brain . J. Neurosci. 16 , 2605 –2611 . 8786436 White F. J. Wang R. Y. (1983 ). Differential effects of classical and atypical antipsychotic drugs on A9 and A10 dopamine neurons . Science 221 , 1054 –1057 . 10.1126/science.6136093 6136093 Wise R. A. Rompre P.-P. (1989 ). Brain dopamine and reward . Annu. Rev. Psychol. 40 , 191 –225 . 10.1146/annurev.psych.40.1.191 2648975 Witten I. B. Steinberg E. E. Lee S. Y. Davidson T. J. Zalocusky K. A. Brodsky M. . (2011 ). Recombinase-driver rat lines: tools, techniques and optogenetic application to dopamine-mediated reinforcement . Neuron 72 , 721 –733 . 10.1016/j.neuron.2011.10.028 22153370 Yavich L. Tanila H. (2007 ). Mechanics of self-stimulation and dopamine release in the nucleus accumbens . Neuroreport 18 , 1271 –1274 . 10.1097/wnr.0b013e328244e576 17632281 Zweifel L. S. Parker J. G. Lobb C. J. Rainwater A. Wall V. Z. Fadok J. P. . (2009 ). Disruption of NMDAR-dependent burst firing by dopamine neurons provides selective assessment of phasic dopamine-dependent behavior . Proc. Natl. Acad. Sci. U S A 106 , 7281 –7288 . 10.1073/pnas.0813415106 19342487

PMC005xxxxxx/PMC5001229.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557078289810.1186/s12864-016-2898-5ResearchLipopolysaccharide treatment induces genome-wide pre-mRNA splicing pattern changes in mouse bone marrow stromal stem cells Zhou Ao aozhou@iu.edu 12Li Meng lvimeng_hrbeu@sina.com 3He Bo bohe_he@yahoo.cn 3Feng Weixing wfeng@compbio.iupui.edu 3Huang Fei huangfei@iu.edu 1Xu Bing xubinghrb@gmail.com 46Dunker A. Keith kedunker@iu.edu 1Balch Curt curt.balch@gmail.com 5Li Baiyan liby@ems.hrbmu.edu.cn 6Liu Yunlong yunliu@iupui.edu 14Wang Yue yuewang@iu.edu 41 Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 2 Bioinformatics Program, Indiana University School of Informatics, Indianapolis, IN 46202 USA 3 College of Automation, Harbin Engineering University, Harbin, Heilongjiang China 4 Department of Medical and Molecular Biology, Indiana University School of Medicine, Indianapolis, IN 46202 USA 5 Bioscience Advising, Indianapolis, IN 46227 USA 6 Department of Pharmacology, Harbin Medical University, Harbin, Heilongjiang China 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.509© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Lipopolysaccharide (LPS) is a gram-negative bacterial antigen that triggers a series of cellular responses. LPS pre-conditioning was previously shown to improve the therapeutic efficacy of bone marrow stromal cells/bone-marrow derived mesenchymal stem cells (BMSCs) for repairing ischemic, injured tissue. Results In this study, we systematically evaluated the effects of LPS treatment on genome-wide splicing pattern changes in mouse BMSCs by comparing transcriptome sequencing data from control vs. LPS-treated samples, revealing 197 exons whose BMSC splicing patterns were altered by LPS. Functional analysis of these alternatively spliced genes demonstrated significant enrichment of phosphoproteins, zinc finger proteins, and proteins undergoing acetylation. Additional bioinformatics analysis strongly suggest that LPS-induced alternatively spliced exons could have major effects on protein functions by disrupting key protein functional domains, protein-protein interactions, and post-translational modifications. Conclusion Although it is still to be determined whether such proteome modifications improve BMSC therapeutic efficacy, our comprehensive splicing characterizations provide greater understanding of the intracellular mechanisms that underlie the therapeutic potential of BMSCs. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2898-5) contains supplementary material, which is available to authorized users. Keywords Alternative splicingLipopolysaccharideMesenchymal stem cellsThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Alternative splicing (AS) is important for gene regulation and is a major source of proteome diversity in mammals [1] through altering the composition of mRNA transcripts by including or excluding specific exons [2]. AS can further modulate organism complexity not only by effectively increasing regulatory and signaling network complexity, but also by doing so in a temporal- and spatial-specific manner, supporting cell differentiation, developmental pathways, and other processes associated with multicellular organisms. Indeed, AS shows a strong relationship with organism complexity, as estimated by the organism’s number of different cell types [3]. The recent ENCODE Project concluded that at least 90 % of human genes express multiple mRNAs through alternative splicing of exons or exon segments [4]. As might be expected, deregulation of this process is associated with numerous diseases [5–10]. Bone marrow-derived mesenchymal stem cells (BMSCs) are adult stem cells capable of self-renewal and differentiation into numerous cell lineages, including osteocytes, adipocytes, and chondrocytes [11]. One promising use of BMSCs is repair of ischemia-damaged cardiac tissue. BMSCs are easy to expand in vitro, can be genetically modified and exhibit significant immunotolerance properties [12–14], making BMSCs an attractive candidate for tissue repair/regeneration therapy. Intramyocardial injection of BMSCs reduces inflammation, fibrosis, infarct size, ventricular remodeling, and therefore, improves cardiac function following tissue insult [15–18]. Because the majority of BMSCs are soon lost during after injection, the observed therapeutic effects likely derive from paracrine effects of bioactive molecules released from these cells [15, 16]. Indeed, BMSC-mediated release of cytoprotective protein factors or transfer of intracellular components (e.g.,mRNAs, microRNAs, and proteins) via cell membrane exosomes, represents a novel mechanism of cell-to-cell communication [19]. To date, however, clinical trials have demonstrated that while effective, delivery of BMSCs to ischemic myocardium results in only modest and short-lived benefits [20, 21]. Therefore, there is a critical need to elucidate the mechanisms by which BMSCs mediate their therapeutic benefits, including identification of their specific paracrine factor(s), and conditions under which their functions can be optimized. Upon injection into damaged heart tissue, BMSCs face a hypoxic, ischemic environment that severely limits their therapeutic efficacy. Thus, preconditioning BMSCs with various growth factors and endogenous or exogenous molecules has been used to improve BMSC therapeutic efficacy [22–24]. Indeed, it has been reported previously that bacterial endotoxin (lipopolysaccharide, LPS) could stimulate BMSCs to release paracrine factors, including angiogenic growth factors, cytokines, and chemokines that facilitate tissue repair [13, 14]. In addition, our previous study suggested that BMSC expression of the LPS receptor, toll-like receptor 4 (TLR4), regulates BMSC paracrine properties and intracellular STAT3 signaling cascades [25]. Moreover, preconditioning of BMSCs with LPS improves their therapeutic efficacy in rodent models of ischemia/reperfusion injury [23]. However, BMSC transcriptomic changes (in particular, alterations in mRNA transcript processing and splicing) that occur following LPS stimulation have been little studied. Besides use as an attractive therapeutic tool for repairing ischemic heart, BMSCs have been used for numerous other diseases, including graft-versus-host disease, Crohn’s disease, stroke, cartilage defects, diabetes and many others [26–31]. With the growing incidence of bacterial endotoxin LPS detected in older or immunocompromised patients with multiple-drug resistant bacteria, diabetes, cancer, indwelling IV catheters, and on complex chemotherapy regimens [32, 33], it is of great importance to study whether the stimulation of these implanted BMSCs by endogenous LPS would alter their therapeutic efficacy. Moreover, because MSCs are present in bone marrow and many other tissues, it merits extensive investigation whether LPS stimulation of these endogenous MSCs would influence the clinical outcomes of complex therapeutic regimens. Despite BMSC’s strong clinical potential, the role(s) of alternative splicing in LPS response has not been fully explored. The recent development of high-throughput sequencing technology has now made transcriptome-wide profiling of splicing isoforms possible. In this study, we used RNA-seq analysis of BMSCs to identify and characterize gene transcripts whose splicing patterns were altered by LPS treatment. Results To investigate LPS-induced transcriptomic changes in BMSCs due to alternative splicing, RNA-seq analysis was conducted on BMSCs before and after LPS treatment, in triplicate. A strand-directed single-end RNA-seq protocol (75 bp reads) was used with the SOLiD 5500xl instrument. The total analysis resulted in 326 million reads, with each of the six samples ranging from 43 to 59 million reads. After removing the reads with low sequencing quality (see Methods) and filtering reads mapped to ribosomal RNAs and repeats, the remaining reads were mapped to the standard mouse reference genome (mm9). The total number of mappable reads in each sample ranged from 29 to 36 million, with an average mapping percentage of 59 %. Among the mappable reads in each sample, 3.8 to 5.0 million are mapped to protein coding exons, and 2.8 to 4.0 million are mapped to splice junctions. Detailed mapping statistics for the six samples are listed in Additional file 1. LPS-induced alternative splicing We applied a MISO (Mixture of Isoform) algorithm [34] to identify alternative splicing events elicited by LPS treatment. Based on a Bayesian inference framework, MISO is a probabilistic framework that quantitates the expression levels of alternatively spliced genes from RNA-Seq data, and identifies differentially regulated exons across samples. MISO computes Percent Spliced In (PSI, or Ψ) values for each alternative splicing event, representing the fraction of a gene’s mRNA that includes the exon. For each event, MISO also calculates a Bayesian Factor (BF) that quantifies the likelihood of the changes. For instance, [BF] = 5 indicates it is five times more likely that a specific alternative splicing event occurred than did not occur. Overall, we identified 197 exons whose splicing patterns differed between control and LPS-treated BMSCs (Bayesian factor [BF] > 5 and |ΔΨ| > 0.05). This number represents 2.32 % of all 8,475 events whose inclusion percentages could be reliably measured from the RNA-seq data; these genes generally had higher expression levels to generate enough read depth for splicing analysis. For genes with lower expression levels, our RNA-seq experiment did not have enough read depth for such analysis. The 197 LPS-induced alternatively spliced events included 82 cassette exons, 28 alternative donor site events (5′-ss), 45 alternative acceptor site events (3′-ss), and 42 intron retention events. Figure 1 demonstrates the magnitude (X-axis) and significance (Y-axis) of LPS-induced splicing pattern changes on all the alternatively exons that could be reliably identified by MISO under both untreated and LPS-treated conditions (Fig. 1). Among these 197 events (red dots in Fig. 1), 117 showed positive ∆Ψ values, indicating that the percentage of transcripts containing the specific exon increased in the LPS-treated samples compared to control samples. Similarly, 80 events showed negative ∆Ψ values, indicating a decrease in the percentage of transcripts containing specific exons. For each of the four types of splicing events (cassette exons, alternative 5′-donor sites, alternative 3′-acceptor sites, and intron retention), we show one Sashimi plot for the exons with the largest LPS-induced changes (either increases or decreases) in percentage of inclusion in the gene product (Fig. 2). The Sashimi plot demonstrates the RNA-seq read densities along exons and junctions, in the context of the structure of the gene’s isoforms. In addition, the distribution and the confidence intervals of the estimated Ψ under both conditions (LPS vs. untreated) are also included.Fig. 1 LPS-induced alternative splicing events. Scatter plot of all the AS events identified in MISO. The X-axis represents ∆Ψ, and the Y-axis represents log (BF). The shape of the dots indicates the type of the events. Specifically, circle indicates cassette exon events; star indicates intron retention events; triangle indicates alternative 5′ splice site events; and diamond indicates alternative 3′ splice site events. Alternatively spliced events with BF ≥ 5 are colored in red Fig. 2 Sashimi plots of four types of AS events. Sashimi plots of four types of AS events were shown, including cassette exon, intron retention, alternative 5′ and 3′ splice site. The red plots represent the LPS treated condition, and the blue ones represent controls. The X-axes indicate genomic locations, and the Y-axes indicate transcription intensity. In each plot, a “sashimi-like” region indicates a heavily transcribed region, in this case, exonic region. The blank regions between exonic regions indicate intronic regions. The “bridges” crossing exons indicate junction reads. The numbers of junction reads are shown on the “bridges”. The exonic structure of each AS event is shown below each Sashimi plot. On the right it displays the estimated Ψ (red line) value and the full posterior distribution (black bars) To validate whether the alternative splicing events were induced by LPS treatment, we performed RNA-seq on BMSCs derived from MyD88−/− animals before and after LPS treatment. MyD88 is a key signaling molecule responsible for LPS response [35]. Among the 197 LPS-induced alternative splicing events in wild-type BMSCs, 189 did not occur following LPS treatment of MyD88−/− BMSCs (Fig. 3). This observation indicates that a large majority of BMSC splicing changes were a direct consequence of LPS induction, and such effects were negated in cells whose LPS response is compromised. It should be noted that in addition to MyD88 pathways, LPS also functions through TRIF pathways [36]; the functions of TRIF pathway is intact in the MyD88−/− cells. This partially explains why some LPS-induced splicing effects remained in MyD88-deficient animals.Fig. 3 LPS-induced splicing changes in wild-type BMSC were repressed in MyD88−/− cells. The X-axis and Y-axis represents ∆PSI in wild type and MyD88 knock out animals respectively. Blue diamond represents LPS induced AS events in wild type only, and red square represents LPS induced AS events in both wild type and MyD88 knock-out cells Among the 197 LPS-induced alternative splicing events, 103 were located in the coding regions of transcripts, and 94 were either in the 5′- or 3′- untranslated regions (UTRs). Among the 103 alternatively spliced coding events, 65 were composed of multiples of three nucleotides, leading to the inclusion or exclusion of specific amino-acid residues in the final protein products. These events could potentially generate multiple viable protein products having the same translation frame. Thirty-eight of the 103 coding exons contained either a premature stop codon, and/or a shift in their translation frames. Such events trigger either nonsense-mediated decay (NMD) mechanisms [37], or a translated protein having a complete different amino acid sequence downstream of the alternatively spliced exon. We then systematically examined the localization and functions of the gene products possessing alternatively spliced exons (Fig. 4 and Additional file 2). Among them, 64 were nuclear proteins, including 17 transcription regulators, 13 enzymes, 2 kinases, 1 peptidase, and 1 ligand-dependent nuclear receptor. The 67 cytoplasmic alternatively spliced gene products included 12 enzymes, 9 kinases, 6 transporters, 3 peptidases, and 2 translation regulators. In addition, we also observed six potentially secreted proteins and 18 plasma membrane-spanning proteins. A detailed list of the genes in each category is provided in Additional file 3. These results strongly suggest that LPS induces splicing changes in highly diverse proteins having a variety of cellular functions.Fig. 4 Distribution of AS genes in different cellular locations To understand the biological functions of genes whose splicing patterns were altered by LPS treatment, we conducted functional annotation analysis using the Database for Annotation, Visualization and Integrated Discovery (DAVID) v6.7 [38]. Three functional terms in the SP_PIR (Swiss-Prot and Protein Information Resources) category showed significant enrichment in our gene list. Among the 161 genes that could be mapped to DAVID gene annotations, 97 categorized as phosphoproteins (p-value = 7.2×10−12, FDR = 6.6×10−10). In addition, 26 genes contained zinc finger domain proteins (p-value = 3.6×10−5, FDR = 2.2×10−3) whose functions range from DNA or RNA binding to protein-protein interactions and membrane association [39]. Furthermore, 35 genes were involved in protein acetylation (p-value = 1.3 ×10−3, FDR = 3.2×10−2). Together, these results suggest that LPS treatment has major effects on the splicing patterns of signaling proteins. Both gene expression levels and splicing patterns may be altered by BMSC responses to LPS treatment. While differential gene expression may lead to changes in the abundance of the entire gene product, alternative spicing modifies the structural composition of a specific protein. To evaluate to what extent the two mechanisms interact, we examined the number of genes present in both differentially expressed and alternatively spliced gene sets. We utilized edgeR [40] to identify genes differentially expressed between LPS-treated and control samples. In total, 416 differentially expressed genes were identified using a false discovery rate ≤ 0.05. Surprisingly, only one gene, Plscr2 (Phospholipid Scramblase 2) was both differentially expressed and alternatively spliced. The expression level of Plscr2 increased 1.77-fold in LPS-induced samples with FDR = 0.01, while the percentage of inclusion of one cassette exon in the 3′-untranslated region (3′-UTR) increased by 0.16. Known protein domains are alternatively spliced in LPS-induced transcripts Alternatively spliced exons residing in known protein domains are more likely to disrupt protein function. Therefore, we systematically searched the overlap between LPS-induced AS events for known protein family domains documented in the pfam database [41]. Among 65 alternatively spliced exons that did not disrupt codon frame, seven overlapped with known protein domains (Table 1). In addition, seven other known domains that overlapped flanking exons had functions ranging from RNA and protein binding, enzymatic activities, methyltransferase activity, phosphopantetheinyl transferase activity, RNA editing, and microRNA processing.Table 1 Alternatively spliced genes containing known protein domains Gene Symbol AS Type Pfam Domain Domain Description Ttc13 cassete exon TPR_11 TPR repeat Rabep1 cassete exon Rabaptin Rabaptin Camk1d cassete exon Pkinase Protein kinase domain Nr1h2 alternative 5′ splice site Hormone_recep Ligand-binding domain of nuclear hormone receptor Adarb1 alternative 5′ splice site A_deamin Adenosine-deaminase (editase) domain Scoc alternative 3′ splice site DUF2205 Predicted coiled-coil protein Ppip5k2 alternative 3′ splice site His_Phos_2 Histidine phosphatase superfamily (branch 2) Alternative splicing in known protein domains may affect protein-protein interactions To examine whether alternatively spliced protein domains modulate protein-protein interactions, we searched for their binding partners based on two criteria: (1) at least one experimental study supporting direct interaction between the partner protein and the alternatively spliced protein in a known protein-protein interaction network [42, 43]; and (2) at least one structural study in the Protein Data Bank (PDB) supporting direct interaction between a domain in the binding partner and the domain modified by alternative splicing. For the first criterion, we merged two datasets of experimentally validated direct interactions [42, 43] and compiled a library of 9,795 protein-coding genes with 80,518 experimentally validated interactions. For the second criterion, we derived the domain interactions in PDB from iPfam [41] and then searched for proteins containing these domains in Pfam [41]. In total, 3,573 interactions with structural evidence were found between 13 alternatively spliced coding transcripts and 3103 binding partners. By joining two interaction tables, we identified eight interactions having both experimental and structural evidence. As shown in Fig. 5, these eight interactions involved three genes with altered splicing domains, Rabep1 (Rab GTPase-binding effector protein 1), Camk1d (Calcium/Calmodulin-Dependent Protein Kinase 1D), and Nr1h2 (nuclear receptor subfamily 1, group H, member 2). The alternatively spliced exons in these genes overlapped with known protein domains, including rabaptin, pkinase, and ligand-binding domain of nuclear hormone receptor.Fig. 5 PPI with both structural and experimental evidences. Ten AS gene products involved in protein-protein interactions. Gene symbols are displayed in white regions, and corresponding protein domains are displayed with gray background. Blue line indicates a gene/protein contains a domain, and a red line indicates an interaction between protein domains The differences in the percentage of inclusion for these three events ranged from 14 to 31 %. The potential protein partners included Rabep1, Gga1 (Golgi-associated, gamma adaptin ear containing, ARF-binding protein 1), Gga2 (Golgi-associated, gamma adaptin ear containing, ARF binding protein 2), Gga3 (Golgi-associated, gamma adaptin ear containing, ARF binding protein 3), Camkk1 (calcium/calmodulin-dependent protein kinase kinase 1, alpha), Nr0b2 (nuclear receptor subfamily 0, group B, member 2), Rxra (retinoid X receptor, alpha), and Rxrb (retinoid X receptor, beta). LPS-induced splicing changes could significantly impact these proteins’ interactions with their partners. Among these putative protein interaction partners, only one protein, Nr0b2 (nuclear receptor subfamily 0, group B, member 2), was not expressed. Intrinsic disorder and molecular recognition features in LPS-induced alternative spliced regions It was previously reported that alternatively spliced regions are enriched with unfolded protein regions (intrinsic disorder) [44]. To examine these features within LPS-induced alternatively spliced regions (cassette exons, alternative 5’/3′ exons and retained introns), we performed disorder prediction on the protein sequences of these regions using VSL2B [45], a bioinformatics algorithm for predicting intrinsically disordered regions based on the biophysical properties of amino acids. Among the alternative regions of 65 protein sequences translated from LPS-induced alternative splicing events, 34 (52.3 %) were predicted to be totally disordered, 21 (32.3 %) partially disordered, and only 10 (15.3 %) totally structured (Fig. 6). These percentages are consistent with previous reports that alternatively spliced exons tend to locate in intrinsically disordered regions [46].Fig. 6 Predicted disorder of AS gene products. a The distribution of 65 non-frameshifting protein coding AS genes in three categories, including totally disordered, partially disordered and structured. b The MoRF containing and non-containing events among partially disordered AS genes. The list of MoRF containing genes is shown on the right. c The same distribution and gene list as panel B but it is for totally disordered AS genes A molecular recognition feature (MoRF) is a region in an RNA that undergoes a disorder–order transformation while bound by another protein. We predicted MoRF regions within the alternative regions using the software tool MoRF2 [44]. As a result, among the 55 alternatively spliced exons in the partial or totally disordered regions, 22 contained regions predicted to be MoRFs (Fig. 6, Additional file 4); these regions could thus be regarded as potential protein-protein interaction sites. Post-translational modification sites within alternatively spliced regions We next annotated post-translational modification (PTM) sites in regions affected by LPS-induced alternative splicing. We searched known PTM sites deposited in UniProt, and we also predicted novel ones using ModPred [47]. Three alternatively spliced exons containing known PTM (phorphorylation) sites localized to three genes, Abi1 (abl-interactor 1), Depdc1a (DEP Domain-Containing 1), and Ybx3 (Y box-binding protein 3). In addition, 13 PTMs were predicted to occur in 29 alternatively spliced regions, including proteolytic cleavage, phosphorylation, amidation, hydroxylation, carboxylation, ADP-ribosylation, O-linked glycosylation, acetylation, GPI anchor amidation, palmitoylation, pyrrolidone carboxylic acid, methylation and ubiquitination (Fig. 7). Proteolytic cleavage sites were the most common PTM sites, appearing in 14 alternative regions. It is possible that LPS affects the signaling activities of these proteins by inclusion or exclusion of the PTM sites in the final protein product (i.e., whether or not it is cleaved).Fig. 7 Predicted PTM sites in AS regions. Column displays different types of PTM sites, and row displays the event types and LPS-induced AS genes. The numbers in the shadowed grids on the crossing of gene A and PTM type B shows how many type B PTM sites fall in the AS region of gene A. The total number of PTM sites in each gene is displayed on the right, and the total number of PTM sites in each type is displayed on the top Characterization of potential splicing regulators We defined 7 regulatory regions for each cassette exon event (Fig. 8), among which Region 1 and 7 are 150 nt constitute exon segments, Region 2, 3, 5 and 4 are 300 nt intronic segments, and Region 4 is the whole cassette exon. We used FIMO [48] to search for CISBP-RNA [49] motifs within the regulatory regions of both up-regulated and down-regulated cassette exon events. With p-value cutoff of 1E-4 and FDR cutoff of 0.1, we identified 29 RBP motifs in the up-regulated events, and 23 in the down-regulated events. BRUNOL5, BRUNOL4 and RBM38 are the most frequently observed RBPs. Their motifs concentrate in Region 2 and 3 for up-regulated events, and in Region 5 for down-regulated events. These three proteins are all known as RNA-splicing related. Motifs of several other RNA-splicing related proteins, including SRSF2, HNRNPL, HNRNPLL, HNRNPH2 and PCBP2, are observed in both up-regulated and down-regulated cassette exon regulatory regions. Some RBPs (SRSF9, RBM5, PCBP3, PCBP1, ZCRB1, NCL, FUSIP1, PABPN1, TARDBP and NOVA2) are found exclusively in up-regulated cassette exon events, and some (KHDRBS3, BRUNOL6, G3BP2, FXR1, SRSF4, SNRNPA, SNRPB2) are found exclusively in down-regulated events.Fig. 8 RNA binding protein (RBP) motifs in regulatory regions of differentially spliced events. RBP names and their occurrences are listed adjacent to corresponding regulatory regions Discussion Lipopolysaccharide (LPS, endotoxin) is a complex associated with the outer membrane of Gram-negative bacteria, capable of triggering a series of cellular responses in many cell types. One promising advance is to use LPS as a pre-conditioning agent to improve BMSC therapeutic efficacy for repairing ischemic, injured tissues [23, 50]. For such application, because LPS is a potent stimulant for the host immune system, BMSCs should be washed using PBS to completely remove any residual endotoxin before administration. We reported previously that BMSCs treated with LPS produced more angiogenic factors VEGF, IGF-1 and HGF [51, 52] which can spur the formation of new blood vessels in ischemic tissue and survival and differentiation of implanted BMSCs. By contrast, with the growing incidence of sepsis, in which free LPS can bind to and activate Toll-like receptor 4 on many cell types, the roles of LPS on endogenous BMSCs and other cell types are worth detailed investigation. Microarray studies have reported that expression levels of hundreds of genes can be altered after LPS treatment in different tissues. In recent years, high-throughput RNA sequencing technology has provided a more accurate and comprehensive measurement of RNA transcript levels and their isoforms than historic array-based methods. This technological advance has enabled measuring not only gene expression level alterations amongst different conditions, but also complicated splicing pattern changes in response to specific cellular perturbations. In this study, we systematically identified alternative splicing changes in mouse bone marrow-derived mesenchymal stem cells (BMSCs) in response to LPS treatment, using RNA-seq technology. We further implemented a series of bioinformatics tools to evaluate the biological functions of alternatively spliced exons and their host genes. We observed strong enrichment in three functional categories amongst the gene products whose splicing patterns were altered by LPS treatment, phosphoproteins, zinc finger proteins, and proteins subject to acetylation. Most of these proteins were signaling proteins, and the subtle differences in their splicing isoforms could affect their function. Among 161 gene products containing AS exons, 97 belonged to phosphoprotein families, five of which contained documented phosphorylation sites in their AS regions found in the UniProt database. These proteins included Kansl2 (KAT8 regulatory NSL complex subunit 2), Depdc1a (DEP domain-containing 1), Abi1 (abl-interactor 1), Ybx3 (Y box-binding protein 3), and UBl4a (Slc10a3-Ubl4 readthrough). The functions of these proteins strongly associate with the functions of BMSCs. For instance, Abi1 contains one cassette of exons whose percentage of inclusion increased by 14 % after LPS induction (ΔΨ = 0.14), with one phosphorylation site in the AS region documented in the UniProt database. Widely expressed with highest levels in bone marrow, spleen, brain, testes, and embryonic brain, Abi1 may negatively regulate cell growth and transformation by interacting with the nonreceptor tyrosine kinases ABL1 and/or ABL2, thus regulating EGF-induced Erk pathway activation and EGFR signaling. In addition to these five proteins, eight other AS regions were predicted to have phosphorylation sites, based on their amino acid contents. These proteins included Usp45 (ubiquitin-specific peptidase 45), Mark3 (MAP/microtubule affinity-regulating kinase 3), Ncor1 (nuclear receptor corepressor 1), Ctnnd1 (cadherin-associated protein, beta 1), Ambra1 (autophagy/beclin-1 regulator 1), Ddx6 (DEAD (Asp-Glu-Ala-Asp) box helicase 6), Ehbp1l1 (EH domain binding protein 1-like 1), and Akt1s1 (AKT1 substrate 1). Overall, LPS may affect the functions of these proteins by including/excluding specific domains amenable to phosphorylation. Among the proteins containing LPS-induced alternative splicing events, 25 contained multiple types of zinc finger domains, including PHD (Plant Homeo Domain), RING (Really Interesting New Gene), and C2H2-type zinc-finger domains. Four proteins, Phf7 (PHD finger protein 7), Phf20 (PHD finger protein 20), Phf20l1 (PHD finger protein 20-like 1), and Phrf1 (PHD and ring finger domains 1), contained PHD-type zinc finger domains known to recognize trimethylated histone lysines (thus possibly influencing chromatin structure). Four other proteins, Rnf14 (ring finger protein 14), Rad18 (RAD18 homolog), Trim28 (tripartite motif-containing 28), and Trim2 (tripartite motif-containing 2), all contain RING-type zinc fingers, known ligases for ubiquitination enzymes and their substrates. It is well documented that both PHD and RING-type domains are usually involved in protein-protein binding [53, 54], and such binding could possibly be disrupted by splicing variations. Overall, the LPS-induced AS genes could be classified into several categories (Fig. 9), including kinases, zinc-finger proteins, transcription, RNA-binding, cytoskeleton, and protein acetylation. Many of these proteins were also phosphoproteins, which play significant roles in cell signaling. Analysis of the relationship between splicing and protein structure has suggested that AS exons play major roles in controlling protein-protein interactions (PPIs) through disrupting either known protein interaction domains or molecular recognition sites, which typically locate in intrinsically disordered regions. Our analysis suggests that LPS-induced alternative splicing could affect PPIs through both mechanisms. In particular, protein interaction domains of three proteins with known PPI partners were disrupted by LPS-induced splicing alterations (Fig. 5). Interestingly, all three interactive domains could self-interact (forming domain-domain interactions with themselves), and one of these domains facilitates homodimerization of Rabep1 (RAB GTPase binding effector protein 1). Expressed in embryonic tissues and most types of stem cells, Rabep1 showed abundant expression in BMSCs (about 30 RPKM). Homo-dimerization of this protein is involved in early endosome fusion [55], an event directly related to the paracrine effects of BMSCs, where small vesicles are released when multivesicular endosomes fuse with the plasma membrane [56, 57]. In addition, Rabep1 also moderates intracellular transportation between lysosomes and the Golgi apparatus [58], and between the Golgi apparatus and endoplasmic reticulum [59]. LPS treatment also increased the inclusion of the interaction domain by 14 %, which could increase either homodimerization or heterodimerization with other interaction partners.Fig. 9 Predicted interaction network among LPS-induced AS genes. Red nodes indicate genes producing phosphoproteins, and gray nodes indicate genes not involved in protein phosphorylation. Genes associated with terms other than phosphoproteins are clustered in corresponding shadowed areas. These terms include acetylation, cytoskeleton, transcription, zinc-finger, RNA-binding and kinase We further evaluated how differences in splicing patterns in transcriptional regulators affected their regulatory activity by assessing gene expression changes of their downstream target genes. NFYA (nuclear transcriptional factor Y) contains an alternative acceptor site whose splicing pattern in BMSCs is altered by LPS treatment; the overall percentage of inclusion of the alternative acceptor site decreased by 31 % (Sashimi plot for NFYA shown in Additional file 5). Moreover, the expression of five downstream target genes of NFYA were enriched for genes found differentially expressed (p-value ≤ 0.01) by LPS treatment (FDR ≤ 0.05), including COL11A1 (collagen, type XI, alpha 1), COL5A3 (collagen, type V, alpha 3), FGFR2 (Fibroblast Growth Factor Receptor 2), PGK1 (phosphoglycerate kinase 1) and RGS4 (regulator of G-protein signaling 4). It was previously reported that NFYA activates transcription levels of COL11A1 and FGFR2 [60]; these two genes were both downregulated by LPS, suggesting inhibition of NFYA function by the removal of 18 nt (or 6 amino acids) after LPS treatment, thus impacting NFYA downstream effectors. Conclusion In summary, we used RNA sequencing to analyze LPS-induced alternative splicing changes in BMSCs. LPS modified the alternative splicing pattern of phosphoproteins, zinc finger proteins, and proteins subject to acetylation. Most of the affected proteins were signaling proteins that could change BMSC biological function. Although it is still to be determined whether such modifications underlie BMSC therapeutic efficacy, our characterizations provide greater understanding of the mechanisms and clinical usage of promising BMSC therapies. Methods Preparation of mouse BMSCs A single-step stem cell purification method was employed as previously described [61]. Briefly, BMSCs were collected from the bilateral femurs and tibias of sacrificed mice by removing the epiphyses and flushing the shaft with complete media, Iscove’s Modified Dulbecco’s Medium (IMDM; Life Technologies) and 10 % fetal bovine serum (Life Technologies), using a syringe with a 26G needle. Cells were disaggregated by vigorous pipetting and passed through a 30-μm nylon mesh to remove any remaining clumps of tissue. Cells were then centrifuged for 5 min at 500 g at 24 °C. The cell pellet was then resuspended and cultured in 75 cm2 culture flasks in complete media at 37 °C with 5 % CO2. Since BMSCs preferentially attach to polystyrene [62], after 48 h, floating non-adherent cells were discarded. Fresh complete media was added and replaced every three or four days thereafter. When the cells reached 90 % confluence, MSC cultures were recovered by the addition of a solution of 0.25 % trypsin-EDTA (Invitrogen) and passaged. Cell passage was restricted to passages 6–10 for the experiments. To purify BMSCs, the cells were subject to fluorescence-activated cell sorting (FACS) analysis, with collection of cells positive for Sca-1 and CD44 [62], but negative for the hematopoietic stem cell and macrophage marker CD45 [25]. RNA sample preparation and RNA-seq assay BMSCs were plated at 1 × 105 cells/well/ml for 24 h and further treated with LPS (200 ng/ml) for another 24 h, and total RNA was extracted before and after LPS treatment, following a standard protocol [25]. Experiments were conducted in triplicate. Standard methods were used for RNA-seq library construction, EZBead preparation, and Next-Gen sequencing, based on the Life Technologies SOLiD 5500xl system. Briefly, 2 μg of total RNA per sample was used for library preparation. The rRNA was first depleted using the standard protocol of RiboMinus Eukaryote Kit for RNA-Seq (Ambion), and rRNA-depleted RNA was concentrated using a PureLink RNA Micro Kit (Invitrogen) with 1 volume of lysis buffer and 2.5 volumes of 100 % ethanol. After rRNA depletion, a whole transcriptome library was prepared and barcoded per sample using the standard protocol of SOLiD Total RNA-seq Kit (Life Technologies). Each barcoded library was quantified by quantitative polymerase chain reaction (qPCR) using SOLiD Library Taqman qPCR Module (Life Technologies) and pooled in equal molarity. EZBead preparation, bead library amplification, and bead enrichment were then conducted using the Life Technologies EZ Bead E80 System. Finally sequencing by ligation was performed using a standard single-read, 5′-3′ strand-specific sequencing procedure (75 nt-read) on SOLiD 5500xl. Bioinformatics analysis for RNA-seq data RNA-seq data analysis included the following steps: quality assessment, sequence alignment, and alternative splicing analysis. The RNA-seq data can be accessed through the Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/) with accession number GSE64568). Data processing and quality assessment We used SOLiD Instrument Control Software and SOLiD Experiment Tracking System software for read quality recalibration. Each sequence read was scanned for low-quality regions, and if a 5-base sliding window had an average quality score less than 20, the read was truncated at that position. Any read < 35 bases was discarded. Our experience suggests that this strategy effectively eliminates low-quality reads, while retaining high-quality regions [63–65]. Sequence alignment We used BFAST (http://sourceforge.net/projects/bfast/) [66] as our primary alignment algorithm due to its high sensitivity for aligning reads on loci containing small insertions and deletions, as compared to the reference genome (mm9). We then used a TopHat-like strategy [67] to align the sequencing reads containing cross-splicing junctions using NGSUtils (http://ngsutils.org/) [63]. After aligning the reads to a filtering index including repeats, ribosome RNA, and other sequences that were not of interest, we conducted a sequence alignment at three levels: genome, known junctions (University of California Santa Cruz Genome Browser), and novel junctions (based on the enriched regions identified in the genomic alignment). We restricted our analysis to uniquely aligned sequences with no more than two mismatches. Alternative splicing analysis We used MISO (mixture of isoforms) [34] to identify alternatively spliced exons whose splicing patterns were altered after LPS treatment. We first used Samtools (v0.1.19) to merge six RNA-seq samples into two BAM files according to their biological conditions, i.e., control vs. LPS-treated samples. We then estimated Percent Spliced In (PSI or Ψ), which indicates the proportion of RNA isoforms containing the alternatively spliced exon (inclusive isoforms) among all isoforms (inclusive plus exclusive isoforms). We also computed a Bayes factor (BF) to describe the likelihood of an AS event between the LPS-treated and control conditions. A BF of 5 means that an AS event is 5 times more likely to be differentially spliced than not. Both Ψ and BF values were computed by the software package MISO [34]. The difference between Ψ s across the two conditions was defined as ΔΨ. We required each AS event to have a BF > 5 and |ΔΨ| > 0.05 to be considered differentially spliced. Ontological annotations The functions and cellular locations of AS genes were annotated by the pathway analysis tool Ingenuity Pathway Analysis (IPA), and the functional and biochemical properties of these genes were further annotated based on SwissProt and PIR keywords with DAVID v6.7 [38]. Protein domains overlapping AS regions Protein domain information was predicted based on the RNA nucleotide sequences of the alternatively spliced exon, and 30-base flanking sequencings of both upstream and downstream exons. These RNA sequences were then translated into peptides, based on open reading frames (ORFs) documented by Ensembl and Refseq, which were then input into Pfam [41] for identification of protein domains overlapping AS regions. Identification of protein-protein interactions (PPI) We also examined whether alternatively spliced exons overlapped with potential protein-protein interaction domains. Based on the protein domains identified in or overlapping AS regions, we retrieved their binding partner domains with iPfam [41], which documents domain-domain interactions in the Protein Data Bank (PDB). We further used Pfam to search for genes encoding partner domains (i.e., potential protein interaction partners). The identified protein interaction partners were verified by two protein-protein interaction databases derived from high-throughput experiments. Other characterizations Protein disorder was predicted with VSL2B [45], a highly regarded protein disorder prediction tool, especially for long regions of disorder [68]. We required the peptides flanking the AS regions to be at least 9 amino acids long for accurate prediction. Potential binding sites were predicted with MoRF2, a software tool that predicts protein-binding sites that undergo a disorder–order transformation while binding another protein molecule [69]. Known post-translational modification (PTM) sites were derived from UniProt, and novel PTM sites were predicted by ModPred [47]. The upstream gene regulator NFYA (Nuclear transcription Factor Y subunit Alpha) [70] was predicted by Ingenuity Pathway Analysis (IPA), based on gene expression data and known regulatory gene interactions. Abbreviations AS, Alternative splicing; BF, Bayesian factor; BMSCs, Bone marrow stromal cells; LPS, Lipopolysaccharide; NMD, Nonsense-mediated decay; PSI, or Ψ, Percent spliced in; PTM, Post-translational modification; PPIs, Protein-protein interactions; UTRs, Untranslated regions Additional files Additional file 1: Statistics of the RNA sequencing experiment. (DOCX 13 kb) Additional file 2: Functions and cellular locations of AS genes. (DOCX 12 kb) Additional file 3: The function and localization of alternatively spliced genes. (DOCX 23 kb) Additional file 4: Alternatively spliced genes containing Molecular Recognition Features (MoRF). (DOCX 12 kb) Additional file 5: Sashimi plot of NFYA. (PDF 121 kb) Acknowledgement The RNA-seq was conducted in the Center for Medical Genomics at Indiana University School of Medicine. The data was analyzed in the Bioinformatics Core at Indiana University School of Medicine. This work was supported in part by the Medical and Molecular Genetics, Indiana University School of Medicine Startup Funds, Showalter Trust Award and by the Indiana Clinical and Translational Sciences Institute, funded in part by grant # TR 000006 from the National Institutes of Health, National Center for Advancing Translational Sciences, Clinical and Translational Sciences Award. Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The RNA-seq data can be accessed through the Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/) with accession number GSE64568). Authors’ contributions AZ, YL and YW designed the study, conducted the analysis, and drafted the manuscript. ML contributed in functional analysis of alternative splicing events. BH, FH and AKD assisted in identification of MoRF in alternatively spliced regions. WF, BX, CB, and BL participated design of the study, and drafting of the manuscript. All the authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Chen L Bush SJ Tovar-Corona JM Castillo-Morales A Urrutia AO Correcting for differential transcript coverage reveals a strong relationship between alternative splicing and organism complexity Mol Biol Evol 2014 31 6 1402 1413 10.1093/molbev/msu083 24682283 2. Stamm S Ben-Ari S Rafalska I Tang Y Zhang Z Toiber D Thanaraj TA Soreq H Function of alternative splicing Gene 2005 344 1 20 10.1016/j.gene.2004.10.022 15656968 3. Chen L Bush SJ Tovar-Corona JM Castillo-Morales A Urrutia AO Correcting for differential transcript coverage reveals a strong relationship between alternative splicing and organism complexity Mol Biol Evol 2014 31 1402 13 10.1093/molbev/msu083 24682283 4. Consortium EP An integrated encyclopedia of DNA elements in the human genome Nature 2012 489 7414 57 74 10.1038/nature11247 22955616 5. Diamond RH Du K Lee VM Mohn KL Haber BA Tewari DS Taub R Novel delayed-early and highly insulin-induced growth response genes. Identification of HRS, a potential regulator of alternative pre-mRNA splicing J Biol Chem 1993 268 20 15185 15192 7686911 6. Du K Peng Y Greenbaum LE Haber BA Taub R HRS/SRp40-mediated inclusion of the fibronectin EIIIB exon, a possible cause of increased EIIIB expression in proliferating liver Mol Cell Biol 1997 17 7 4096 4104 10.1128/MCB.17.7.4096 9199345 7. Huang R Xing Z Luan Z Wu T Wu X Hu G A specific splicing variant of SVH, a novel human armadillo repeat protein, is up-regulated in hepatocellular carcinomas Cancer Res 2003 63 13 3775 3782 12839973 8. Kanai Y Saito Y Ushijima S Hirohashi S Alterations in gene expression associated with the overexpression of a splice variant of DNA methyltransferase 3b, DNMT3b4, during human hepatocarcinogenesis J Cancer Res Clin Oncol 2004 130 11 636 644 10.1007/s00432-004-0586-3 15490234 9. Kurahashi H Takami K Oue T Kusafuka T Okada A Tawa A Okada S Nishisho I Biallelic inactivation of the APC gene in hepatoblastoma Cancer Res 1995 55 21 5007 5011 7585543 10. Saito Y Kanai Y Sakamoto M Saito H Ishii H Hirohashi S Overexpression of a splice variant of DNA methyltransferase 3b, DNMT3b4, associated with DNA hypomethylation on pericentromeric satellite regions during human hepatocarcinogenesis Proc Natl Acad Sci U S A 2002 99 15 10060 10065 10.1073/pnas.152121799 12110732 11. Pevsner-Fischer M Morad V Cohen-Sfady M Rousso-Noori L Zanin-Zhorov A Cohen S Cohen IR Zipori D Toll-like receptors and their ligands control mesenchymal stem cell functions Blood 2007 109 4 1422 1432 10.1182/blood-2006-06-028704 17038530 12. Gnecchi M Danieli P Cervio E Mesenchymal stem cell therapy for heart disease Vasc Pharmacol 2012 57 1 48 55 10.1016/j.vph.2012.04.002 13. van den Akker F de Jager SC Sluijter JP Mesenchymal stem cell therapy for cardiac inflammation: immunomodulatory properties and the influence of toll-like receptors Mediat Inflamm 2013 2013 181020 14. Lu X Liu T Gu L Huang C Zhu H Meng W Xi Y Li S Liu Y Immunomodulatory effects of mesenchymal stem cells involved in favoring type 2 T cell subsets Transpl Immunol 2009 22 1–2 55 61 10.1016/j.trim.2009.08.002 19695330 15. Gnecchi M He H Liang OD Melo LG Morello F Mu H Noiseux N Zhang L Pratt RE Ingwall JS Paracrine action accounts for marked protection of ischemic heart by Akt-modified mesenchymal stem cells Nat Med 2005 11 4 367 368 10.1038/nm0405-367 15812508 16. Gnecchi M He H Noiseux N Liang OD Zhang L Morello F Mu H Melo LG Pratt RE Ingwall JS Evidence supporting paracrine hypothesis for Akt-modified mesenchymal stem cell-mediated cardiac protection and functional improvement Faseb J 2006 20 6 661 669 10.1096/fj.05-5211com 16581974 17. Kocher AA Schuster MD Szabolcs MJ Takuma S Burkhoff D Wang J Homma S Edwards NM Itescu S Neovascularization of ischemic myocardium by human bone-marrow-derived angioblasts prevents cardiomyocyte apoptosis, reduces remodeling and improves cardiac function Nat Med 2001 7 4 430 436 10.1038/86498 11283669 18. Steele A Steele P Stem cells for repair of the heart Curr Opin Pediatr 2006 18 5 518 523 10.1097/01.mop.0000245352.75830.1d 16969166 19. Tomasoni S Longaretti L Rota C Morigi M Conti S Gotti E Capelli C Introna M Remuzzi G Benigni A Transfer of growth factor receptor mRNA via exosomes unravels the regenerative effect of mesenchymal stem cells Stem Cells Dev 2013 22 5 772 780 10.1089/scd.2012.0266 23082760 20. Abdel-Latif A Bolli R Tleyjeh IM Montori VM Perin EC Hornung CA Zuba-Surma EK Al-Mallah M Dawn B Adult bone marrow-derived cells for cardiac repair: a systematic review and meta-analysis Arch Intern Med 2007 167 10 989 997 10.1001/archinte.167.10.989 17533201 21. Dai W Hale SL Martin BJ Kuang JQ Dow JS Wold LE Kloner RA Allogeneic mesenchymal stem cell transplantation in postinfarcted rat myocardium: short- and long-term effects Circulation 2005 112 2 214 223 10.1161/CIRCULATIONAHA.104.527937 15998673 22. Gnecchi M Zhang Z Ni A Dzau VJ Paracrine mechanisms in adult stem cell signaling and therapy Circ Res 2008 103 11 1204 1219 10.1161/CIRCRESAHA.108.176826 19028920 23. Yao Y Zhang F Wang L Zhang G Wang Z Chen J Gao X Lipopolysaccharide preconditioning enhances the efficacy of mesenchymal stem cells transplantation in a rat model of acute myocardial infarction J Biomed Sci 2009 16 74 10.1186/1423-0127-16-74 19691857 24. Herrmann JL Wang Y Abarbanell AM Weil BR Tan J Meldrum DR Preconditioning mesenchymal stem cells with transforming growth factor-alpha improves mesenchymal stem cell-mediated cardioprotection Shock 2010 33 1 24 30 10.1097/SHK.0b013e3181b7d137 19996917 25. Wang Y Abarbanell AM Herrmann JL Weil BR Manukyan MC Poynter JA Meldrum DR TLR4 inhibits mesenchymal stem cell (MSC) STAT3 activation and thereby exerts deleterious effects on MSC-mediated cardioprotection PLoS One 2010 5 12 e14206 10.1371/journal.pone.0014206 21151968 26. Ma A Jiang L Song L Hu Y Dun H Daloze P Yu Y Jiang J Zafarullah M Chen H Reconstruction of cartilage with clonal mesenchymal stem cell-acellular dermal matrix in cartilage defect model in nonhuman primates Int Immunopharmacol 2013 16 3 399 408 10.1016/j.intimp.2013.02.005 23499511 27. El-Badri N Ghoneim MA Mesenchymal stem cell therapy in diabetes mellitus: progress and challenges J Nucleic Acids 2013 2013 194858 10.1155/2013/194858 23762531 28. Wang LQ Lin ZZ Zhang HX Shao B Xiao L Jiang HG Zhuge QC Xie LK Wang B Su DM Timing and dose regimens of marrow mesenchymal stem cell transplantation affect the outcomes and neuroinflammatory response after ischemic stroke CNS Neurosci Ther 2014 20 4 317 326 10.1111/cns.12216 24393245 29. Uccelli A Moretta L Pistoia V Mesenchymal stem cells in health and disease Nat Rev Immunol 2008 8 9 726 736 10.1038/nri2395 19172693 30. Kurtzberg J Prockop S Teira P Bittencourt H Lewis V Chan KW Horn B Yu L Talano JA Nemecek E Allogeneic human mesenchymal stem cell therapy (remestemcel-L, Prochymal) as a rescue agent for severe refractory acute graft-versus-host disease in pediatric patients Biol Blood Marrow Transplant 2014 20 2 229 235 10.1016/j.bbmt.2013.11.001 24216185 31. Dalal J Gandy K Domen J Role of mesenchymal stem cell therapy in Crohn’s disease Pediatr Res 2012 71 4 Pt 2 445 451 10.1038/pr.2011.56 22430380 32. Martin GS Sepsis, severe sepsis and septic shock: changes in incidence, pathogens and outcomes Expert Rev Anti-Infect Ther 2012 10 6 701 706 10.1586/eri.12.50 22734959 33. Balk RA Severe sepsis and septic shock. Definitions, epidemiology, and clinical manifestations Crit Care Clin 2000 16 2 179 192 10.1016/S0749-0704(05)70106-8 10768078 34. Katz Y Wang ET Airoldi EM Burge CB Analysis and design of RNA sequencing experiments for identifying isoform regulation Nat Methods 2010 7 12 1009 1015 10.1038/nmeth.1528 21057496 35. Rhee SH Hwang D Murine TOLL-like receptor 4 confers lipopolysaccharide responsiveness as determined by activation of NF kappa B and expression of the inducible cyclooxygenase J Biol Chem 2000 275 44 34035 34040 10.1074/jbc.M007386200 10952994 36. Fitzgerald KA Rowe DC Barnes BJ Caffrey DR Visintin A Latz E Monks B Pitha PM Golenbock DT LPS-TLR4 signaling to IRF-3/7 and NF-kappaB involves the toll adapters TRAM and TRIF J Exp Med 2003 198 7 1043 1055 10.1084/jem.20031023 14517278 37. Chang YF Imam JS Wilkinson MF The nonsense-mediated decay RNA surveillance pathway Annu Rev Biochem 2007 76 51 74 10.1146/annurev.biochem.76.050106.093909 17352659 38. da Huang W Sherman BT Lempicki RA Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources Nat Protoc 2009 4 1 44 57 10.1038/nprot.2008.211 19131956 39. Laity JH Lee BM Wright PE Zinc finger proteins: new insights into structural and functional diversity Curr Opin Struct Biol 2001 11 1 39 46 10.1016/S0959-440X(00)00167-6 11179890 40. Robinson MD McCarthy DJ Smyth GK EdgeR: a bioconductor package for differential expression analysis of digital gene expression data Bioinformatics 2010 26 1 139 140 10.1093/bioinformatics/btp616 19910308 41. Finn RD Miller BL Clements J Bateman A IPfam: a database of protein family and domain interactions found in the protein data bank Nucleic Acids Res 2014 42 Database issue D364 373 10.1093/nar/gkt1210 24297255 42. Bossi A Lehner B Tissue specificity and the human protein interaction network Mol Syst Biol 2009 5 260 10.1038/msb.2009.17 19357639 43. Wang X Wei X Thijssen B Das J Lipkin SM Yu H Three-dimensional reconstruction of protein networks provides insight into human genetic disease Nat Biotechnol 2012 30 2 159 164 10.1038/nbt.2106 22252508 44. Romero PR Zaidi S Fang YY Uversky VN Radivojac P Oldfield CJ Cortese MS Sickmeier M LeGall T Obradovic Z Alternative splicing in concert with protein intrinsic disorder enables increased functional diversity in multicellular organisms Proc Natl Acad Sci U S A 2006 103 22 8390 8395 10.1073/pnas.0507916103 16717195 45. Peng K Radivojac P Vucetic S Dunker AK Obradovic Z Length-dependent prediction of protein intrinsic disorder BMC Bioinformatics 2006 7 208 10.1186/1471-2105-7-208 16618368 46. Buljan M Chalancon G Dunker AK Bateman A Balaji S Fuxreiter M Babu MM Alternative splicing of intrinsically disordered regions and rewiring of protein interactions Curr Opin Struct Biol 2013 23 3 443 450 10.1016/j.sbi.2013.03.006 23706950 47. Pejaver V Hsu WL Xin FX Dunker AK Uversky VN Radivojac P The structural and functional signatures of proteins that undergo multiple events of post-translational modification Protein Sci 2014 23 8 1077 1093 10.1002/pro.2494 24888500 48. Grant CE Bailey TL Noble WS FIMO: scanning for occurrences of a given motif Bioinformatics 2011 27 7 1017 1018 10.1093/bioinformatics/btr064 21330290 49. Ray D Kazan H Cook KB Weirauch MT Najafabadi HS Li X Gueroussov S Albu M Zheng H Yang A A compendium of RNA-binding motifs for decoding gene regulation Nature 2013 499 7457 172 177 10.1038/nature12311 23846655 50. Wang ZJ Zhang FM Wang LS Yao YW Zhao Q Gao X Lipopolysaccharides can protect mesenchymal stem cells (MSCs) from oxidative stress-induced apoptosis and enhance proliferation of MSCs via Toll-like receptor (TLR)-4 and PI3K/Akt Cell Biol Int 2009 33 6 665 674 10.1016/j.cellbi.2009.03.006 19376254 51. Crisostomo PR Wang Y Markel TA Wang M Lahm T Meldrum DR Human mesenchymal stem cells stimulated by TNF-alpha, LPS, or hypoxia produce growth factors by an NF kappa B- but not JNK-dependent mechanism Am J Physiol Cell Physiol 2008 294 3 C675 682 10.1152/ajpcell.00437.2007 18234850 52. Wang M Tan J Wang Y Meldrum KK Dinarello CA Meldrum DR IL-18 binding protein-expressing mesenchymal stem cells improve myocardial protection after ischemia or infarction Proc Natl Acad Sci U S A 2009 106 41 17499 17504 10.1073/pnas.0908924106 19805173 53. Pawson T Nash P Assembly of cell regulatory systems through protein interaction domains Science 2003 300 5618 445 452 10.1126/science.1083653 12702867 54. Pena PV Davrazou F Shi X Walter KL Verkhusha VV Gozani O Zhao R Kutateladze TG Molecular mechanism of histone H3K4me3 recognition by plant homeodomain of ING2 Nature 2006 442 7098 100 103 16728977 55. Stenmark H Vitale G Ullrich O Zerial M Rabaptin-5 is a direct effector of the small GTPase Rab5 in endocytic membrane fusion Cell 1995 83 3 423 432 10.1016/0092-8674(95)90120-5 8521472 56. Harding C Heuser J Stahl P Receptor-mediated endocytosis of transferrin and recycling of the transferrin receptor in rat reticulocytes J Cell Biol 1983 97 2 329 339 10.1083/jcb.97.2.329 6309857 57. Johnstone RM Adam M Hammond JR Orr L Turbide C Vesicle formation during reticulocyte maturation. Association of plasma membrane activities with released vesicles (exosomes) J Biol Chem 1987 262 19 9412 9420 3597417 58. Hirst J Lui WW Bright NA Totty N Seaman MN Robinson MS A family of proteins with gamma-adaptin and VHS domains that facilitate trafficking between the trans-Golgi network and the vacuole/lysosome J Cell Biol 2000 149 1 67 80 10.1083/jcb.149.1.67 10747088 59. Valsdottir R Hashimoto H Ashman K Koda T Storrie B Nilsson T Identification of rabaptin-5, rabex-5, and GM130 as putative effectors of rab33b, a regulator of retrograde traffic between the Golgi apparatus and ER FEBS Lett 2001 508 2 201 209 10.1016/S0014-5793(01)02993-3 11718716 60. Pallai R Simpkins H Chen J Parekh HK The CCAAT box binding transcription factor, nuclear factor-Y (NF-Y) regulates transcription of human aldo-keto reductase 1C1 (AKR1C1) gene Gene 2010 459 1–2 11 23 10.1016/j.gene.2010.03.006 20338228 61. Peister A Mellad JA Larson BL Hall BM Gibson LF Prockop DJ Adult stem cells from bone marrow (MSCs) isolated from different strains of inbred mice vary in surface epitopes, rates of proliferation, and differentiation potential Blood 2004 103 5 1662 1668 10.1182/blood-2003-09-3070 14592819 62. Secunda R Vennila R Mohanashankar AM Rajasundari M Jeswanth S Surendran R Isolation, expansion and characterisation of mesenchymal stem cells from human bone marrow, adipose tissue, umbilical cord blood and matrix: a comparative study Cytotechnology 2014 67 793 807 10.1007/s10616-014-9718-z 24798808 63. Breese MR Liu Y NGSUtils: a software suite for analyzing and manipulating next-generation sequencing datasets Bioinformatics 2013 29 4 494 496 10.1093/bioinformatics/bts731 23314324 64. Juan L, Wang G, Radovich M, Schneider BP, Clare SE, Wang Y, Liu Y. Potential roles of microRNAs in regulating long intergenic noncoding RNAs. BMC Med Genomics 2013;6 Suppl 1:S7. https://bmcmedgenomics.biomedcentral.com/articles/10.1186/1755-8794-6-S1-S7. 65. Todd AG Lin H Ebert AD Liu Y Androphy EJ COPI transport complexes bind to specific RNAs in neuronal cells Hum Mol Genet 2013 22 4 729 736 10.1093/hmg/dds480 23175440 66. Homer N Merriman B Nelson SF BFAST: an alignment tool for large scale genome resequencing PLoS One 2009 4 11 10.1371/journal.pone.0007767 19907642 67. Trapnell C Pachter L Salzberg SL TopHat: discovering splice junctions with RNA-Seq Bioinformatics 2009 25 9 1105 1111 10.1093/bioinformatics/btp120 19289445 68. Peng ZL Kurgan L Comprehensive comparative assessment of in-silico predictors of disordered regions Curr Protein Pept Sci 2012 13 1 6 18 10.2174/138920312799277938 22044149 69. Cheng YG Oldfield CJ Meng JW Romero P Uversky VN Dunker AK Mining alpha-helix-forming molecular recognition features with cross species sequence alignments Biochemistry 2007 46 47 13468 13477 10.1021/bi7012273 17973494 70. Zhu J Zhang Y Joe GJ Pompetti R Emerson SG NF-Ya activates multiple hematopoietic stem cell (HSC) regulatory genes and promotes HSC self-renewal Proc Natl Acad Sci U S A 2005 102 33 11728 11733 10.1073/pnas.0503405102 16081537

PMC005xxxxxx/PMC5001230.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27535125279510.1186/s12864-016-2795-yResearchPredicting diabetes mellitus genes via protein-protein interaction and protein subcellular localization information Tang Xiwei tangxiwei2010@gmail.com 123Hu Xiaohua xh29@drexel.edu 26Yang Xuejun 3Fan Yetian 4Li Yongfan 1Hu Wei 1Liao Yongzhong 1Zheng Ming cai 1Peng Wei 5Gao Li 61 School of Information Science and Engineering, Hunan First Normal University, Changsha, 410205 China 2 College of Computing and Informatics, Drexel University, Philadelphia, PA 19104 USA 3 College of Computer, National University of Defense Technology, Changsha, 410073 China 4 School of Mathematical Sciences, Dalian University of Technology, Dalian, 116023 China 5 Computer Center, Kunming University of Science and Technology, Kunming, 650500 China 6 School of Computer, Central China Normal University, Hubei, 430079 China 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.433© Tang et al. 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Diabetes mellitus characterized by hyperglycemia as a result of insufficient production of or reduced sensitivity to insulin poses a growing threat to the health of people. It is a heterogeneous disorder with multiple etiologies consisting of type 1 diabetes, type 2 diabetes, gestational diabetes and so on. Diabetes-associated protein/gene prediction is a key step to understand the cellular mechanisms related to diabetes mellitus. Compared with experimental methods, computational predictions of candidate proteins/genes are cheaper and more effortless. Protein-protein interaction (PPI) data produced by the high-throughput technology have been used to prioritize candidate disease genes/proteins. However, the false interactions in the PPI data seriously hurt computational methods performance. In order to address that particular question, new methods are developed to identify candidate disease genes/proteins via integrating biological data from other sources. Results In this study, a new framework called PDMG is proposed to predict candidate disease genes/proteins. First, the weighted networks are building in terms of the combination of the subcellular localization information and PPI data. To form the weighted networks, the importance of each compartment is evaluated based on the number of interacted proteins in this compartment. This is because the very different roles played by different compartments in cell activities. Besides, some compartments are more important than others. Based on the evaluated compartments, the interactions between proteins are scored and the weighted PPI networks are constructed. Second, the known disease genes are extracted from OMIM database as the seed genes to expand disease-specific networks based on the weighted networks. Third, the weighted values between a protein and its neighbors in the disease-related networks are added together and the sum is as the score of the protein. Last but not least, the proteins are ranked based on descending order of their scores. The candidate proteins in the top are considered to be associated with the diseases and are potential disease-related proteins. Various types of data, such as type 2 diabetes-associated genes, subcellular localizations and protein interactions, are used to test PDMG method. Conclusions The results show that the proteins/genes functionally exerting a direct influence over diabetes are consistently placed at the head of the queue. PDMG expands and ranks 445 candidate proteins from the seed set including original 27 type 2 diabetes proteins. Out of the top 27 proteins, 14 proteins are the real type 2 diabetes proteins. The literature extracted from the PubMed database has proved that, out of 13 novel proteins, 8 proteins are associated with diabetes. IEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Diabetes mellitus (often also known as diabetes) is a set of metabolic disorders. The latest data from World Health Organization (WHO) (http://www.who.int/diabetes/en/) shows that 9 % of adults worldwide are affected with diabetes. In 2012, 1.5 million people died of the disease. WHO points out that diabetes will become the No. 7 pestilence of threatening the human survival in 2030. It is estimated that America spent $245 billion treating diabetes in 2012 (http://www.diabetes.org/). Among these costs, $176 billion is directly allocated for medical expenditures, while the remaining funding is used for increasing productivity. Therefore, Diabetes mellitus has evoked great concern in the public health. In diabetes mellitus, blood sugar levels cannot be reasonably adjusted by the body [1]. For a person with diabetes, the pancreas fails to make sufficient insulin, improperly uses the insulin, or both. In the fast-flowing blood, insulin and glucose work together. The former helps the laster to come into cells of the body and produce energy. Sugar is unable to enter the cells if the insulin does not function properly. This results in the amount of glucose in the blood to go steadily up until generating the high concentration of blood sugar, and causing the cells in the lack of fuel. Typically, diabetes can be classified into three categories: type 1 diabetes, type 2 diabetes (T2D) and gestational diabetes. When beta cells in the pancreas are destroyed and unable to produce, store, and release the hormone insulin, type 1 diabetes(formerly known as insulin-dependent) occurs [2]. In people with type 1 diabetes, the levels of the blood sugar have not been properly controlled due to the deficient insulin production. The patients with type 1 diabetes often have to regularly inject insulin which help to control their blood sugar. In type 2 diabetes (referred to as non-insulin-dependent), beta cells are able to secrete enough insulin but the body cannot use the insulin effectively and attempts to compensate by making a higher quantity of insulin [3], causing insulin resistance. The production of hepatic glucose cannot be suppressed because of hepatic insulin resistance, and the ability to absorb peripheral glucose is impaired by peripheral insulin resistance. The two factors lead to fasting and postprandial hyperglycemia. The report by World Health Organization (WHO) reveals that 90 % of diabetics worldwide have T2D. In the past three decades the number of persons with T2D has increased sharply in countries of all income levels (http://www.who.int/diabetes/en/). Gestational diabetes mellitus is a condition where women without prior history of diabetes develop glucose intolerance and high concentration of blood sugar during pregnancy (usually in the third trimester) [4]. Women who had been attacked by gestational diabetes are more likely to develop type 2 diabetes later in life. Diabetes is caused by various factors. The inherited factors, i.e., genetically determined abnormalities of insulin action play an important role. The scope of metabolic abnormalities related to variations of the insulin receptor may cover hyperinsulinaemia and mildly high blood sugar levels to symptomatic diabetes [5–7]. For example, certain mutations of some genes like HLA-DQA1, HLA-DQB1 and HLA-DRB1 raise the risk of causing type 1 diabetes. A few vital proteins in the immune system are generated according to the instructions from these genes [8–10]. Predicting diabetes-associated proteins is very important to understand how diabetes develops since most diabetes-associated variations have an impact on the function of proteins. Linkage studies are often used to determine the genomic intervals which are linked to the disease of interest [11]. Prioritizing a mass of candidate genes via experimental technologies is so expensive and time-consuming that it becomes often impossible to detect the real disease genes by analyzing the list of genes belonged to the interval. Consequently, computational methods have been becoming a prominent option to address such problems. A lot of computational methods have been developed to sequence and predict the most likely disease-related genes by combining various types of data from different sources, for instance, gene expressing profiles [12, 13], functional annotation information [13–17] and sequence-based features [18]. Meanwhile, huge amounts of protein-protein interactions produced by high-throughput technologies play an important role in the disease identification since they offer functional information in a network environment [19]. Furthermore, the proteins coded by genes which are linked to a specific or familiar disease phenotype tend to stay together and form clusters in the protein-protein interaction network [20]. In 2006, it was reported that exploiting protein-protein interactions brings prediction of positional candidate disease genes much closer to the possibility. A large scale usage of PPIs can predict novel candidate proteins [21]. Many methods and frameworks based on the protein-protein interaction networks have been proposed to rank or identify potential disease candidate genes to understand genetic diseases. For Alzheimer Disease, a list of candidate genes/proteins are prioritized by a computational method in terms of the public human protein-protein interaction networks (PPINs) [22]. In the paper of Erten et al., the topological similarity in the human PPINs is employed to prioritize candidate disease genes [23]. Nevertheless, using the PPINs is a risky choice since false interactions made by high-throughput experiments have a negative impact on the disease gene prioritization [24–28]. In order to mitigate that particular risk, new methods are developed to identify candidate disease genes or proteins via integrating biological data from other sources. In the work of Wu et al., the gene expressing data are integrated with the PPI data to identify cancer-related genes [29]. The functional similarity of Gene Ontology is combined with protein protein interactions (PPIs) to prioritize candidate cardiomyopathy genes [30]. However, these methods neglect the fact that proteins are unable to conduct the desired functions until they take up the correct subcellular compartments. More specifically, a protein can interact with another one only if they are localized at the same subcellular compartments [31, 32]. In this article, we propose a method, i.e., Predicting Diabetes Mellitus Genes (PDMG), to rank candidate diabetes mellitus genes by incorporating protein subcellular localization information into the protein-protein interaction networks. First, the protein subcellular localization data are incorporate into the PPINs and the weighted networks are built. Second, we collect the gene records of diabetes from Online Mendelian Inheritance in Man (OMIM) and extract seed genes from these records. Only the genes of T2D are retained since the genes of other diabetes subtypes in OMIM are rare. Then T2D-specific PPINs are constructed by utilizing seed proteins and their interacting neighbors (candidate proteins) from the weighted PPINs. Subsequently, we compute the disease-associated score for each protein in the T2D-specific networks and sort them in descending order. Finally, we discuss the top 27 candidate proteins. Methods In this section, the PDMG method is introduced in detail (see Fig. 1). We first give a general description of sequencing problem of the disease genes. Subsequently, the technology incorporating the subcellular localization information into the PPINs is discussed. Furthermore, we elaborate the method of building disease-specific networks starting from the known disease genes/proteins. Finally, we describe the prioritizing approach of candidate diabetes genes/proteins based on the disease-specific networks. Fig. 1 The schematic of PDMG algorithm for sequencing candidate proteins of the disorders. Our method is mainly comprised of three steps, building weighted PPINs, producing T2D-specific networks and prioritizing candidate proteins Disease gene prioritizing problem In bioinformatics, the predicting problem of genes that have a close relationship with complex diseases is actually converted into node prioritization problem. The nodes representing the candidate genes/proteins will be scored in accordance with one or more strategies. Then the scores are used to rank them. There is an interesting phenomenon in the biological networks, i.e., the ’guilt-by-association’ principle. It depends on the assumption that the genes/proteins leading to diseases tend to have the similar or same properties [33]. In term of the principle, people can extract a group of disease-causing genes from the disease databases as the original seed proteins and then quantify the associations between the candidate genes and seed genes. Consequently, the candidate genes are sorted according to the associations [13, 34]. Let D indicate a disease of interest. S is a seed gene set in which the genes are associated with D. The candidate protein/gene set, represented C, is mechanistically associated with D. The sets S and C constitute the inputs of the disease gene prioritizing algorithm. The known genes in S related to D serve as a starting point for prioritizing the proteins/genes in C. Next, capturing the relationships between the genes in C and the genes in S becomes a critical step. This need to use the topological characteristics of human PPINs. The PPINs, denoted G=(V,E,w), consisting of a group of proteins V and undirected interactions E among the proteins. Meanwhile, uv∈E indicates the interaction between u∈V and v∈V. Due to the false positive rate of the protein-protein interaction data, it is necessary to assigned a weight value to each interaction uv∈E. The confidence scores represent the reliability of the interactions between u and v. In this article, protein subcellular localization data is used to calculate the confidence scores between proteins. The candidate gene products are sorted based on the scores. Scoring PPIs The eukaryotic cells are elaborately organized into functionally-distinct intracytoplasmic "inclusions" or compartments enclosed within membranes, such as a nucleus and other organelles. The compartments specialize in performing all types of biological functions. The micro-environments have significant influence over protein functions since they control access to and availability of various interacting proteins. In essence, the interactions strongly converge among proteins which are located in the same area of the cells(one-sided binomial test with P < 0.001), but the degree of concentration widely depends on the compartments [35]. For instance, the interactions between cytoplasmic proteins are 1.3-fold converged above the threshold. Instead, the interactions among microtubule proteins are 56-fold converged above the threshold. This suggests that the compartment shared by two interacting proteins in the microtubule cytoskeleton better explains the physical and functional interaction than the area of the cytoplasm in which the proteins interact [35]. The fact demonstrates that the significance of different compartments is different in cell activities. After investigating the associations between subcellular localizations and PPIs, Peng et al. find that the farmer is helpful for identifying essential proteins [36, 37]. They give us the motivation of using subcellular localizations to predicting candidate disease genes. Moreover, my research suggests that over half of the T2D genes code essential proteins. Thus, we reason that subcellular localization information can improve the methods of prioritizing candidate disease genes. Peng et al. report that the significance of a compartment is not out of proportion to the number of interacted proteins in this compartment [36]. In order to score the compartments, the number of the proteins in each compartment is counted. For every compartment, its score is described as the number of interacted proteins in the compartment, denoted by CX, divided by the number of proteins in the largest size compartment (consisting of the largest number of interacted proteins), represented by CM. The score SC is calculated by using 1 SC(I)=CX(I)CM, From the formulation, the value of SC ranges from 0 to 1, where I ∈ {1, 2, …, 11}. According to the scores of compartments, the interactions between proteins in the PPINs can be weighted. The different scores of the compartments mean that some compartment are more important than other ones. The phenomenon leads to the importance of PPIs taken place in different compartments should also be different. Consider a set of compartments Loc(u) where protein u is localized. For the two proteins of an interaction (u,v), each protein might be annotated by multiple subcellular localizations. It is reasonable that the interacted protein pairs are localized at the same compartment. Therefore, the interaction (u,v) can be annotated by the shared compartments, i.e., SLoc(u,v)=Loc(u)⋂Loc(v). Furthermore, the score of the interaction (u,v) is defined as 2 W(u,v)=max(SC(I)),ifSLoc(u,v)≠ΦSC(CN),otherwise If SLoc(u,v)≠Φ, the score of the interaction (u,v) is assigned with the maximum value of score of the shared compartments. Since the subcellular localization information of some proteins may be missing, for the interactions with SLoc(u,v)=Φ, the scores of these interactions are assigned with the minimum value of SC(I) among compartments. In Eq. 2, CN is the compartment with the smallest size. Disease-specific networks The OMIM database (http://www.omim.org/) severs as the starting point to extract an initial collection of disease-associated genes, i.e., the seed set S. With the seed genes and weighted PPIs, we derive a disease-specific networks in terms of the nearest-neighbor expansion approach. In other words, the disease-related networks consist of the seed proteins and their direct neighbors. Prioritizing candidate disease gene products In this subsection, we score the proteins in the disease-specific networks and rank them based on descending order of their scores. In order to score every candidate protein, we employ the weighted degree centrality (WDC) [38], relying on the scored disease-specific networks. Specifically, The score of each candidate disease protein, denoted by SPD, is computed in terms of the scored interaction between a protein and its direct neighbor. It can be expressed as 3 SPD(u)=∑vNuWu,v where Nu refers to the set including total neighbors of the protein u and Wu, v represents the weighted value of edge between the protein u and its neighbor v. All proteins in the disease-specific networks are ranked in descending order of SPD. Results and discussions In this section, we evaluate the ability of PDMG to rank candidate disease genes using the known T2D-gene, subcellular localization and PPI information. The datasets used in the experiments are first described. Next, the diabetes-related networks are discussed. Finally, we analyze the novel diabetes genes predicted by PDMG. Data sources Known T2D genes. To form the interaction networks linked to the disease and to detect gene-disease associations from the networks characters, an original set of seed genes known to be associated with the disease is as starting point. We obtain the disease-associated genes of T2D mellitus from OMIM. In OMIM, human genes involved in inherited diseases are recorded in a mini-review format. They are enclosed some information like the gene functions, molecular pathways, and other disease-associated information. To extract a group of T2D-associated genes, we conduct a search of the OMIM database and traverse each OMIM gene record where the term “Diabetes mellitus” is consisted of the “description” field. As a result, 84 OMIM gene records were retrieved. T2D-related entries are shown as Table 1. Based on the HUGO Gene Nomenclature Committee (HGNC) database (http://www.genenames.org/), we replace these genes with their corresponding standard symbols and obtain the seed proteins which correspond to these seed genes. We get 27 proteins coded by the known T2D genes, i.e., GPD2, NEUROD1, IRS1, CAPN10, PPARG, SLC2A2, IGF2BP2, WFS1, CDKAL1, HMGA1,ENPP1, GCK, TCF7L2, KCNJ11, ABCC8, MAPK8IP1, UCP3, MTNR1B, HNF1A, TBC1D4, IRS2, LIPC, HNF1B, GCGR, RETN, AKT2 and HNF4A. Table 1 T2D-related gene records Number Gene/Locus Phenotype 1 Gpd2 Diabetes, type 2, susceptibility to 2 Neurod1 Diabetes mellitus, noninsulin-dependent 3 Irs1 Diabetes mellitus, noninsulin-dependent 4 Capn10 Diabetes mellitus, noninsulin-dependent 1 5 Pparg Diabetes, type 2 6 Slc2a2 Diabetes mellitus, noninsulin-dependent 7 Igf2bp2 Diabetes mellitus, noninsulin-dependent, susceptibility to 8 Wfs1 Diabetes mellitus, noninsulin-dependent, association with 9 Cdkal1 Diabetes mellitus, noninsulin-dependent, susceptibility to 10 Hmga1-rs1, Diabetes mellitus, noninsulin-dependent, Hmga1 susceptibility to 11 Enpp1 Diabetes mellitus, non-insulin-dependent, susceptibility to 12 Gck Diabetes mellitus, noninsulin-dependent, late onset 13 Pax4 Diabetes mellitus, type 2 14 Slc30a8 Diabetes mellitus, noninsulin-dependent, susceptibility to 15 Tcf7l2 Diabetes mellitus, type 2, susceptibility to 16 Kcnj11 Diabetes mellitus, type 2, susceptibility to 17 Abcc8 Diabetes mellitus, noninsulin-dependent 18 Mapk8ip1 Diabetes mellitus, noninsulin-dependent 19 Ucp3 Obesity, severe, and type II diabetes 20 Mtnr1b Diabetes mellitus, type 2, susceptibility to 21 Hnf1a Diabetes mellitus, noninsulin-dependent, 2 22 Pdx1 Diabetes mellitus, type II, susceptibility to 23 Tbc1d4 Diabetes mellitus, noninsulin-dependent, 5 24 Irs2 Diabetes mellitus, noninsulin-dependent 25 Lipc Diabetes mellitus, noninsulin-dependent 26 Hnf1b Diabetes mellitus, noninsulin-dependent 27 Gcgr Diabetes mellitus, noninsulin-dependent 28 Retn Diabetes mellitus, noninsulin-dependent, susceptibility to 29 Akt2 Diabetes mellitus, type II 30 Hnf4a Diabetes mellitus, noninsulin-dependent Protein subcellular localizations. The protein subcellular localization data comes from the COMPARTMENTS database [39]. The resource is obtained by integrated a variety of subcellular localization evidences in terms of high-throughput screens, manually curated annotations and sequence-based identification with automatic text mining for all major model organisms. In the COMPARTMENTS database, the different compartments are labeled as: Nucleus, Golgi apparatus, Cytosol, Cytoskeleton, Peroxisome, Lysosome, Endoplasmic reticulum, Mitochondrion, Endosome, Extracellular space and Plasma membrane. Protein-protein interactions. In the experiments, the human protein-protein interactions are downloaded from BioGrid database(Release version BIOGRID-3.2.111) [40]. The human PPINs include 16, 275 proteins and 143, 611 interactions. T2D-specific networks The nearest-neighbor expansion technology is used to construct the T2D-specific protein interaction subnetworks based on the T2D-associated proteins mentioned above and the global PPINs weighted by subcellular localization information. Here, we employ 27 known proteins associated with T2D as the seed diabetes set. The proteins in the weighted PPINs, interacting with the proteins in the seed diabetes set, are pulled out and constitute the candidate T2D protein set. Each interaction between the seed protein and candidate protein composes the diabetes-interaction-set. The two types of proteins (we call them as diabetes-protein-set) and interactions in the diabetes-interaction-set form T2D-specific networks. In the work, the diabetes-protein-set and diabetes-interaction-set contains 445 human proteins and 543 interactions, respectively. Novel proteins predicted by PDMG PDMG is used to calculate the relevance score for each protein in the T2D-specific PPINs. We rank them based on descending order of their scores. Table 2 list top 27 T2D candidate proteins containing 14 known T2D-associated proteins and 13 novel proteins. The 13 novel proteins are not initially retrieved from OMIM database based on the term "diabetes mellitus". The results show that our prioritizing technology demonstrates very high specificity: out of 27 top-ranking proteins, 14 proteins are known T2D-related proteins in terms of OMIM annotation. Meanwhile, it can be found that the scores of all known proteins but two ones (HNF1B and GCK) are larger than those of other candidate proteins. Furthermore, to examine PDMG’s ability to predict novel diabetes-associated proteins, we use literature study method to determine if the predicted proteins are associated with diabetes. The retrieve results display that out of 13 novel proteins, 8 proteins have been proved to be diabetes-related proteins by literature in the PubMed database (http://www.ncbi.nlm.nih.gov/pubmed). The 8 novel proteins are presented as follows. Table 2 Top 27 rank-ordered T2D relevant proteins Rank Protein Score Description Diabetes relevance 1 PPARG 85.83 peroxisome proliferator activated Known receptor gamma, T2D, susceptibility to 2 HMGA1 63.99 high mobility group AT-hook 1, Diabetes, Known noninsulin-dependent, susceptibility to 3 HNF4A 60.08 hepatocyte nuclear factor 4 alpha, Known Diabetes mellitus, noninsulin-dependent 4 IRS1 45.12 insulin receptor substrate 1, Diabetes, Known noninsulin-dependent 5 HNF1A 24.21 HNF1 homeobox A, Diabetes, Known noninsulin-dependent, 2 6 AKT2 23.28 v-akt murine thymoma viral Known oncogene homolog 2, Diabetes, type II 7 TCF7L2 20.03 transcription factor 7 like 2, Known Diabetes, type 2, susceptibility to 8 IGF2BP2 17.77 insulin like growth factor 2 Known mRNA binding protein 2, Diabetes, noninsulin-dependent, susceptibility to 9 MAPK8IP1 14.52 mitogen-activated protein kinase 8 Known interacting protein 1, Diabetes, noninsulin-dependent 10 IRS2 12.78 insulin receptor substrate 2, Known Diabetes, noninsulin-dependent 11 NEUROD1 7.03 neurogenic differentiation 1, Known Diabetes, noninsulin-dependent 12 UBC 6.52 ubiquitin C Novel 13 HNF1B 6 HNF1 homeobox B, Diabetes, Known noninsulin-dependent 14 EP300 4 E1A binding protein p300 Novel 15 CREBBP 4 CREB binding protein Novel 16 ESR1 3.47 estrogen receptor 1 Novel 17 AKT1 3.02 v-akt murine thymoma Novel viral oncogene homolog 1 18 NRF1 3.02 NFKB repressing factor Novel 19 PCBD1 3 pterin-4 alpha-carbinolamine dehydratase 1 Novel 20 SP1 3 Sp1 transcription factor Novel 21 HDAC4 3 histone deacetylase 4 Novel 22 YWHAB 2.49 tyrosine 3-monooxygenase/tryptophan Novel 5-monooxygenase activation protein beta 23 EGFR 2.47 epidermal growth factor receptor Novel 24 GCK 2.45 glucokinase, Diabetes, noninsulin-dependent, Known late onset 25 ELAVL1 2.45 ELAV like RNA binding protein 1 Novel 26 APP 2.29 amyloid beta precursor protein Novel 27 SUMO2 2.01 small ubiquitin-like modifier 2 Novel CREBBP #15 Rende et al. find that CREB binding protein (CREBBP) plays suggestive roles in linking Type 2 diabetes [41]. Their study reveals that heterozygous CREBBP defect leads to raised effects of hormones like leptin and adiponectin, insulin resistance and preventing obesity. Manabe et al. observe that the mRNA expression of CREBBP is reduced in the uteri of ovariectomized STZ-treated diabetic mice [42]. A recent literature report [43] shows that, compared with healthy conditions, the expressing of histone acetyltransferases CREBBP in latent autoimmune diabetes in adults patients is downregulated. ESR1 #16 Linner et al. conclude that the rs2207396 mutation in ESR1 suggests the risk of type 2 diabetes in hypogonadal men [44]. By investigating the relationship between single nucleotide polymorphisms (SNPs) of the candidate gene and the quantitative traits related to metabolic syndrome in Han Chinese type 2 diabetes, Wei et al. [45] find that Rs722208 of ESR1 is associated with fasting plasma glucose (FPG)(P= 0.045). AKT1 #17 Devaney et al. [46] report that AKT1 is a risk factor for metabolic syndrome and insulin resistance which is one of the five essential endophenotypes linked to T2D. Hami et al. find a significant bilateral downregulation of AKT1 gene expression in the hippocampus of pups born to diabetic mothers [47]. NRF1 #18 By researching defect of Nuclear factor-erythroid 2-related Factor 1 (NRF1) in beta-cells, Zheng et al. discover that Nrf1 acts as an essential regulator of mitochondrial function, glucose metabolism and insulin secretion [48]. Specifically, Nrf1 inactivation in beta-cells results in a pre-T2D phenotype because of impairment of insulin secretion and disruption of glucose metabolism [48]. In the study from Hirotsu et al., Nrf1 over-expression has a negative impact on both glucose utilization and production in the liver by suppressing the genes related to both glycolysis and gluconeogenesis [49]. PCBD1 #19 The findings from Ferre et al. suggest that a PCBD1 deficiency may cause hypomagnesemia and diabetes [50]. Simaite et al. observe an abundant expression of Pcbd1 in the developing pancreas of both mouse and Xenopus embryos [51]. The genetic evidence obtained by them displays that PCBD1 variations can lead to early-onset nonautoimmune diabetes with characteristics like dominantly inherited HNF1A-diabetes. YWHAB #22 YWHAB interacts with GCGR, a type 2 diabetes-related protein. To examine the effect of YWHAB on GCGR function, Han et al. investigate glucose production in primary mouse hepatocytes. They discover that YWHAB is overexpressed in mouse hepatocytes. In other words, YWHAB inhibits glucose production [52]. Studies show that YWHAB may plays a critical role in glucose metabolism. YWHAB actually regulates the activity of ChREBP (glucose responsive transcription factor), carbohydrate response element-binding protein, which has important influence on the glucose-mediated induction of proteins associated with hepatic glycolysis and lipogenesis [53]. Besides, YWHAB also controls the activity of AKT, which mediates insulin signaling [54]. EGFR #23 Chen et al. suggest that EGFR (epidermal growth factor receptor) mediates TGF-b-induced renal fibrosis and is inhibited by the EGFR inhibitor, erlotinib, in STZ-induced diabetic mice [55]. More recently, they also report the resistance of podocyte-specific EGFR knockout mice to the development of diabetes-associated podocyte damage [56]. SUMO2 #27 The transcriptional activity of T and B cells is negatively regulated by the mouse SUMO2 [57, 58]. The mouse SUMO2 in T cells is overexpressed, which inhibits the production of both Th1 and Th2 cytokines [57, 58]. This means that the mouse SUMO2 plays a more complex role in the progression of autoimmune diabetes. The early literature [59] also shows that SUMO is related to NF-kB activation and may thus be linked to type 1 diabetes with apoptosis in pancreatic beta cells. Conclusions With the available PPI data increasing rapidly, a unprecedented opportunity for predicting disease-associated genes/proteins at the network level is appear. The PPINs have been widely adopted by many state of the art algorithms to address the gene prioritization problem. They are based on the principle that the genes/proteins causing similar diseases tend to cluster together in the network. However, the high false positive rates and false negative rates of the available PPI data have a negative influence on the accuracy of methods identifying disease genes/proteins only by the topological properties of the networks. To improve the prediction, researchers develop all kinds of new approaches to predict candidate disease genes via combining other data from different sources with PPINs. But these methods neglect an obvious fact proteins don’t perform their desired functions unless they are localized at the appropriate subcellular compartments. In this work, subcellular localization data are integrated with PPINs. The combination is achieved by building disease-specific PPINs and employing them in the prioritization. Specifically, OMIM is used to obtain seed genes/proteins of type 2 diabetes. With these seed proteins, we produce T2D-specific PPINs from the weighted PPINs based on the nearest-neighbor expansion approach. And then the scores of candidate T2D proteins are calculated by WDC method. Finally, we rank the proteins based on descending order of their scores. In order to prove PDMG’s ability to predict potential disease-related proteins, we employ the literature review method to analyze the novel proteins/genes predicted by PDMG. The results show that PDMG has predicted 13 novel proteins in top 27 candidate proteins. Out of the 13 novel proteins, 8 proteins CREBBP, ESR1, AKT1, NRF1, PCBD1, YWHAB, EGFR, SUMO2 are associated with diabetes in literature. The evidences display that the 8 novel proteins are recovered from the interaction data and subcellular localization information analysis although they are not retrieved from OMIM database. Therefore, PDMG method can make up for the false negatives (to an extent) of PPINs. Besides, according to the ranked candidate proteins, one may gain many new biological suppositions about the new protein functions in the context of protein interaction networks out of scope of this work. From IEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA. 9–12 November 2015 Declarations This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4. Authors’ contributions XT constructs PDMG framework, carries out the corresponding algorithm and writes the article. XH and XY conceived of the study. YL, YF and YL provide a number of valuable suggestions in development of the algorithm. WP and WH perform the data collection and analysis. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Acknowledgements Publication of this article was partially funded by the National Natural Science Foundation of China under Grant Nos. 61472133, 61502214, 31560317, Hunan Provincial Natural Science Foundation of China Nos. 15JJ2038, 15JJ2037, Research Foundation of Education Bureau of Hunan Province Nos. 14A027, [2015]118, [2013]532, Hunan Key Laboratory no. 2015TP1017. ==== Refs References 1 Mellitus D Diagnosis and classification of diabetes mellitus Diabetes care 2005 28 S37 S5 S10 2 Davies JL Kawaguchi Y Bennett ST A genome-wide search for human type 1 diabetes susceptibility genes Nature 1994 371 6493 130 6 10.1038/371130a0 8072542 3 Butler AE Bonner-Weir S Janson, J Diabetes 2003 52 1 102 10 10.2337/diabetes.52.1.102 12502499 4 Buchanan TA Xiang AH Gestational diabetes mellitus J Clin Invest 2005 115 3 485 91 10.1172/JCI200524531 15765129 5 Marx J Unraveling the causes of diabetes Science 2002 296 5568 686 10.1126/science.296.5568.686 11976439 6 Notkins AL The causes of diabetes Sci Am 1979 241 5 62 10.1038/scientificamerican1179-62 228390 7 Loeken MR Advances in understanding the molecular causes of diabetes-induced birth defects J Soc Gynecologic Invest 2006 13 1 2 10 10.1016/j.jsgi.2005.09.007 8 Nguyen C Varney MD Harrison LC Definition of high-risk type 1 diabetes HLA-DR and HLA-DQ types using only three single nucleotide polymorphisms Diabetes 2013 62 6 2135 40 10.2337/db12-1398 23378606 9 Hu X Deutsch AJ Lenz TL Additive and interaction effects at three amino acid positions in HLA-DQ and HLA-DR molecules drive type 1 diabetes risk Nat Genet 2015 47 8 898 905 10.1038/ng.3353 26168013 10 Chen LM Association of the HLA-DQA1 and HLA-DQB1 Alleles in Type 2 Diabetes Mellitus and Diabetic Nephropathy in the Han Ethnicity of China Exp Diabetes Res 2013 2013 1 5 11 Glazier AM Nadeau JH Aitman TJ Finding Genes That Underlie Complex Traits Science 2002 298 5602 2345 9 10.1126/science.1076641 12493905 12 Lage K Karlberg E A human phenome-interactome network of protein complexes implicated in genetic disorders Nat Bio 2007 25 3 309 16 10.1038/nbt1295 13 Aerts S Lambrechts D Gene prioritization through genomic data fusion Nat Biotech 2006 24 5 537 44 10.1038/nbt1203 14 Adie E Adams R SUSPECTS:enabling fast and effective prioritization of positional candidates Bioinformatics 2006 22 6 773 4 10.1093/bioinformatics/btk031 16423925 15 Turner F Clutterbuck D Semple C POCUS: mining genomic sequence annotation to predict disease genes Genome Biology 2003 4 11 R75 10.1186/gb-2003-4-11-r75 14611661 16 Masotti D Nardini C TOM: enhancement and extension of a tool suite for in silico approaches to multigenic hereditary disorders Bioinformatics 2008 24 3 428 9 10.1093/bioinformatics/btm588 18048394 17 Chen J Bardes EE ToppGene Suite for gene list enrichment analysis and candidate gene prioritization Nucleic Acids Res 2009 37 suppl 2 W305—11 19465376 18 Adie EA Adams RR Speeding disease gene discovery by sequence based candidate prioritization BMC Bioinformatics 2005 6 55 1 13 10.1186/1471-2105-6-S3-P1 15631638 19 Stelzl U Wanker EE The value of high quality protein-protein interaction networks for systems biology Curr Opin Chem Biol 2006 10 551 8 10.1016/j.cbpa.2006.10.005 17055769 20 Gandhi TKB Zhong J Analysis of the human protein interactome and comparison with yeast, worm and fly interaction datasets Nat Genet 2006 38 285 93 10.1038/ng1747 16501559 21 Oti M Snel B Huynen MA Predicting disease genes using proteinCprotein interactions J Med Genet 2006 43 8 691 8 10.1136/jmg.2006.041376 16611749 22 Chen JY Shen C Sivachenko AY Mining Alzheimer disease relevant proteins from integrated protein interactome data Pac Symp Biocomput 2006 11 367 78 17094253 23 Erten S Bebek G Disease gene prioritization based on topological similarity in protein-protein interaction networks Res Comput Mol Biol 2011 2011 54 68 10.1007/978-3-642-20036-6_7 24 Sprinzak E Sattath S Margalit H How Reliable are Experimental Protein-Protein Interaction Data J Mol Biol 2003 327 5 919 23 10.1016/S0022-2836(03)00239-0 12662919 25 Chen J Yuan B Detecting Functional Modules in the Yeast Protein-Protein Interaction Network Bioinformatics 2006 22 18 2283 90 10.1093/bioinformatics/btl370 16837529 26 Bader GD Hogue CWV Analyzing yeast protein-protein interaction data obtained from different sources Nat Biotechnol 2002 20 10 991 7 10.1038/nbt1002-991 12355115 27 Batada N Hurst LD Tyers M. Evolutionary and physiological importance of hub proteins PLoS Comp Bio 2006 2 e88 10.1371/journal.pcbi.0020088 28 Collins SR Kemmeren P Zhao XC Toward a comprehensive atlas of the physical interactome of Saccharomyces cerevisiae Mol Cell Proteomics 2007 6 3 439 50 10.1074/mcp.M600381-MCP200 17200106 29 Wu C Zhu J Zhang X Integrating gene expression and protein-protein interaction network to prioritize cancer-associated genes BMC Bioinformatics 2012 13 1 182 10.1186/1471-2105-13-182 22838965 30 Li W Chen L He W Prioritizing Disease Candidate Proteins in Cardiomyopathy-Specific Protein-Protein Interaction Networks Based on “Guilt by Association” Analysis PloS one 2013 8 8 e71191 10.1371/journal.pone.0071191 23940716 31 Kumar A Agarwal S Subcellular localization of the yeast proteome Genes and development 2002 16 6 707 19 10.1101/gad.970902 11914276 32 de Lichtenberg U Jensen LJ Dynamic complex formation during the yeast cell cycle. science. 2005 307 5710 724 7 15692050 33 Altshuler D Daly M Kruglyak L Guilt by association Nat Genet 2000 26 2 135 8 10.1038/79839 11017062 34 Kohler S Bauer S Horn D Walking the interactome for prioritization of candidate disease genes Am J Hum Genet 2008 82 4 949 58 10.1016/j.ajhg.2008.02.013 18371930 35 Huh WK Falvo JV Global analysis of protein localization in budding yeast Nature 2003 425 6959 686 91 10.1038/nature02026 14562095 36 Peng X Wang J An efficient method to identify essential proteins for different species by integrating protein subcellular localization information Bioinformatics Biomed 2015 2015 277 80 37 Peng X Wang J Rechecking the Centrality-Lethality Rule in the Scope of Protein Subcellular Localization Interaction Networks PloS one 2015 10 6 1 22 38 Tang X Wang J Predicting essential proteins based on weighted degree centrality IEEE/ACM Trans Comput Biol Bioinformatics 2014 11 2 407 18 10.1109/TCBB.2013.2295318 39 Binder JX, Pletscher-Frankild S, et al. COMPARTMENTS: unification and visualization of protein subcellular localization evidence. Database. 2014; 2014. bau012. 40 Stark C Breitkreutz BJ Biogrid: a general repository for interaction datasets Nucleic Acids Res 2006 34 1 D535—9 16381927 41 Rende D Baysal N Kirdar B Complex disease interventions from a network model for type 2 diabetes PloS One 2013 8 6 e65854 10.1371/journal.pone.0065854 23776558 42 Manabe Y Tochigi M Insulin-like growth factor 1 mRNA expression in the uterus of streptozotocin-treated diabetic mice J Reprod Dev 2013 59 4 398 404 10.1262/jrd.2012-169 23719562 43 Liu X Xu J Reduced Histone H3 Acetylation in CD4 Disease Markers 2015 2015 1 8 44 Linner C Svartberg J Giwercman A Estrogen receptor alpha single nucleotide polymorphism as predictor of diabetes type 2 risk in hypogonadal men Aging Male 2013 16 2 52 7 10.3109/13685538.2013.772134 23506158 45 Wei FJ Cai CY Quantitative candidate gene association studies of metabolic traits in Han Chinese type 2 diabetes patients Genet Mol Res GMR 2015 14 4 15471 10.4238/2015.November.30.25 26634513 46 Devaney JM Gordish-Dressman H AKT1 polymorphisms are associated with risk for metabolic syndrome Hum Genet 2011 129 2 129 39 10.1007/s00439-010-0910-8 21061022 47 Hami J Kerachian MA Effects of streptozotocin-induced type 1 maternal diabetes on PI3K/AKT signaling pathway in the hippocampus of rat neonates J Receptors Signal Transduction 2015 2015 1 7 48 Zheng H Fu J CNC-bZIP protein Nrf1-dependent regulation of glucose-stimulated insulin secretion Antioxidants Redox Signal 2015 22 10 819 31 10.1089/ars.2014.6017 49 Hirotsu Y Higashi C Transcription factor NF-E2-related factor 1 impairs glucose metabolism in mice Genes Cells 2014 19 8 650 65 10.1111/gtc.12165 25041126 50 Ferre S, de Baaij JHF, et al. Mutations in PCBD1 cause hypomagnesemia and renal magnesium wasting. J Am Soc Nephrol. 2013; 2013. ASN 2013040337. 51 Simaite D Kofent J Recessive mutations in PCBD1 cause a new type of early-onset diabetes Diabetes 2014 63 10 3557 64 10.2337/db13-1784 24848070 52 Han J Zhang M The Identification of Novel Protein-Protein Interactions in Liver that Affect Glucagon Receptor Activity PloS one 2015 10 6 e0129226 10.1371/journal.pone.0129226 26075596 53 Sakiyama H Wynn RM Regulation of Nuclear Import/Export of Carbohydrate Response Element-binding Protein (ChREBP) INTERACTION OF AN alpha-HELIX OF ChREBP WITH THE 14–3-3 PROTEINS AND REGULATION BY PHOSPHORYLATION J Biol Chem 2008 283 36 24899 908 10.1074/jbc.M804308200 18606808 54 Somanath PR 14-3-3 beta-Rac1-p21 activated kinase signaling regulates Akt1-mediated cytoskeletal organization, lamellipodia formation and fibronectin matrix assembly J Cell Physiol 2009 218 2 394 404 10.1002/jcp.21612 18853424 55 Chen J Chen JK EGFR signaling promotes TGF-dependent renal fibrosis J Am Soc Nephrol 2012 23 2 215 24 10.1681/ASN.2011070645 22095949 56 Chen J Chen JK Harris RC EGF receptor deletion in podocytes attenuates diabetic nephropathy J Am Soc Nephrol 2015 26 5 1115 25 10.1681/ASN.2014020192 25185988 57 Hwang KW Won TJ Erratum to “Characterization of the regulatory roles of the SUMO” Diabetes/metabolism Res Rev 2012 28 2 196 202 10.1002/dmrr.2273 58 Hwang KW Won TJ Characterization of the regulatory roles of the SUMO Diabetes/metabolism Res Rev 2011 27 8 854 61 10.1002/dmrr.1261 59 Owerbach D Pina L Gabbay KH A 212-kb region on chromosome 6q25 containing the TAB2 gene is associated with susceptibility to type 1 diabetes Diabetes 2004 53 7 1890 3 10.2337/diabetes.53.7.1890 15220215

PMC005xxxxxx/PMC5001231.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556416291210.1186/s12864-016-2912-yResearchA matrix rank based concordance index for evaluating and detecting conditional specific co-expressed gene modules Han Zhi hanzhi@nankai.edu.cn 123Zhang Jie jie.zhang@osumc.edu 34Sun Guoyuan sunguoyuan@mail.nankai.edu.cn 12Liu Gang liugang56@mail.nankai.edu.cn 12Huang Kun kun.huang@osumc.edu 341 College of Computer and Control Engineering, Nankai University, Tianjin, China 2 College of Software, Nankai University, Tianjin, China 3 Department of Biomedical Informatics, The Ohio State University, Columbus, OH USA 4 The CCC Biomedical Informatics Shared Resource, The Ohio State University, Columbus, OH USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.519© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background  Gene co-expression network analysis (GCNA) is widely adopted in bioinformatics and biomedical research with applications such as gene function prediction, protein-protein interaction inference, disease markers identification, and copy number variance discovery. Currently there is a lack of rigorous analysis on the mathematical condition for which the co-expressed gene module should satisfy. Methods In this paper, we present a linear algebraic based Centralized Concordance Index (CCI) for evaluating the concordance of co-expressed gene modules from gene co-expression network analysis. The CCI can be used to evaluate the performance for co-expression network analysis algorithms as well as for detecting condition specific co-expression modules. We applied CCI in detecting lung tumor specific gene modules. Results and Discussion Simulation showed that CCI is a robust indicator for evaluating the concordance of a group of co-expressed genes. The application to lung cancer datasets revealed interesting potential tumor specific genetic alterations including CNVs and even hints for gene-fusion. Deeper analysis required for understanding the molecular mechanisms of all such condition specific co-expression relationships. Conclusion The CCI can be used to evaluate the performance for co-expression network analysis algorithms as well as for detecting condition specific co-expression modules. It is shown to be more robust to outliers and interfering modules than density based on Pearson correlation coefficients. The International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Gene co-expression network analysis (GCNA) is widely adopted in bioinformatics and biomedical research. It has many biomedical applications such as gene function prediction [1–4], protein-protein interaction inference [1, 5–7], disease markers identification [3, 8], and copy number variance discovery [9, 10]. Many GCNA algorithms have been developed to identify gene modules of strongly co-expressed genes [3, 7, 11–15]. These gene modules can be used to further infer co-regulation mechanisms such as common transcription factors as well as genetic mutations such as copy number alterations in specific chromatin regions. Mathematically, the co-expression of the genes is often measured using correlation metrics with Pearson correlation coefficient being the most widely used one [1, 7, 11]. However, there has been a lack of rigorous analysis on the mathematical condition for which the co-expressed gene module should satisfy. As to be shown in this paper, the mathematical condition is a rather straightforward linear algebra based condition. And the condition can lead to an effective metric for characterizing the concordance of the gene expression profiles in the module. With the rigorous treatment and the effective metric, we can evaluate each module as well as the algorithm. In addition, this metric can be used to detect condition specific co-expressed gene modules. Condition specific gene co-expression is an interesting problem and many methods have been developed to detect it [16–20]. However, most of the methods are based on first detecting differential correlation at gene-pair level such as the Fisher’s transformation and the Expected Conditional F-statistic (ECF) developed in [17]. Instead, using the new metric we developed here and the randomized test for this metric, we can detect condition specific gene co-expression holistically at the gene module level instead of just gene pairs. As demonstrated in our example on lung cancer data, this can lead to new candidates on different mechanisms for co-expression and discovery of potential new genetic variants associated with diseases such as cancers. Our preliminary results suggest that there is rich biological information contained in the co-expression relationships and the condition specificity that needs to be uncovered by deeper analysis and biological validations. Methods Rank condition of the expression profile data matrix for co-expressed genes Given a set of n (≥2) genes and their expression levels over N (≥3) samples, the expression profiles can be expressed using a matrix 1 G=g11⋯g1N⋮⋱⋮gn1⋯gnN=g1⋮gn∈ℜn×N, with N-dimensional row vector gi = [gi1, gi2, …, giN] be the expression profile for the i-th gene across the samples (i = 1, 2, …, n). If two genes i and j are perfectly co-expressed, ie, |ρ(gi, gj)| = 1 where ρ(⋅,⋅) is the Pearson correlation coefficient between two vectors, then given the linear relationship between the two vectors, we have 2 gik=αij⋅gjk+βijk=1,2,…,N for some constants αij and βij and 3 gi=αij⋅gj+βij⋅1NT, where 1N = [1, 1, …, 1]T ∈ ℜN is N-dimensional. Therefore the matrix G can be re-written as 4 G=g1g2⋮gn=g1α12g1+β12⋅1NT⋮α1ng1+β1n⋅1NT=g1α12g1⋮α1ng1+0β12⋅1NT⋮β1n⋅1NT=1α12⋮α1n⋅g1+0β12⋮β1n⋅1NT. Thus the matrix G can be decomposed as the sum of two matrices each has rank 1. Since it has been well established in linear algebra that given two matrices A and B of the same size, rank(A + B) ≤ rank(A) + rank(B) [21], we have the following proposition: Proposition 1 If the absolute value of the Pearson correlation coefficient (PCC) between every pair of rows of a matrix G (G ∈ ℜn × N, n ≥ 2, N ≥ 3) is 1, then rank(G) ≤ 2. Furthermore, if any of the rows of G is shifted or scaled (e.g., gik' = λ ⋅ gik + ε), the PCC value between them will still have absolute value 1 and thus the Proposition 1 still holds. SVD based methods for estimating the rank of G Given the matrix G, its singular value decomposition (SVD) is G = USVT where U ∈ ℜn × n, V ∈ ℜN × N are both orthonormal matrices and S is a diagonal matrix with all the elements being zero except for the ones on the diagonal line which are non-negative and sorted in descending order (ie, S11 ≥ S22 ≥ … ≥ SKK ≥ 0, where K = min(n, N)). In addition, let ‖G‖ be the Frobenius norm of G such that G2=∑i=1n∑j=1Ngij2, then it is well known that 5 G2=∑i=1n∑j=1Ngij2=∑i=1KSKK2,K=minn,N. If G satisfies the condition of Proposition 1, then the rank of G is 2 which implies S33 = S44 = … = SKK = 0. Thus 6 R12=S112+S222G2=1. In reality, given the expression profile matrix of a set of co-expressed genes, the perfect PCC value can never be reached and thus G is never really rank 2, but instead it can be approximated with a rank 2 matrix. Thus in theory, given an expression profile matrix G, we can examine if the genes (row vectors) are co-expressed by testing if the ratio R12 defined in (6) is close to 1. We refer R12 as the concordance index. Data transformation and centralized concordance index While the concordance index R12 can be used as a potential indicator for the concordance of the rows of G and thus for evaluating co-expressed modules, it is difficult to set a hard threshold for it. This is even more challenging for real data due to noise, batch effects, and background signals that may skew the PCC values. In addition, since SVD is based on the L2 -norm, it can be biased by any specifically large outlier or just a few genes with high expression levels. Thus the data needs to be transformed before processing. The transformation of the data we proposed involves two steps: centralization and standardization. First, each row of G needs to be centralized by subtracting its average such that 7 G¯=g11−g1¯⋯g1N−g1¯⋮⋱⋮gn1−gn¯⋯gnN−gn¯=g1c⋮gnc,wheregi¯=∑k=1NgikNfori=1,2,…,n. Next, each row of G¯ is standardized to have norm 1, ie 8 G^=g1cg1c⋮gncgnc=g^1⋮g^n,whereg^k=gkcgkc,k=1,2,…,n. The centralization step aims to mitigate the background signal while the standardization step avoids bias towards any particularly highly expressed genes. Interestingly, since the Pearson correlation coefficient between gi and gj is defined as 9 ρgi,gj=∑k=1Ngik−g¯i⋅gjk−g¯j∑k=1Ngik−g¯i2⋅∑k=1Ngjk−g¯i2=gic⋅gjcTgic⋅gjc=g^i⋅g^jT and ‖ĝk‖ = 1 (k = 1, 2, …, n), therefore |ρ(gi, gj)| = 1 implies ĝk = ĝ1 or ĝk = − ĝ1. In other words, G^=α1⋮αnĝ1,whereαi∈+1,−1fori=1,2,…,n. Therefore we have the following proposition: Proposition 2 If the absolute value of the Pearson correlation coefficient (PCC) between every pair of rows of a matrix G (G ∈ ℜn × N, n ≥ 2, N ≥ 3) is 1 and Ĝ is the centralized matrix of G with each row standardized as in (8), then rank(Ĝ) = 1. In addition, the inner product between every pair of rows equals the Pearson correlation coefficient of the two rows. Thus the singular value decomposition (SVD) for Ĝ is G^=U^S^V^T where U^∈ℜn×n,V^∈ℜN×N are both orthonormal matrices and Ŝ is a diagonal matrix with all the elements being zero except for the ones on the diagonal line which are non-negative and sorted in descending order (ie, Ŝ11 ≥ Ŝ22 ≥ … ≥ ŜKK ≥ 0, where K = min(n, N)). In fact, given that Ĝ is rank 1 and Ĝ = n, we have 11 S^112=nandS^22=…=S^KK=0. We therefore define a centralized concordance index (CCI) as CCI=S^112n. Estimate the significance of the CCI We examine two approaches for determining if the observed CCI is significantly large to reflect co-expression relationship among the n genes over the entire whole genome dataset. First, we randomly permute the entries of every row of Ĝ and calculate CCIp. This process is repeated M times (usually we choose M = 1000). Then we set ppermute=#CCIp≥CCI/M. Conceptually this gives a measurement on how significant is the observed concordance index in the background of the same data distribution. Next, we randomly choose n genes from the whole genome gene expression data and calculate the CCIr. This process is repeated M times. We then calculate the z-score ZCCI for the CCI based on the random sampling such that ZCCI=CCI−meanCCIrstdCCIr. The significance is then estimated from the z-score. This gives a measurement on how significant is the observed CCI for the tested gene module in the entire genome. We choose z-score instead of the percentile of the CCI due to three reasons: 1) simulation and tests on real data shows that CCIr follow a bell-shaped distribution which can be reasonably approximated by a Gaussian distribution as shown in Figs. 1 and 2) even with 1000 times sampling, it is still relatively small comparing to all the possible combinations, thus sometimes although CCI is larger than all CCIr, it is not reasonable to assume that the p-value (significance) is extremely small, instead z-score gives a reasonable estimate on the deviation of the observed CCI from the random samples; and 3) last but not the least, one of our goals is to use the metric to compare results from different conditions, z-scores are standardized with the same distribution and thus allows us to compare with different conditions.Fig. 1 Simulation on the distribution of CCI and its relationship with noise in the data. Top: Relationship between CCI and noise level. The x-axis reflects the effects of the noise on the centralized matrix. Middle: The distribution of CCI calculated from 1000 randomly selected gene lists (with 220 genes) in the 41 lung cancer tumor samples (using GSE18842). Bottom: The distribution of CCI calculated from 1000 random permutation of the data from the correlated gene module Fig. 2 Comparison between CCI and density metrics. a The CCI versus density metrics with the increases of number of outliers under two different noise levels. b The boxplots for the two metrics with different number of outliers and noise levels. The values are normalized to the values with zero outlier. c The CCI versus density metrics with the increasing size of the interfering module. d The boxplots for the two metrics with different number of outliers and noise levels with the values normalized to the values without interfering module Simulation To evaluate the performance of the concordance index, we generate a matrix of 50 × 100 with absolute value of PCC between every pair of rows being 1. The base vector is generated as a 100-dimensional row vector using uniform distribution from 0 to 1. The scaling factors (α) and shifts (β) are also generated using uniform distribution from 0 to 1. The matrix G is calculated using Eq.(4). Then Gaussian noises with zero mean at different levels (σ = 0.01, 0.02, 0.05, 0.07, 0.1, 0.15, 0.2, 0.3, 0.5, 1) are added to the matrix and corresponding concordance indices are calculated. This process is repeated 1000 times for each noise level. In addition, for each test the centralized matrix Ĝr was compared with the original Ĝ using ratio RF=G^r−G^FG^F, where ‖ ⋅ ‖F is the Frobenius norm of the matrix. Comparison with density metric Since a focus of mining co-expression network is to identify densely connected gene modules, the metric density defined for network mining is often used. For a module with n in weighted network, its density is defined as d=∑i=1n−1∑j=i+1nwijnn−1/2. For co-expression networks, the weight wij is often defined as the correlation coefficient |ρ(gi, gj)| or its transformation. In order to examine the relationship between CCI and density, we compare CCI with the density defined using |ρ(gi, gj)| as weights. Specifically, we consider two scenarios. The first is to test the responses of the metrics to outliers. We first generate the simulated matrix G as described above. Then outlier (independently generated vectors) will be added. We calculate both metrics under different number of outliers and different noise levels for G. The second scenario is to consider the possibility that two modules sometimes can be erroneously linked together. To test this, we generate two gene modules and test the responses of the two metrics with respect to different sizes and noise levels of the modules. Each test is repeated 100 times. Datasets We test the concordance index and its significance using a large gene expression dataset. The dataset was downloaded from NCBI Gene Expression Omnibus (GEO). The dataset is GSE18842 containing gene expression microarray data from 46 non-small cell lung cancer (adenocarcinoma) tumor samples and 45 non-cancer control tissue samples [22]. The GSE18842 dataset was generated using Affymetrix HU133 2.0Plus GeneChip. The normalization of the dataset was verified by inspecting the boxplot and data distributions. We also tested some of the findings using TCGA non-small cell lung cancer adenocarcinoma [23] and squamous cell tumor data [24] from cBioPortal (http://www.cbioportal.org/). Weighted co-expression network analysis While the R package WGCNA developed by the Horvath’s group is a widely adopted co-expression gene module discovery tool, it has some limitation as it is based on hierarchical clustering algorithm that does not allow overlap between modules and does not control the density of the detected modules [11]. In this paper, we apply a recently developed algorithm called Normalized lmQCM [15]. This algorithm takes a network mining approach allowing overlaps between modules and also is guaranteed to have a lower bound on the density of the detected modules. Using CCI to detect condition specific modules The concordance index and its significance evaluation provide us a means to evaluate if a co-expressed gene module (CGM) in one condition is also strongly co-expressed in another condition. We first apply the Normalized lmQCM to each conditions (normal and disease) in both datasets using selected parameters (to be discussed in the Results section). For each gene module, we then calculated two CCIs, one using the data from the condition it was generated and one using the data from the opposite condition. For instance, if the module was generated from the Parkinson’s disease patients, CCIs for the same gene module is calculated for both Parkinson’s disease samples and the control sample. Then the ppermute and ZCCI are calculated for both conditions too. Gene modules that are significant (ZCCI ≤ τ) in one condition but not significant (ZCCI > τ) in the other condition are reported for further analysis. The threshold τ is determined based on the significance requirement. For instance, τ is often chosen such that the one-tail p-value for the ZCCI is less than 0.05 for single gene module or 0.05/m if m gene modules are being tested (for multiple test compensation). Enrichment analysis for gene modules For the reported modules, we carry out enrichment analysis using TOPPGene (https://toppgene.cchmc.org/enrichment.jsp). We specifically pay attention to significantly enriched Gene Ontology (GO) terms, cytobands, transcription factor binding sites, and human/mouse phenotypes. Results Simulation on the relationship between CCI As described in the Methods section, we generated the matrix with correlated rows. The matrix G was then calculated using Eq. (4). Then Gaussian noises with zero mean at different levels (σ = 0.01, 0.02, 0.05, 0.07, 0.1, 0.15, 0.2, 0.3, 0.5, 1) were added to the matrix and corresponding concordance indices were calculated. This process is repeated 1000 times for each noise level. In addition, for each test the centralized matrix Ĝr was compared with the original Ĝ using ratio RF=G^r−G^FG^F, where ‖ ⋅ ‖F is the Frobenius norm of the matrix. The relationship between the CCI and the difference between the noisy matrix with the original matrix is plotted in Fig. 1 Top. We then tested the distribution of the CCI in random gene lists using real data. As shown in Fig. 1 Middle, 1000 randomly selected gene lists with 220 genes (based on a real module with CCI 0.4957 generated from the co-expression analysis) in 41 lung cancer tumor samples from GSE18842 were generated and the distribution of the CCI follow a bell shaped curve with a mean of 0.1974 and standard deviation of 0.0117. Thus zCCI is 25.49. In addition we carried out 1000 times of random permutation of the data from the co-expressed gene module with 220 genes and the distribution is shown in Fig. 1 Bottom. The permutation results follow a tight distribution with mean of 0.0482 and standard deviation of 0.00176. While this clearly shows that the observed CCI (0.4957) is not associated with the data distribution, the fact that these permutation results are much lower than randomly picked gene sets from the original dataset (as shown in Fig. 1 Middle) suggests the permutation test practically cannot provide new information regarding the significance of the modules. Therefore in the rest of the paper we focus on the z-score based approach from random gene list to evaluate the modules. Similar distributions were observed in multiple datasets with different number of genes or sample sizes (data not shown). Comparison with density metric As described in the Methods section, we first consider the scenario when the different numbers of “outlier vectors” were added to the correlated matrix G with 50 rows and 100 columns. Figure 2a shows two instances of the simulation for different choices of the noise level. In both cases, the metrics (CCI and density) decrease as the number of outliers increases. However, it can be seen that the curve for the CCI is smoother than the curve for the density, suggesting that CCI is more robust in response to outliers. This is further confirmed in Fig. 2b when the ranges of the values for both metrics over the 100 times simulation are plotted. In Fig. 2b, the values of the metrics are normalized according to the value of zero outlier. It is clear that the ranges for CCI are always tight when the density values span a wide range. Similar results are observed for the two-module scenario as shown in Fig. 2c and d. In addition, it is clear that with the increase of size of the interfering module, the density is no long sensitive when the size of the interfering module is more than half of the original module while the CCI consistently decreases. Identify tissue specific co-expressed modules in lung tumor We first carried out weighted gene co-expression network mining using the normalized lmQCM algorithm on the lung tumor data (41 samples) using parameter γ = 0.4. γ is a major parameter for the normalized lmQCM algorithm. The larger its value, the higher is the density of the identified gene modules. Our previous study suggested γ = 0.4 is a reasonable values for such dataset. The algorithm yielded 168 gene modules with at least five genes (ranging from five to 891). We then calculated the CCI and z-score based on 1000 randomly selected gene lists of the same size for every module. Then we calculated the CCI and z-score for the same set of gene lists in the control samples. We selected the threshold for z-score to be 3.433 such that the one-tail p-value is less than 0.05/168 = 0.000298. Among the 168 gene modules, all the z-scores derived from the tumor samples are higher than the threshold while 15 of the gene modules have z-scores lower than the threshold in control samples. Figure 3 shows the heatmaps of three examples of the gene modules. Two (Figure 3 Top and Middle) have high z-scores in tumor samples but lower than threshold z-scores in control samples. This is also reflected in the heatmap. In the tumor samples (Fig. 3 Left), the expression levels of the samples show clear consistent patterns across the samples while there is no clear pattern in the control samples. The last module in Fig. 3 (bottom) has high CCIs and z-scores in both tumor and control samples and it is clear the expression levels of the genes show consistent patterns in both cohorts.Fig. 3 Examples of the gene modules in tumor samples (left column) and control samples (right column). The top two modules show significant difference in co-expression between control and tumor samples with high CCIs and z-scores in tumor and low CCIs as well as low z-scores in control samples. The bottom module has high CCIs and z-scores in both tumor and control samples These 15 gene modules are further analyzed for enriched biological processes, cytobands and transcription factor binding sites. Table 1 summarizes the findings from these 15 gene modules.Table 1 Enrichment analysis of the 15 gene modules that are specific to tumor samples in lung cancer Module Size GO BP term (p-value) Cytoband (p-value) TF (p-value) 4 162 Epidermis development (p = 8.762E-23); 1q21-q22 (p = 3.354E-06); AP1 (p = 2.177E-04, 22 genes); 2p24.3 (p = 5.021E-06) AREB6 (p = 9.900E-04, 8 genes) 5 159 neuron differentiation (p = 1.516E-07); 6q14.2 (p = 1.796E-04); PAX4 (p = 1.046E-05, 9 genes); generation of neurons (p = 2.105E-07) 5q33 (p = 2.686E-04) MSX1 (p = 2.088E-04, 7 genes) 9 98 Neurogenesis (p = 1.490E-06); MSX1 (p = 2.875E-06, 7 genes); central nervous system development (p = 2.543E-06) RNGTGGGC UNKNOWN (p = 3.383E-05, 11 genes) 17 62 meiotic nuclear division (p = 1.035E-04); 7p15.3-p15.1 (p = 1.676E-03); meiotic cell cycle (p = 1.349E-04) 12q22-q24.1 (p = 1.676E-03) 19 55 cellular glucuronidation (p = 2.634E-05); 4q13 (p = 2.202E-04); uronic acid metabolic process (p = 3.005E-05) 4q31.3-q32 (p = 1.474E-03) 25 48 glutamate decarboxylation to succinate (p = 4.577E-06); 4q21.22 (p = 1.646E-06); glutamate catabolic process (p = 9.526E-05) 8p11.22 (p = 1.485E-04) 38 36 calcium ion export (p = 2.626E-05) 7q21.3 (p = 8.602E-06); 9p21.3 (p = 2.647E-04) 44 33 vasodilation of artery involved in baroreceptor response to increased systemic arterial blood pressure (p = 2.381E-06); 7p12.2 (p = 1.729E-05); RORA1 (p = 6.429E-03, 3 genes); baroreceptor response to increased systemic arterial blood pressure (p = 1.424E-05) 11p15.2-p15.1 (p = 9.241E-04) ERR1 (p = 7.357E-03, 3 genes) 50 30 fatty acid derivative metabolic process (p = 1.751E-07); 4q28-q32 (p = 1.673E-03) WGTTNNNNNAAA UNKNOWN  (p = 2.278E-03, 4 genes); icosanoid metabolic process (p = 1.751E-07) FOXO4 (p = 1.168E-02, 6 genes) 66 21 4p16.3 (p = 2.197E-9, 5 genes); 13 genes on 4p13-16 E2F1 (p = 9.854E-4, 3 genes) 67 21 9q21.33 (p = 5.973E-05); RACTNNRTTTNC UNKNOWN  (p = 3.031E-05, 3 genes) 9q22.32 (p = 1.652E-04); 18 genes on 9q21-34 70 20 1q22-q23.2 (p = 5.196E-04) 8q22-q23 (p = 1.038E-03) 81 18 21q22.3 (p = 2.655E-05) 84 18 4q31.23 (p = 6.329E-06); 4q31 (p = 1.242E-05); 12 genes on 4q23-31 CREB (p = 1.568E-04, 3 genes) 116 13 Xp11.23 (p = 6.422E-04) MEIS1 (p = 1.089E-03, 3 genes) Discussion One important issue is the biological mechanism leading to the differences in co-expression structures between the tumor and the control samples. As shown in Table 1, it is clear that there are multiple possible mechanisms. From the functional point of view, the first gene module (Module 4) is highly enriched in epidermis development function. This is consistent with the fact that lung cancer is an epithelial cancer. However the molecular mechanism for such difference is still not clear. While it is often expected that such difference may be due to difference in transcription factors (TFs) which co-regulate the co-expressed genes, our analysis (data not shown) on the enriched TFs shown in Table 1 did not reveal any statistically significant increase in level of the TFs in tumor samples. Another possible mechanism of co-expression is that the genes may lie on the same cytoband with copy number variations (CNV) among the tumor samples. We have indeed observed a few such gene modules including modules 66 (13 genes on 4p13-13), 67 (18 genes on 9q21-34), and 84 (12 genes on 4q23-31). The difference between the tumor and control samples implies that the potential CNV may be specific to the tumor. We tested the module 66 on TCGA lung cancer data using cBioPortal. In addition to the lung adenocarcinoma data with 230 patients, we also tested on the lung squamous cell data with 178 patients. Figure 4 shows the OncoPrint plots for the distribution of different types of mutations on the genes in module 66 in the patients.Fig. 4 The OncoPrint plots for different types of mutations on the genes in module 66 in the lung adenocarcinoma patients. Top: OncoPrint for genes in Module 66 (with 21 genes) in the lung adenocarcinoma study in TCGA generated generated by cBioPortal. Bottom: Oncoprint for the same gene module in lung squamous cell tumors in TCGA. The genes circled in red are all on cytobands 4p13-16 and the ones circled in blue are on cytoband 8p11.23 As shown in Fig. 4, the majority of the genes identified in Module 66 on cytobands 4p13-16 showed consistent CNV in lung cancer patients of both types. However, they are all amplifications in adenocarcinoma while mostly deletion in squamous cell tumors. To verify the relationship between the CNV and gene expression levels, we examined the correlations between the copy number measurements and the gene expression levels (measured using RNA-seq) of these genes and they all show positive correlations with an example (for the MRFAP1 genes) showing Fig. 5.Fig. 5 Correlations between the copy number measurements and the gene expression levels (measured using RNA-seq) of gene MRFAP1. Top: The box plot for the expression levels of gene MRFAP1 with respect to inferred copy number variation. Bottom: The correlation between the expression levels of MRFAP1 with the measurement for copy number values is 0.726 (PCC) In addition, the genes which are on close cytobands with similar CNV distribution in patients show strong co-expression as shown in Fig. 6 while the ones not on the same cytobands do not (data not show, the correlation ranges from 0.3 to less than 0). These observations suggest that the expression levels and co-expression of the genes on these cytobands are strongly associated with the CNV status of these bands. However, we also observed difference in correlation in the original dataset GSE18842 and the testing TCGA dataset. This could be partially due to difference in sample selections and measurement methods (GSE18842 data were generated using Affymetrix genechips while TCGA expression data were generated using RNA-seq).Fig. 6 Examples of co-expressed genes on the same cytobands from the same gene module. Top: The correlation between expression levels of MRFAP1 and GRPL1 is 0.650 (PCC). Bottom: The correlation between the expression levels of SLBP and GRPL1 is 0.606 (PCC) An additional interesting observation is that in both lung adenocarcinoma and squamous cell tumor samples, two genes from cytoband 8p11.23 show consistent copy number aberrations in the patients. While the mechanism for their co-expression with the ones on cytoband 4p1 is not clear, literature review shows that the gene TACC3 in Module 66 on cytoband 4p16 is known to have a gene fusion with FGFR1 gene in 3 % of glioblastoma multiforme patients [25]. FGFR1 gene happens to locate on 8p11.23-22. It is of great interest for future research to investigate if the relationship between the 4p16 and 8p11.23 is partially due to a gene fusion event. Conclusion In summary, we have developed a linear algebraic based index CCI for evaluating the concordance of co-expressed gene modules from gene co-expression network analysis. The CCI can be used to evaluate the performance for co-expression network analysis algorithms as well as for detecting condition specific co-expression modules. It is shown to be more robust to outliers and interfering modules than density based on Pearson correlation coefficients. We applied CCI in detecting lung tumor specific gene modules. The application revealed interesting potential tumor specific genetic alterations including CNVs and even hints for gene-fusion. Deeper analysis required for understanding the molecular mechanisms of all such condition specific co-expression relationships. Acknowledgement This work is partially supported by NCI (U01 CA188547 to KH), the National Natural Science Foundation of China (61572265 to ZH) and the Ohio Supercomputer Center. Declarations The publication costs for this article were funded by NCI U01 CA188547 grant, the National Natural Science Foundation of China (61572265 to ZH) and the OSU Startup Grant to KH. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials All the datasets used in this study were obtained from public sources as described in the Methods section. Authors’ contributions KH and JZ conceived of the study and collected the data. ZH performed the computational coding and implementation. ZH, JZ, GS and GL conducted data analysis. KH drafted the manuscript, JZ and ZH edited the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Pujana MA Han J-DJ Starita LM Stevens KN Tewari M Ahn JS Rennert G Moreno V Kirchhoff T Gold B Assmann V Elshamy WM Rual J-F Levine D Rozek LS Gelman RS Gunsalus KC Greenberg Ra Sobhian B Bertin N Venkatesan K Ayivi-Guedehoussou N Solé X Hernández P Lázaro C Nathanson KL Weber BL Cusick ME Hill DE Offit K Network modeling links breast cancer susceptibility and centrosome dysfunction Nat Genet 2007 39 1338 1349 10.1038/ng.2007.2 17922014 2. Kais Z Barsky SH Mathsyaraja H Zha A Ransburgh DJR He G Pilarski RT Shapiro CL Huang K Parvin JD KIAA0101 interacts with BRCA1 and regulates centrosome number Mol Cancer Res 2011 9 1091 1099 10.1158/1541-7786.MCR-10-0503 21673012 3. Zhang J Lu K Xiang Y Islam M Kotian S Kais Z Lee C Arora M Liu H-W Parvin JD Huang K Weighted frequent gene co-expression network mining to identify genes involved in genome stability PLoS Comput Biol 2012 8 e1002656 10.1371/journal.pcbi.1002656 22956898 4. Kotian S Banerjee T Lockhart A Huang K Catalyurek UV Parvin JD NUSAP1 influences the DNA damage response by controlling BRCA1 protein levels Cancer Biol Ther 2014 15 533 543 10.4161/cbt.28019 24521615 5. Bhardwaj N Lu H Co-expression among constituents of a motif in the protein-protein interaction network J Bioinform Comput Biol 2009 7 1 17 10.1142/S0219720009003959 19226657 6. Sun Y Li H Liu Y Mattson MP Rao MS Zhan M Evolutionarily conserved transcriptional co-expression guiding embryonic stem cell differentiation PLoS One 2008 3 e3406 10.1371/journal.pone.0003406 18923680 7. Hu H Yan X Huang Y Han J Zhou XJ Mining coherent dense subgraphs across massive biological networks for functional discovery Bioinformatics 2005 21 i213 i221 10.1093/bioinformatics/bti1049 15961460 8. Zhang J Xiang Y Ding L Keen-Circle K Borlawsky TB Ozer HG Jin R Payne P Huang K Using gene co-expression network analysis to predict biomarkers for chronic lymphocytic leukemia BMC Bioinformatics 2010 11 Suppl 9 S5 10.1186/1471-2105-11-S9-S5 9. Zhang J Ni S Xiang Y Parvin JD Yang Y Zhou Y Huang K Gene Co-expression analysis predicts genetic aberration loci associated with colon cancer metastasis Int J Comput Biol Drug Des 2013 6 60 71 10.1504/IJCBDD.2013.052202 23428474 10. Xu Y, Duanmu H, Chang Z, Zhang S, Li Z, Li Z, Liu Y, Li K, Qiu F, Li X. The application of gene co-expression network reconstruction based on CNVs and gene expression microarray data in breast cancer. Mol Biol Rep. 2011;39(2):1627–1637. 11. Langfelder P Horvath S WGCNA: an R package for weighted correlation network analysis BMC Bioinformatics 2008 9 559 10.1186/1471-2105-9-559 19114008 12. Li A Horvath S Network module detection: affinity search technique with the multi-node topological overlap measure BMC Res Notes 2009 2 142 10.1186/1756-0500-2-142 19619323 13. Yongbin Ou C-QZ A new multimembership clustering method J Ind Manag Optim 2007 3 619 624 10.3934/jimo.2007.3.619 14. Xiang Y Zhang C-Q Huang K Predicting glioblastoma prognosis networks using weighted gene co-expression network analysis on TCGA data BMC Bioinformatics 2012 13 Suppl 2 S12 22536863 15. Zhang J, Huang K: Normalized lmQCM : an algorithm for detecting weak quasi-clique modules in weighted graph with application in functional gene cluster discovery in cancer. Cancer Inform. 2016. In press. 16. Ma H Schadt EE Kaplan LM Zhao H COSINE: COndition-SpecIfic sub-NEtwork identification using a global optimization method Bioinformatics 2011 27 1290 8 10.1093/bioinformatics/btr136 21414987 17. Lai Y Wu B Chen L Zhao H A statistical method for identifying differential gene-gene co-expression patterns Bioinformatics 2004 20 3146 55 10.1093/bioinformatics/bth379 15231528 18. Li K-C Genome-wide coexpression dynamics: theory and application Proc Natl Acad Sci U S A 2002 99 16875 16880 10.1073/pnas.252466999 12486219 19. Abu-Jamous B Fa R Roberts DJ Nandi AK UNCLES: method for the identification of genes differentially consistently co-expressed in a specific subset of datasets BMC Bioinformatics 2015 16 184 10.1186/s12859-015-0614-0 26040489 20. Kalluru V Machiraju R Huang K Identify condition-specific gene co-expression networks Int J Comput Biol Drug Des 2013 6 50 9 10.1504/IJCBDD.2013.052201 23428473 21. Marsaglia G Bounds For The Rank Of The Sum Of Two Matrices 1964 22. Sanchez-Palencia A Gomez-Morales M Gomez-Capilla JA Pedraza V Boyero L Rosell R Fárez-Vidal ME Gene expression profiling reveals novel biomarkers in nonsmall cell lung cancer Int J Cancer 2011 129 355 64 10.1002/ijc.25704 20878980 23. Collisson EA Campbell JD Brooks AN Berger AH Lee W Chmielecki J Beer DG Cope L Creighton CJ Danilova L Ding L Getz G Hammerman PS Neil Hayes D Hernandez B Herman JG Heymach JV Jurisica I Kucherlapati R Kwiatkowski D Ladanyi M Robertson G Schultz N Shen R Sinha R Sougnez C Tsao M-S Travis WD Weinstein JN Wigle Da Comprehensive molecular profiling of lung adenocarcinoma Nature 2014 511 543 50 10.1038/nature13385 25079552 24. The Cancer Genome Atlas Network Comprehensive genomic characterization of squamous cell lung cancers Nature 2012 489 519 25 10.1038/nature11404 22960745 25. Singh D Chan JM Zoppoli P Niola F Sullivan R Castano A Liu EM Reichel J Porrati P Pellegatta S Qiu K Gao Z Ceccarelli M Riccardi R Brat DJ Guha A Aldape K Golfinos JG Zagzag D Mikkelsen T Finocchiaro G Lasorella A Rabadan R Iavarone A Transforming fusions of FGFR and TACC genes in human glioblastoma Science (80-) 2012 337 1231 1235 10.1126/science.1220834

PMC005xxxxxx/PMC5001232.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27535545279810.1186/s12864-016-2798-8ResearchComputing energy landscape maps and structural excursions of proteins Sapin Emmanuel 1Carr Daniel B. 2De Jong Kenneth A. 13Shehu Amarda amarda.@gmu.edu 1451 Department of Computer Science, George Mason University, 4400 University Drive, Fairfax, 22030 VA USA 2 Department of Statistics, George Mason University, 4400 University Drive, Fairfax, 22030 VA USA 3 Krasnow Institute for Advanced Study, George Mason University, 4400 University Drive, Fairfax, 22030 VA USA 4 Department of Bioengineering, George Mason University, 4400 University Drive, Fairfax, 22030 VA USA 5 School of Systems Biology, George Mason University, 10900 University Boulevard, Manassas, 20110 VA USA 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.546© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Structural excursions of a protein at equilibrium are key to biomolecular recognition and function modulation. Protein modeling research is driven by the need to aid wet laboratories in characterizing equilibrium protein dynamics. In principle, structural excursions of a protein can be directly observed via simulation of its dynamics, but the disparate temporal scales involved in such excursions make this approach computationally impractical. On the other hand, an informative representation of the structure space available to a protein at equilibrium can be obtained efficiently via stochastic optimization, but this approach does not directly yield information on equilibrium dynamics. Methods We present here a novel methodology that first builds a multi-dimensional map of the energy landscape that underlies the structure space of a given protein and then queries the computed map for energetically-feasible excursions between structures of interest. An evolutionary algorithm builds such maps with a practical computational budget. Graphical techniques analyze a computed multi-dimensional map and expose interesting features of an energy landscape, such as basins and barriers. A path searching algorithm then queries a nearest-neighbor graph representation of a computed map for energetically-feasible basin-to-basin excursions. Results Evaluation is conducted on intrinsically-dynamic proteins of importance in human biology and disease. Visual statistical analysis of the maps of energy landscapes computed by the proposed methodology reveals features already captured in the wet laboratory, as well as new features indicative of interesting, unknown thermodynamically-stable and semi-stable regions of the equilibrium structure space. Comparison of maps and structural excursions computed by the proposed methodology on sequence variants of a protein sheds light on the role of equilibrium structure and dynamics in the sequence-function relationship. Conclusions Applications show that the proposed methodology is effective at locating basins in complex energy landscapes and computing basin-basin excursions of a protein with a practical computational budget. While the actual temporal scales spanned by a structural excursion cannot be directly obtained due to the foregoing of simulation of dynamics, hypotheses can be formulated regarding the impact of sequence mutations on protein function. These hypotheses are valuable in instigating further research in wet laboratories. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2798-8) contains supplementary material, which is available to authorized users. Keywords Protein equilibrium dynamicsMulti-state proteinMulti-basin energy landscapeEnergy landscape mapSample-based representationEvolutionary algorithmStructural excursionMechanical workNearest-neighbor graphLow-cost pathsIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Experimental, theoretical, and computational studies have shown that protein function is the result of a complex yet precise relationship between protein structure and dynamics [1–3]. Long gone are the days when proteins were viewed as rigid molecules [4], with the atom nuclei frozen in specific positions in the three-dimensional (3D) structural models captured by X-ray crystallography [5]. Nowadays, wet-laboratory techniques based on single-molecule fluorescence spectroscopy provide irrefutable evidence of proteins as macromolecules in perpetual motion [2], even catching proteins in the act of switching between different structures to bind different molecular partners [6]. The ability of a protein to switch between different structures under physiological conditions (at equilibrium) is key to biomolecular recognition and function modulation [7]. This finding warrants characterizing the equilibrium structural dynamics of a protein as a means of exposing the range of activities of a protein in the cell [8]. While significant advances have been made in the wet laboratory [6, 9–12], existing techniques are in principle limited by the disparate spatio-temporal scales involved in protein dynamics; proteins undergo small (sub-angstrom) and large (>10Å) structural changes at different temporal scales, spanning from a few femto-seconds to milli-seconds and more [13]. Dwell times of proteins in specific structural states may be too short to be detected in the wet laboratory. Computational methods that simulate the constrained dynamics of the bonded atomic particles in a protein molecule via iterative application of Newton’s laws of motions are appealing. By following motions of atoms along the negative gradient of a molecular mechanics force field, these methods, also known as Molecular Dynamics (MD) methods, directly simulate structural excursions of a protein [14]. Since energy landscapes are highly multi-dimensional [15] (directly related to both independent and concerted motions of thousands or more atoms comprising a protein molecule), MD methods have to be operated in a random-restart fashion to sufficiently explore the structure space accessed by a protein at equilibrium. Typical computational efforts can exceed several weeks on large (several-hundred core) supercomputers [16] for medium-size proteins (100−300 amino-acids long), though advances in hardware and specialized architectures are beginning to broaden the scope and scale of MD methods to larger macromolecular assemblies and even viral capsids [17, 18]. The challenges regarding characterization of equilibrium protein dynamics are better understood from a protein energy landscape perspective, which links protein structure, dynamics, and function [19]. Briefly, measuring the extent to which a structure satisfies the (physical) constraints that atoms place on one another allows one to associate an energy landscape with the structure space of a protein. Structural excursions of a protein at equilibrium correspond to hops between energy basins in the landscape [20]. A basin, visually corresponding to a valley in the energy landscape, contains structures with similar energies. The set of structures mapped to the same basin represent a particular protein state. These states can be thermodynamically-stable or semi-stable, depending on the width and depth of the corresponding basin. A protein may spend more time in a wider and deeper basin (the state is stable) than in a narrower and shallower basin (the state is semi-stable) [20]. Energy barriers between basins regulate the time it takes for a protein to switch between basins [7, 20]. The interested reader is referred to works in [1, 21] for detailed reviews of energy landscapes and motions of proteins. The energy landscape view clarifies why a complete and detailed account of protein equilibrium dynamics is a non-trivial task. In principle, the task requires a comprehensive characterization of both the protein structure space and the underlying energy landscape that governs the accessibility of structures and excursions between them at equilibrium. While wet-laboratory studies may not catch semi-stable structural states (due to insufficiently long dwell times), computational approaches that simulate protein dynamics quickly become computationally intractable for even moderate-size proteins. In this paper we present a novel computational methodology that takes a complementary approach to the MD-based approach. While the goal remains to elucidate the equilibrium dynamics of a protein by computing structural excursions at equilibrium, the dynamics are not simulated directly. Instead, a two-step approach is followed. First, stochastic optimization (randomized/stochastic search) is employed to explore a protein’s structure space and construct a map of the energy landscape relevant for equilibrium dynamics. Second, the map is analyzed and queried for paths of intermediate structures that link two structural states of interest, effectively yielding an ensemble of paths that provide an on-demand view of the equilibrium dynamics relevant to a specific structural excursion. This two-step approach foregoes any direct information on the temporal scales involved, as there is no notion of physical time in the computed structures and paths connecting them. However, by doing so, the computational demands become much more reasonable; for instance, investigations of medium-size proteins of 150 or more amino acids can be conducted on small clusters (of no more than 16 CPUs) in no more than a few days (ranging from 7 to 15). Moreover, during this process, close to a million structures are generated, embedded in multi-dimensional maps of energy landscapes, and available to answer queries about energetically-feasible structural excursions at equilibrium between any two structures of interest. The advantages of stochastic over systematic search to explore high-dimensional variable spaces have been demonstrated in various domains. In protein structure modeling, algorithms that navigate the structure space of a protein via the Monte Carlo (MC) approach have been shown to have higher exploration capability than MD-based ones [22]. Furthermore, evolutionary algorithms (EAs) have been shown to provide significant improvements over MC-based algorithms [23, 24]. Specifically, for de novo protein structure prediction, EAs with domain-specific insight have been shown to rapidly locate the global minimum and reproduce the native structure [25, 26]. However, when the focus is on multi-state proteins with complex multi-basin energy landscapes, the objective goes beyond rapidly locating one structural state and requires an exploration of the breadth of the structure space. Recent evolutionary search techniques have advanced efforts in this direction [27–29]. A key starting point of recent work is the increasingly rich set of structural data for both the wildtype (WT) and variants of multi-state proteins being deposited by wet-laboratory scientists in the Protein Data Bank (PDB) [30]. Work in [31] has shown that a statistical characterization of this structural information provides important and useful information about the structure space of a particular protein. A simple, generational EA operationalizes this idea in [27]. Work in [28] enhances the exploration capability of the EA while still operating under the classic optimization setting of finding the global minimum (via a decentralized selection operator that delays population take-over by the most fit individuals). Work in [29] finally switches from the classic setting to that of obtaining a comprehensive map of a (multi-state) protein’s multi-basin energy landscape via the concept of the hall of fame. The resulting EA is shown able to map complex, multi-basin fitness landscapes beyond the protein modeling domain via careful combination of local and global search [32]. These collective algorithmic developments have now made it possible to build comprehensive and detailed maps of energy landscapes of medium-size proteins with a modest computational budget. The EA we employ for mapping protein energy landscapes in the methodology proposed here builds on all these previously-published evolutionary search techniques to effectively and efficiently map the structure space available to a protein at equilibrium. The “Methods” Section summarizes this EA for the sake of completeness, paying particular attention to those aspects that give it the ability to efficiently map energy landscapes of medium-size proteins. Analysis of maps computed to represent energy landscapes is non-trivial. Even when the focus is simply to locate basins, the analysis involves several hundred thousands of multi-dimensional data points that reside in a highly non-linear landscape. Past work [28, 29, 33] has relied on visual analysis of 2D projections of all structures ever computed during the execution of an EA or only those structures in the hall of fame/map. We make the case in the “Results” Section that such analysis is informative, but the projection can sacrifice possibly interesting energetic features in the multi-dimensional map. Hence, in this paper we utilize additional graphical techniques to visualize and analyze the computed multi-dimensional maps. The techniques reveal not only basins already captured in the wet laboratory, but also new energetic features indicative of interesting, unknown thermodynamically-stable and semi-stable regions of the equilibrium structure space. Mapping the energy landscape of a protein provides an opportunity to extract information on its equilibrium dynamics in much in the same way the map of a city allows extracting information on routes connecting landmarks. In previous work [28, 29, 33, 34], we have relied on qualitative summarizations of protein dynamics based on the location of energy barriers and other features of a mapped landscape, and how these features differ in the variant forms of a protein. Here we propose a procedure to extract information on the equilibrium dynamics of a protein by computing structural excursions between structures of interest. The procedure builds on ideas utilized in robotics-inspired methods for protein motion computation [35–37]. In these methods, structures are embedded in a nearest-neighbor graph (referred to as a roadmap), which is then queried for a path connecting a start to a goal structure structure. In this paper, the structures are those produced by an EA mapping process. That is, they constitute a comprehensive and detailed map of the energy landscape. Care has to be taken to embed them in a nearest-neighbor graph and utilize them for path queries. Moreover, unlike related work in robotics-inspired modeling, where the focus is typically on one path, the procedure proposed here reveals an ensemble of energetically-similar paths. This focus is warranted in order to obtain a broader view of the stochastic but energy-driven nature of protein structural excursions (and equilibrium dynamics). The methodology proposed in this paper to build maps, analyze them, and then query them for structural excursions is applied to several proteins of importance in human biology and disease. In addition, detailed comparison of the maps and path ensembles is conducted on the WT and 7 variant sequences of an oncogenic protein. This comparative setting evaluates the ability of the proposed methodology to explain the impact of mutations on protein equilibrium dynamics and in turn on misfunction. These results are presented in the “Results” Section, and a discussion of how they reproduce, explain, or further existing knowledge is provided in the “Discussion” Section. While the actual temporal scales spanned by a modeled structural excursion cannot be obtained by the proposed methodology due to the foregoing of simulation of dynamics, specific hypotheses can be formulated nonetheless regarding the impact of sequence mutations on function. These hypotheses are valuable in instigating further research on structure-function studies in wet laboratories. The advantages and disadvantages of the proposed methodology, as well as possible directions of further research, are summarized in the “Conclusions” Section, which concludes this paper. Methods The input to the proposed methodology consists of a protein sequence α, a set of structures SPDB representing stable structural states for sequences no more than 3 amino acids different from α, and (a pair of start and goal) structures of interest for a possible excursion. The methodology first performs a principal component analysis (PCA) on the structures in SPDB in order to define a low-dimensional representation of the protein structure space. An evolutionary algorithm (EA) is then applied to this PCA-defined space to construct a map representing the all-atom energy landscape of α. Finally, a path-finding algorithm uses this map to compute energetically-feasible paths realizing structural excursions of interest. The methodology is shown in pseudocode in Algorithm 1. Below we first relate details on the principle that allows utilizing structures in SPDB to define the (reduced) variable space underlying the structure space of a protein sequence of interest α, as well as describes the technique employed to do so (lines 1–2 in Algorithm 1). The EA that explores this variable space to build a multi-dimensional map of the all-atom energy landscape of α (line 3 in Algorithm 1) is then described. The graphical statistical techniques utilized to analyze a computed multi-dimensional map and reveal interesting energetic features, such as energy basins, are related afterwards. A description of the algorithm employed to build and query the map for energetically-feasible excursions of the target protein sequence α between two structures of interest (line 5 in Algorithm 1) concludes this section. Extracting variable axes to define a reduced protein structure space As mentioned in the “Background” Section, a key starting point that has recently allowed EAs to explore complex structures spaces of multi-state proteins is the ability to define variable spaces of reasonable dimensionality to represent protein structure spaces. These variable spaces are extracted based on a statistical characterization of the increasingly rich structural information available in the PDB for a protein sequence α and other (variant) sequences similar to it. The characterization is rooted in the principle of conformational selection, summarized next. Utilization of structures and the principle of conformational selection: Let us suppose a structure has been captured for a sequence β of a protein in the wet laboratory. This structure represents a thermodynamically-stable state for β. If β is a variant of a given protein (that is, within a few amino-acid mutations of some neighboring sequence α), then the structure that is stable for β may possibly be of low-energy in the structure space of some similar sequence α. This is in effect the principle of conformational selection [38], under which perturbations such as sequence mutations do not change a protein’s structure space but rather the probabilities (which in turn are related to energies) with which a given sequence is expected to populate the various structural states; in other words, even a structure detected for a variant is expected to be assumed by the WT (and vice-versa) but possibly with a different probability at equilibrium. In summary, known structures of different sequence variants of a protein represent stable and semi-stable structural states of a target sequence. Extracting variable axes via multivariate statistical analysis Structures in the set SPDB are first “converted” into structures of α (line 1 in Algorithm 1). The structures are stripped down to CA atoms (effectively discarding all atoms except the central carbon atom – CA atom – of each amino acid in the amino-acid/protein chain). A structure stripped down to the CA atoms is referred to as a CA trace. Since the CA traces corresponding to the set SPDB come from sequences possibly different (within a few mutations) from α, the amino-acid identities of the CA atoms are replaced with those in the target sequence α in each CA trace. The resulting traces are then subjected to a multivariate statistical analysis, PCA, originally described in [27], to extract new variable axes; these are the principle components (PCs) obtained from the PCA (line 2 in Algorithm 1). In summary, PCA yields new variable axes via an optimal rotation of the original axes that maximizes variance of the data along the new axes [39]s. Ordering of the new axes (PCs) by the variance of the data when projected onto them allows extracting a subset m that is typically much less than the original dimensionality of the data, if PCA is indeed effective. Work in [28] shows this to be the case for many multi-state proteins with multi-basin landscapes; with the top two PCs one captures more than 45 % of the variance (which means they can be employed for data visualization) and anywhere between 10−25 PCs allow capturing more than 90 % of the variance. The latter is a reduction by more than ten-fold, as the original structures are of proteins with more than 100 amino acids; stripping them down to their CAs prior to PCA exposes more than 300 Cartesian coordinates on which PCA operates to reveal no more than 25 PCs/coordinates that still capture more than 90 % of the variance. The variance-ordered PCs are used as variables ({PC1,…,PCm}) through which to represent a structure. As described in [27], a structure can be represented as an m-dimensional point whose coordinates are projections over the m axes (obtained via essentially a dot-product operation with each of the axes). The reverse is also possible. Given an m-dimensional point, a process that essentially depends on a linear combination of the axes yields 3D coordinates of the CA atoms of the structure corresponding to the point. Going back and forth between the variable space and the structure space of a given protein sequence makes it computationally feasible to map and query the structure space of a protein by techniques that operate on the variable space. Next we describe the EA that explores this variable space to build a PCA-based map of the all-atom energy landscape of a given protein sequence α. The map is analyzed and queried for paths by techniques that also operate on the variable space. EA building of a multi-dimensional energy landscape map The EA employed here to map a protein energy landscape is the result of a series of recent works [27–29, 32, 34] that carefully and gradually investigate the impact of various design and implementation decisions regarding the exploration versus exploitation capability of EA-based stochastic search in multi-basin protein energy landscapes. At a conceptual level, the EA evolves a fixed-size population of individuals over generations towards better-fit individuals. Individuals are points in an m-dimensional space whose variable axes are the top variance-ordered PCs obtained as described above. The fitness of an individual in the EA is evaluated via the Rosetta score12 energy function, which measures the all-atom energy of the 3D protein structure corresponding to the individual. The EA is memetic, as an offspring individual obtained by varying a parent individual is subjected to improvement. This is particularly important for individuals that represent molecular structures in order to reduce the number of constraints violated in offspring. An improved offspring is then considered for addition to the map, which is thus dynamically updated during the evolutionary process. Algorithm 2 summarizes the EA in pseudocode. Rather than specifying a budget in terms of a total number of generations, the algorithm exhausts a total number of fitness or energy evaluations (line 4 in Algorithm 2), as these are the most computationally-demanding step of any algorithm manipulating molecular structures. Once the budget is exhausted, the map is outputted. For completeness, we provide more details of the EA in what follows, paying particular attention to the shaded boxes in Algorithm 2 that constitute the main functional units of the EA. It is worth noting that these units make use of various parameters. In the interest of clarity, these parameters are not listed in Algorithm 2, but we describe them in context and list their values when relating implementation details. Initialization mechanism to seed the EA Proper initialization is key to exploration. As mentioned above, the CA traces extracted from SPDB and “threaded” onto the sequence of interest α are the first to be added to the initial population (line 3 in Algorithm 2); the traces are first projected onto the m PCs so as to obtain individuals corresponding to them in the variable space. Prior work has considered various strategies to fill in the rest of the population; typically, a higher exploration capability is obtained as the population size increases from 500 to 2,000 individuals (we use 2,000 in this work), and the number of PDB-obtained structures can be significantly smaller than this target population size. In [27, 28], the rest of the population is filled by individuals obtained as offspring of the CA traces via the variation operator (described below). A comprehensive analysis in [29] compares this strategy to two others, one where the rest of the population is filled by individuals drawn at random in the space of the m PCs, and another where the initial population does not make use of any of the experimentally-known structures but consists of only individuals drawn at random in the variable space. Comparison on the average fitness and average diversity (measured via Euclidean distance in the variable/PC space) of a population over generations demonstrates that the strategy where the initial population consists of individuals derived from the experimentally-known traces and individuals drawn at random provides a better balance between exploitation (improvement in average fitness over generations) and exploration (retainment of diversity over generations). In the results described in the “Results” Section, this strategy is employed to seed the EA and obtain the energy landscape maps of various protein sequences. Obtaining offspring via a variation operator As line 6 in Algorithm 2 indicates, each parent p yields an offspring c. Variation is introduced in each population through a variation operator (line 7 in Algorithm 2) described in detail in [27, 28]. In summary, a vector is first defined in the PC space; its elements are magnitudes of movement along each of the m PCs. The magnitude of the movement along the top PC (that captures the most variance) is sampled uniformly at random in the segment [ −s, s], where s is a user parameter. The magnitudes of the movements along the other PCs respect their variance relative to the variance captured by the first/top PC. While the shape of the space is preserved, the boundaries of the m-dimensional embedding of the wet-laboratory traces are not observed, as the ultimate goal is to generate individuals that represent new structures not captured in the wet laboratory for the target sequence α. Fixed versus variable budget improvement operator The obtained offspring c is subjected to an improvement operator to obtain a better offspring c~ (line 8 in Algorithm 2). The process consists of three steps. First, the offspring, which is a point in the m-dimensional PC space, is converted into a set of backbone atoms with coordinates in 3D. This step consists of recovering the CA trace via simple algebra operations (detailed in [28], and then recovering the backbone skeleton from the CA trace via the BBQ backbone reconstruction protocol [40]. The next step subjects the backbone skeleton to the Rosetta relax protocol [41]. This protocol is open-source and written in C/C++, which allows easy integration in the EA. The protocol repeatedly guesses coordinates for the side-chain atoms (utilizing the target sequence α in the backbone structure fed to it as input) and improves them via a simulated annealing MC search. The result is a 3D structure for all atoms (backbone and side chains) that corresponds to a local minimum in the (Rosetta score12) all-atom energy surface of α. In the third step, the improved individual c~ corresponding to the resulting structure is obtained. The CA trace is extracted from the structure, and the trace is projected back onto the space of PCs to obtain c~. The all-atom Rosetta score (score12) is recorded and associated with the c~. The fact that it is the improved offspring c~ and not c that is added to the set of offspring in line 13 in Algorithm 2 is what makes the EA shown in Algorithm 2 a Lamarckian EA. In prior work [27–29], a fixed number NrImprovItersMax of iterations of the MC search have been utilized in the improvement operator. Since each iteration exhausts one energy evaluation, the budget of energy evaluations can be effectively wasted by attempting to improve sub-optimal offspring. Recent work in [34] introduces a variable-budget improvement operator, which allocates iterations/energy evaluations based on the promise of an offspring for further improvement. The improvement operator spends only one iteration at a time on improving an offspring c until a maximum NrImprovItersMax has been reached on the lineage from a parent to the currently improved offspring. The neighborhood of the currently improved offspring in the Map is analyzed and compared in terms of average fitness to the fitness/energy of the offspring, and a determination is made (via an empirically-determined relationship) on whether the improvement should be terminated prior to reaching the maximum number of iterations. The relationship also determines whether the improved offspring ought to be considered for addition to the Map or not (lines 9–10 in Algorithm 2). If not, the lineage is penalized, as well, so as to remember that this specific region in the variable space ought not to considered further. Lines 11–12 in Algorithm 2) show that the parent of the terminated offspring is replaced with a new individual. While work in [34], generates the new individual at random in the variable space, here we consider an alternative strategy; two parent individuals are selected at random and crossed over (utilizing one-point crossover) to obtain the new individual. These two different strategies are compared in the first set of results related in the “Results” Section. A Sample-based map of a protein energy landscape via a hall of fame A large population is critical to capture a possibly large set of local minima in a rugged energy landscape. Maintaining all individuals ever generated in memory is not practical; nor is it effective, as many individuals generated during the execution of the EA may be highly structurally-similar. What is needed is a map with a tunable resolution. Work in [29] proposes utilizing the concept of a hall of fame to serve as a dynamically-updated, resolution-tunable map of a protein’s energy landscape. The hall of fame is an evolutionary strategy to equip an EA with memory. The algorithm invoked to update it is shown in pseudocode in Algorithm 3. As Algorithm 3 shows in lines 1–2, if the fitness f(c) of the individual c considered for addition to the map is not below a threshold fitThreshold, c is not considered (reflecting the objective to update the map with fit individuals). Otherwise, c is considered (line 3) and then compared to neighboring individuals C in the map (line 4). If a neighboring individual C whose Manhattan distance (in the space of m top PCs) falls below the threshold distThreshold but has higher fitness than the fitness of c, then the individual is replaced by c (lines 5–11). If c is similar but does not reside deeper in the local minimum containing C (lines 8–9), c does not replace C. Note that if c is not similar to any other individual in the map, it is added, as it represents a new region not currently present in the map. The idea is to update the map with individuals that may represent the same region in the variable space but allow further exploitation of a local minimum and with fit individuals representing novel regions. The distThreshold represents a resolution, as the map is a set of distinct local minima individuals (obtained after improvement) separated by at least the defined threshold distThreshold in the space of PCs. Increasing distThreshold makes the map sparser. Lowering it, provides more detail but also increases the number of individuals in the map. Selection operator Line 15 in Algorithm 2 invokes the selection operator, where offspring compete with parents for survival. A comparative analysis of various implementations in [28] suggests that a local/decentralized selection operator, where each offspring competes only with parents in a given neighborhood, stalls take-over of the population by most-fit individuals, thus delaying premature convergence. The neighborhood captures the notion of structural similarity, so that offspring only replace structurally-similar parents if they lower Rosetta score12 energy. Structural similarity is determined efficiently by embedding individuals in an explicit 2D grid over the top two PCs. Cell width is also a user-defined parameter, and values employed here for the construction of the grid are those suggested to be optimal by the comparative analysis in prior work [28]. In recent work [29], a modification is proposed to the local selection operator, which we employ here in applications of the EA analyzed in the “Results” Section. If an offspring does not have any parent individuals in its neighborhood, it survives and is included in the population for the next generation; in prior work [28], such an offspring would compete with all parents. Analysis of a multi-dimensional map via graphical statistical techniques Projections of the multi-dimensional maps onto 2D, while informative (as related in the “Results” Section, may hide interesting energetic features that only appear along the remaining axes. Graphical techniques for visualization of multi-dimensional data are employed here to elucidate interesting energetic features hidden along the different dimensions of the variable space explored by the EA. In all the proteins investigated here, the top 4 PCs capture about 80 % of the dynamics. Therefore, hidden energetic features are sought on at most 4D projections of the computed maps (PC1-PC2-PC3-PC4) by way of two-way conditioned plots. Two-way conditioned plots provide a way to obtain insight in data patterns related to a 4D domain. Such graphics have a substantial history and are alternatively referred to as multi-window displays, casement displays and co-plots [42–45]. The basic idea is to focus on plots of two variables at a time, conditioning on the other two variables so the basic view is a function of the other variables (or not). Let us refer to the former the primary variables, and the latter as the conditioned-upon variables. Since the PCs are ordered by variance, PC1 and PC2 are used as the primary variables, leaving PC3 and PC4 to be the conditioned-upon variables. In the two-way conditioned plots we employ to visualize the map along essentially the top 4 PCs, the data is partitioned in 16 subsets that are quartile intervals for PC3 and PC4. Let us consider a specific quartile, Q i for PC3 and Q j for PC4. The m-dimensional individuals in the map are then visualized as follows. All coordinates of an individual along PC5 and on are discarded, and the only individuals retained are those whose third coordinate falls in Q i of PC3 and fourth coordinate falls in Q j of PC4. This subset resides in a 4D space. In effect, considering the fitness value of each individual adds a fifth dimension. These individuals are visualized in a 2D plot as follows. They are binned in hexagon bins, a popular idea in visualization of multi-dimensional data introduced in [46]. Only the lowest-energy (best) individual is then visualized per bin, plotting it as a 2D point along PC1 and PC2, and color-coding it based on its energy. A blue-to-red color-scheme is employed corresponding to low-to-high energy values. It is worth noting that the conditioned-plot approach to multi-dimensional data visualization sacrifices much of the resolution of the conditioning (partitioning) variables while retaining much of the resolution for the variables used in the plots. The comparison of juxtaposed plots, however, provides valuable insight into the impact of the conditioning variables. As the results in the “Results” Section relate, a layout of 16 color-coded, hexagon-binned, two-way conditioned plots (16 by combination of each of the quartiles of PC3 with quartiles of PC4) provides an effective way to visualize a 5D view of the maps of energy landscapes constructed by the above-described EA. In particular, the layout allows visualizing how basins elongate along the other dimensions, and where along these dimensions they populate regions not captured in wet laboratories. Graph-based query of map for energetically-feasible structural excursions The map that the EA described above constructs to represent an energy landscape is essentially a (hall of fame) list of (multi-dimensional) individuals with associated fitness values (Rosetta score12 energies). In order to query the map for ensembles of energetically-feasible structural excursions between any two structures of interest for a protein at hand, the map is first converted into a nearest-neighbor graph. Details are related below. After the map is essentially equipped with connectivity information, any informed graph search algorithm can be employed to query the map for energetically-feasible paths. Below we describe how the nearest-neighbor graph representation of the map is provided with energy-based weights, and how Dijkstra’s shortest-path algorithm [47] is then employed to extract a lowest-cost path connecting two structures of interest from a graph of close to a million vertices. Finally, the rest of the “Methods” Section describes how Dijkstra is employed in an iterative fashion to obtain an ensemble of low-cost paths in order to provide a broader picture of energetically-feasible structural excursions. A nearest-neighbor graph representation of the energy landscape map The map is converted into a nearest-neighbor graph G=(V,E) as follows. The individuals in the map populate V. Each vertex is then connected via edges to k other vertices that are its nearest neighbors. Euclidean distance in the m-dimensional variable space is used to measure the proximity between two vertices/individuals. The computation of nearest neighbors can be potentially a time-consuming step, but nearest-neighbor search data structures, such as a kd-tree [48], provide a remedy, particularly when the number N of data points is much larger than 2m (that is, N>>2m), where m is the dimensionality of the variable space [49]. We employ a process similar to how the kd-tree organizes data points to support fast nearest-neighbor queries. Specifically, Euclidean distance calculations are terminated earlier than considering all variable axes if the distance already surpasses a dynamic threshold (the latter is updated as neighbors are found). Since the set V can be very large (recall that the distThreshold parameter in the map construction can allow for a highly-detailed map with millions of individuals), the number of nearest neighbors for a vertex is limited to k=8. That is, the branching factor for the graph is limited to 8. The graph is directed; a vertex v may be among the k-nearest neighbors of a vertex u, but u may not be among the k nearest-neighbors of v. While edges are added based on essentially a proximity relationship between vertices, weights or costs associated with them are based on the following energy-/fitness-based relationship: Cost(e=(u,v))=max{score12(v)−score12(u),0}. The idea behind this is as follows: If the directed edge e=(u,v) lowers the energy of a protein hopping from u to v, then this particular u→v excursion does not require additional energy, as it is a down-hill movement in the landscape. Down-hill movements occur instantaneously per thermodynamics; no build-up of energy is needed to allow the excursion to take place. On the other hand, an up-hill movement, where score12(v)>score12(u), requires the system to build enough energy in order to cross what is essentially an energy barrier. This way of associating costs with edges is based on the principle of mechanical work, as the cost that would be tallied up with a path of edges would essentially sum only up-hill movements in the landscape; that is, only keep track of the total amount of external work that needs to be performed to give energy to fund such movements. This way of associating costs with edges is shown to assess the relevance of a lowest-cost path as a representative of a structural excursion better than an alternative approach based on the integral cost along the path [50] (and has been used by us before on robotics-inspired protein motion computation [35]). Querying nearest-neighbor graph for low-cost paths Given two structures Sstart and Sgoal, the nearest-neighbor graph can be queried for a path as follows. First, the two structures are projected onto the m-dimensional variable space, and their projections vstart and vgoal are added to the vertex set. The vertex set is then inspected to find the nearest neighbor ustart to vstart and the nearest neighbor ugoal to ugoal. The directed edges (vstart,ustart) and (ugoal,vgoal) are then added to the set of edges in the graph, with weights are defined above. A path in the enhanced nearest-neighbor graph is then an ordered list of vertices 〈vstart,ustart,…,ugoal,vgoal〉. Dijkstra’s shortest path algorithm is used to compute the lowest-cost path vstart⇝vgoal. The following modification is carried out in order to produce a physically-realistic lowest-cost path. Since the vertices correspond to individuals obtained via randomized search, the sampling of the structure space is non-uniform; while some structures may have nearest neighbors in very close proximity, this cannot be guaranteed over all structures computed by the EA. Indeed, the most densely-sampled regions will be those in basins due to the nature of EAs. The unintended consequence of non-uniform sampling is that a structure (vertex) may be connected via an edge to a structure (vertex) far away in the structure space. Such connections are valid in the nearest-neighbor graph construction, but they do not provide physically-realistic information regarding structural transitions. Rather than place additional proximity constraints among a vertex and its k-nearest neighbors in the construction of the edge list of the graph, such constraints are imposed when querying the graph for paths; that is, the neighbors of a vertex are a subset of its k neighbors in the graph subjected to an additional proximity constraint. A user parameter is considered for this purpose, max_nn_dist (maximum nearest-neighbor distance), and values for this are generated by dividing the Euclidean distance between the individuals corresponding to the start and goal structures by values in the set {15,10,7.5}. The latter can be considered path resolutions, and in the “Results” Section we demonstrate the implication of the resulting different values for max_nn_dist. In summary, a large value allows making large hops in the variable space and associating non-credible costs, such as would be obtained by directly connecting two nearby basins without considering the energetic barrier in between the basins (the equivalent of tunneling through an invisible mountain). A small value is conservative, making much smaller hops and effectively is impacted by the ruggedness of the energy landscape. Very small values of max_nn_dist may result in no paths at all, as no nearest neighbors can be found to meet a very conservative distance criterion. Dijkstra’s algorithm can be run in an iterative manner to produce more than the lowest-cost path. Once the lowest-cost path is computed, the intermediate vertices (excluding start and goal) in the path are removed from the graph, together with their edges. The remaining graph is queried again for the lowest-cost path, and this process is continued, removing intermediate vertices after identifying a path, until no more paths can be found; that is, the start and goal are now in different connected components. The result of this iterative process is an ensemble of low-cost paths, which are analyzed in the “Results” Section to obtain summary statistics regarding energetically-similar structural excursions of a protein. Implementation details The algorithms for map building and querying are implemented in C/C++, whereas the graphical techniques for analysis of a built map are implemented in R. The EA is run until the budget of 1,000,000 Rosetta score12 evaluations is exhausted. Population size in the EA is 2,000 individuals. A preliminary analysis in [29] also shows that this population size, combined with the initialization strategy described in above, injects greater diversity in the initial population. The target cumulative variance to obtain m PCs is set at 90 %, as in prior work. The step size s in the variation operator is set to 1, and NrImprovItersMax in the improvement operator is set to 5. In the map update, fitThreshold is set to 0 Rosetta Energy Units (REUs) for most proteins. For CaM, where Rosetta heavily penalizes non-compact structures, fitThreshold is set to 250 REUs. Also in the map update, distThreshold is set to be twice the minimum Manhattan distance between two wet-laboratory structures of a protein under consideration. In the variable-budget improvement operator, neighbors of an offspring in the map are individuals no more than 1 unit away in Manhattan distance. Prior work on the selection operator indicates that C25 and C49 are reasonable choices that delay premature convergence [28]. Similarly, reasonable values for the grid cell width vary from 1−2 for small proteins less than 100 amino acids and 10 for other longer proteins. The EA is run on a 16 core red hat Linux box with 3.2 GhZ HT Xeon CPU and 8GB RAM. The cores are employed to parallelize offspring improvements. This results in significant time savings. The experiments reported here are carried out on a 16-core platform, but, since the distribution is embarrassingly parallel, more time savings can be obtained with more cores. Results Test cases and experimental setup The proposed methodology is applied to 10 protein sequences, and performance is evaluated both in terms of running time and quality of the maps and structural excursions modeled on each sequence in relation with existing wet- and dry-laboratory evidence on known features of the energy landscapes and equilibrium dynamics. Test cases The selected test cases are proteins of importance in human biology and with a significant number of structures in the PDB [30]. They are the the superoxide dismutase [Cu-Zn] (SOD1), Calmodulin (CaM), and the WT and disease-related variant forms of the catalytic domain of uncomplexed H-Ras (to which we refer as H-Ras from now on). SOD1 is a 150 amino-acid long protein whose mutations have been linked to familial Amyotrophic lateral sclerosis (ALS) [51]. CaM is an enzyme 148 amino acids long that binds calcium and regulates over 100 target proteins, including kinases, phosphodiesterases, calcium pumps, and motility proteins [52–54]. H-Ras is a 166 amino-acid long protein that mediates signaling pathways that control cell proliferation, growth and development. H-Ras switches between two distinct structural states to regulate its biological activity [55]. Sequence mutations are implicated in various human cancers and other developmental disorders [56], and we study here several single and double mutants (7 variant sequences in all). Data collection and preparation Due to the implication of these proteins in various critical human diseases, ample structural data of their WT and mutated (variant) sequences exist in the PDB. Only X-ray structures are collected for H-Ras, whereas NMR structures are additionally included for SOD1 and CaM to enrich these datasets. The WT sequence of each of these proteins is obtained from UniProt [57]. Structures obtained from the PDB whose sequence changes by more than 3 amino acid from the WT sequence are discarded. Structures with missing internal amino acids are also discarded. Remaining structures are cropped at the termini, if necessary, so their lengths match the length of the WT. This protocol results in 186 wet-laboratory structures collected for SOD1 from the PDB, 86 for H-Ras, and 697 for CaM. As described in the “Methods” Section, application of PCA to these datasets yields a cumulative variance of 90 % at m=25, m=10, and m=10 PCs for SOD1, CaM, and H-Ras, respectively. A cumulative variance of 45−50 % is captured by the top two PCs on each of these proteins. In the interest of space, the cumulative variance profiles are not shown here, but they have been presented in prior work on analysis of the PC spaces for each of these proteins [27, 28]. Experimental setup The proposed methodology is applied to SOD1 (WT), CaM (WT), and 8 different sequences of H-Ras. The breakdown of the run time of the methodology on each of its components (map building, nearest-neighbor graph computation, and map querying) is shown via pie charts in Fig. 1. The run time of the EA and the size of the maps built on each test case are listed in Table 1. Analysis of the impact of the two different strategies described in the “Methods” Section on how to restart an unpromising lineage is carried out over 3 independent runs of the EA and is related first. Fig. 1 Run time profiling. The break-down of the run time along each of the three main components of the methodology is shown for SOD1, CaM, and H-Ras WT. The path query time refers to the proportion of running time spent on computing the lowest-cost path Table 1 Map build run time across SOD1, CaM, and H-Ras sequences Sequence |Map | Time (CPU Days) SOD1 669,102 13 CAM 170,570 9.5 H-Ras WT 890,391 9 H-Ras G12S 699,265 7 H-Ras G12C 704,610 10 H-Ras G12D 694,739 8 H-Ras G12V 649,006 7 H-Ras Q61L 602,893 7 H-Ras C32YS118C 693,567 8 H-Ras R164AQ165V 559,862 7 The analysis then focuses on maps and paths computed on each of the test cases. The analysis on SOD1 and CaM is conducted on color-coded 2D projections of the maps built for each protein and the structural excursions computed for each of them via map queries. This is related next. The rest of the analysis is on H-Ras, on which there is a wealth of structure data and disease-related mutations. The graphical techniques summarized in the “Methods” Section are applied to the multi-dimensional map generated for H-Ras WT to reveal in detail energetic features that are lost in a 2D projection. Path ensembles, computed as described in the “Methods” Section, are then visualized and analyzed for H-Ras WT and several single- and double-mutant variants. Summary statistics are juxtaposed to supplement the visual comparison of maps and paths. More results are related in the Additional files accompanying this paper. The “Discussion” Section summarizes all results presented on H-Ras to reconcile existing literature and further our understanding of the role of equilibrium structural dynamics on the link between mutations and misfunction in H-Ras variants. Exploration versus exploitation: restarting failed lineages with individuals generated at random or via crossover Two different settings are investigated to restart a failed lineage, generating a new individual at random in the variable space versus generating it via crossover of two parent individuals selected at random in the current population. The EA with each setting is run 3 times, and two measurements are tracked over generations. The first, the average fitness of the growing map (average over fitness values of individuals in the map at a given generation) estimates the exploitation power of the resulting EA. The second, the average diversity among individuals in the growing map (average value over all pairwise Euclidean distances over individuals in the map at a given generation). Figure 2 shows these two measurements for the H-Ras WT map over generations; the 99 % confidence intervals are also shown. Fig. 2 Crossover evaluation. The top panel shows the average fitness/energy among individuals in the hall of fame list (map), as the map is updated over generations. The bottom panel shows the average diversity among individuals in the hall of fame (measured via Euclidean distance) as the map is updated over generations Figure 2 shows that EA with crossover lowers both the average fitness and the average diversity of a growing map faster; that is, the crossover enhances exploitation but lowers exploration. The impact on exploitation is smaller, however, than the impact on exploration. Taken together, this analysis suggests that the EA, where a failed lineage is restarted with an individual generated at random, will be as effective in exploitation and more effective in exploration than when the individual restarting a lineage is generated via crossover. It is worth noting that the differences are not significant; this is expected, as crossover of two individuals that correspond to protein structures is likely to result in similar constraint violations as an individual generated at random in the variable space. The rest of the analysis on the proteins studied here employs the EA where failed lineages are restarted with individuals drawn at random in the variable space. Projection-based visualization and analysis of computed energy landscape maps and structural excursions for SOD1 and CaM One way to visualize computed multi-dimensional maps of energy landscapes is to project individuals in a map onto the top two PCs and color-code the projections based on the Rosetta score12 energy values; effectively, the 2D projection of a computed map is a 2D projection of the explored score12 all-atom energy landscape of a protein. Color-coded 2D projections of all individuals ever generated or individuals in a map have been employed by us before to conclude that low-energy regions of an explored protein energy landscape are co-located with projections of experimentally-known structures of a protein [29]; thus, suggesting the ability of a mapping EA operating in a reduced variable space to produce reliable maps of multi-basin energy landscapes. In the following, we show such projections for maps built for SOD1 and CaM. A lowest-cost path is also shown for each protein to demonstrate the ability of the proposed methodology to model structural excursions. Analysis of computed map and basin-basin excursions of SOD1 Figure 3 shows the color-coded 2D projection of the map built for SOD1 (WT sequence). The map contains two well-delineated basins. This two-basin feature is related to the phosphorylation event [58], grouping the experimentally-known structures (their PC1-PC2 projections are drawn as black dots) into one of the two basins. The map is queried for a structural excursion between the two basins. Two structures, one residing in each basin, are selected and provided as start and goal to the map querying algorithm in the proposed methodology. The lowest-cost path is computed with a value of max_nn_dist corresponding to about 5.43Å /15 (where 5.43Å is the least-root-mean-squared-deviation – lRMSD – between the two structures, and 15 relates to the sought path resolution (as described in the “Methods” Section. The query is successful; the succession of structures in the path is shown in Fig. 3 by projecting each of the structures onto the top two PCs. The computation of the lowest-cost path points to numerous structures computed by the EA that allow connecting the two basins despite such a conservative (subangstrom) max_nn_dist value. The path also goes nearby various experimentally-known structures in the projection of the energy landscape, which lends more credibility to its validity. Taken altogether, the path demonstrates the ability of SOD1 to undergo structural changes related to the phosphorylation event, effectively switching between two structural states (that separate the experimentally-known structures) during phosphorylation. Fig. 3 Visualization in 2D of map and a lowest-cost path computed for SOD1. The computed map for SOD1 is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Black dots show projections of experimentally-known structures. A lowest-cost path connecting the two visible basins is additionally drawn. Its cost (in REUs) is also listed Analysis of computed map and basin-basin excursions of CaM The ability of the proposed methodology to compute both maps and structural excursions is additionally illustrated on CaM. The color-coded 2D projection of the map is shown in Fig. 4. The map has a characteristic shape, with a hollow region in the middle, indicating the inability of the EA to find low-energy structures in this region. A broad and deep basin is found, populated by many experimentally-known structures, whose PDB ids are annotated. A long narrow strip of low-energy structures is also found. Figure 4 additionally shows the lowest-cost path computed to capture a structural excursion from a compact, closed structure of CaM (PDB id 1XFZ) to the calcium-bound structure (PDB id 1CLL); The lRMSD between the CA atoms of these two structures is 9.5Å, and the shown path is computed with a value of max_nn_dist corresponding to about 9.5Å /10; effectively limiting structural changes between any two successive structures in the path to subangstrom values. Fig. 4 Visualization in 2D of map and a lowest-cost path computed for caM. The computed map for CaM is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Black dots show projections of experimentally-known structures. A lowest-cost path connecting two experimentally-known structures is also drawn; its cost (in REUs) is listed. The PDB ids of the structures the path connects, as well as other experimentally-known structures of interest, are additionally shown where these structures project onto the top two PCs As Fig. 4 shows, the path goes through the calcium-free structure (PDB id 1CFD), passes through compact structures with which CaM binds proteins and peptides (PDB ids 1NWD and 2F3Y) to then reach a structure representative of the calcium-bound state (Ca(2+)-CaM) and in the state bound to myosin light chain kinase (CaM-MLCK) (PDB id 2KOE) just before terminating in the calcium-bound state (PDB id 1CLL). The path confirms that in the succession of structures from the compact state to the calcium-bound state, the domain collapse, re-arrangement, and partial unfolding of the helix linker in CaM are gradual. The succession of structures in the path points to a rearrangement of the domains in the compact state that is needed for CaM to then open up, before populating a semi-open state with a partially-unfolded linker that then further allows it to adopt the open, calcium-bound state. This detailed observation is in agreement with other studies, both those employing MD [59] and others employing robotics-inspired approaches [35]. Multi-dimensional visualization of computed map of H-Ras WT Prior work in [29, 34] has analyzed the 2D color-coded projection the H-Ras WT energy landscape in great detail and has concluded that the EA mapping the H-Ras WT energy landscape reproduces the two, large basins corresponding to the two states, On and Off, between which H-Ras switches to regulate its activity in the cell. In addition, the map contains novel low-energy regions not probed in the wet laboratory for H-Ras WT, some of which we analyze in detail here. However, we now do so by considering more than two dimensions. While important features can be preserved (and thus analyzed and subjected to interpretation) in a 2D projection, other features can be hidden by the projection. Note that, though the top two PCs capture around 45−50 % of the variance of the experimentally-known structures for each protein, essentially 50 % of the dynamics is hidden when projecting the computed maps onto two dimensions. Moreover, the ruggedness of the energy landscape requires careful preparation of the large number of points in a computed map when visualizing them after projection. The above projections for SOD1 and CaM, for instance, are visualized after ordering the points from high to low-energy, so that the low-energy ones are plotted on top of the high-energy ones to prevent occlusion. Below we relate conditioned plots for the H-Ras WT multi-dimensional map computed by the EA; the plots are constructed as described in the “Methods” Section. The projections are along PC1 and PC2, and the data are conditioned on each of the 4 quartiles of PC3 and PC4. It is worth noting that the top 4 PCs capture more than 75 % of the variance, and thus almost all of the dynamics of H-Ras. The quartile intervals for PC3 and PC4 each have roughly 222,580 data points for H-Ras WT. Table 2 shows the number of cases in common to a chosen quartile of PC3 and a chosen quartile of PC4. The left top panel of Fig. 5 shows a hexagon bin plot along PC1 and PC2 conditioned on the first quartile of PC3 and the first quartile of PC4 (containing 70,908 individuals, as related in Table 2). The color scheme uses color thresholds based on the binned quantiles of cell minimum-energy distributions without subsetting. Quantiles of {0,20,60,99,100} % correspond to Rosetta score12 values of {−374,−348,−321,−115,−18} REUs. The corresponding color-scheme is {dark blue, light blue, gray, pink }; yellow is reserved to show projections of the experimentally-known structures collected for H-Ras. The right bottom panel of Fig. 5 shows the shift of the minimum-energy patterns, as different subsets of the data are inspected per the 16 two-way conditioned plot layout. In particular, several interesting observations can now be drawn regarding the location of energy basins that the above 2D color-coded projections of the maps did not allow. Fig. 5 Visualization via conditioning plots of map computed for H-Ras WT. The map computed for the H-Ras WT is visualized via conditioning plots, which plot projections of the hall of fame on PC1 and PC2 conditional on projections on PC3 and PC4. Sixteeen plots are generated, considering combinations of all quartiles of PC3 with all quartiles of PC4. The 4D space in each subplot is discretized via hexagons, plotting for each hexagon only the projection of the lowest-energy individual in the hexagon. The blue-to-red color-coding scheme follows the low to blue energy range. Dots in yellow show PC1-PC2 projections of experimentally-known structures of H-Ras WT Table 2 Distribution of H-Ras WT individuals along PC3 and PC4 quartiles PC3 PC4 Q1 Q2 Q3 Q4 Q1 70,908 57,038 54,913 39,723 222,582 Q2 59,644 57,896 54,655 50,387 222,582 Q3 51,601 56,742 56,261 57,971 222,575 Q4 40,430 50,908 56,748 74,494 222,580 Q5 222,583 222,584 222,577 222,575 890,391 Since the hexagonal binning effectively smooths the ruggedness of the mapped energy landscape, two distinct basins can clearly be seen without the noise due to the ruggedness. The basins are most visible on the PC1-PC2 scatter plots along the second quartile of PC3 and the second or third quartile of PC4 (the [PC3:Q2; PC4:Q2-3] views). The basins reach deep in the energy landscape, as some of the conditioned plots show (for instance, along PC3:Q2 and PC4:Q2-3). The On basin (the dark blue region on the right) persists along all quartiles of PC4 (see first column of the 16-plot layout Fig. 5) but disappears quickly after the second quartile of PC3. No basins are visible on the third and onwards quartiles of PC3 and PC4. The Off basin (dark blue region on the left) is located (and is most visible) on [PC3:Q2; PC4:Q2-3] views. The experimentally-known structures appear on different quartiles of PC3 and PC4. Specifically, the majority can be found no further than the second quartiles of PC3 and PC4. This observation is particularly interesting, as the portion of the On basin that continues onto the third quartile of PC4 (and second quartile of PC3) does not contain any experimentally-known structures in it. This portion of the On basin is in effect a novel region of the H-Ras WT energy landscape not currently probed in the wet laboratory. As such, the structures in this region constitute a novel stable region that is worth pursuing further in the wet laboratory, particularly in the context of designing drug inhibitors for H-Ras. Similar observations can be drawn regarding portions of the Off basin along specific quartiles, where no experimentally-known structures reside. Comparison of maps and basin-basin excursions of H-Ras WT and variants Maps and structural excursions computed by the proposed methodology on H-Ras are now investigated in greater detail. A comparative setting is pursued to understand dfferences between H-Ras WT and 7 disease-related variants, five of which are single mutants, and two are double mutants. The H-Ras sequences are listed in column 1 in Table 3. The standard naming convention [Code1][Position][Code2] for a single-mutant variant relates that the amino acid named ‘Code1’ (using one-letter amino-acid codes) at position ‘Position’ in the WT is replaced with the amino acid named ‘Code2’ in this particular variant. In other variants, the additional mutations are joined in order of positions; e.g. Y32CC118S. Table 3 Comparison of lowest-cost on →off path across H-Ras WT and variants Sequence m a x_n n_d i s t (Å) Path cost (REU) Highest Energy (REU) Nr. Edges WT 1.45/10 266 –251 90 1.45/7.5 108 –299 56 G12S 1.45/10 — — — 1.45/ /7.5 130 –230 26 G12C 1.45/10 — — — 1.45/ /7.5 232 –150 66 G12D 1.45/10 — — — 1.45/ /7.5 119 –277 51 G12V 1.45/10 — — — 1.45/ /7.5 109 –276 72 Q61L 1.45/10 — — — 1.45/ /7.5 85 –263 62 Y32CC118S 1.45/10 — — — 1.45/ /7.5 131 –265 54 R164AQ165V 1.45/10 — — 1.45/ /7.5 131 –277 72 The EA described in the “Methods” Section is employed to obtain maps for each of the 8 H-Ras sequences. The maps are then queried to compute the lowest-cost paths and other low-cost paths (as described in the “Methods” Section) connecting a structure representative of the On state (PDB id 1QRA) to a structure representative of the Off state (PDB id 4Q21) in each of these 8 H-Ras sequences, effectively modeling the On →Off structural excursion. Two values of max_nn_dist are considered, corresponding to 1.45Å /10 and 1.45Å /7.5, where 1.45Å is the lRMSD between the CA atoms of the structures selected to represent the On and Off states. Summary statistics for the lowest-cost path and the ensemble of low-cost paths computed on each of the 8 H-Ras sequences are shown in Tables 3 and 4. The low-cost paths (and maps) are also shown on color-coded 2D projections for selected sequences (WT in Fig. 6, G12C in Fig. 7, Q61L in Fig. 8, and Y32CC118S in Fig. 9). Color-coded 2D projections of maps and paths computed for the other variants are related in the Additional files 1, 2, 3 and 4. Summary statistics on paths modeling the (reverse) Off →On structural excursion are related in the Additional files 5 and 6. Fig. 6 Visualization in 2D of map and paths computed for H-Ras WT. The computed map for H-Ras WT is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants Fig. 7 Visualization in 2D of map and paths computed for H-Ras G12C. The computed map for the H-Ras G12C variant is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants Fig. 8 Visualization in 2D of map and paths computed for H-Ras Q61L. The computed map for the H-Ras Q61L variant is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants Fig. 9 Visualization in 2D of map and paths computed for H-Ras Y32CC118S. The computed map for the H-Ras Y32CC118S variant is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants Table 4 Comparison of ensemble of low-cost on →off paths across H-Ras WT and variants Sequence m a x_n n_d i s t (Å) (μ,σ)Cost (REU) (μ,σ)Highest Energy (REU) (μ,σ)Nr. Edges WT 1.45/10 (418.3, 106.) (–203.5, 25.1) (101.3, 16.1) 1.45/7.5 (127.6, 7.90) (–277.3, 15.2) (64.2, 15.6) G12S 1.45/10 — — — 1.45/7.5 (143, 40.9) (–259, 92) (73, 22) G12C 1.45/10 — — — 1.45/7.5 (266.4, 18.2) (–139.3, 24.4) (60, 9.8) G12D 1.45/10 — — — 1.45/7.5 (140.4, 15.5) (–253.9, 15.9) (54.6, 8.6) G12V 1.45/10 — — — 1.45/7.5 (132.3, 13.5) (–236.3, 83.4) (64.6, 13.) Q61L 1.45/10 — — — 1.45/7.5 (161.3, 45.2) (–240.7, 21.8) (64.1, 9.2) Y32CC118S 1.45/10 — — — 1.45/7.5 (158.2, 18.9) (–257.5, 18.1) (63.9, 10.2) R164AQ165V 1.45/10 — — — 1.45/7.5 (159.7, 21.3) (–245.6, 23.8) (65.9, 7.1) Table 3 compares the lowest-cost On →Off path for each of the 8 H-Ras sequences. The cost of the path, the highest energy among structures in the path, and the number of edges in the path are listed in columns 3–5. The lowest-cost path on each H-Ras sequence has been queried off the EA-built map under the two different values for max_nn_dist listed above. The lower value makes it harder to find paths, as indicated by the higher costs and the lack of paths on any sequence but the WT in Table 3. The higher value allows finding more paths, and even lower-cost paths, as the ruggedness of the energy landscape within a ball of radius max_nn_dist is effectively ignored. Since the higher setting of max_nn_dist still corresponds to a very small distance between two successive structures (1.45Å /7.5) and allows obtaining low-cost paths on both WT and variants, the paths shown on 2D projections of the computed maps are those computed for max_nn_dist set to 1.45Å /7.5. Additional file 7 shows the paths that are obtained on H-Ras WT on the lower, more stringent value of 1.45Å /10 for max_nn_dist. The paths are higher in cost, as described above, but they navigate similar regions in the landscape as the paths computed at the less stringent distance of 1.45Å /7.5. Comparison of the lowest-cost path found for each of the 8 H-Ras sequences at the less stringent distance allows drawing the following conclusion: The majority of the single mutants (with the exception of Q61L and G12V) incur a significantly higher energetic cost for the On →Off structural excursion. This points to a higher energetic barrier separating the On and Off states, which is also visible on many of the 2D projections of the maps built for these variant sequences. The latter is particularly prominent for the G12C variant and can additionally be qualitatively confirmed by comparing the color-coded 2D projection of the H-Ras WT map in Fig. 6 to the 2D projection of the H-Ras G12C map in Fig. 7. While the results related in Table 3 are informative, they do not take into account the stochasticity of protein motions. Summary statistics on the ensemble of low-cost paths, computed as described in the “Methods” Section, are listed in Table 4 for each of the 8 H-Ras sequences. The comparison of the average cost and average highest-energy along structures in paths generally preserves the ordering of the variants on the lowest-cost paths above. The only variant where this is not the case is Q61L, where a lowest-cost path even lower than in the H-Ras WT can be found, but this path is an outlier compared to the ensemble. The rest of the low-cost paths found for Q61L are much higher in cost, contributing to an average statistic of 161.3 REUs, which is among the highest (the highest average cost is obtained on the G12C variant) when compared to the WT and other variants. This conclusion is in line with qualitative observations made in [33] and similar ones based on visualization of the 2D projection of the energy landscape map in Fig. 8; a high energy barrier between the On and Off basins in the Q61L variant contributes to a structural rigidity in Q61L that effectively causes Q61L to be constitutively activated (always on). The same mechanism is observed on the majority of the variants of H-Ras here. H-Ras variants where the mutation has a profound impact on the cost of the On →Off structural excursion are those where G12 is mutated to S, C, or D. The higher average path costs over the H-Ras WT for these variants can be also be confirmed by the color-coded 2D projections of the computed maps. For instance, Fig. 7 shows that the entire landscape is elevated in G12C, as many structures become more costly; the On −Off barrier is also higher than in the WT, contributing to the higher average cost for the On →Off excursion. This observation holds on G12S and G12D, as well. In particular, in the G12S variant, whose 2D projection of the map and paths are shown in the Additional file 1, the On basin is very deep, effectively trapping this variant in the On/GTP-binding state. The G12C is also trapped in the On state, but that is due to everything else in the landscape being much more energetically costly. G12V is the only G12* variant where the average cost (and the landscape) is not significantly different from the WT (the paths and the landscape are shown in the Additional file 2). This result is in agreement with an earlier study, where the G12V mutation is proposed to have a subtle effect more on the binding than the energy landscape of the uncomplexed H-Ras variant [33]. Visualization of the maps via color-coded 2D projections reveals an additional interesting energetic feature. G12C, G12S, and the double mutants Y32CC118S and R164AQ165V populate two more regions, distinct from the On and Off basins, with lower-energy structures than the WT, G12V, G12D, and Q61L (the maps for G12S/D and R164AQ165V are provided in the Additional files 1, 3, and 4, respectively). Preliminary evidence of these regions was related by us in prior work on analysis of a first-generation version of our EA on H-Ras WT, G12V, and Q61L [33]. However, in [33], these regions were not exploited as well. These regions, dubbed Conf1 and Conf2 in [33] (Conf1 corresponds to PC1 in [ −3,0] and PC2 in [ 36], and Conf2 corresponds to PC1 in [ −9,−6] and PC2 in [ 1215] in the 2D projections), are populated with very low-energy structures by the EA employed here in the H-Ras G12C, G12S, Y32CC118S, and R164AQ165V variants. The regions constitute new basins, effectively, in these variants. It is interesting that the Conf1 basin emerges only on the G12C/S mutations and not on the G12V mutation, particularly considering that the structure caught in the wet laboratory for the G12V variant projects to this region of the structure space. This is a novel finding of our methodology and suggests that perhaps the relationships regarding shared molecular function profiles between the G12* variants and these double mutants ought to be investigated in greater detail in the wet laboratory. Finally, it is worth noting that the Conf1 region is populated well by the double mutants, as well. In particular, the Conf1 basin is deeper in the Y32CC118S variant (see Fig. 9), as expected, given that this region contains projections of wet-laboratory structures caught for this variant (thus representing a stable state). This basin is also deep in the R164AQ165V variant (see the Additional file 4). However, both double mutants have a higher energy barrier and a shallower off basin than the WT (see Fig. 9 for the Y32CC118S variant, and the Additional file 4 for the R164AQ165V variant), which results in higher-cost On →Off excursions, as related in Table 4, effectively rigidifying these variants. The latter explains the loss of GTP-binding activity noted for the R164AQ165V variant. Discussion The results presented here suggest that an increasingly detailed picture is emerging of the H-Ras energy landscape. The two-basin feature of the H-Ras energy landscape has been elucidated in both wet and dry laboratories; extensive computational studies by McCammon and colleagues via MD methods have both verified the existence of these two basins and the energy barrier separating them [60]. The two-basin characteristic has also been reproduced via prior versions of the EA algorithm employed here that did not make use of a map but rather analyzed all structures ever generated. The graphical techniques employed in this paper to analyze the map constructed by the proposed methodology provide for the first time a highly detailed view of the multi-dimensional H-Ras energy landscape. In particular, Fig. 5 shows not only how the On and Off basins elongate along the third and fourth dimensions, but also clarify which regions of this multi-dimensional space provide interesting new energetic features not captured in other laboratories. For instance, as described in detail in the “Results” Section, a significant portion of the On basin that continues onto PC4:Q3 (and PC3:Q2) does not contain any experimentally-known structures. Effectively, this represents a new region of the H-Ras energy landscape that is reported to be associated with the stable on structural state by the EA employed here but has yet to be captured in the wet laboratory. The graphical techniques employed here also allow making comparative observations regarding the depth and width of the On and Off basins. The layout of the 16 two-way conditioned plots in Fig. 5 shows that the On basin is both wider and deeper than the Off basin (this observation can also be made, though less reliably, on the 2D projection of the energy landscape in Fig. 6). The [PC3:Q3-4; PC4:Q3-4] views in Fig. 5 join the two basins, effectively showing the landscape at the higher energy levels. As one proceeds deeper in the landscape, the regions separate to yield the distinct On and Off basins; energy barriers appear along [PC3:Q1-2; PC4-Q*]. The higher width of the On basin points to the higher stability of this basin; that is, the temporal scale of structural excursions of H-Ras from the On to the Off state will be dominated by diffusions within the deep and broad On basin. The juxtaposition of maps and On →Off structural excursions for the H-Ras WT and the 7 single- and double-mutant variants in the “Results” Section elucidates, among other things, that two new basins emerge on the landscapes of some single- and double mutants, referred to as Conf1 and Conf2. In particular, these are observed to be richly populated in G12C, G12S, and the double mutants, but poorly populated on the other variants and H-Ras WT. Figure 10 provides a 3D view of the lowest-energy structures (falling in the 1st percentile of the energy distribution) in the map computed for H-Ras WT and the experimentally-known structures by projecting them onto the top three PCs. Picking a lower percentile loses the range of the PCs, which we want to retain in order to show projections of all the experimentally-known structures (drawn as red spheres). The 3D space is partitioned into truncated octahedron cells, as advocated by Carr in [61], and one sphere is drawn at the centroid of each cell. The color and size of a sphere is based on the minimum energy value in the corresponding cell. Three energy intervals are observed for H-Ras WT in this way: [−37.442−367.286] REUs (large blue spheres), (−367.286−350.180] REUs (smaller green-blue spheres), and (−350.180−330.102] REUs (small violet spheres). The interval boundaries correspond to the 0, 0.005, 0.02, and 1 % percentiles. The approximate locations of the On, Off, Conf1, and Conf2 basins are delineated in blue in Fig. 10. The PDB ids of selected experimentally-known structures are also annotated. Fig. 10 Visualization in 3D of map computed for H-Ras WT. The lowest-energy structures in the map computed for H-Ras WT are shown projected onto the top 3 PCs. Projections of the experimentally-known structures are also drawn, as red spheres of a larger radius. The PDB ids of some of these structures are also shown. The four basins that emerge on the WT and the various variants are also delineated and named per the convention described in the main text Comparison of Figs. 5 and 10 shows that the Conf1 and the On basins are merged together by structures with slightly higher energy values (a few REUs in score12). In [33], where an early version of the EA is employed (with narrow initialization, no map, and a budget-fixed improvement operator), these structures effectively merging the On and Conf1 basin in the WT are not reported, as the earlier EA has lower exploitation capability. In contrast, the 2D maps of the G12C and the double mutants show Conf1 to be separated by an energy barrier from the On state rather than merged into the On state as in the WT, and to also protrude deeper in the energy landscape than in the WT. The experimentally-known structure with PDB id 1LF0 sits in the region of the structure space corresponding to the Conf1 basin in the variants and the elongated On basin in the WT (see Fig. 10). This structure has been captured for the H-Ras A59G variant in the active/On state [62]. A 20-ns unbiased MD simulation in [63] has noted that this structure may mediate the On →Off switching in the A59G variant. The intermediate role of this structure is confirmed by the EA here, as this structure is reported to be low-energy for the H-Ras WT and part of the elongated On basin. However, none of the low-cost paths computed for the WT directly employ this structure, as the work-based cost does not promote diffusing in a basin. The in-basin diffusion may explain why this structure has not been captured as an intermediate for the WT during the On →Off excursion in the wet laboratory; it is only in H-Ras variants that an energy barrier gives rise to the distinct Conf1 basin. This barrier may trap variants in Conf1 long enough for this structure to be caught in wet laboratories. Interestingly, another structure, with PDB id 1LF5 (residing in the Off basin in Fig. 10), has been caught for A59G in the Off state. Taken together, the comparative analysis suggests that the wide On basin retreats in the variants, and an energy barrier splits it into two basins, a narrower On basin and Conf1. The H-Ras WT, once outside the wide On basin, may switch to the stable Off basin or a semi-stable basin observed most clearly in the [PC3:Q1; PC4-Q1] view. This basin sits at the top of the map, in between the On and Off basins, and is referred to as the Conf2 basin. Conf2 is not populated by the lowest-energy structures, but it does contain low-energy structures and two experimentally-known ones. The latter are reported in the PDB under ids 1Q21 and 2Q21. The structure with PDB id 1Q21 is reported as active/On for the WT, whereas that with PDB id 2Q21 is reported as active/On for the G12V variant [64]. The structures are very similar, as noted in [64], and differ mainly in the configuration of the side-chain at position 12, confirming the proximity of these two structures in the PC variable space in Fig. 10 (found at [ −9,−6] in PC1 and [12, 15] in PC2). The work in this paper again confirms that these two structures are functional for the WT from a thermodynamic availability point of view, but perhaps difficult to access within physiological temporal scales due to the high-energy barriers that surround the Conf2 basin. The juxtaposition of the H-Ras WT to the variants in the “Results” Section shows that the Conf2 basin is richly populated in G12C, G12S, and the double mutants. In particular, it is wider and protrudes deeper in the energy landscape for G12C and G12S but not G12V. This is an interesting finding that points to further work in the wet laboratory, as it suggests a novel function regulation mechanism that can be modulated via inhibitors. The comparison of landscapes and path ensembles across the H-Ras variants provides observations that not only validate and reconcile wet-laboratory findings but may also be useful to further investigation in the wet-laboratory on understanding mutations and designing inhibitors to disrupt aberrant activity [65]. For instance, in addition to the analysis above, a conclusion can be reached regarding the structure with PDB id 6Q21; the asymmetric unit (chain D) of this structure is projected and shown in the 3D view in Fig. 10. This structure is reported for the H-Ras WT in [66]. This unit is in a slightly different structure than the canonical on state (PDB id 1QRA), providing in [66] the earliest evidence of the structural flexibility of H-Ras WT. Figure 10 shows that the structure captured for the WT in PDB id 6Q21 is in a region of the energy landscape populated by low-energy structures part of the elongated On basin in H-Ras WT. The increases in costs reported here associated with structural excursions of H-Ras variants correspond to increases in the time it takes to undergo the excursion at equilibrium. Since molecular recognition events occur at carefully-calibrated temporal scales, any disruption to temporal scales is consequential for molecular recognition events, and thus normal biological activity in the cell. Conclusions This paper introduces a novel methodology to map a protein’s energy landscape and model equilibrium dynamics. Rather than simulate the dynamics of the covalently-bound network of atoms in a protein molecule, the proposed methodology relies on stochastic search to obtain a sample-based representation of the constrained structure space relevant for the dynamics, and then employs discrete search structures to summarize the dynamics. An EA is employed to map the multi-dimensional energy landscape of a protein, and a nearest-neighbor graph representation of the map is then queried to reveal energetically-feasible successions of structures mediating structural excursions of interest. Analysis of applications on several proteins of importance to human biology and disease suggests the proposed methodology is useful in understanding the relationship between protein structure, dynamics, and function with a practical computational budget. While obtaining a detailed characterization of protein equilibrium dynamics remains a challenge in silico, the work here exploits the wealth of structure data and novel randomized search strategies to enhance exploration of the thermodynamically-available structure space. The exploitation of structure data is a powerful and timely mechanism to map the structure space of a protein. The availability of wet-laboratory structures representing semi-stable and stable structural states for many proteins allows formulating algorithms that can map energy landscapes within a reasonable computational budget, as demonstrated here. The work presented here opens up several promising directions for future research. One direction concerns lowering the dependency of the methodology on sufficient structure data, as well as expanding its applicability to systems where experimentally-known structure reside in a non-linear low-dimensional space. The first can be addressed via techniques such as Normal Mode Analysis, already integrated with some success in robotics-inspired modeling of protein motions [67–70]. The second can be addressed via linear dimensionality reduction techniques. Another direction of future research concerns improving the predictions of the locations and depths of mapped basins by employing various energy functions. This direction aims to increase the reliability of in-silico predictions. Considering multiple energy functions remains challenging, however, as considerable recoding efforts are required to efficiently integrate such functions in in-house code. All data obtained by the proposed methodology and analyzed here are available to the research community upon request. Similarly, any components of the proposed methodology can be shared as linux binaries. Additional files Additional file 1 Visualization in 2D of Map and Paths Computed for H-Ras G12S. The computed map for the H-Ras G12S variant is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants. (PDF 1597 kb) Additional file 2 Visualization in 2D of map and paths computed for H-Ras G12V. The computed map for the H-Ras G12V variant is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants. (PDF 1772 kb) Additional file 3 Visualization in 2D of map and paths computed for H-Ras G12D. The computed map for the H-Ras G12D variant is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants. (PDF 1823 kb) Additional file 4 Visualization in 2D of map and paths computed for H-Ras R164AQ165V. The computed map for the H-Ras R164AQ165V variant is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants. (PDF 1648 kb) Additional file 5 Comparison of lowest-cost off →on path across H-Ras WT and variants. Column 1 lists the different H-Ras sequences investigated. The two different values used in the query of the map are listed in column 2. The cost of the path, the highest energy among structures in the path, and the number of edges in the path are listed in columns 3–5. (PDF 13 kb) Additional file 6 Comparison of ensemble of low-cost off →on paths across H-Ras WT and variants. Column 1 lists the different H-Ras sequences investigated. The two different values used in the query of the map are listed in column 2. Columns 3–5 show summary statistics, such as mean and standard deviation, are reported for path cost, highest energy over structures in a path, and the number of edges in a path. (PDF 17 kb) Additional file 7 Visualization in 2D of map and (finer-resolution) paths computed for H-Ras WT. The computed map for H-Ras WT is projected onto 2D and projections are color-coded by Rosetta score12 energy values. Low-cost paths (costs in REUs are listed) modeling the On →Off structural excursion are also drawn, now using a more stringent distance criterion for two successive structures in the path. The projections of experimentally-known structures on the top two PCs are related by showing whether the structures are captured in the wet laboratory for the WT or variants. (PDF 2109 kb) From IEEE International Conference on Bioinformatics and Biomedicine 2015Washington, DC, USA. 9-12 November 2015 Acknowledgements The authors thank the Shehu and De Jong laboratory members for useful feedback on this research. Declarations This publication and the work reported in it are supported in part by NSF CCF Grant No. 1421001 and NSF IIS CAREER Award No. 1144106. This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4. Authors’ contributions ES, AS, and KD designed the evolutionary algorithm and the path computation techniques and designed the experiments. ES implemented the evolutionary algorithm and the path computation techniques. DC designed and implemented the visual statistical analysis. ES, DC, KD, and AS wrote the manuscript. KD and AS edited the final version. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. ==== Refs References 1 Frauenfelder H Sligar SG Wolynes PG The energy landscapes and motion on proteins Science 1991 254 5038 1598 603 10.1126/science.1749933 1749933 2 Jenzler-Wildman K Kern D Dynamic personalities of proteins Nature 2007 450 964 72 10.1038/nature06522 18075575 3 Hub JS de Groot BL Detection of functional modes in protein dynamics PLoS Comp Biol 2009 5 8 1000480 10.1371/journal.pcbi.1000480 4 Schroedinger E What Is Life? 1944 New York Cambridge University Press 5 Kendrew JC Dickerson RE Strandberg BE Hart RG Davies DR Phillips DC Shore VC Structure of myoglobin: A three-dimensional fourier synthesis at 2å resolution Nature 1960 185 4711 422 7 10.1038/185422a0 18990802 6 Gall A Ilioaia C Krüger TP Novoderezhkin VI Robert B van Grondelle R Conformational switching in a light-harvesting protein as followed by single-molecule spectroscopy Biophys J 2015 108 11 2713 20 10.1016/j.bpj.2015.04.017 26039172 7 Boehr DD Nussinov R Wright PE The role of dynamic conformational ensembles in biomolecular recognition Nat Chem Biol 2009 5 11 789 96 10.1038/nchembio.232 19841628 8 Shehu R Nussinov R Computational methods for exploration and analysis of macromolecular structure and dynamics PLoS Comput Biol 2015 11 10 1004585 10.1371/journal.pcbi.1004585 9 Zhu G NMR of Proteins and Small Biomolecules. Topics in Current Chemistry, vol. 326 2012 Berlin-Heidelberg Springer 10 Fenwick RB van den Bedem H Fraser JS Wright PE Integrated description of protein dynamics from room-temperature X-ray crystallography and NMR Proc Natl Acad Sci USA 2014 111 4 445 54 10.1073/pnas.1323440111 24335589 11 Greenleaf WJ Woodside MT Block SM High-resolution, single-molecule measurements of biomolecular motion Annu Rev Biophys Biomol Struct 2007 36 171 90 10.1146/annurev.biophys.36.101106.101451 17328679 12 Hohlbein J Craggs TD Cordes T Alternating-laser excitation: single-molecule FRET and beyond Chem Soc Rev 2014 43 1156 71 10.1039/C3CS60233H 24037326 13 Lewandowski JR Halse ME Blackledge M Emsley L Protein dynamics. Direct observation of hierarchical protein dynamics Science 2015 348 6234 578 81 10.1126/science.aaa6111 25931561 14 Karplus M Kuriyan J Molecular dynamics and protein function Proc Natl Acad Sci USA 2005 102 19 6679 85 10.1073/pnas.0408930102 15870208 15 Becker OM Karplus M The topology of multidimensional potential energy surfaces: Theory and application to peptide structure and kinetics J Chem Phys 1997 106 4 1495 517 10.1063/1.473299 16 Dubrow A What got done in one year at NSF’s Stampede supercomputer Comput Sci Eng 2015 17 2 83 8 10.1109/MCSE.2015.40 17 Zhao G Perilla JR Yufenyuy EL Meng X Chen B Ning J Ahn J Gronenborn AM Schulten K Aiken C Zhang P Mature HIV-1 capsid structure by cryo-electron microscopy and all-atom molecular dynamics Nature 2013 497 7451 643 6 10.1038/nature12162 23719463 18 Perilla JR Goh BC Cassidy CK Liu B Bernardi RC Rudack T Yu H Wu Z Schulten K Molecular dynamics simulations of large macromolecular complexes Curr Opin Struct Biol 2015 31 64 74 10.1016/j.sbi.2015.03.007 25845770 19 Tsai C Kumar S Ma B Nussinov R Folding funnels, binding funnels, and protein function Protein Sci 1999 8 6 1181 90 10.1110/ps.8.6.1181 10386868 20 Okazaki K Koga N Takada S Onuchic JN Wolynes PG Multiple-basin energy landscapes for large-amplitude conformational motions of proteins: Structure-based molecular dynamics simulations Proc Natl Acad Sci USA 2006 103 32 11844 49 10.1073/pnas.0604375103 16877541 21 Wales DJ Energy landscapes: some new horizons Curr Opin Struct Biol 2010 20 1 3 10 10.1016/j.sbi.2009.12.011 20096562 22 Jorgensen WL Tirado-Rives J Monte carlo vs molecular dynamics for conformational sampling J Phys Chem 1996 100 34 14508 13 10.1021/jp960880x 23 Morris GM Goodsell DS Halliday RS Huey R Hart WE Belew RK Olson AJ Automated docking using a lamarckian genetic algorithm and an empirical binding free energy function J Comput Chem 1998 19 14 1639 62 10.1002/(SICI)1096-987X(19981115)19:14<1639::AID-JCC10>3.0.CO;2-B 24 Unger R The genetic algorithm approach to protein structure prediction Struct Bond 2004 110 153 75 10.1007/b13936 25 Olson B De Jong KA Shehu A Off-lattice protein structure prediction with homologous crossover Conf on Genetic and Evolutionary Computation (GECCO) 2013 New York ACM 26 Olson B Shehu A Multi-objective stochastic search for sampling local minima in the protein energy surface ACM Conf on Bioinf and Comp Biol (BCB) 2013 Washington ACM 27 Clausen R Shehu A A multiscale hybrid evolutionary algorithm to obtain sample-based representations of multi-basin protein energy landscapes ACM Conf on Bioinf and Comp Biol (BCB) 2014 Newport Beach ACM 28 Clausen R Shehu A A data-driven evolutionary algorithm for mapping multi-basin protein energy landscapes J Comp Biol 2015 22 9 844 60 10.1089/cmb.2015.0107 29 Sapin E, De Jong KA, Shehu A. Evolutionary search strategies for efficient sample-based representations of multiple-basin protein energy landscapes. In: IEEE Intl Conf Bioinf and Biomed (BIBM). IEEE: 2015. p. 13–20. 30 Berman HM Henrick K Nakamura H Announcing the worldwide Protein Data Bank Nat Struct Biol 2003 10 12 980 10.1038/nsb1203-980 14634627 31 Clausen R Shehu A Exploring the structure space of wildtype ras guided by experimental data ACM Conf on Bioinf and Comp Biol Workshops (BCBW) 2013 Washington ACM 32 Sapin E, De Jong KA, Shehu A. A novel ea-based memetic approach for efficiently mapping complex fitness landscapes. In: ACM Conf on Genet Evol Comput (GECCO). ACM: 2016. p. 85–92. 33 Clausen R Ma B Nussinov R Shehu A Mapping the conformation space of wildtype and mutant h-ras with a memetic, cellular, and multiscale evolutionary algorithm PLoS Comput Biol 2015 11 9 1004470 10.1371/journal.pcbi.1004470 34 Sapin E, De Jong KA, Shehu A. From optimization to mapping: an evolutionary algorithm for protein energy landscapes. IEEE/ACM Trans Bioinf and Comp Biol. 2016. [in preprint]. 35 Maximova T, Plaku E, Shehu A. Computing transition paths in multiple-basin proteins with a probabilistic roadmap algorithm guided by structure data. In: IEEE Intl Conf Bioinf and Biomed (BIBM). IEEE: 2015. p. 35–42. 36 Molloy K, Shehu A. A general, adaptive, roadmap-based algorithm for protein motion computation. IEEE Trans NanoBioScience. 2015. in press. 37 Molloy K, Clausen R, Shehu A. A stochastic roadmap method to model protein structural transitions. Robotica. 2015. in press. 38 Ma B Kumar S Tsai C Nussinov R Folding funnels and binding mechanisms Protein Eng 1999 12 9 713 20 10.1093/protein/12.9.713 10506280 39 Luenberger DG Linear and Nonlinear Programming 1984 Norwell Addison-Wesley 40 Gront D Kmiecik S Kolinski A Backbone building from quadrilaterals: a fast and accurate algorithm for protein backbone reconstruction from alpha carbon coordinates J Comput Chem 2007 28 29 1593 7 10.1002/jcc.20624 17342707 41 Kaufmann KW Lemmon GH DeLuca SL Sheehan JH Meiler J Practically useful: What the rosetta protein modeling suite can do for you Biochemistry 2010 49 14 2987 98 10.1021/bi902153g 20235548 42 Carr DB Nicholson WL Littlefield RJ Hall DL Wegman EJ DePriest DJ Interactive color display methods for multivariate data Statistical Image Processing and Graphics 1986 New York Marcel Decker 43 Cleveland WS Visualizing Data 1993 Summit Hobart Press 44 Dawkins BP investigating the geometry of a P-dimensional data sets J Amer Stat Assoc 1995 90 429 350 9 45 Carr DB Scanning a 4-D domain for local minima: A protein folding example Top Sci Vis 1995 6 2 9 12 46 Carr DB Olsen AR White D Hexagon mosaic maps for display of univariate and bivariate geographical data Cartogr Geogr Inf Syst 1992 19 4 228 36 10.1559/152304092783721231 47 Dijkstra EW A note on two problems in connexion with graphs Numer Math 1959 1 269 71 10.1007/BF01386390 48 Friedman JH Bentley JL Finkel RA An algorithm for finding best matches in logarithmic expected time ACM Trans Math Softw (TOMS) 1977 3 3 209 26 10.1145/355744.355745 49 Goodman JE O’Rourke J Indyk P Goodman JE O’Rourke J Nearest neighbors in high-dimensional spaces Handbook of Discrete and Computational Geometry 2004 Boca Raton CRC Press 50 Jaillet L Cortés J Siméon T Sampling-based path planning on configuration-cost costmaps IEEE Trans Robot 2010 26 4 635 46 10.1109/TRO.2010.2049527 51 Conwit RA Preventing familial ALS: a clinical trial may be feasible but is an efficacy trial warranted? J Neurol Sci 2006 251 1-2 1 2 10.1016/j.jns.2006.07.009 17070848 52 Gsponer J Christodoulou J Cavalli A Bui JM Richter B Dobson CM Vendruscolo M A coupled equilibrium shift mechanism in calmodulin-mediated signal transduction Structure 2008 16 5 736 46 10.1016/j.str.2008.02.017 18462678 53 Yap KL Yuan T Mal TK Vogel HJ Ikura M Structural basis for simultaneous binding of two carboxy-terminal peptides of plant glutamate decarboxylase to calmodulin J Mol Biol 2003 328 1 193 204 10.1016/S0022-2836(03)00271-7 12684008 54 Zhang BW Jasnow D Zuckermann DM Efficient and verified simulation of a path ensemble for conformational change in a united-residue model of calmodulin Proc Natl Acad Sci USA 2007 104 46 18043 48 10.1073/pnas.0706349104 17984047 55 Karnoub AE Weinberg RA Ras oncogenes: split personalities Nat Rev Mol Cell Biol 2008 9 7 517 31 10.1038/nrm2438 18568040 56 Fernández-Medarde A Santos E Ras in cancer and developmental diseases Genes Cancer 2011 2 3 344 58 10.1177/1947601911411084 21779504 57 Magrane M the UniProt consortium UniProt knowledgebase: a hub of integrated protein data Database 2011 2011 bar009 1 13 58 Wilcox KC Zhou L Jordon JK Huang Y Yu Y Redler RL Chen X Caplow M Dokholyan NV Modifications of superoxide dismutase (sod1) in human erythrocytes a possible role in amyotrophic lateral sclerosis J Biol Chem 2009 284 20 13940 47 10.1074/jbc.M809687200 19299510 59 Gsponer J Christodoulou J Cavalli A Bui JM Richter B Dobson CM Vendruscolo M A coupled equilibrium shift mechanism in calmodulin-mediated signal transduction Structure 2008 16 5 736 46 10.1016/j.str.2008.02.017 18462678 60 Grant BJ Gorfe AA McCammon JA Ras conformational switching: Simulating nucleotide-dependent conformational transitions with accelerated molecular dynamics PLoS Comput Biol 2009 5 3 1000325 10.1371/journal.pcbi.1000325 61 Carr DB Buja A Tukey P Looking at large data sets using binned data plots Computing and Graphics in Statistics 1991 New York Springer 62 Hall BE Bar-Sagi D Nassar N The structural basis for the transition from Ras-GTP to Ras-GDP Proc Natl Acad Sci USA 2002 99 19 12138 42 10.1073/pnas.192453199 12213964 63 Lukman S Grant BJ Gorfe AA Grant GH McCammon JA The distinct conformational dynamics of K-Ras and H-Ras A59G PLoS Comput Biol 2010 6 9 e1000922 10.1371/journal.pcbi.1000922 20838576 64 Tong LA de Vos AM Milburn MV Kim SH Crystal structures at 2.2 A resolution of the catalytic domains of normal ras protein and an oncogenic mutant complexed with GDP J Mol Biol 1991 217 3 503 16 10.1016/0022-2836(91)90753-S 1899707 65 Grant BJ Lukman S Hocker HJ Sayyah J Brown JH McCammon JA Gorfe AA Novel allosteric sites on Ras for lead generation PLoS ONE 2011 6 10 25711 10.1371/journal.pone.0025711 66 Milburn MV Tong L deVos AM Brünger A Yamaizumi Z Nishimura S Kim SH Molecular switch for signal transduction: structural differences between active and inactive forms of protooncogenic ras proteins Science 1990 247 4945 939 45 10.1126/science.2406906 2406906 67 Kirillova S Cortés J Stefaniu A Siméon T An nma-guided path planning approach for computing large-amplitude conformational changes in proteins Protein Struct Funct Bioinf 2008 70 1 131 43 10.1002/prot.21570 68 Al-Bluwi I Vaisset M Siméon T Cortés J Modeling protein conformational transitions by a combination of coarse-grained normal mode analysis and robotics-inspired methods BMC Struct Biol 2013 13 S2 1 69 Kim MK Chirikjian GS Jernigan RL Elastic models of conformational transitions in macromolecules J Mol Graph Model 2002 21 2 151 60 10.1016/S1093-3263(02)00143-2 12398345 70 Schuyler Ad Jernigan RL Wasba PK Ramakrishnan B Chirikjian GS Iterative cluster-nma (icnma): a tool for generating conformational transitions in proteins Protein Struct Funct Bioinf 2009 74 3 760 76 10.1002/prot.22200

PMC005xxxxxx/PMC5001233.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556805289610.1186/s12864-016-2896-7ResearchRead-Split-Run: an improved bioinformatics pipeline for identification of genome-wide non-canonical spliced regions using RNA-Seq data Bai Yongsheng Yongsheng.Bai@indstate.edu 12Kinne Jeff jkinne@cs.indstate.edu 3Donham Brandon bdonham@cs.indstate.edu 3Jiang Feng fjiang@sycamores.indstate.edu 3Ding Lizhong lding@sycamores.indstate.edu 1Hassler Justin R. jhassler@sbpdiscovery.org 4Kaufman Randal J. rkaufman@sbpdiscovery.org 41 Department of Biology, Terre Haute, USA 2 The Center for Genomic Advocacy, Indiana State University, 600 Chestnut Street, Terre Haute, IN 47809 USA 3 Department of Mathematics and Computer Science, Indiana State University, 200 North Seventh Street, Terre Haute, IN 47809 USA 4 Sanford-Burnham-Prebys Medical Discovery Institute, La Jolla, California 92037 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.503© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Most existing tools for detecting next-generation sequencing-based splicing events focus on generic splicing events. Consequently, special types of non-canonical splicing events of short mRNA regions (IRE1α targeted) have not yet been thoroughly addressed at a genome-wide level using bioinformatics approaches in conjunction with next-generation technologies. During endoplasmic reticulum (ER) stress, the gene encoding the RNase Ire1α is known to splice out a short 26 nt region from the mRNA of the transcription factor Xbp1 non-canonically within the cytosol. This causes an open reading frame-shift that induces expression of many downstream genes in reaction to ER stress as part of the unfolded protein response (UPR). We previously published an algorithm termed “Read-Split-Walk” (RSW) to identify non-canonical splicing regions using RNA-Seq data and applied it to ER stress-induced Ire1α heterozygote and knockout mouse embryonic fibroblast cell lines. In this study, we have developed an improved algorithm “Read-Split-Run” (RSR) for detecting genome-wide Ire1α-targeted genes with non-canonical spliced regions at a faster speed. We applied the RSR algorithm using different combinations of several parameters to the previously RSW tested mouse embryonic fibroblast cells (MEF) and the human Encyclopedia of DNA Elements (ENCODE) RNA-Seq data. We also compared the performance of RSR with two other alternative splicing events identification tools (TopHat (Trapnell et al., Bioinformatics 25:1105–1111, 2009) and Alt Event Finder (Zhou et al., BMC Genomics 13:S10, 2012)) utilizing the context of the spliced Xbp1 mRNA as a positive control in the data sets we identified it to be the top cleavage target present in Ire1α+/− but absent in Ire1α−/− MEF samples and this comparison was also extended to human ENCODE RNA-Seq data. Results Proof of principle came in our results by the fact that the 26 nt non-conventional splice site in Xbp1 was detected as the top hit by our new RSR algorithm in heterozygote (Het) samples from both Thapsigargin (Tg) and Dithiothreitol (Dtt) treated experiments but absent in the negative control Ire1α knock-out (KO) samples. Applying different combinations of parameters to the mouse MEF RNA-Seq data, we suggest a General Linear Model (GLM) for both Tg and Dtt treated experiments. We also ran RSR for a human ENCODE RNA-Seq dataset and identified 32,597 spliced regions for regular chromosomes. TopHat (Trapnell et al., Bioinformatics 25:1105–1111, 2009) and Alt Event Finder (Zhou et al., BMC Genomics 13:S10, 2012) identified 237,155 spliced junctions and 9,129 exon skipping events (excluding chr14), respectively. Our Read-Split-Run algorithm also outperformed others in the context of ranking Xbp1 gene as the top cleavage target present in Ire1α+/− but absent in Ire1α−/− MEF samples. The RSR package including source codes is available at http://bioinf1.indstate.edu/RSR and its pipeline source codes are also freely available at https://github.com/xuric/read-split-run for academic use. Conclusions Our new RSR algorithm has the capability of processing massive amounts of human ENCODE RNA-Seq data for identifying novel splice junction sites at a genome-wide level in a much more efficient manner when compared to the previous RSW algorithm. Our proposed model can also predict the number of spliced regions under any combinations of parameters. Our pipeline can detect novel spliced sites for other species using RNA-Seq data generated under similar conditions. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2896-7) contains supplementary material, which is available to authorized users. Keywords Alternative splicingNon-canonicalRNA-SeqXbp1ENCODEThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background In metazoans, during endoplasmic reticulum (ER) stress, the endoribonuclease (RNase) Inositol Requiring Enzyme 1a (Ire1α) initiates removal of a 26 nt region from the mRNA encoding the transcription factor Xbp1 via an non-canonical mechanism (atypically within the cytosol). This causes a transitional open reading frame-shift to produce a potent transcription factor, Xbp1s, that induces expression of numerous downstream genes in response to ER stress as part of the unfolded protein response (UPR) [1, 2]. In addition, spliceosome-independent cytoplasmic splicing, as a part of the unfolded protein response pathway, has been described in yeast [3] where HAC1p was found to be the sole splicing substrate of Ire1. The mechanism of Ire1α-mediated RNA-splicing is likely conserved in all eukaryotes as well [4]. In recent years, many popular methods have been developed to identify novel splice sites in RNA-Seq data, including TopHat [5] and Alt Event Finder [6]. A detailed review on the limitations of several other tools for identification of alternative splicing events (TrueSight [7], Splicing-Compass [8], and PASTA [9]) was described previously [10]. In short, indeed none of these existing tools were suitably designed for detecting the type of non-canonical sometimes called non-canonical splice sites generated by Ire1α-targeted Xbp1 mRNA splicing. Given that non-canonical splicing events of short mRNA regions occurring within the cytosol have not yet been investigated using next-generation technologies at a genome-wide level, cutting-edge bioinformatics methods of detecting such targets are needed to quickly discover such splicing events in a patient-specific manner in order to derive future therapeutic value. In order to supply the medical and scientific fields with such a tool we previously developed a novel bioinformatics pipeline method, named Read-Split-Walk [10] for detecting non-canonical, short, splicing regions using RNA-Seq data. We applied the method to ER stress-induced Ire1α heterozygous and knockout mouse embryonic fibroblast (MEF) cell lines to identify Ire1α targets of which the 26 nt non-canonical splice site in Xbp1 was detected as the most prominent splice target by our initial RSW pipeline in heterozygous (Het) samples, not mapped in the negative control Ire1α knockout (KO) samples for both Thapsigargin (Tg) and Dithiothreitol (Dtt) treated experiments. In our previous study, we also compared the Xbp1 results from our approach with results using the alignment program BWA [11], Bowtie2 [12], STAR [13], Exonerate [14] and the Unix “grep” command. Although our previous RSW method gave better results overall than the above-mentioned approaches, we realized that RSW’s running speed needed to be further improved in order to handle the massive amount of data in other experiments (human ENCODE project: https://www.encodeproject.org). In addition, we wanted to test, under different combinations of parameters, how and where reported spliced regions would differ. Therefore, we have designed a newer algorithm which we call “Read-Split-Run” (RSR) that can process RNA-Seq data in a more efficient manner with flexible parameters. We also proposed a linear regression equation under the assumption of the Generalized Linear Model for RSR parameters that can automatically predict the number of spliced regions given any parameter settings for a particular experiment. We compared our RSR algorithm with the above-mentioned alternative splicing events detection tools using metrics of how each tool ranks Xbp1 as the top cleavage target and its presence and absence in Ire1α+/− and Ire1α−/− MEF samples. We have also compared our RSR pipeline and other tools to process a human ENCODE dataset and reported their statistics of running performance and sensitivity (the number of spliced junctions identified). Results The web interface features of RSR In addition to providing the source code for download, the current web site of RSR (http://bioinf1.indstate.edu/RSR) allows users to use RSR by using a web form to upload data to the RSR server. After a job is submitted the server runs the pipeline and sends an email with a download link when the results are ready. The web form allows selection of a flexible combination of parameters. For example, users can select “Mode (Comparative or Non-comparative)”, “Reads Type (Single or Paired-end reads)”, “Experimental Replicates (1, 2, 3, …)”. The input files must be in FASTQ format. Based on the user’s initial selection, the pre-processing step will automatically reflect the number of input files needed. Users also have the options of checking the quality encoding and read length for short read input sequence files. The pipeline moves to the next step only if the read lengths for all input files are confirmed to be equal. A screenshot for the RSR web interface is shown in Fig. 1.Fig. 1 A screenshot for the Read-Split-Run web interface The spliced regions detected by the RSR pipeline for mouse MEF RNA-Seq data The identified spliced regions by the RSR pipeline for five cases with different combinations of parameters when processing Ire1α(+/−) and Ire1α(−/−) samples in both experiments (Tg and Dtt treated) are shown in Table 1. The detailed information of spliced regions identified by RSR for Tg and Dtt treated samples under different parameter settings are reported in Additional files 1 and 2. The attached Additional files 3 and 4 give statistics from running the new pipeline on mouse Tg and Dtt data with various parameter settings for MS, MD, and BB on a system with two Intel Xeon E5-2650 2.6GHz processors (a total of 16 cores and 32 hyperthreads).Table 1 Comparison of total number of junctions identified by RSR for five cases from Tg and Dtt treated samples Parameter Case 1 Case 2 Case 3 Case 4 Case 5 Variable Name Minimum split size 8 11 11 15 18 Maximum candidate distance 40,000 40,000 50,000 50,000 50,000 Read mapping region boundary buffer size 5 5 5 8 8 Minimum candidate distance 2 2 2 2 2 Minimum number of supporting reads 2 2 2 2 2 Maximum good alignment allowed per read 8 11 11 15 18 Tg Het Total number of junctions identified 122 140 143 135 141 Tg KO Total number of junctions identified 153 177 177 170 177 Dtt Het Total number of junctions identified 6496 6614 6661 6673 5683 Dtt KO Total number of junctions identified (Novel/Known) 6135 6247 6285 6308 5687 The bolded numbers show parameters with different test settings The spliced regions detected by the RSR pipeline for human ENCODE data The total number of spliced regions and running time of RSR on each chromosome of human ENCODE RNA-Seq data are shown in Fig. 2. The analysis for Chr14 was not performed due to memory constraints on the machine running the RSR pipeline. Additional file 5 shows the number of spliced regions identified by our RSR algorithm for each chromosome of the human ENCODE RNA-Seq dataset.Fig. 2 Number of junctions and clock time reported by RSR for the human ENCODE RNA-Seq sample (Separated by chromosomes). Blue bar: junctions; red bar: clock time; purple bar: reads processed (million) Comparison of detected spliced regions between RSR and other tools We also compared our RSR algorithm with other NGS based alternative splicing events detection tools (TopHat and Alt Event Finder). Our Read-Split-Run algorithm outperformed the other two software in the context of ranking Xbp1 gene as the top cleavage target present in Ire1α+/− but absent in Ire1α−/− MEF samples. In particular, we have ran TopHat and Alt Event Finder on both Tg and Dtt samples of MEF cell line RNA-Seq data. TopHat identified 23 reads supporting Xbp1 in Tg Het and 86 reads for Tg KO samples. For Dtt samples, TopHat reported a total of 59 (Het) and 289 (KO). Although the number of reads supporting Xbp1 reported by TopHat in Tg Het sample is slightly higher than our RSR method (23 vs 21), our RSR accomplished a better turnaround (173) when compared to TopHat (59) in the Dtt dataset. Surprisingly, TopHat also reported reads supporting Xbp1 in KO samples (86 in Tg and 289 in Dtt), which are false positive reads. Alt Event Finder did not identify any reads in supporting 26 nt Xbp1 spliced regions. The comparison results between RSR and other tools in running RNA-Seq data from MEF cell line is listed in Table 2.Table 2 Number of reads for supporting Xbp1 26 nt spliced regions reported by RSR and other tools 500 nM Thapsigargin (Tg) 1 mM Dithiothreitol (Dtt) Software Het (Ire1α +/−) KO (Ire1α −/−) Het (Ire1α +/−) KO (Ire1α −/−) Read-Split-Run (RSR) 21 0 173 0 TopHat 23 86 59 289 BWA 0 0 67 0 Bowtie2 0 0 171 0 STAR 0 0 0 0 Alt Event Finder 0 0 0 0 When applying these tools on human ENCODE RNA-Seq data, we found that the running speed and splicing regions detected by these tools are different. Due to the memory constraint of our server, we ran RSR by splitting the genome alignment files into individual chromosomes. It took RSR less than half an hour to run most chromosomes except for chromosome 9 and 12. In contrast, it took TopHat roughly 20 h to run the ENCODE RNA-Seq dataset. Alt Event Finder only required 3 h time to process the dataset, but the result was not informative for identifying short non-canonical spliced regions. Indeed, we have designed the web-based interface to and results reporting of our RSR pipeline to be as user-friendly as possible. Figure 3 shows comparison results between RSR and other tools in identifying number of spliced regions for the human ENCODE RNA-Seq dataset that we have tested.Fig. 3 Comparison results in identifying number of spliced regions between RSR, Alt Event Finder, TopHat tools. Blue bar: TopHat; red bar: RSR; green bar: AltEventFinder The general linear model for RSR algorithm We generated a General Linear Model for both Tg and Dtt samples in the context of the total number of unique spliced regions (i.e. present in Het and not in KO samples) identified by the RSR algorithm (Fig. 4). The model is derived from the results based on many pairs of Het and KO test cases (100 for Tg and 64 for Dtt samples) with different combinations of three RSR algorithm parameters (minimum split size (MS), maximum candidate distance (MD), and read mapping region boundary buffer size (BB)). We therefore obtained two linear regression equations (one for Tg and the other for Dtt) as shown in Fig. 4. It is clear that the numbers of spliced regions identified by RSR decreases as MS values increase. This is true because the numbers of split pairs fed into the second step of bowtie decrease when MS values increase. The parameter of MD plays less critical roles as we expected. The BB parameter seems to follow the correlation (the numbers of spliced regions increase as BB values increase) for the Tg dataset, but not for the Dtt dataset. We would like to increase test cases for the Dtt sample to see whether the trend will change.Fig. 4 Linear regression equations for mouse MEF Tg and Dtt experiments Discussion Parameters consideration in RSR pipeline Our proposed model of taking different parameter combinations to run the bowtie aligner and RSR algorithm could be applied to other species. However, different parameter combinations would predict different outcomes (i.e. number of spliced regions) for different species, even for different experiments. In our pipeline, we chose three parameters (MS, MD, and BB) to test how different combinations affect the prediction outcome. Specifically, we have run a number of test cases (100 for Tg and 64 for Dtt) with different combinations and used the results to generate linear regression equations. To increase the robustness of our RSR algorithm, it is ideal to perform a large-scale simulation study in order to look for the optimal combination. Typical questions remain to be answered: What would be the trade-off between lower/upper bound split size vs alignment sensitivity? What is the optimal consolidation slip/buffer size? More or fewer supporting reads will be reported if different cut-off criteria are applied, and these can be adjusted to achieve the desired balance between sensitivity and specificity in specific applications. We chose three parameters (minimum split size (MS), maximum candidate distance (MD), and read mapping region boundary buffer size (BB)) for the RSR algorithm because the number of reads supporting spliced regions could be different given different combination values of these parameters. Our GLM was generated by running a number of test cases. A smaller MS and bigger MD and BB could increase the number of junctions reported. Empirically MS could be set to approximately 1/3 of the read length, and should also be larger than 8 bp to ensure the split half is not too short to be mapped accurately. An estimated MD value could be determined according to the average gene length of species tested. Therefore, users could choose their customized MD value according to the species on which their experiments were performed. This criterion is determined according to the assumption that the split pairs supporting the junction are often mapped onto the same gene. Finally, BB could be set to a number that should not reflect a large boundary variation (5 or less would be reasonable). The file deletion of our RSR algorithm We noticed, by splitting reads into multiple read half pairs, the size of the result files substantially increased when the human ENCODE dataset was processed. To reduce the hard drive storage of large data files, we automatically delete files throughout the pipeline as they are no longer needed. For example, if the split step only needs to utilize the unmapped read datasets, the alignment files generated from the first step of bowtie are deleted. After the second step of bowtie is finished, all unmapped files will be deleted and only the alignment file will be kept for the next RSR step. The RSR pipeline can also compare spliced regions between samples and output reported regions side-by-side. RSR running speed, sensitivity, and specificity In this study, we have developed a newer pipeline (RSR) of RSW with an improved running speed and proposed a General Linear Model for the algorithm. We used RSR to process different combinations of running parameters for the MEF and human ENCODE RNA-Seq data. We have compared our RSR algorithm with two popular NGS based alternative splicing events detection tools (TopHat and Alt Event Finder) and reported the spliced regions detection results. Neither of these two tools achieved better sensitivity (Number of junctions identified) than our RSR algorithm in identifying reads supporting the Xbp1 26 bp spliced region. This can be explained in part due to the fact that Alt Event Finder processes the mapped reads to report splice regions yet does not consider unmapped reads in the analysis input. Moreover, the current version of Alt Event Finder focuses on identifying exon skipping events only. TopHat reported canonical exon-exon splice sites as well. The spliced junctions identified by TopHat and our RSR are reported in Table 3. TopHat reported more junctions than our RSR. But many of them were known junctions or false positive ones. It is clear that the common junctions detected by both tools or overlapping rate is low. Indeed, the overlapping rate is even smaller for results from the human ENCODE dataset.Table 3 The overlapping spliced junctions identified by TopHat and our RSR mouse-Het mouse-KO mouse-Het mouse-KO human Software 500 nM Thapsigargin (Tg) 1 mM Dithiothreitol (Dtt) ENCODE TopHat 956 923 8897 7847 237,155 RSR 144 183 6727 6343 32,597 Common 38 41 2398 2128 314 Common/TopHat 3.97 % 4.44 % 26.95 % 27.12 % 0.13 % Common/RSR 26.39 % 22.40 % 35.65 % 33.55 % 0.96 % In our original RSW paper, we also compared the Xbp1 results from our approach with results using the alignment program BWA [11], Bowtie2 [12], STAR [13], Exonerate [14] and the Unix “grep” command. Although our RSW method gave better results overall than the above-mentioned approaches, comparison results also suggested that reads supporting removal of the 26 nt intron from Xbp1 mRNA were not fully acknowledged. A study using in vitro cleavage assay combined with microarray analysis reported 13 additional mRNAs as Ire1α cleavage targets [15]. The discovery shed light on the existence of other possible targets. A future version of the algorithm will focus on rescuing these false negative reads in order to achieve a better sensitivity. Applying RSR on human ENCODE RNA-Seq data The discovery of a new set of non-canonical splicing events in humans is important not only because of the obvious potential for novel alteration of targeted transcript function, but also the potential for the resulting excised sequences to function as silencing RNAs associated with particular disease states. In addition, the frequency of these novel splicing events could be subject to altered regulation in some individuals, resulting in identifiable splicing profiles associated with the risk of certain diseases. We used ENCODE RNA-Seq datasets to train our RSR algorithm and hope to identify additional targets and elucidate their splicing patterns. Results should eventually provide unique insight of elucidating how short non-canonical spliced sequences act their biological functions in the context of relevant biological processes and diseases. Conclusions The positive control for our application, the Xbp1 26 nt non-canonical splice site, was clearly detected in Het samples but not in the KO control samples from Tg and Dtt treated MEF experiments, and was reported again as the top cleavage target for an Ire1α target splice site. Although we have tested the RSR using human ENCODE datasets, our algorithm could also be easily extended for prediction of spliced regions for other species under any given parameter settings. Methods The RNA-Seq read sequence data The mouse test data were downloaded from NCBI Gene Expression Omnibus (GEO) under the accession number GSE54631. Mouse embryonic fibroblast cells (MEF) that were heterozygous for Ire1α (Ire1α(+/−)) and cells which had Ire1α knocked out (Ire1α(−/−)) treated for 4 h with either 500nM Thapsigargin (Tg) or 1 mM Dithiothreitol (Dtt). Both RNA-Seq experiments are single end reads and had no experimental replicates performed. The human test data are ENCSR000CUR which were downloaded from the ENCODE project (https://www.encodeproject.org/experiments/ENCSR000CUR/). The data were paired-end RNA-Seq experiments performed on human skin melanocytes primary whole cells (NHEM-M2) and sequenced using Illumina HiSeq 2000. There were two biological replicates (adult 52 years old and adult 55 years old) and no technical replicates used in this experiment. The reference genome for Read-Split-Run We downloaded the mouse (mm9) and human (hg19) genome reference sequences from the University of California Santa Cruz (UCSC) genome browser (http://genome.ucsc.edu). We also downloaded respective UCSC gene files (knownGene.txt) from the UCSC genome browser. The splice junction file was created by setting the sequence entry on each side of the junction site to 4 bp shorter than the read length using a RNA-Seq software python script (getsplicefa.py) from ERANGE version 3.1 (http://woldlab.caltech.edu/~alim/RNAseq/). The original reference genome and splice junction site file were merged together to form an expanded genome. The algorithm of Read-Split-Run We first recall the basic pipeline of the earlier work [10] before highlighting areas of improvement. Pseudocode is given for the pipeline in Fig. 5. The pipeline takes as input an RNA-seq file containing many short reads. The bowtie sequence aligner, version 1.0.1, [16] is first invoked, and unmapped reads are passed to the next stage of the pipeline as possible candidates resulting from the splicing. If a given non-aligned read sequence S did result from the splicing, the splice point could be at any position within S. The next stage of the algorithm splits each non-aligned sequence S into pairs (S1, S2) in all ways so that both parts are at least some minimum size (a parameter we denote MS, with a typical value between 8 and 1/3 of length of the original read sequences). Bowtie is invoked again, this time on each sub-sequence that resulted from splitting a non-aligned sequence from the first invocation of bowtie. Alignments of the sub-sequences are scanned for sub-sequences that were (i) derived as split pairs from the same original non-aligned read sequence, and (ii) aligned at positions on the same chromosome that are not too far apart (a parameter we denote MD, with a typical value of around 40,000). These conditions are consistent with a splicing event, and we save all pairs of alignments that satisfy the conditions, which we call “matched pairs”. The final stage of the pipeline scans all matched pairs to determine for each matched pair how many other matched pairs are likely a result of the same splice location; one matched pair “supports” another if the spliced region between the two ends is the same length and at a position on the same chromosome that is very close (a parameter we denote BB, with a typical value of between 2 and 5). The most interesting splice junctions are those with the highest number of matched pairs that support them.Fig. 5 Pseudocode for Read-Split-Run algorithm. The junctions output by step 6 of the algorithm can optionally be restricted to those which are supported by some minimum number of sequences The present work began by porting the previous pipeline from being written in Perl to C++ (compiled with g++ 4.8.3 using optimization parameter –O4 on Linux). Porting to C++ resulted in a speedup by a factor of roughly two to three. Other than the sequence alignment using bowtie, most of the running time in the pipeline is in comparing aligned sub-sequences to determine the set of matched pairs, and comparing matched pairs to determine which support each other. The previous pipeline compared all pairs in each step, resulting in a running time that is quadratic in the number of sub-sequences coming out of the second bowtie step. We improve this step drastically so that the running time is quadratic only in the number of sub-sequences that were derived from the same initial non-aligned read sequence (typically less than a few hundred, whereas there may be millions of sub-sequences coming out of the second bowtie step). We obtain a similar improvement in the step that scans matched pairs to determine which support each other. Methods of running Read-Split-Run on mouse MEF RNA-Seq data The running parameter values employed for both Tg and Dtt samples are shown in Table 4. The steps of finding matched pairs and scanning matched pairs for supporting reads for Dtt samples were performed on a separate system running two Intel Xeon E5-2680 2.8Gz processors (a total of 20 cores and 40 hyperthreads). The highest running times are for tests with lower values of MS – this increases the number of sub-sequences that must be considered. A larger value for the length of initial read sequences also increases running time; the Dtt tests had higher running times because they consist of 77 bp (as opposed to 33 bp for Tg tests) and because the Dtt tests had roughly three times as many read sequences to begin with.Table 4 The running parameter values employed for both Tg and Dtt samples Sample Minimum split size (MS) Maximum candidate distance (MD) Read mapping region boundary buffer size (BB) Tg 8, 11, 12, 16 10000, 20000, 30000, 40000, 50000 1, 3, 5, 7, 9 DTT 11, 16, 20, 24 10000, 20000, 40000, 50000 3, 5, 7, 9 Methods of running Read-Split-Run on human ENCODE RNA-Seq dataset The phases: bowtie, splitting, and second step of bowtie were performed on same hardware mentioned above as the Dtt and Tg sets, whereas the RSR program was run on the “compute node,” described above. The parameters for this run were: MS - 33, MD - 50,000, and BB - 5. Before we could run the split-pairs portion of the pipeline, the output from bowtie (phase 2) had to be split into individual chromosomes so that they could fit into memory. Even in doing so, chromosome 14 had so much data (611Gb) that it could not be run. No junctions were identified on chromosome M. Comparison with other tools We compared our RSR algorithm with a couple of other NGS based alternative splicing events detection tools (TopHat [5] and Alt Event Finder [6]). We applied these tools on RNA-Seq data from a mouse embryonic fibroblast (MEF) cell line to check which of these tools can identify and rank Xbp1 as the top cleavage target and its presence and absence in Ire1+/− and Ire1−/− MEF samples and extended the analysis to the ENCODE RNA-Seq datasets. We ran TopHat v2.0.13 using options: −I 3000000, −g 10, --coverage-search, −microexon-search, to generate the “accepted_hits.bam” file for RNA-Seq data for each experiment condition from MEF cell line. Alt Event Finder v0.1 was ran by taking the “transcript.gtf” file generated from Cufflinks [17–20] and the “accepted_hits.bam” file generated by TopHat. Other metrics (i.e. running speed and usability) of these tools were also examined. A general linear model for RSR We proposed a modified General Linear Model (Fig. 6) for RSR. The variables (parameters) considered in the model are: minimum split size, maximum candidate distance, and read mapping region boundary buffer size. 80 test cases for Het and KO samples of both Tg and Dtt experiments were run on mouse MEF datasets to produce the General Linear Model equation.Fig. 6 A modified General Liner Model for the Read-Split-Run algorithm Abbreviations BB, boundary buffer size; Dtt, Dithiothreitol; ENCODE, Encyclopedia of DNA Elements; ER, endoplasmic reticulum; GLM, General Linear Model; Het, heterozygote; KO, knock-out; MD, maximum candidate distance; MEF, mouse embryonic fibroblast; MS, minimum split size; RSR, Read-Split-Run; RSW, Read-Split-Walk; Tg, Thapsigargin; UPR, unfolded protein response Additional files Additional file 1: Spliced regions identified by RSR for the Tg sample. This file contains all supplementary results for five test cases for the Tg sample. (XLSX 105 kb) Additional file 2: Spliced regions identified by RSR for the Dtt sample. This file contains all supplementary results for five test cases for the Dtt sample. (XLSX 2828 kb) Additional file 3: Statistics reported by RSR for 100 test cases for the Tg sample. This file contains all supplementary results for 100 test cases for the Tg sample. (XLSX 92 kb) Additional file 4: Statistics reported by RSR for 64 test cases for the Dtt sample. This file contains all supplementary results for 64 test cases for the Dtt sample. (XLSX 79 kb) Additional file 5: Total number of spliced regions identified by RSR in the human ENCODE RNA-Seq dataset. This file includes all supplementary results for number of supporting reads, splice length, range of supporting reads (spliced regions) identified by RSR in the human ENCODE RNA-Seq dataset. (XLSX 1411 kb) Acknowledgements This research was supported by ISU start-up funds to YB. RJK acknowledges support from NIH grants DK088227, DK042394, DK103183 and CA128814 and the Hevery Foundation. The authors thank The Center for Genomic Advocacy (TCGA) and Department of Mathematics and Computer Science at Indiana State University for computing servers. Authors also thank reviewers for comments and Gary Stuart for helping biological interpretation. Declarations The publication charges for this article have been funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and material The datasets supporting the conclusions of this article are included within the article and its additional files. The RSR source and accompanying examples are freely available for academic use at https://github.com/xuric/read-split-run under the Apache License, Version 2.0 license. Authors’ contributions YB designed and supervised the project, performed the analysis, provided biological interpretation, and wrote the manuscript. JK wrote the software code, performed the analysis, and wrote the manuscript. BD wrote the software code, ran the pipeline, and provided the human data results. FJ and LD participated in result comparisons between RSR and other tools. JRH performed the experiments and provided biological interpretation. RJK designed the experiments and wrote the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Calfon M Zeng H Urano F Till JH Hubbard SR Harding HP Clark SG Ron D IRE1 couples endoplasmic reticulum load to secretory capacity by processing the XBP-1 mRNA Nature 2002 415 6867 92 96 10.1038/415092a 11780124 2. Yoshida H Matsui T Yamamoto A Okada T Mori K XBP1 mRNA is induced by ATF6 and spliced by IRE1 in response to ER stress to produce a highly active transcription factor Cell 2001 107 7 881 891 10.1016/S0092-8674(01)00611-0 11779464 3. Ruegsegger U Leber JH Walter P Block of HAC1 mRNA translation by long-range base pairing is released by cytoplasmic splicing upon induction of the unfolded protein response Cell 2001 107 1 103 114 10.1016/S0092-8674(01)00505-0 11595189 4. Back SH Lee K Vink E Kaufman RJ Cytoplasmic IRE1alpha-mediated XBP1 mRNA splicing in the absence of nuclear processing and endoplasmic reticulum stress J Biol Chem 2006 281 27 18691 18706 10.1074/jbc.M602030200 16644724 5. Trapnell C Pachter L Salzberg SL TopHat: discovering splice junctions with RNA-Seq Bioinformatics 2009 25 9 1105 1111 10.1093/bioinformatics/btp120 19289445 6. Zhou A Breese MR Hao Y Edenberg HJ Li L Skaar TC Liu Y Alt Event Finder: a tool for extracting alternative splicing events from RNA-seq data BMC Genomics 2012 13 Suppl 8 S10 10.1186/1471-2164-13-S8-S10 23281921 7. Li Y Li-Byarlay H Burns P Borodovsky M Robinson GE Ma J TrueSight: a new algorithm for splice junction detection using RNA-seq Nucleic Acids Res 2013 41 4 e51 10.1093/nar/gks1311 23254332 8. Aschoff M Hotz-Wagenblatt A Glatting KH Fischer M Eils R Konig R SplicingCompass: differential splicing detection using RNA-seq data Bioinformatics 2013 29 9 1141 1148 10.1093/bioinformatics/btt101 23449093 9. Tang S Riva A PASTA: splice junction identification from RNA-sequencing data BMC Bioinformatics 2013 14 116 10.1186/1471-2105-14-116 23557086 10. Bai Y Hassler J Ziyar A Li P Wright Z Menon R Omenn GS Cavalcoli JD Kaufman RJ Sartor MA Novel bioinformatics method for identification of genome-wide non-canonical spliced regions using RNA-Seq data PLoS One 2014 9 7 e100864 10.1371/journal.pone.0100864 24991935 11. Li H Durbin R Fast and accurate short read alignment with Burrows-Wheeler transform Bioinformatics 2009 25 14 1754 1760 10.1093/bioinformatics/btp324 19451168 12. Langmead B Salzberg SL Fast gapped-read alignment with Bowtie 2 Nat Methods 2012 9 4 357 359 10.1038/nmeth.1923 22388286 13. Dobin A Davis CA Schlesinger F Drenkow J Zaleski C Jha S Batut P Chaisson M Gingeras TR STAR: ultrafast universal RNA-seq aligner Bioinformatics 2013 29 1 15 21 10.1093/bioinformatics/bts635 23104886 14. Slater GS Birney E Automated generation of heuristics for biological sequence comparison BMC Bioinformatics 2005 6 31 10.1186/1471-2105-6-31 15713233 15. Oikawa D Tokuda M Hosoda A Iwawaki T Identification of a consensus element recognized and cleaved by IRE1 alpha Nucleic Acids Res 2010 38 18 6265 6273 10.1093/nar/gkq452 20507909 16. Langmead B Trapnell C Pop M Salzberg SL Ultrafast and memory-efficient alignment of short DNA sequences to the human genome Genome Biol 2009 10 3 R25 10.1186/gb-2009-10-3-r25 19261174 17. Roberts A Pimentel H Trapnell C Pachter L Identification of novel transcripts in annotated genomes using RNA-Seq Bioinformatics 2011 27 17 2325 2329 10.1093/bioinformatics/btr355 21697122 18. Roberts A Trapnell C Donaghey J Rinn JL Pachter L Improving RNA-Seq expression estimates by correcting for fragment bias Genome Biol 2011 12 3 R22 10.1186/gb-2011-12-3-r22 21410973 19. Trapnell C Hendrickson DG Sauvageau M Goff L Rinn JL Pachter L Differential analysis of gene regulation at transcript resolution with RNA-seq Nat Biotechnol 2013 31 1 46 53 10.1038/nbt.2450 23222703 20. Trapnell C Williams BA Pertea G Mortazavi A Kwan G van Baren MJ Salzberg SL Wold BJ Pachter L Transcript assembly and quantification by RNA-Seq reveals unannotated transcripts and isoform switching during cell differentiation Nat Biotechnol 2010 28 5 511 515 10.1038/nbt.1621 20436464

PMC005xxxxxx/PMC5001234.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557423278910.1186/s12864-016-2789-9ResearchEfficient sequential and parallel algorithms for finding edit distance based motifs Pal Soumitra 1Xiao Peng 1Rajasekaran Sanguthevar rajasek@engr.uconn.edu 21 Department of Computer Science and Engineering, University of Connecticut, 371 Fairfield Road,, Storrs, 06269 CT USA 2 Department of Computer Science and Engineering, University of Connecticut, 371 Fairfield Road,, Storrs, 06269 CT USA 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.465© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Motif search is an important step in extracting meaningful patterns from biological data. The general problem of motif search is intractable and there is a pressing need to develop efficient, exact and approximation algorithms to solve this problem. In this paper, we present several novel, exact, sequential and parallel algorithms for solving the (l,d) Edit-distance-based Motif Search (EMS) problem: given two integers l,d and n biological strings, find all strings of length l that appear in each input string with atmost d errors of types substitution, insertion and deletion. Methods One popular technique to solve the problem is to explore for each input string the set of all possible l-mers that belong to the d-neighborhood of any substring of the input string and output those which are common for all input strings. We introduce a novel and provably efficient neighborhood exploration technique. We show that it is enough to consider the candidates in neighborhood which are at a distance exactly d. We compactly represent these candidate motifs using wildcard characters and efficiently explore them with very few repetitions. Our sequential algorithm uses a trie based data structure to efficiently store and sort the candidate motifs. Our parallel algorithm in a multi-core shared memory setting uses arrays for storing and a novel modification of radix-sort for sorting the candidate motifs. Results The algorithms for EMS are customarily evaluated on several challenging instances such as (8,1), (12,2), (16,3), (20,4), and so on. The best previously known algorithm, EMS1, is sequential and in estimated 3 days solves up to instance (16,3). Our sequential algorithms are more than 20 times faster on (16,3). On other hard instances such as (9,2), (11,3), (13,4), our algorithms are much faster. Our parallel algorithm has more than 600 % scaling performance while using 16 threads. Conclusions Our algorithms have pushed up the state-of-the-art of EMS solvers and we believe that the techniques introduced in this paper are also applicable to other motif search problems such as Planted Motif Search (PMS) and Simple Motif Search (SMS). Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2789-9) contains supplementary material, which is available to authorized users. Keywords MotifEdit distanceTrieRadix sortIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Motif search has applications in solving such crucial problems as identification of alternative splicing sites, determination of open reading frames, identification of promoter elements of genes, identification of transcription factors and their binding sites, etc. (see e.g., Nicolae and Rajasekaran [1]). There are many formulations of the motif search problem. A widely studied formulation is known as (l,d)-motif search or Planted Motif Search (PMS) [2]. Given two integers l,d and n biological strings the problem is to find all strings of length l that appear in each of the n input strings with atmost d mismatches. There is a significant amount of work in the literature on PMS (see e.g., [1, 3–5], and so on). PMS considers only point mutations as events of divergence in biological sequences. However, insertions and deletions also play important roles in divergence [2, 6]. Therefore, researchers have also considered a formulation in which the Levenshtein distance (or edit distance), instead of mismatches, is used for measuring the degree of divergence [7, 8]. Given n strings S(1),S(2),…,S(n), each of length m from a fixed alphabet Σ, and integers l,d, the Edit-distance-based Motif Search (EMS) problem is to find all patterns M of length l that occur in atleast one position in each S(i) with an edit distance of atmost d. More formally, M is a motif if and only if ∀i, there exist k∈ [ l−d,l+d],j∈ [ 1,m−k+1] such that for the substring Sj,k(i) of length k at position j of S(i), EDSj,k(i),M≤d. Here ED(X,Y) stands for the edit distance between two strings X and Y. EMS is also NP-hard since PMS is a special case of EMS and PMS is known to be NP-hard [9]. As a result, any exact algorithm for EMS that finds all the motifs for a given input can be expected to have an exponential (in some of the parameters) worst case runtime. One of the earliest EMS algorithms is due to Rocke and Tompa [7] and is based on Gibbs Sampling which requires repeated searching of the motifs in a constantly evolving collection of aligned strings, and each search pass requires O(nl) time. This is an approximate algorithm. Sagot [8] gave a suffix tree based exact algorithm that takes O(n2mld|Σ|d) time and O(n2m/w) space where w is the word length of the computer. Adebiyi and Kaufmann [10] proposed an exact algorithm with an expected runtime of O(nm+d(nm)(1+pow(ε)) lognm) where ε=d/l and pow(ε) is an increasing concave function. The value of pow(ε) is roughly 0.9 for protein and DNA sequences. Wang and Miao [11] gave an expectation minimization based heuristic genetic algorithm. Rajasekaran et al. [12] proposed a simpler Deterministic Motif Search (DMS) that has the same worst case time complexity as the algorithm by Sagot [8]. The algorithm generates and stores the neighborhood of every substring of length in the range [l−d,l+d] of every input string and using a radix sort based method, outputs the neighbors that are common to atleast one substring of each input string. This algorithm was implemented by Pathak et al. [13]. Following a useful practice for PMS algorithms, Pathak et al. [13] evaluated their algorithm on certain instances that are considered challenging for PMS: (9,2), (11,3), (13,4) and so on [1], and are generated as follows: n=20 random DNA/protein strings of length m=600, and a short random string M of length l are generated according to the independent identically distributed (i.i.d) model. A separate random d-hamming distance neighbor of M is “planted” in each of the n input strings. Such an (l,d) instance is defined to be a challenging instance if l is the largest integer for which the expected number of spurious motifs, i.e., the motifs that would occur in the input by random chance, is atleast 1. The expected number of spurious motifs in EMS are different from those in PMS. Table 1 shows the expected number of spurious motifs for l∈ [ 5,21] and d upto max{l−2,13}, n=20, m=600 and Σ={A,C,G,T} [see Additional file 1]. The challenging instances for EMS turn out to be (8,1), (12,2), (16,3), (20,4) and so on. To compare with [13], we consider both types of instances, specifically, (8,1), (9,2), (11,3), (12,2), (13,4) and (16,3). Table 1 Expected number of spurious motifs in random instances for n=20,m=600. Here, ∞ represents value ≥1.0e+7 l d=0 1 2 3 4 5 6 7 8 9 10 11 12 13 5 0.0 1024.0 1024.0 ∞ 6 0.0 4096.0 4096.0 ∞ ∞ 7 0.0 14141.8 16384.0 ∞ ∞ ∞ 8 0.0 225.8 65536.0 65536.0 ∞ ∞ ∞ 9 0.0 0.0 262144.0 262144.0 ∞ ∞ ∞ ∞ 10 0.0 0.0 1047003.6 1048576.0 ∞ ∞ ∞ ∞ ∞ 11 0.0 0.0 1332519.5 4194304.0 ∞ ∞ ∞ ∞ ∞ ∞ 12 0.0 0.0 294.7 1.678e+07 1.678e+07 ∞ ∞ ∞ ∞ ∞ ∞ 13 0.0 0.0 0.0 6.711e+07 6.711e+07 ∞ ∞ ∞ ∞ ∞ ∞ ∞ 14 0.0 0.0 0.0 2.517e+08 2.684e+08 ∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞ 15 0.0 0.0 0.0 2.749e+07 1.074e+09 ∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞ 16 0.0 0.0 0.0 139.1 4.295e+09 4.295e+09 ∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞ 17 0.0 0.0 0.0 0.0 1.718e+10 1.718e+10 ∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞ 18 0.0 0.0 0.0 0.0 3.965e+10 6.872e+10 ∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞ 19 0.0 0.0 0.0 0.0 1.226e+08 2.749e+11 2.749e+11 ∞ ∞ ∞ ∞ ∞ ∞ ∞ 20 0.0 0.0 0.0 0.0 35.8 1.100e+12 1.100e+12 ∞ ∞ ∞ ∞ ∞ ∞ ∞ 21 0.0 0.0 0.0 0.0 0.0 4.333e+12 4.398e+12 ∞ ∞ ∞ ∞ ∞ ∞ ∞ The instances in bold represents challenging instances The sequential algorithm by Pathak et al. [13] solves the moderately hard instance (11,3) in a few hours and does not solve the next difficult instance (13,4) even after 3 days. A key time-consuming part of the algorithm is in the generation of the edit distance neighborhood of all substrings as there are many common neighbors. Contributions In this paper we present several improved algorithms for EMS that solve instance (11,3) in less than a couple of minutes and instance (13,4) in less than a couple of hours. On (16,3) our algorithm is more than 20 times faster than EMS1. Our algorithm uses an efficient technique (introduced in this paper) to generate the edit distance neighborhood of length l with distance atmost d of all substrings of an input string. Our parallel algorithm in the multi-core shared memory setting has more than 600 % scaling performance on 16 threads. Our approach uses following five ideas which can be applied to other motif search problems as well: Efficient neighborhood generation: We show that it is enough to consider the neighbors which are at a distance exactly d from all possible substrings of the input strings. This works because the neighbors at a lesser distance are also included in the neighborhood of some other substrings. Compact representation using wildcard characters: We represent all possible neighbors which are due to an insertion or a substitution at a position by a single neighbor using a wildcard character at the same position. This compact representation of the candidate motifs in the neighborhood requires less space. Avoiding duplication of candidate motifs: Our algorithm uses several rules to avoid duplication in candidate motifs and we prove that our technique generates neighborhood that is nearly duplication free. In other words, our neighborhood generation technique does not spend a lot of time generating neighbors that have already been generated. Trie based data structure for storing compact motifs: We use a trie based data structure to efficiently store the neighborhood. This not only simplifies the removal of duplicate neighbors but also helps in outputting the final motifs in sorted order using a depth first search traversal of the trie. Modified radix-sort for compact motifs: Our parallel algorithm stores the compact motifs in an array and uses a modified radix-sort algorithm to sort them. Use of arrays instead of tries simplifies updating the set of candidate motifs by multiple threads. Methods In this section we introduce some notations and observations. An (l,d)-friend of a k-mer L is an l-mer at an exact distance of d from L. Let Fl,d(L) denote the set of all (l,d)-friends of L. An (l,d)-neighbor of a k-mer L is an l-mer at a distance of atmost d from L. Let Nl,d(L) denote the set of all (l,d)-neighbors of L. Then 1 Nl,d(L)=∪t=0dFl,t(L). For a string S of length m, an (l,d)-motif of S is an l-mer at a distance atmost d from some substring of S. Thus an (l,d)-motif of S is an (l,d)-neighbor of atleast one substring Sj,k=SjSj+1…Sj+k−1 where k∈[l−d,l+d]. Therefore, the set of (l,d)-motifs of S, denoted by Ml,d(S), is given by 2 Ml,d(S)=∪k=l−dl+d∪j=1m−k+1Nl,d(Sj,k). For a collection of strings S={S(1),S(2),…,S(m)}, a (common) (l,d)-motif is an l-mer at a distance atmost d from atleast one substring of each S(i). Thus the set of (common) (l,d)-motifs of S, denoted by Ml,d(S), is given by 3 Ml,d(S)=∩i=1nMl,d(S(i)). One simple way of computing Fl,d(L) is to grow the friendhood of L by one distance at a time for d times and to select only the friends having length l. Let G(L) denote the set of strings obtained by one edit operation on L and G({L1,L2,…,Lr})=∪t=1rG(Lt). If G1(L)=G(L), and for t>1, Gt(L)=G(Gt−1(L)) then 4 Fl,d(L)={x∈Gd(L):|x|=l}. Using Eqs. (1), (2), (3) and (4), Pathak et al. [13] gave an algorithm that stores all possible candidate motifs in an array of size |Σ|l. However the algorithm is inefficient in generating the neighborhood as the same candidate motif is generated by several combinations of the basic edit operations. Also, the O(|Σ|l) memory requirement makes the algorithm inapplicable for larger instances. In this paper we mitigate these two limitations. Efficient neighborhood generation We now give a more efficient algorithm to generate the (l,d)-neighborhood of all possible k-mers of a string. Instead of computing (l,t)-friendhood for all 0≤t≤d, we compute only the exact (l,d)-friendhood. Lemma1. Ml,d(S)=∪k=l−dl+d∪j=1m−k+1Fl,d(Sj,k). Proof. Consider the k-mer L=Sj,k. If k=l+d then we need d deletions to make L an l-mer. There cannot be any (l,t)-neighbor of L for t<d. Thus 5 ∪t=0dFl,t(Sj,l+d)=Fl,d(Sj,l+d). Suppose k<l+d. Any (l,d−1)-neighbor B of L is also an (l,d)-neighbor of L′=Sj,k+1 because ED(B,L′)≤ED(B,L)+ED(L,L′)≤(d−1)+1=d. Thus ∪t=0dFl,t(Sj,k)⊆Fl,d(Sj,k)⋃∪t=0dFl,t(Sj,k+1) which implies that 6 ∪r=kk+1∪t=0dFl,t(Sj,r)=Fl,d(Sj,k)⋃∪t=0dFl,t(Sj,k+1). Applying (6) repeatedly for k=l−d,l−d+1,…,l+d−1, along with (5) in (1) and (2) gives the result of the lemma. We generate Fl,d(Sj,k) in three phases: we apply δ deletions in the first phase, β substitutions in the second phase, and α insertions in the final phase, where d=δ+α+β and l=k−δ+α. Solving for α,β,δ gives max{0,q}≤δ≤(d+q)/2, α=δ−q and β=d−2δ+q where q=k−l. In each of the phases, the neighborhood is grown by one edit operation at a time. Compact motifs The candidate motifs in Fl,d(Sj,k) are generated in a compact way. Instead of inserting each character in Σ separately at a required position in Sj,k, we insert a new character ∗∉Σ at that position. Similarly, instead of substituting a character σ∈Sj,k by each σ′∈Σ∖{σ} separately, we substitute σ by ∗. The motifs common to all strings in S is determined by using the usual definition of union and the following definition of intersection on compact strings A,B∈(Σ∪{∗})l in (3): 7 A∩B=∅if∃js.t.Aj,Bj∈Σ,Aj≠Bjσ1σ2…σlelse, whereσj=bjifaj=∗ajifbj=∗. Trie for storing compact motifs We store the compact motifs in a trie based data structure which we call a motif trie. This helps implement the intersection defined in (7). Each node in the motif trie has atmost |Σ| children. The edges from a node u to its children v are labeled with mutually exclusive subsets label(u,v)⊆Σ. An empty set of compact motifs is represented by a single root node. A non-empty trie has l+1 levels of nodes, the root being at level 0. The trie stores the l-mer σ1σ2…σl, all σj∈Σ, if there is a path from the root to a leaf where σj appears in the label of the edge from level j−1 to level j. For each string S=S(i) we keep a separate motif trie M(i). Each compact neighbor A∈Fl,d(Sj,k) is inserted into the motif trie recursively as follows. We start with the root node where we insert A1A2…Al. At a node u at level j where the prefix A1A2…Aj−1 is already inserted, we insert the suffix AjAj+1…Al as follows. If Aj∈Σ we insert A′=Aj+1Aj+2…Al to the children v of u such that Aj∈label(u,v). If label(u,v)≠{Aj}, before inserting we make a copy of subtrie rooted at v. Let v′ be the root of the new copy. We make v′ a new child of u, set label(u,v′)={Aj}, remove Aj from label(u,v), and insert A′ to v′. On the other hand if Aj=∗ we insert A′ to each children of u. Let T=Σ if Aj=∗ and T={Aj} otherwise. Let R=T∖∪vlabel(u,v). If T≠∅ we create a new child v′ of u, set label(u,v′)=R and recursively insert A′ to v′. Figure 1 shows examples of inserting into the motif trie. Fig. 1 Inserting into motif trie for Σ={A,C,G,T} and l=3. a After inserting ∗G T into empty trie. b After inserting another string A∗C We also maintain a motif trie ℳ for the common compact motifs found so far, starting with ℳ=M(1). After processing string S(i) we intersect the root of M(i) with the root of ℳ. In general a node u2∈M(i) at level j is intersected with a node u1∈ℳ at level j using the procedure shown in Algorithm 1. Figure 2 shows an example of the intersection of two motif tries. Fig. 2 Intersection of motif tries. a Trie for A G∗∪C∗T. b Intersection of trie in Fig. 1 b and trie in Fig. 2 a The final set of common motifs is obtained by a depth-first traversal of ℳ outputting the label of the path from the root whenever a leaf is traversed. An edge (u,v) is traversed separately for each σ∈label(u,v). Efficient compact neighborhood generation A significant part of the time taken by our algorithm is in inserting compact neighbors into the motif trie as it is executed for each neighbor in the friendhood. Our efficient neighborhood generation technique and the use of compact neighbors reduce duplication in neighborhood but do not guarantee completely duplication free neighborhood. In this section, we design few simple rules to reduce duplication further. Later we will see that these rules are quite close to the ideal as we will prove that the compact motif generated after skipping using the rules, are distinct if all the characters in the input string are distinct. To differentiate multiple copies of the same compact neighbor, we augment it with the information about how it is generated. This information is required only in the proof and is not used in the actual algorithm. Formally, each compact neighbor L of a k-mer Sj,k is represented as an ordered tuple 〈Sj,k,T〉 where T denotes the sequence of edit operations applied to Sj,k. Each edit operation in T is represented as a tuple 〈p,o〉 where p denotes the position (as in S) where the edit operation is applied and o∈{D,R,I} denotes the type of the operation – deletion, substitution and insertion, respectively. At each position there can be one deletion or one substitution but one or more insertions. The tuples in T are sorted lexicographically with the natural order for p and for o, D<R<I. The rules for skipping compact neighbors are given in Table 2. Rule 1 applies when Sj,k is not the rightmost k-mer and the current edit operation deletes the leftmost base of Sj,k, i.e., Sj. Rule 2 applies when the current edit operation substitutes a base just after a base that was already deleted. Rule 3 skips the neighbor which is generated from a k-mer except the rightmost by deleting a base and substituting all bases before it. Rules 4–9 apply when the current operation is an insertion. Rule 4,6 apply when the insertion is just before a deletion and a substitution, respectively. Rule 5 applies when the insertion is just after a deletion. Rule 7,8 apply when the k-mer is not the leftmost. Rule 7 applies when the insertion is at the leftmost position and Rule 8 applies when all bases before the position of insertion are already substituted. Rule 9 applies when the k-mer is not the rightmost and the insertion is at the right end. The first in each pair of the figures in Fig. 3 illustrates the situation where the corresponding rule applies. Fig. 3 Construction of L ′ under different rules in the proof of Lemma 2. Insertions are shown using arrows, deletions using − and substitutions using ∗. Rule 5 case (i) is similar to Rule 4 case (i) Table 2 Conditions for skipping motif L=〈M,S j,k,T〉 Rule Conditions (in all rules t≥0) 1 (j+k≤m)∧〈j,D〉∈T 2 {〈j+t,D〉,〈j+t+1,R〉}⊆T 3 (j+k≤m)∧{〈j,R〉,〈j+1,R〉,…,〈j+t,R〉,〈j+t+1,D〉}⊆T 4 {〈j+t,D〉,〈j+t,I〉}⊆T 5 {〈j+t,D〉,〈j+t+1,I〉}⊆T 6 {〈j+t,R〉,〈j+t,I〉}⊆T 7 (j>1)∧〈j,I〉∈T 8 (j>1)∧{〈j,R〉,〈j+1,R〉,…,〈j+t,R〉,〈j+t+1,I〉}⊆T 9 (j+k≤m)∧〈j+k,I〉∈T Let M¯l,d(S) denote the multi-set of tuples for the compact motifs of S that were not skipped by our algorithm using the rules in Table 2 and Ml, d(S) be the set of compact motifs generated by (3). Let Γ(〈Sj, k,T〉) be the resulting string when the operations in T are applied to Sj, k and Γ(Z)=∪L∈ZΓ(L). Lemma2. Γ(M¯l,d(S))=Ml,d(S). Proof. By construction, Γ(M¯l,d(S))⊆Ml,d(S). We show Ml,d(S)⊆Γ(M¯l,d(S)) by giving a contradiction when Ml,d(S)∖Γ(M¯l,d(S))≠∅. We define an order on the compact neighbors L1=〈Sj1,k1,T1〉 and L2=〈Sj2,k2,T2〉 as follows: L1<L2 if Γ(L1)<Γ(L2) and L2<L1 if Γ(L2)<Γ(L1). When Γ(L1)=Γ(L2) we have L1<L2 if and only if (k1<k2)∨((k1=k2)∧(p1<p2))∨((k1=k2)∧(p1=p2)∧(o1<o2)) where 〈p1,o1〉∈T1,〈p2,o2〉∈T2 are the leftmost edit operations where T1,T2 differ. Suppose M∈Ml,d(S)∖Γ(M¯l,d(S)). Let L=〈Sj, k,T〉 be the largest (in the order defined above) tuple skipped by our algorithm such that Γ(L)=M. For each r=1,…,9 we show a contradiction that if L is skipped by Rule r then there is another L′=〈Sj′,k′,T′〉 with the same number of edit operations and Γ(L′)=M but L<L′. Figure 3 illustrates the choice of L′ under different rules. Rule 1. Here j+k≤m and 〈j,D〉∈T. Consider T′=(T∖{〈j,D〉})∪{j+k,D}, and j′=j+1,k′=k. Rule 2. Consider T′=T∖{〈j+t,D〉,〈j+t+1,R〉}∪{〈j+t,R〉,〈j+t+1,D〉}, and j′=j,k′=k. Rule 3. T′=T∖{〈j,R〉,〈j+t+1,D〉}∪{〈j+t+1,R〉,〈j+k,D〉}, j′=j+1,k′=k. Rule 4. For this and subsequent rules k<l+d as there is atleast one insertion and hence k′ could possibly be equal to k+1. We consider two cases. Case (i) j+k≤m: T′=T∖{〈j+t,D〉,〈j+t,I〉}∪{〈j+t,R〉,〈j+k,D〉}, j′=j,k′=k+1. Case (ii) j+k=m+1: Here deletion of Sj is allowed by Rule 1. T′=T∖{〈j+t,D〉,〈j+t,I〉}∪{〈j−1,D〉,〈j+t,R〉}, j′=j−1,k′=k+1. Rule 5. The same argument for Rule 4 applies considering 〈j+t+1,I〉 instead of 〈j+t,I〉. Rule 6. T′=T∖{〈j+t,I〉}∪{〈j+t+1,I〉}, and j′=j,k′=k. Rule 7. T′=T∖{〈j,I〉}∪{〈j−1,R〉}, j′=j−1,k′=k+1. Rule 8. T′=T∖{〈j+t,I〉}∪{〈j−1,R〉}, j′=j−1,k′=k+1. Rule 9. T′=T∖{〈j+k,I〉}∪{〈j+k,R〉}, j′=j,k′=k+1. Consider two compact motifs M1=〈Sj1,k1,T1〉 and M2=〈Sj2,k2,T2〉 in M¯l,d(S). For q∈{1,2}, let pq(1),oq(1),pq(2),oq(2),…,pq(d),oq(d) be the sequence of edit operations in Tq arranged in the order as the neighbors are generated by our algorithm, and the intermediate neighbors be Lq(h)=Sjq,kq,pq(1),oq(1),pq(2),oq(2),…,pq(h),oq(h) for all h=1,2,…,d. We also denote the initial k-mer as a neighbor Lq(0)=〈Sjq,kq,∅〉. Lemma3. If Sjs are all distinct and ΓL1(h)=ΓL2(h) for 1≤h≤d then p1(h),o1(h)=p2(h),o2(h) and ΓL1(h−1)=ΓL2(h−1). Proof. To simplify the proof, we use pq,oq,Lq to denote pq(h),oq(h),Lq(h), respectively, for all q∈{1,2}. Without loss of generality we assume p1≤p2. As p1,p2 are positions in S, it would be enough to prove 〈p1,o1〉=〈p2,o2〉 because that would imply ΓL1(h−1)=ΓL2(h−1). If 〈p1,o1〉≠〈p2,o2〉 then either (a) o1=o2 and p1<p2 or (b) o1≠o2 and p1≤p2, giving us the following 9 possible cases. We complete the proof by giving a contradiction in each of these 9 cases: Case o 1 o 2 cond. Case o 1 o 2 cond. Case o 1 o 2 cond. 1 D D p 1<p 2 4 R D p 1≤p 2 7 I D p 1≤p 2 2 D R p 1≤p 2 5 R R p 1<p 2 8 I R p 1≤p 2 3 D I p 1≤p 2 6 R I p 1≤p 2 9 I I p 1<p 2 Cases 2, 3, 4, 7 Our algorithm applies edit operations in phases: first deletions, followed by substitutions and finally insertions. In all these cases, one of Γ(L1),Γ(L2) does not have any ∗ because only deletions have been applied so far and the other has at least one ∗ because a substitution or an insertion has been applied. This implies Γ(L1)≠Γ(L2), a contradiction. Case 1 L2 has Sp2 deleted. As Γ(L1)=Γ(L2), Sp2 must have been deleted in some operation prior to reaching L1. As the deletions are applied in order, left to right, we must have p1=p2 which is a contradiction. Case 5 This case has been illustrated in Fig. 4a. L1 has no substitution at a position >p1 and no insertion at all. The ∗ at p2 in L2 must be matched with the ∗ at p1 in L1 and as the characters in S are distinct, all of Sp1+1,…,Sp2 cannot appear in L1 and hence must be deleted in L1. Fig. 4 Proof of uniqueness (Lemma 2). Subfigures a,b,c,d illustrates the cases 5,6,7,8,9 respectively Now for each t<p1, right to left, and y=t+p2−p1, we have the following: Sy is either deleted or substituted in L1, which implies that Sy must be substituted in L2 as the deletion of Sy in L2 is prohibited by Rule 2, and finally to match this ∗ in L2, St must be substituted in L1 as St cannot be deleted in L1, again by Rule 2. But this makes Rule 3 applicable to L1 and L1 must have been skipped. This is a contradiction. Case 6 By Rule 9 the insertion in L2 cannot be at the rightmost position and hence L2 must have at least one character after the insertion. By Rules 4 and 6, Sp2 must not be deleted or substituted in L2 and hence it must not be deleted or substituted in L1 either. Thus p1<p2. There cannot be any insertion or substitution at a position >p1 in L1. Thus the ∗ due to the insertion at p2 in L2 must be matched by the ∗ due to the substitution at p1 in L1 and all of Sp1+1,…,Sp2−1 must be deleted in L1. By Rule 7, Sp2 cannot be the leftmost in Sj2,k2. So we consider Sp2−1 in L1,L2. It is either deleted or substituted in L1 and hence by Rule 5, it must be substituted in Sp2 (there can be multiple insertions at p2 in L2 but that does not affect this argument). To match this ∗, Sp1−1 must be substituted in L1. Using a similar argument as in Case 5, St must be substituted in L1 for each t<p1−1. But this again makes Rule 3 applicable to L1 and L1 must have been skipped, which is not possible. This case has been illustrated in Fig. 4b. Case 8 Due to Rules 4, 6 and 9, Sp1 must not be deleted or substituted in L1 and hence it must not be deleted or substituted in L2 either. Thus p1<p2. The ∗ due to the insertion in L1 must be matched by a substitution at p3<p1 such that all of Sp3+1,…,Sp1−1 are deleted in L2. By Rule 7, p1 cannot be the leftmost in L1. For each t<p1, right to left, and y=t+p3−p1, we have the following: Sy is substituted in L1 because as the deletion of Sy in L1 is prohibited by Rules 2 and 5, Sy must be substituted in L2 again by Rule 2, and to match this ∗, St must be substituted in L1. But this makes Rule 8 applicable to L1 and L1 must have been skipped which is not possible. This case has been illustrated in Fig. 4c. Case 9 This case has been illustrated in Fig. 4d. Due to Rules 4, 6 and 9, Sp1,Sp2 must not be deleted or substituted in L1,L2. The insertion at p2 in L2 must be matched by a substitution at a position p3 in L1 such that p1<p3<p2 and all of Sp3+1,…,Sp2−1 must be deleted in L1. Now for each position y, from right to left, where p1<y<p2, Sy is either deleted or substituted in S1, Sy cannot be deleted in L2 by Rules 2 and 5 and hence must be substituted in L2, which again must be matched by a substitution at a position t in L1 such that p1<t<p3. However this is impossible as the number of possible ys is larger than the number of possible ts. If all Sjs are distinct and Γ(M1)=Γ(M2) then applying Lemma 3 repeatedly for h=d,d−1,…,0 gives us the fact that starting k-mers Sj1,k1,Sj2,k2 as well as the corresponding edit operations in T1,T2 for M1,M2 must be the same. This is another way of stating the following theorem. Theorem1. If Sjs are all distinct then M¯l,d(S) is duplication free. In general Sjs are not distinct. However, as the input strings are random, the duplication due to repeated characters are limited. On instance (11,3) our algorithm generates each compact motif, on an average, 1.55 times using the rules compared to 3.63 times without the rules (see Fig. 5). Fig. 5 Histogram of number of times a motif is repeated with and without using the skipping rules 1–9 Implementation To track the deleted characters, instead of actually deleting we substitute them by a new symbol − not in Σ′. We populate the motif trie M(i) by calling genAll(S(i)) given in Algorithm 2. Rules 1–8 are incorporated in G(L,j,δ,β,α), H(L,j,β,α) and I(L,j,α) which are shown in Algorithms 3, 4, and 5, respectively where sub(L,j,σ) substitutes Lj by σ and ins(L,j,σ) inserts σ just before Lj. Modified radix-sort for compact motifs A simpler data structure alternative to tries for storing compact motifs could be an array. However, it becomes difficult to compute the intersection in (3) as defined in (7) when the compact motifs are stored in arrays. One straight-forward solution is to first expand the ∗s in the compact motifs, then sort the expanded motifs and finally compute the intersection by scanning through the two sorted arrays. This, to a great extent, wipes out the advantage using the ∗s in the compact motifs. However, we salvage execution time by executing a modified radix-sort that simultaneously expands and sorts the array of compact motifs: Compact-Radix-Sort(A,l) where the first parameter A represents the array of compact motifs and the second parameter represents the number of digits of the elements in A which is equal to the number of base positions l in a motif. As in the standard radix-sort, our algorithm uses l phases, one for each base position in the motif. In the ith phase it sorts the motifs using bucket sort on the ith base of the motifs. However, in case of compact motifs, for each ∗ at a base position, the bucket counters for all σ∈Σ are incremented. While reordering the motifs as per the bucket counts, if there is a ∗ at ith base position of a motif, |Σ| copies of the motif are created and they are placed at appropriate locations in the array after finalizing the correct σ for the ∗. The details are given in Algorithm 6. In each phase a bucket counter B and a cumulative counter C are used. The temporary array T stores the partially expanded motifs from the current phase. Discussion We did an experiment to compare the time taken by the two approaches – (i) using the expanded motifs, i.e., without using the wildcard character, and (ii) using compact motifs and sorting them using Compact-Radix-Sort. For a single input string of instance (16,3), the first approach generated in 24.4 s 198,991,822 expanded motifs in which 53,965,581 are unique. The second approach generated in 13.7 s 11,474,938 compact motifs with the same number of unique expanded motifs. This shows the effectiveness of the second approach. Parallel algorithm We now give our parallel algorithm in the multi-core shared memory setting. To process each input sequence S(i) the algorithm uses p+1 threads. The main thread first prepares the workload for other p threads. A workload involves the generation of the neighborhood for a k-mer of S(i), where l−d≤k≤l+d. There are total ∑k=l−dl+d(m−k+1)=(2d+1)(m−l+1) workloads. The number of neighbors generated in the workloads are not the same due to the skipping of some neighbors using rules 1–9. For load balancing, we randomly and evenly distribute workloads to threads. Each thread first generates all the compact motifs in its workloads and then sort them using Compact-Radix-Sort. If i>2 then it removes all neighbors not present in M(i−1) which is the set of common motifs of S(1),S(2),…,S(i−1). The master thread then merges the residue candidate motifs from all the p threads to compute M(i). The merging takes place in log2p phases in a binary tree fashion where the jth phase uses 2log2p−j threads each merging two sorted arrays of motifs. Results and discussion We implemented our algorithms in C++ and evaluated on a Dell Precisions Workstation T7910 running RHEL 7.0 on two sockets each containing 8 Dual Intel Xeon Processors E5-2667 (8C HT, 20 MB Cache, 3.2 GHz) and 256 GB RAM. For our experiments we used only one of the two sockets. We generated random (l,d) instances according to Pevzner and Sze [2] and as described in the background section. For every (l,d) combination we report the average time taken by 4 runs. We compare the following four implementations: EMS1: A modified implementation of the algorithm in [13] which considered the neighborhood of only l-mers whereas the modified version considers the neighborhood of all k-mers where l−d≤k≤l+d. EMS2: A faster implementation of our sequential algorithm which uses tries for storing candidate motifs where each node of the trie stores an array of pointers to each children of the node. However, this makes the space required to store a tree node dependent on the size of the alphabet Σ. EMS2M: A slightly slower but memory efficient implementation of our sequential algorithm where each node of the trie keeps two pointers: one to the leftmost child and the other to the immediate right sibling. Access to the other children are simulated using the sibling pointers. EMS2P: Our parallel algorithm which uses arrays for storing motifs. We experimented with p=1,2,4,8,16 threads. We run the four algorithms on the challenging instances (8,1), (12,2), (16,3) and on the instances (9,2), (11,3), (13,4) which are challenging for PMS and have been used for experimentation in [13]. We report the runtime and the memory usage of the four algorithms in Table 3. Table 3 Comparison between EMS1 and three implementations of EMS2 Instance Metric EMS1 EMS2 EMS2M EMS2P threads 1 2 4 8 16 (8,1) time 0.11 s 0.13 s 0.12 s 0.09 s 0.08 s 0.06 s 0.05 s 0.06 s memory 2.69 MB 4.25 MB 3.17 MB 2.67 MB 3.20 MB 3.55 MB 6.02 MB 7.99 MB (12,2) time 19.87 s 15.60 s 16.62 s 2.71 s 1.94 s 1.44 s 0.89 s 0.55 s memory 34.28 MB 210.47 MB 126.91 MB 84.98 MB 104.60 MB 125.18 MB 142.82 MB 150.23 MB (16,3) time 1.74 h 23.73 m 26.79 m 3.73 m 2.32 m 1.38 m 48.58 s 36.93 s memory 8.39 GB 11.62 GB 6.97 GB 8.55 GB 10.21 GB 10.53 GB 9.84 GB 9.91 GB (9,2) time 10.84 s 1.72 s 3.02 s 1.12 s 0.96 s 0.78 s 0.49 s 0.35 s memory 3.44 MB 26.67 MB 17.04 MB 42.86 MB 57.76 MB 54.77 MB 59.85 MB 66.53 MB (11,3) time 33.48 m 1.91 m 3.57 m 45.85 s 30.78 s 19.68 s 13.49 s 9.78 s memory 92.86 MB 477.12 MB 313.33 MB 2.27 GB 2.63 GB 2.65 GB 2.55 GB 2.60 GB (13,4) time - 1.08 h 1.76 h 44.03 m 26.16 m 14.51 m 8.62 m 6.82 m memory - 8.26 GB 5.58 GB 149.60 GB 179.66 GB 180.13 GB 168.80 GB 172.74 GB Time is in seconds (s), minutes (m) or hours (h). An empty cell implies the algorithm did not complete in the stipulated 72 h Our efficient neighborhood generation enables our algorithm to solve instance (13,4) in less than two hours which EMS1 could not solve even in 3 days. The factor by which EMS2 takes more memory compared to EMS1 gradually decreases as instances become harder. As EMS2 stores 4 child pointers for A,C,G,T in each node of the motif trie whereas EMS2M simulates access to children using only 2 pointers, EMS2 is faster. Memory reduction in EMS2M is not exactly by a factor 2(=4/2) because we also keep a bit vector in each node to represent the subset of {A,C,G,T} a child corresponds to. The memory reduction would be significant for protein strings. Our parallel algorithm EMS2P using one thread is significantly faster than the sequential algorithms EMS2 and EMS2M but uses more memory. This space-time trade off is due to the fact that the arrays are faster to access but the tries use lesser memory. Moreover, the repeated motifs are uniquely stored in a single leaf node in the trie but stored separately in the array. The scaling performance using multiple threads are shown in Fig. 6 where we plot the ratio of time taken by p threads to the time taken by a single thread on the Y-axis. The time required for handling 16 threads turns out to be costlier than actually processing the motifs in the smallest instance (8,1). We observe speed up consistent across other bigger instances. For example, instance (16,3) takes about 224 s using 1 thread and 37 s using 16 threads. This gives more than 600 % scaling performance using 16 threads. Fig. 6 Scaling performance of our parallel algorithm Conclusions We presented several efficient sequential and parallel algorithms for the EMS problem. Our algorithms use some novel and elegant rules to explore the candidate motifs in such a way that only a small fraction of the candidate motifs are explored twice or more. In fact, we also proved that these rules are close to ideal in the sense that no candidate motif is explored twice if the characters in the input string are all distinct. This condition may not be practical and ideas from [14] can be used when the characters in the input string are repeated. Nevertheless, the rules help because the instances are randomly generated and the k-mers in the input string are not much frequent. The second reason for the efficiency of our sequential algorithms is the use of a trie based data structure to compactly store the motifs. Our parallel algorithm stores candidate motifs in an array and uses a modified radix-sort based method for filtering out invalid candidate motifs. Our algorithms pushed up the state-of-the-art of EMS solvers to a state where the challenging instance (16,3) is solved in slightly more than half a minute using 16 threads. Future work could be to solve harder instances, including those involving protein strings, and possibly using many-core distributed algorithms. Additional file Additional file 1 Expected number of spurious motifs. This file gives the expression for the expected number of spurious (l,d)-motifs in n random strings of length m from the alphabet Σ. (PDF 143 kb) This work has been supported in part by the NIH grant R01-LM010101 and NSF grant 1447711. Declarations Publication of this article was funded by the NIH grant R01-LM010101 and NSF grant 1447711. This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at https://github.com/soumitrakp/ems2.git. Availability A C++ based implementation of our algorithm can be found at the following github public repository: https://github.com/soumitrakp/ems2.git. Authors’ contributions SP and SR conceived the study. SP implemented the algorithms and PX carried out the experiments. SP and SR analyzed the results and wrote the paper. All authors reviewed the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. ==== Refs References 1 Nicolae M, Rajasekaran S. qPMS9: An Efficient Algorithm for Quorum Planted Motif Search. Nat Sci Rep. 2015;5. doi:10.1038/srep07813. 2 Floratou A, Tata S, Patel JM. Efficient and Accurate Discovery of Patterns in Sequence Data Sets. IEEE Trans Knowl Data Eng. 2011; 23(8):1154–68. http://doi.ieeecomputersociety.org/10.1109/TKDE.2011.69. 3 Nicolae M Rajasekaran S Efficient Sequential and Parallel Algorithms for Planted Motif Search BMC Bioinformatics 2014 15 1 34 10.1186/1471-2105-15-34 24479443 4 Tanaka S Improved Exact Enumerative Algorithms for the Planted (l,d )-motif Search Problem IEEE/ACM Trans Comput Biol Bioinformatics (TCBB) 2014 11 2 361 74 10.1109/TCBB.2014.2306842 5 Yu Q Huo H Zhang Y Guo H PairMotif: A new pattern-driven algorithm for planted (l,d ) DNA motif search PloS One 2012 7 10 48442 10.1371/journal.pone.0048442 6 Karlin S Ost F Blaisdell BE Waterman MS Patterns in DNA and Amino Acid Sequences and Their Statistical Significance Mathematical Methods for DNA Sequences 1989 Boca Raton, FL, USA CRC Press Inc 7 Rocke E Tompa M An Algorithm for Finding Novel Gapped Motifs in DNA Sequences Proceedings of the Second Annual International Conference on Computational Molecular Biology 1998 New York, NY, USA ACM 8 Sagot MF Spelling Approximate Repeated or Common Motifs using a Suffix Tree LATIN’98: Theoretical Informatics 1998 Brazil Springer 9 Lanctot JK Li M Ma B Wang S Zhang L Distinguishing string selection problems Inform Comput 2003 185 1 41 55 10.1016/S0890-5401(03)00057-9 10 Adebiyi EF Kaufmann M Guigó R Gusfield D Extracting Common Motifs under the Levenshtein Measure: Theory and Experimentation Algorithms in Bioinformatics: Second International Workshop, WABI 2002 Rome, Italy, September 17–21, 2002 Proceedings 2002 Berlin, Heidelberg Springer Berlin Heidelberg 11 Wang X Miao Y GAEM: A Hybrid Algorithm Incorporating GA with EM for Planted Edited Motif Finding Problem Curr Bioinformatics 2014 9 5 463 9 10.2174/1574893609666140901222327 12 Rajasekaran S Balla S Huang CH Thapar V Gryk M Maciejewski M Schiller M High-performance Exact Algorithms for Motif Search J Clin Monitoring Comput 2005 19 4–5 319 28 10.1007/s10877-005-0677-y 13 Pathak S Rajasekaran S Nicolae M EMS1: An Elegant Algorithm for Edit Distance Based Motif Search Int J Foundations Comput Sci 2013 24 04 473 86 10.1142/S0129054113500159 14 Knuth DE The Art of Computer Programming, Volume 4, Generating All Tuples and Permutations, Fascicle 2 2005 New Jersey, USA Addison Wesley

PMC005xxxxxx/PMC5001235.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556157290210.1186/s12864-016-2902-0ResearchIntegrative analysis of somatic mutations and transcriptomic data to functionally stratify breast cancer patients Zhang Jie Jie.Zhang@osumc.edu Abrams Zachary Zachary.Abrams@osumc.edu Parvin Jeffrey D. Jeffrey.Parvin@osumc.edu Huang Kun Kun.Huang@osumc.edu Department of Biomedical Informatics, The Ohio State University, Columbus, OH 43210 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.513© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Somatic mutations can be used as potential biomarkers for subtyping and predicting outcomes for cancer patients. However, cancer patients often carry many somatic mutations, which do not always concentrate on specific genomic loci, suggesting that the mutations may affect common pathways or gene interaction networks instead of common genes. The challenge is thus to identify the functional relationships among the mutations using multi-modal data. We developed a novel approach for integrating patient somatic mutation, transcriptome and clinical data to mine underlying functional gene groups that can be used to stratify cancer patients into groups with different clinical outcomes. Specifically, we use distance correlation metric to mine the correlations between expression profiles of mutated genes from different patients. Results With this approach, we were able to cluster patients based on the functional relationships between the affected genes using their expression profiles, and to visualize the results using multi-dimensional scaling. Interestingly, we identified a stable subgroup of breast cancer patients that are highly enriched with ER-negative and triple-negative subtypes, and the somatic mutation genes they harbor were capable of acting as potential biomarkers to predict patient survival in several different breast cancer datasets, especially in ER-negative cohorts which has lacked reliable biomarkers. Conclusions Our method provides a novel and promising approach for integrating genotyping and gene expression data in patient stratification in complex diseases. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2902-0) contains supplementary material, which is available to authorized users. Keywords Distance correlationBreast cancer patient stratificationFunctional analysis of somatic mutationIntegrative analysisThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background The initiation, development, and metastasis of cancers are complicated processes involving multi-cell, multi-tissue interactions and communications. Most cancers confer heterogeneity among patients that lead to different clinical outcomes such as survival time and response to treatment. With recent rapid advancement in next generation sequencing (NGS) technologies and computing capacity for processing and storing large data, more and more human cancer genomes have been characterized in a systematic way, bringing great opportunities for researchers to carry out integrative analysis to identify potential molecular markers for stratifying patients into subtypes with different predicted clinical outcomes [1]. Currently The Cancer Genome Atlas (TCGA) project harbors comprehensive data ranging from genomic sequences, genetic variants, transcriptomic and proteomic data to clinical data for multiple types of human cancer tissues as well as normal tissues. It is a great source for scientists to integrate data from different levels and mine the buried interaction among them, which will shed light on the understanding of cancer subtyping, prognosis as well as the cancer initiation and development [2–4]. In TCGA database, we often observe patients with a lot of somatic mutations that can significantly alter corresponding protein structures or functions of the genes they reside on (we named the affected gene as significantly mutated gene, or SMG). SMGs are the results of splice-site-change, nonsense, non-stop or frame-shift mutations. The prevalence of SMGs in almost all cancer types let us postulate that they may be potentially used as signatures for subtyping and outcome prediction, or as starting point to elucidate the tumorigenesis process. However, there is a big challenge in using SMGs for cancer patient stratification — the overlaps between the SMGs from different patients are usually small and the lists are usually not converging to common pathways [1, 5]. For instance, the breast cancer (BRCA) project in TCGA has identified three commonly mutated genes TP53, GATA3, and PI3KC but every patient has a much larger number of somatic mutations which cannot be easily summarized and compared even at the pathway level [1]. Therefore, it is of great interest in identifying the potential relationships between the mutated genes from different patients. In this paper, instead of directly working on the gene lists, we propose to examine the functional relationships of the SMGs between different patients based on functional genomics data. One of such functional measurements is gene expression profile obtained from microarray or RNA-seq experiments, which has already been curated in TCGA. Specifically, given two sets of SMGs from two patients, we develop a method to establish the relationship between them based on expression profiles of the two gene lists. Given a list of genes with their expression profiless measured in a cohort of patients, one way to characterize their roles is to examine how these genes lead to separation of the patients. In other words, we can establish a “patient network” using the difference of the expression levels of the genes as distance metric. Then given two gene lists, we can compare the similarity between the patient networks established by each of the lists. The similarity will provide pivotal information on the similarity between the roles of these two gene lists among the patients. Mathematically, such similarity between patient networks can be computed using a recently developed metric called distance correlation [6]. Therefore in this paper, we develop a workflow for establishing the functional similarity among SMGs from different patients based on distance correlation. Our goal is trying to reveal the yet unknown links between different SMG, which indicate their functional relationships in the context of human gene interaction network, and use this relationship to stratify patients with different subtypes. While we demonstrate our approach using a breast cancer study, our method provides a novel promising approach of integrating genotype and gene expression data in patient stratification in complex diseases. Methods In this paper, we obtained whole genome exome-seq data (WES) from TCGA for the patients with breast cancers and derived the SMG list for each patient. The list of SMGs from each patient were used as features for this patient. We then computed distance correlation of every pair of SMG lists to obtain the functional relationships between the affected genes in different patients based on the gene expression profiles. The process yielded the distance correlation matrix across the patients. Then we visualized the patients by multi-dimensional scaling, and further clustered the patients into different groups. Our workflow is summarized in Fig. 1.Fig. 1 Workflow of identifying functional gene relationships using variants and transcriptomic data The key component in this workflow is to compute the distance correlation between a pair of gene lists (in this case, expression profiles of two SMG lists from two patients). The intuition behind distance correlation can be considered as following: A gene list can be used to cluster the patient cohort of a heterogeneous disease, generating a clustering result. Two different gene lists will generate two results, and the results may be similar if the two gene lists play similar functional roles in the disease phenotype. The distance correlation measures the similarity of the two results. In our case, we used the gene expression data (RNA-seq) of the entire cohort to compute the distance correlation, although theoretically, any gene expression dataset of a cohort with similar disease diversity can be used, and from a more general point of view, any type of data which present deep enough functional relationship among genes, even on normal people, can be used. After we obtained the distance correlation matrix of any two SMG lists in the context of gene expression, which represents the functional relationship of any two sets of SMGs in the breast cancer disease gene expression, we use this matrix to cluster the entire breast cancer cohort, and the results should show a group of patients grouped by their common underlying perturbation resulted from seemingly different SMG lists. Datasets The Cancer Genome Atlas (TCGA http://www.cancergenome.nih.gov) level-3 breast cancer patients’ somatic mutation derived from WES and RNA-seq data were downloaded from TCGA data portal in July, 2013. Among all 876 available patients at the time of download, 445 have matching SMG and RNA-seq data. The data from these patients were chosen for further analysis. 83 normal breast sample RNA-seq level 3 data were also obtained from TCGA. SMG selection Somatic mutations derived from WES of the TCGA breast cancer patients were screened for significant mutation genes (SMG). SMG was defined as genes with frame-shift Indels, splice site change, non-stop mutation, or nonsense mutation. The mutation of mismatch, silent, RNA and in-frame indel were not included in SMG. For a specific group of patients, the number of SMG refers to the union of SMGs in that group of patients. For all the patients we analyzed in this study, their corresponding SMGs were listed in Additional file 1: Table S2. Computing distance correlation Distance correlation is a recently developed metric with two advantages [6]. First, it can be used to calculate the “correlation” between two matrices instead of just two vectors. Essentially it calculates the similarity of effects of two “feature sets” on separating the same set of samples. Secondly, unlike Pearson correlation that is based on a linear model, it can respond to nonlinear relationships. These properties make it a good candidate for our purpose when comparing relationships between two gene lists. In this project, the distance correlation was computed using Matlab as described in [6]. Given two lists of SMGs ga and gb with na and nb genes respectively, we first extract their gene expression matrices across N patients as Ea=e1a⋯eNa∈ℜna×NandEb=e1b⋯eNb∈ℜnb×N, where eji (i ∈ {a, b}, j ∈ {1, 2, …, N}) are ni- dimensional column vectors representing the expression profiles for the j-th patient over the i-th SMG list. The distance matrices among the patients for the two sets of SMGs can be calculated as Da=djka∈ℜN×NandDb=djkb∈ℜN×N with djki = ‖eji − eki‖, i ∈ {a, b}, j, k = 1, 2, …, N. Let dj,⋅i¯ and d⋅,ki¯ be the average of the j-th row and k-th column for the matrix Di (i ∈ {a, b}) respectively. Also set d⋅,⋅i¯ be the grand average of all entries of Di (i ∈ {a, b}). Then set the centralized distance matrices to be Di¯=djki¯=djki−dj,⋅i¯−d⋅,ki¯+dj⋅ki¯∈ℜN×Nwithi∈ab. Then the distance covariance between two distance matrices can be computed as dCovEa,Eb=1N2∑j,k=1Ndjka¯⋅djkb,¯ and the distance correlation is defined as dCorEa,Eb=dCovEaEbdCovEaEa⋅dCovEb,Eb. For the 445 SMG lists obtained from the 445 patients, we compute the distance correlation matrix DdCor=dCorEiEj∈ℜ455×455,i,j=1,2,…,445. Multidimensional scaling and clustering In order to visualize the distribution of the patients with the proximity measurements defined by the distance correlation matrix, we applied multidimensional scaling (MDS) to embed the data points (each point represents a patient) in 3D space. Specifically we used Matlab function cmdscale() with its default settings. The distance correlation matrix was first transformed to a dissimilarity matrix (using 1 − DdCor) before MDS. K-means clustering was performed upon the patients using data using the same dissimilarity matrix. It was carried out using Matlab k-means function with default square-Euclidean distance and replicates of 50, K = 3 or 5. Jaccard index computing SMGs for every pair of patients in TCGA BRCA cohort were used to calculate the similarity between the two SMG lists using Jaccard index (J), which is defined as: J=A∩BA∪B, where A and B are the two groups of SMGs from any pair of patients in the TCGA BRCA cohort. A∩B is the set of overlapping genes within the two SMG groups A and B, and A∪B is the union of these two groups. Survival analysis For validation, NCBI GEO breast cancer dataset GSE1456 (containing 318 patients of mixed types) [7] as well as Netherlands Kanker Instituut (NKI) NKI-295 dataset (containing 295 patients of mixed types) were used [8]. These microarray datasets (and their specific subtypes) contain gene expression data and matching survival time (years) that are needed for  survival analysis. Log-rank test was performed to determine the significance of difference in survival time between two patient groups and Kaplan-Meier curves were plotted. Pathway analysis and gene query in TCGA database Ingenuity Pathway Analysis (IPA) was used to analyze enriched biological functions and pathways in the identified SMGs. The prevalence of SMGs on other cancer types in TCGA database was generated using the cBioPortal online tools (http://www.cbioportal.org) [9]. Results We applied the above described workflow to analyze 445 breast cancer patients with matching SMG and RNA-seq data from TCGA. The distance correlation matrix was calculated and transformed. After MDS, the patients were imbedded into a 3D space for visualization, as shown in Fig. 2, with each point representing a patient.Fig. 2 K-means clustering on the embedded patients , revealing a subtype of breast cancer patients enriched with triple-negative patients. a: K = 3, Red: Group 1, Blue: Group 2, Green: Group 3. b: K = 5, Group 2 from panel A was further clustered into three groups (blue, magenta and red) When the patients were clustered using the K-means clustering algorithm, we observed a distinctive group of patients as highlighted by the red circle in Fig. 2. The number of clusters is tested by checking the silhouette values and plots for different choice of K. The silhouette value reaches its high peak at K = 5 (data not shown) but this group is stable even when the number of clusters changed (e.g., K = 3 vs. 5). In addition, we inspected the silhouette plots and found that the clusters are more separated when K = 3. Thus we use K = 3 for most the rest analysis. In order to test if the clustering of patients can be achieved using other methods or could be an artifact, we carried out three tests. First, we directly used the SMGs as features for the patients and the similarity among the patients were established by calculating the Jaccard indicies between every pair of patients. Out of all the 98,790 patient pairs, 96.2 % are zeros, which means they do not share any common genes. Thus using SMGs cannot effectively separate the patients. Secondly, we tested if using non-cancer gene expression data can lead to the same observation. As shown in Fig. 3a, there is no clear separation among the patients and the clusters obtained from K-means algorithm do not have any enrichment of specific subtypes of breast cancers when we used 83 normal breast tissue samples RNA-seq data instead of breast cancer data. Finally, we tested randomly selected “pseudo-SMGs” for the patients. Basically for each patient, we randomly select the same number of genes as her SMGs, applying the same workflow and the result is shown in Fig. 3b. Similar results are observed as in Fig. 3a.Fig. 3 The results of the distance correlation workflow on control data. a: Applying the workflow using normal breast gene expression data. The three groups from K-means clustering do not enrich specific subtypes of breast cancers. b: Applying the workflow on randomly selected “pseudo-SMGs”. No subtype enriched patient cluster can be observed In order to gain insight on this distinctive group of patients, we examined the status of the known molecular markers for breast cancers, namely estrogen receptor (ER), progesterone receptor (PR), and HER2. Statistical analysis revealed that this group is significantly enriched with ER-negative, PR-negative, HER2-negative, or triple negative breast cancer (ER-, PR-, HER2- or TNBC) patients. Specifically, while it contains 41 patients consisting only 9.2 % of the total cohort, it includes 34 % of the total TNBC patients (Table 1). To examine if this group can be differentiated easily from the cohort using other genes, we repeated the process using randomly selected “pseudo SMGs” of the same sizes for every patient. The clustering result was not able to separate the patients into groups with such enrichment of the ER- or TNBC patients. Since both ER- and TNBC patients are known to have worse prognosis then the ER+ patients, our further analysis was focused on this specific group, and we refer it as the “Group 1” in the rest of the paper.Table 1 Statistical tests on the patient subtypes enriched in each group from K = 3 clustering results. No statistic test was performed for HER2 (and TN) status, due to the fact that more than 25 % patients do not contain HER2 status Total ER+ ER- χ2 adj-P value PR+ PR- χ2 adj-P value HER2+ HER2- Triple - Group1 41 14 26 0.00075 13 26 0.00658 8 23 16 Group2 304 233 68 0.700 205 95 0.488 55 158 28 Group3 100 85 12 0.0619 68 30 0.6476 15 53 5 Total (with sig mutation and matching RNAseq) 445 332 106 286 151 78 235 49 Total in TCGA 876 634 187 548 267 136 447 90 Group 1 contains 201 SMGs that are specifically present in Group 1 patients (Fig. 4, Additional file 1: Table S1). Enrichment and pathway analysis using IPA showed that these SMGs are highly enriched with cancer-related genes, genes for embryonic development, cell morphology and organ development, indicating this group of genes are more involved in the early stage of cancer cell development and differentiation process (Fig. 5). Several upstream regulator drugs are found to regulate multiple genes in this group, among them, Ethinyl estradiol, an orally bioactive estrogen, regulates ABCB11, CCR7, CD97, CYP2D6, CYP7B1, SGK1, suggesting although being ER-negative, estrogen may still play a role in this group of patients; the drug, which is used to treat myelodsyplastic syndromes and acute myeloid leukemia, regulates BMP4, CCR7, MAGEC1, METAP2, MGMT, RARB, RARRES1, SGK1, SNRPN, and TGFBR2 [10, 11]. This may be a direction for future therapeutic research on this specific subtype of triple-negative breast cancer. Interestingly, the narcotic substance amphetamine regulates BMP4, DCC, SGK1, and TGFBR2.Fig. 4 Venn diagram showing the genes shared/unique among the three groups from K = 3 clustering results Fig. 5 Pathway analysis showing the top 10 biological functions enriched in the genes specifically to Group 1 isolated from K = 3 clustering In addition, analysis using cBioPortal shows that the group of 201 SMG genes is found frequently altered (mutated, or contain copy number variance) in almost all types of cancers available in TCGA database (Fig. 6).Fig. 6 Group 1 specific genes are altered in multiple cancer types (TCGA data). AML: acute myeloid leukemia; ACC: adenoid cystic carcinoma; BC: bladder cancer; BUC: bladder urothelial carcinoma; BLGG: brain lower grade glioma; BIC: breast invasive carcinoma; CSCC &EAC: cervical squamous cell carcinoma & endocervical adenocarcinoma; GBM: glioblastoma multiforme; HNSCC: head & neck squamous cell carcinoma; KRCCC: kidney renal clear cell carcinoma; KRPCC: kidney renal papillary cell carcinoma; LAC: lung adenocarcinoma; LSCC: lung squamous cell carcinoma; OSCC: ovarian serous cystadenocarcinoma; Prostate AC: prostate adenocarcinoma; SCM: skin cutaneous melanoma; SAC: stomach adenocarcinoma; TC: thyroid carcinoma; UCEC: uterine corpus endometrial carcinoma; LHC: liver hepatic carcinoma; Pancreatic AC: pancreatic adenocarcinoma We further tested if this unique group of 201 SMGs (Additional file 1: Table S1) or its subsets is associated with patient outcome (survival time to be specific in this paper) using multiple publicly available breast cancer gene expression data. The results are shown in Fig. 7. The subsets were selected based on the IPA pathway annotation. Our test on NKI data suggested that the 201 SMGs are able to separate patients (based on K-means algorithm with K = 2) into two groups with significant survival time difference but cannot effectively separate the ER-negative patients. The 201 SMGs can be clustered into several functional/pathway groups based upon gene enrichment analysis using Ingenuity Pathway Analysis (IPA®). Among these groups, we found that the group of 27 genes with embryonic development functions performed the best, which can separate the ER-negative breast cancer patients into two groups with significantly different survival times (Fig. 6 Middle). In addition, this 27-gene set can also separate patients in the other dataset (GSE1456) as shown in Fig. 7 Right. Given the high enrichment of ER-negative patients in the Group 1, these results suggest that the 27 genes may form the core of the Group 1 SMGs. As a comparison, the SMGs unique to Group 2 were not able to separate the ER-negative patients with significantly different survival outcomes, and it does not performs as good as Group 1 SMGs on general population survival test (data not shown).Fig. 7 Survival analysis using the Group 1 specific genes and its subset on separate breast cancer microarray data. Left: one NKI all cohort; Middle: on NKI ER-negative cohort; Right: on GSE1456 all cohort Discussion With recently rapid development in next-generation sequencing technology and computing capacity, huge amount of data in different modalities for cancer specimens have been accumulated in an amazing speed in public databases. Therefore, integrating and mining these data becomes a major challenge in the bioinformatics field currently. In this work, we developed a novel approach to integrate genomic, transcriptomic and clinical data of cancer patients, specifically to compare somatic mutations of patients based on their functional relationships in the context of gene expression profiles, thus tackling the challenge of low overlapping of mutated genes among cancer patients. By introducing the distance correlation metric to directly measure the relationship between two sets of genes affected by somatic mutations, we not only can cluster the patients into different groups with different clinical subtypes, but also visualize the clusters and identify group specific mutations. The power of using distance correlation freed us from comparing only gene pairs, but directly comparing gene list to list. The distance correlation captures not only linear relationship of the two lists as Pearson correlation does, but also reveals non-linear relationship as well, which covers the biological interaction in far more and deeper extent. Applying this approach on TCGA breast cancer patients reveals a group of patients who are mostly negative with one or more of the three breast cancer biomarkers (ER, PR, HER2) [12], and one third of the group are triple-negative subtype. Triple-negative breast cancer (TNBC) composes of 12–20 % of breast cancer patients [13]. It progresses more aggressively and does not respond well to hormone therapy. The rapid and aggressive progress of the disease course makes the prognosis of TNBC very poor [14] and the prediction difficult. After examining the group of patients we identified here, they harbor SMGs tightly interlinked each other and enriched with early stage cancer development. Among them, the 27 embryonic development genes form tight interaction networks as shown in the Fig. 8, and those genes can be used for breast cancer survival prognosis, especially for the poorly understood ER-negative cohort. TCGA database has not been curated long enough for this subtype of patients, therefore we did not test our findings on TCGA data. Instead, we chose two older GEO breast cancer microarray datasets. Unfortunately, the GEO datasets we tested does not contain enough TNBC patients, so we only tested on ER-negative cohort. The clustering results indicated that a portion of the triple negative patients maybe fundamentally different from the rest of the breast cancer patients due to the somatic mutations they harbor. Many of their genes shared common upstream regulators such as the drug for acute myeloid leukemia or estrogen, suggesting this group of people may benefit from other type of treatments that have not been administrated to TNBC patients. We suggested that the common upstream regulators and drugs interacting with these genes can provide insight on the development and treatment of TNBC patients. In addition, while among the 27 genes some of them are known to be associated with other cancers such as AFF1 [15], BMP4 [16], and TRIM24 [17], others such as MED27 is not widely know to be associated with cancers. Thus our work also generated new hypothesis on cancer related genes.Fig. 8 Group 1 genes enriched with embryonic development, organ development and morphology function (IPA) Conclusions In summary, a common challenge in studying complex diseases such as cancers is the lack of common genetic mutations among the patients. Besides pursuing commonly affected pathways, we provide a complementary approach for integrating the genotype data with transcriptome data to study the relationships between the genetic mutations at the functional level. While our main goal is on exploring the functional relationships of mutated gene groups, the identified genes may also serve as potential biomarkers for different subtypes of cancers. Currently due to the limitation of the data, we focus on the protein coding genes from the WES experiments. In the near future, we plan to apply the same workflow to other cancer datasets in TCGA to further test the effectiveness of this method as well as identifying diseases in which such functional relationship can lead to meaningful stratification of the patients. With the cost of whole genome sequencing decreasing dramatically, it is expected that more somatic mutations on the non-coding regions and regulatory regions can be made available and the approach need to be expanded to accommodate such mutations. Abbreviations BRCA, breast cancer; ER, estrogen receptor; GEO, gene expression omnibus; HER2, human epidermal growth factor receptor 2; IPA, ingenuity pathway analysis; MDS, multi-dimensional scaling; NCBI, National Center for Biotechnology Information; NGS, next generation sequencing; PR, progesterone receptor; RNA-seq, ribonucleic acid sequencing; SMG, significant mutant gene; TCGA, the cancer genome atlas; TNBC, triple-negative breast cancer; WES, whole genome exome sequencing Additional file Additional file 1: Supplementary tables. This file contain two tables, the first table contain the SMGs in group 1 patients and their mutation frequencies among group 1 patients. The second table contain the patient IDs and their corresponding SMGs from TCGA BRCA. (DOCX 117 kb) Acknowledgement We thank Ohio Supercomputer Center for computing support. This work was partially funded by NCI U01 CA188547 grant. ZA was funded by NLM fellowship. Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials All datasets used in this study were publicly available from the website described in the Methods section. Authors’ contributions KH conceived of the study. JZ, ZA and JP collected the data. JZ performed the computational coding and conducted data analysis. JZ and KH drafted the manuscript, JP participated the design of the method. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Cancer Genome Atlas Network. Comprehensive molecular portraits of human breast tumours. Nature. 2012, 490(7418): 61–70. 2. Ritchie MD Holzinger ER Li R Pendergrass SA Kim D Methods of integrating data to uncover genotype–phenotype interactions Nat Rev Genet 2015 16 2 85 97 10.1038/nrg3868 25582081 3. Kristensen VN Lingjærde OC Russnes HG Vollan HKM Frigessi A Børresen-Dale A-L Principles and methods of integrative genomic analyses in cancer Nat Rev Cancer 2014 14 5 299 313 10.1038/nrc3721 24759209 4. Wang C Machiraju R Huang K Breast cancer patient stratification using a molecular regularized consensus clustering method Methods 2014 67 3 304 12 10.1016/j.ymeth.2014.03.005 24657666 5. Jia P Zheng S Long J Zheng W Zhao Z dmGWAS: dense module searching for genome-wide association studies in protein-protein interaction networks Bioinformatics 2011 27 1 95 102 10.1093/bioinformatics/btq615 21045073 6. Székely GJ Rizzo ML Bakirov NK Measuring and testing dependence by correlation of distances Ann Stat 2007 35 6 2769 2794 10.1214/009053607000000505 7. Pawitan Y Bjöhle J Amler L Borg A-L Egyhazi S Hall P Han X Holmberg L Huang F Klaar S Liu ET Miller L Nordgren H Ploner A Sandelin K Shaw PM Smeds J Skoog L Wedrén S Bergh J Gene expression profiling spares early breast cancer patients from adjuvant therapy: derived and validated in two population-based cohorts Breast Cancer Res 2005 7 6 R953 64 10.1186/bcr1325 16280042 8. van ’t Veer LJ Dai H van de Vijver MJ He YD Hart AAM Mao M Peterse HL van der Kooy K Marton MJ Witteveen AT Schreiber GJ Kerkhoven RM Roberts C Linsley PS Bernards R Friend SH Gene expression profiling predicts clinical outcome of breast cancer Nature 2002 415 6871 530 6 10.1038/415530a 11823860 9. Gao J Aksoy BA Dogrusoz U Dresdner G Gross B Sumer SO Sun Y Jacobsen A Sinha R Larsson E Cerami E Sander C Schultz N Integrative analysis of complex cancer genomics and clinical profiles using the cBioPortal Sci Signal 2013 6 269 11 10.1126/scisignal.2004088 10. Putnik M Zhao C Gustafsson J-Å Dahlman-Wright K Global identification of genes regulated by estrogen signaling and demethylation in MCF-7 breast cancer cells Biochem Biophys Res Commun 2012 426 1 26 32 10.1016/j.bbrc.2012.08.007 22902638 11. Banerjee S Bacanamwo M DNA methyltransferase inhibition induces mouse embryonic stem cell differentiation into endothelial cells Exp Cell Res 2010 316 2 172 180 10.1016/j.yexcr.2009.08.011 19715692 12. Brenton JD Carey LA Ahmed AA Caldas C Molecular classification and molecular forecasting of breast cancer: ready for clinical application? J Clin Oncol 2005 23 29 7350 60 10.1200/JCO.2005.03.3845 16145060 13. Anders CK Carey LA Biology, metastatic patterns, and treatment of patients with triple-negative breast cancer Clin Breast Cancer 2009 Suppl 2 S73 81 10.3816/CBC.2009.s.008 19596646 14. Wahba HA El-Hadaad HA Current approaches in treatment of triple-negative breast cancer Cancer Biol Med 2015 12 2 106 16 26175926 15. Srinivasan RS Nesbit JB Marrero L Erfurth F LaRussa VF Hemenway CS The synthetic peptide PFWT disrupts AF4-AF9 protein complexes and induces apoptosis in t(4;11) leukemia cells Leukemia 2004 18 8 1364 72 10.1038/sj.leu.2403415 15269783 16. Montesano R Sarközi R Schramek H Bone morphogenetic protein-4 strongly potentiates growth factor-induced proliferation of mammary epithelial cells Biochem Biophys Res Commun 2008 374 1 164 8 10.1016/j.bbrc.2008.07.007 18625198 17. Ignat M Teletin M Tisserand J Khetchoumian K Dennefeld C Chambon P Losson R Mark M Arterial calcifications and increased expression of vitamin D receptor targets in mice lacking TIF1alpha Proc Natl Acad Sci U S A 2008 105 7 2598 603 10.1073/pnas.0712030105 18287084

PMC005xxxxxx/PMC5001236.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556417289410.1186/s12864-016-2894-9ResearchImproving alignment accuracy on homopolymer regions for semiconductor-based sequencing technologies Feng Weixing fengweixing@hrbeu.edu.cn 1Zhao Sen zhaosen@hrbeu.edu.cn 1Xue Dingkai xuedingkai@hrbeu.edu.cn 1Song Fengfei songfengfei@hrbeu.edu.cn 1Li Ziwei liziwei@hrbeu.edu.cn 1Chen Duojiao chenduojiao@hrbeu.edu.cn 1He Bo bohe@hrbeu.edu.cn 1Hao Yangyang haoyan@umail.iu.edu 2Wang Yadong ydwang@hit.edu.cn 3Liu Yunlong yunliu@iupui.edu 121 Automation College, Harbin Engineering University, Harbin, Heilongjiang 150001, People’s Republic of China 2 Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 3 School of Computer Science and Technology, Harbin Institute of Technology, Harbin, Heilongjiang 150001, People’s Republic of China 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.521© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Ion Torrent and Ion Proton are semiconductor-based sequencing technologies that feature rapid sequencing speed and low upfront and operating costs, thanks to the avoidance of modified nucleotides and optical measurements. Despite of these advantages, however, Ion semiconductor sequencing technologies suffer much reduced sequencing accuracy at the genomic loci with homopolymer repeats of the same nucleotide. Such limitation significantly reduces its efficiency for the biological applications aiming at accurately identifying various genetic variants. Results In this study, we propose a Bayesian inference-based method that takes the advantage of the signal distributions of the electrical voltages that are measured for all the homopolymers of a fixed length. By cross-referencing the length of homopolymers in the reference genome and the voltage signal distribution derived from the experiment, the proposed integrated model significantly improves the alignment accuracy around the homopolymer regions. Conclusions Besides improving alignment accuracy on homopolymer regions for semiconductor-based sequencing technologies with the proposed model, similar strategies can also be used on other high-throughput sequencing technologies that share similar limitations. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2894-9) contains supplementary material, which is available to authorized users. Keywords HomopolymerIon Torrent/ProtonBayesianAlignmentThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background The rapid development of high-throughput sequencing technologies leads to appearances of many innovative sequencing platforms [1, 2]. Ion Torrent and Ion Proton are semiconductor-based sequencing platforms that are primarily designed for personal genome sequencing [3, 4]. Different from sequencing techniques enriched with substitution errors [5, 6], Ion semiconductor sequencing platforms suffer from the inaccuracy in detecting the length of homopolymers repeats of the same nucleotide [7, 8]. These homopolymer errors often lead to the inaccurate local alignment results, and become a critical barrier against accurate detection of genomic variations [9–11] (http://www.broadinstitute.org/gatk/media/docs/Samtools.pdf). The sequencing chemistry for the Ion semiconductor-based technology is that the incorporation of a deoxyribonucleotide (dNTP) into a strand of DNA couples with the release of a hydrogen ion, which changes the pH of the solution and then leads to the electronic voltage pulse in the ion sensor. Multiple identical bases on the DNA strand often result in the detection of multiple times of the baseline voltage corresponding to the measurements at mononucleotide loci [12]. The difficulty on the homopolymer length identification mainly results from the inaccurate measurement on the magnitude of the voltage pulse, which follows a signal distribution that can be dependent on multiple factors including the type of nucleotide, the length of homopolymer, and the relative position in the DNA template. Thus far, several algorithms have been proposed in correcting the inaccurate homopolymer length identification, based on the raw data from the detected voltage signals for Ion semiconductor sequencing technologies. Lysholm designed a flow-space FAAST tool, where flowpeak information retrieved from detected voltage signals is utilized to improve accuracy of Smith-Waterman-Gotoh local alignment through correction of likely sequencing errors and thus obtain optimized homopolymer length [8]. However, since dedicatedly designed for the naïve Smith-Waterman-Gotoh algorithm, the method undertakes a heavy computing burden and limits the further application with other alignment programs. In addition, the parameter selection in the algorithm is ad hoc, and was not designed for maximizing the performance. Zeng designed a PyroHMMsnp algorithm, where a hidden Markov model (HMM) is built to recognize overcall or undercall status of homopolymers in a realignment process, and is used to deduce the most possible homopolymer lengths [7]. Similar with other refined alignment algorithms, this approach uses an EM-based strategy, which assumes the variant pattern on most of the reads at one specific loci follows the same distribution; this assumption maybe invalid for certain biological applications, such as the variant identification in cancer somatic tissues. In addition, PyroHMMsnp design does not have hidden state for mismatches, and therefore tends to mistakenly convert mismatches into INDELs. In this project, we aim to develop a simple computational strategy for improving the alignment accuracy by using the voltage signals, and relying only on the measurements of individual sequencing read. In addition to the measured electrical voltage signals, it is evident that the reference genome contains significant amount of prior information that are not adequately considered by other methods. This is under the assumption that only a small percentage of nucleotides are different between two individuals; for human, it is about 1 % of whole genome nucleotides [13, 14]. Based on such consideration, we proposed a Bayesian-based integrated model to merge these two information sources to improve performance of homopolyer length identification. We demonstrate that our algorithm significantly outperformed Torrent Suite, the software package coupled with Ion Torrent and Proton Sequencers for accurately identifying the length of the homopolymer repeats, and therefore improved sequence alignment accuracy. Methods Ion Torrent sequencing Different from imaging-based sequencing platforms, Ion semiconductor technology detects nucleotide composition using electronic sensors. During the sequencing process, the sensor detects released hydrogen ions when nucleotide incorporation occurs. The sensor then detects the pH change caused by hydrogen release, and translates such chemical signal to electrical voltage signal, which is proportional to the number of captured ions. Since one type of nucleotides is sequenced in one machine cycle, if homopolymer exists, the detected voltage level should reflect the length of homopolymer. Despite this simple principle, practically, however, the detected electrical voltage follows a distribution, and in many cases, may not accurately recapitulate the length of homopolymer. In order to design a bioinformatics strategy for correcting the length of homopolymers, we first systematically evaluate the signal distribution of the detected electrical voltage for all the nucleotide positions that share the same homopolymer length, same homopolymer nucleotide type (A, C, G, or T), and similar positions in the sequence reads. The original voltage signals for different nucleotides were extracted from the SFF file, which is exported from the Torrent Suite package. Bayesian inference of homopolymer length We design a Bayesian-based model to infer the length of homopolymer based on the local genomic sequence context, including the homopolymer nucleotide type (Ni = A, C, G, or T), detected electrical voltage (V), and the nucleotide position in the sequencing reads (Pj). In the current model, nucleotide position were classified into several categories. 1 PL|Ni,Pj,V=PV|Ni,Pj,L∗PL/PV=PV|Ni,Pj,L∗PL/∑i,jPV|Ni,Pj,L*PNiPjL In the equation, P(V|Ni,Pj,L) is the prior possibility of occurrence of a specific voltage V if given homopolymer length L under situation of nucleotide Ni and read position Pj, while P(L) and P(V) respectively represent the probability of a specific homopolymer length L, and the probability of a specific voltage V. Both these two probabilities can be statistically derived from the entire sequencing data. In summary, P(L|Ni,Pj,V) is the probability of occurrence of a specific homopolymer length L if given sequencing voltage V under sequencing context of nucleotide Ni and read position Pj. Integrated model to identify homopolymer length The performance of statistical-based inference model highly relies on fully understanding the sources of detection error, and their intervened relationships. Additional biological information can be used to increase the detection accuracy. For most of the biological applications, it is reasonable to assume that only a small percentage of the nucleotide positions represent true variants comparing to the reference genome. Therefore, combining the homopolymer length in the reference genome with the statistically-inferred homopolymer length can potentially improve the detection accuracy. We therefore construct an integrated model by defining a score S for the homopolymer length at a specific homopolymer loci: 2 S=W*logPL|Ni,Pj,V+1−W∗PenL|Seq_ref In the model, Pen(L|Seq_ref) is a penalty value when mismatch occurs between the reference genome sequence and the deduced Ion Torrent sequence for a given homopolymer length L. The penalty value is defined as 0 for perfect match, −1 for substitution, and −2 for insertions/deletions. In order to ensure that the two types of measurements staying in a similar scale, Bayesian posterior probability P(L|Ni,Pj,V) is converted into logarithmic form. In Eq. 2, W is weighting factor to balance the contribution of the Bayesian model-derived score, and reference genome-derived penalty. For one homopolymer, its length L can be determined as the candidate with the largest score Si: 3 L=argmaxiSi,i=2,3,4,5,6,… For a specific assay, the weighting factor w will be determined by minimizing the identification error for the homopolymers whose length is known, such as samples also detected using other technologies. Results and discussion Data preparing We have tested our model on one HapMap human dataset, NA11881, of which both Ion Torrent data and Illumina sequencing data is available. The availability of such dataset enables training and testing a statistical model for refining the identification of homopolymer length. The Ion Torrent dataset was generated in the Center for Medical Genomics at Indiana University, of which a targeted genomic region of 59 genes were sequenced. The overall targeted genomic area covers 90,918 basepairs. To derive the length of the homopolymer repeats, the electrical voltage signal for each influx nucleotide machine cycle was retrieved from the SFF file, where the type of the nucleotide (A, C, G, or T) is determined. Among 452,161 Ion Torrent sequencing reads that passed quality control, our assay detected 1,430,986 homopolymers with >1.5 voltage units; these regions are defined as homopolymer candidates that are used in further analysis. In order to further characterize the homopolymer profiles being identified in our dataset, we further examine the nucleotide composition of all the detected homopolymers, and their relative loci in the sequenced reads (Fig. 1). We observed enrichment of A and T homopolymers in our dataset, and evenly distributed homopolymer locations (except for the last location due to the varying lengths of the Ion Torrent reads).Fig. 1 Profile of retrieved homopolymers. Profile of retrieved homopolymers according to (a) nucleotide type and (b) position in the sequencing reads Illumina sequencing data from the same individual, NA11881, is downloaded from the 1000 Genomes database (http://www.1000genomes.org/data). Due to the chemistry differences, Illumina technology is more accurate in detecting homopolymer lengths. We therefore use the dataset from the Illumina platform as gold standard when refining the length of homopolymer repeats. Distribution of detected voltage signals in homopolymer repeat regions We examine the three factors that affect the distribution of the voltage signals on the homopolymer regions, the length of the homopolymer repeats, the types of homopolymer nucleotides, and the relative positions of the homopolymer repeats within a read. The homopolymer positions were classified into four zones depending on their distance from the beginning of the reads, Z1: 1–75 bp, Z2: 76–150 bp, Z3: 151–225 bp, and Z4: 226–300 bp. For the homopolymers with A nucleotides and appear in the first 75 bases, the retrieved signal distributions for each homopolymer length was demonstrated in Fig. 2. The ground truth for the homopolymer is derived from the Illumina dataset, which do not have apparent homopolymer issues. In Fig. 2, the horizontal axis is of voltage level and the vertical axis is of probability density for all the homopolymers of a fixed size. From left to right, there are five curves which correspond to the homopolymers with 2, 3, 4, 5, and 6 nucleotides. Here, the probability density of the voltages are fitted as in Gaussian distributions, where the mean values are 1.85, 2.78, 3.68, 4.64 and 5.57 respectively. It is observed that the standard deviation increases with homopolymer length. It increases from 0.14 for 2-base homopolymers to 0.38 for 6-base ones. A similar trend has been reported elsewhere [7]. This shift clearly suggests that the voltage signals become less specific with the homopolymer length increases. It is critical to consider these factors in the model for accurately inferring the homopolymer length. This is especially important for the sequencing reads with longer homopolymers.Fig. 2 Prior possibilities of the detected voltages. Prior possibilities of the detected voltages when nucleotide type is A and position in the sequencing reads belongs to Z1 Besides homopolymer length, we also observed differences in signal distribution for homopolymers with different nucleotide composition (A, C, G, or T) and their positions in the sequencing reads. As shown in Fig. 3a, when fixing the homopolymer length (n = 4) and homopolymer position zone (Z1, position 1–75 in the sequencing reads), we observed slightly different signal distribution for the homopolymers with different nucleotide compositions. Specifically, the C homopolymers tend to have higher signal values with mean signal intensity at 3.74, as comparing to other three nucleotides with average value at 3.68. In addition, the standard deviation for the C homopolymers (stdev = 0.30) are also slightly larger than the other three types (stdev = 0.24). Similar inconsistency was also observed for homopolymers that locate at different positions zones in the sequencing reads (Fig. 3b). Using all the AAAA as an example, the average signals tend to be higher in the beginning of the reads, and decrease toward the end of the reads. The average signal for Z1 to Z4 is 3.68, 3.57, 3.54, and 3.57 respectively. All these results suggest that the derived voltage signal is dependent on the homopolymer nucleotide composition and its relative positions in the sequencing reads, and should be considered while inferring the length of the homopolymers.Fig. 3 Other factors in Identification of homopolymer length. Other factors in identification of homopolymer length as (a) nucleotide type when homopolymer length is 4 and position in the sequencing reads belongs to Z1 and (b) position in the sequencing reads when homopolymer length is 4 and nucleotide type is A Bayesian inference of homopolymer length Motivated by these observations, we develop a Bayesian-based model in inferring the length of homopolymer based on the homopolymer length, their relative positions in the reads, and the detected voltage signal. Since the nucleotide composition includes A, C, G, T and the homopolymer positions are classified into four zones (Z1, Z2, Z3, Z4), in total, 16 Bayesian inference models are built based on the aforementioned prior signal distributions. In each model, the homopolymer length is identified if given a specific voltage level under a particular nucleotide type and position in sequencing read. In fact, after calculation of prior signal distributions of different kinds of homopolymer lengths, the length of homopolymer can be simply decided using naïve counting from the measured electrical voltage or the k-nearest neighbors algorithm. That is to identify the length of one homopolymer according to its nearest distance to the mean values of different prior signal distributions. In such way, the number of identification errors is 169,212, or 11.82 % of the whole 1, 430,986 homopolymers. Comparing to k-nearest neighbors algorithm, with our designed Bayesian inference models, the number of identification errors decrease to 71, 460, or 4.99 % of the whole homopolymers. However, our Bayesian inference result cannot outperform that from the Torrent Suite, where the number of identification errors is 29,623, or 2.07 % of the whole homopolymers. This is due to the fact that significant training has been included the Torrent Suite algorithm, which is proprietary, and uses a large amount of genomic features. Identification of homopolymer length with Bayesian and reference genome information Despite the superior performance of the Bayesian model comparing to naïve counting from the measured electrical voltage, both our model and output from the Torrent Suite, experience significant inconsistency based on our dataset with ground truth. Since genetic variants should only occur in a small percentage of genomic loci. We therefore hypothesize that using a combination of voltage signal with the guidance of the standard reference genome will significantly increase the detection accuracy. Using our proposed integrated model with Bayesian and reference genome information, we try to identify homopolymer length. In the integrated model, Eq 2, the weight parameter W was firstly optimized when the best identification result acquired (five cross validation) comparing to the results from the Illumina sequencing results. In Fig. 4a, the process of weight optimization is presented for the model under situation of nucleotide A and position Z1. When the weight is equal to 0, only reference genome information is referred in identification, while the weight equaling to 1 means only Bayesian inference information is used. Finally, the best weight value is equal to 0.28 when the least identification errors were found. The distribution of these errors is presented in Fig. 4b. Since the exact lengths of homopolymers were measured through Illumina platform, among 144,230 homopolymers under situation of nucleotide A and position Z1, lengths of 143,870 homopolymers were successfully identified by our proposed method with 360 errors. This is significant improvement comparing to using Bayesian model only. The performance also improved comparing to relying only on the reference genome, which enables to identify homopolymer-related variants from the sequencing data.Fig. 4 Identification result of homopolymer lengths. Identification result of homopolymer lengths when nucleotide type is A and position in the sequencing read belongs to Z1. The result is presented as (a) frequency of identification errors and (b) distribution of identification result All the optimized weights and corresponding identification errors are listed in Table 1. In Table 1, comparing with other methods, the best identification result is obtained with our proposed approach, which is also presented in Fig. 5.Table 1 Identification errors of homopolymer length with different methods No Nt Pos Count Errors (%) KNN Torrent suite Bayesian Reference Proposed approach Weight Errors 1 A 1–75 144230 7.002 1.119 2.296 0.298 0.28 0.250 2 A 76–150 112776 12.121 1.651 4.722 0.489 0.34 0.453 3 A 151–225 97568 18.733 2.926 8.150 0.423 0.14 0.421 4 A 226–300 48033 22.292 4.655 10.259 0.535 0.24 0.510 5 C 1–75 88732 6.534 1.843 2.779 0.034 0.14 0.027 6 C 76–150 77650 10.382 2.489 4.595 0.556 0.36 0.121 7 C 151–225 63658 18.581 3.187 6.383 0.545 0.28 0.542 8 C 226–300 35736 17.910 4.600 6.159 0.926 0.30 0.923 9 G 1–75 97493 4.141 1.422 1.826 0.609 0.30 0.376 10 G 76–150 78192 14.874 1.623 3.864 0.322 0.32 0.152 11 G 151–225 64680 16.868 2.273 5.683 1.062 0.14 1.062 12 G 226–300 34116 18.754 2.492 7.985 0.147 0.12 0.147 13 T 1–75 156550 5.186 1.106 2.504 0.076 0.14 0.054 14 T 76–150 152034 11.446 1.571 5.780 0.342 0.30 0.297 15 T 151–225 111090 14.720 2.331 7.290 0.419 0.32 0.362 16 T 226–300 68448 13.912 3.315 8.240 0.723 0.28 0.599 “Count” means the number of each class of homopolymers. “KNN” means the method of K nearest neighbors. “Reference” means only reference information is used in the designed model(Weight = 0) Fig. 5 Comparison of identification results among different identification methods. Comparison of identification results among different identification methods according to (a) all methods and (b) two methods of only using reference information and the proposed method To show robustness of our proposed method, we also conducted analysis on one Ion Proton data(HapMap human dataset, NA12878) with the same pipeline and obtained the similar result (Additional file 1). Since more homopolymers retrieved in the Ion Proton data, their positions were classified into five zones depending on their distance from the beginning of the reads, Z1: 1–50 bp, Z2: 51–100 bp, Z3: 101–150 bp, Z4: 151–200 bp and Z5: 201–250 bp. Conclusions As an important category of sequencing platform, Ion semiconductor-based technology has been widely utilized due to its good performance of faster and cheaper sequencing. However, the technology is far from perfect and suffers from the problem of homopolymer uncertain length. With Bayesian inference and reference genome information, an integrated model was designed to resolve such a problem. Bayesian inference of homopolymer length was first calculated from detected voltage signals. Merged with reference genome sequences information, the homopolymer length was eventually deduced. Compared to several known algorithms, the proposed method presents a greatly improved performance. It should be noted that the proposed method is designed for refining the sequencing alignment based on individual sequencing read information. This is different from other approaches that rely on the coordinated information from all the reads that align to the same genomic region. Our strategy enables mapping the reads that contain variants in only a small percentage of DNA fragments, such as cancer genome. The general framework of our method can also be used for other sequencing technologies that contain significant amount of sequencing error around homopolymer regions, such as nanopore technology. Additional file Additional file 1: Supplementary results. This file contains all supplementary results that are not covered in the manuscript, including 5 figures and 1 table on Ion Proton data. Figure S1. is about profile of retrieved homopolymers according to (a) nucleotide type and (b) position in the sequencing reads. Figure S2. is about prior possibilities of the detected voltages when nucleotide type is A and position in the sequencing reads belongs to Z1. Figure S3. is about other factors in identification of homopolymer length as (a) nucleotide type when homopolymer length is 4 and position in the sequencing reads belongs to Z1 and (b) position in the sequencing reads when homopolymer length is 4 and nucleotide type is A. Figure S4. is about identification result of homopolymer lengths when nucleotide type is A and position in the sequencing read belongs to Z1. The result is presented as (a) frequency of identification errors and (b) distribution of identification result. Figure S5. is about comparison of identification results among different identification methods according to (a) all methods and (b) two methods of only using reference information and the proposed method. Table S1. is about identification errors of homopolymer length with different methods. (PDF 396 kb) Acknowledgement This work was supported in part by grants from National Natural Science Foundation of China (61471139, 61403092), the National High-Tech Research and Development Program (863) of China 2015AA020101 and Fundamental Research Funds for the Central Universities (HEUCFT1302, HEUCFX41303). Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The sequencing data used here can be found in 1000 Genomes database. Authors’ contributions BH and YL designed the study. WF and YW designed the model. SZ, DX, FS, ZL, DC performed the computational coding and implementation. YH conducted data analysis. WF and YL drafted the manuscript. All the authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Schuster SC Next-generation DNA, sequencing Nat Methods 2008 5 1 11 21 10.1038/nmeth1154 18175410 2. Ansorge WJ Next-generation DNA, sequencing techniques N Biotechnol 2009 25 4 195 203 10.1016/j.nbt.2008.12.009 19429539 3. Merriman B Ion Torrent R&D Team Rothberg JM Progress in Ion Torrent semiconductor chip based sequencing Electrophoresis 2012 33 3397 3417 10.1002/elps.201200424 23208921 4. Rothberg JM Leamon JH The development and impact of 454 sequencing Nat Biotechnol 2008 26 10 1117 1124 10.1038/nbt1485 18846085 5. Dohm JC Substantial biases in ultra-short read data sets from high-throughput DNA sequencing Nucleic Acids Res 2008 36 16 e105 10.1093/nar/gkn425 18660515 6. Smith AD Xuan Z Zhang M Using quality scores and longer reads improves accuracy of Solexa read mapping BMC Bioinf 2008 9 128 10.1186/1471-2105-9-128 7. Zeng F Jiang R Chen T PyroHMMsnp: an SNP caller for Ion Torrent and 454 sequencing data Nucleic Acids Res 2013 41 13 e136 10.1093/nar/gkt372 23700313 8. Lysholm F Andersson B Persson B FAAST: Flow-space Assisted Alignment Search Tool BMC Bioinformatics 2011 12 293 10.1186/1471-2105-12-293 21771335 9. Li H Ruan J Durbin R Mapping short DNA sequencing reads and calling variants using mapping quality scores Gen Res 2008 18 11 1851 1858 10.1101/gr.078212.108 10. Whiteford N Haslam N Weber G Bradley M Neylon C An analysis of the feasibility of short read sequencing Nucleic Acids Res 2005 33 19 171 10.1093/nar/gni170 15647500 11. Ratan A Comparison of sequencing platforms for single nucleotide variant calls in a human sample PLoS One 2013 8 2 e55089 10.1371/journal.pone.0055089 23405114 12. Margulies M Egholm M Altman WE Attiya S Bader JS Bemben LA Genome sequencing in micro fabricated high-density picolitre reactors Nature 2005 437 7057 376 380 16056220 13. Venter JC The sequence of the human genome Science 2001 291 5507 1304 1351 10.1126/science.1058040 11181995 14. Wheeler DA The complete genome of an individual by massively parallel DNA sequencing Nature. 2008 452 872 876 10.1038/nature06884 18421352

PMC005xxxxxx/PMC5001237.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557108289910.1186/s12864-016-2899-4ResearchGenomic data mining reveals a rich repertoire of transport proteins in Streptomyces Zhou Zhan zhanzhou@zju.edu.cn 123Sun Ning 21007079@zju.edu.cn 2Wu Shanshan swodylm@zju.edu.cn 1Li Yong-Quan lyq@zju.edu.cn 12Wang Yufeng yufeng.wang@utsa.edu 31 College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, 310058 People’s Republic of China 2 Zhejiang Provincial Key Laboratory of Microbial Biochemistry and Metabolism Engineering, Zhejiang University, Hangzhou, 310058 People’s Republic of China 3 Department of Biology and South Texas Center for Emerging Infectious Diseases, University of Texas at San Antonio, San Antonio, TX 78249 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.510© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Streptomycetes are soil-dwelling Gram-positive bacteria that are best known as the major producers of antibiotics used in the pharmaceutical industry. The evolution of exceptionally powerful transporter systems in streptomycetes has enabled their adaptation to the complex soil environment. Results Our comparative genomic analyses revealed that each of the eleven Streptomyces species examined possesses a rich repertoire of from 761-1258 transport proteins, accounting for 10.2 to 13.7 % of each respective proteome. These transporters can be divided into seven functional classes and 171 transporter families. Among them, the ATP-binding Cassette (ABC) superfamily and the Major Facilitator Superfamily (MFS) represent more than 40 % of all the transport proteins in Streptomyces. They play important roles in both nutrient uptake and substrate secretion, especially in the efflux of drugs and toxicants. The evolutionary flexibility across eleven Streptomyces species is seen in the lineage-specific distribution of transport proteins in two major protein translocation pathways: the general secretory (Sec) pathway and the twin-arginine translocation (Tat) pathway. Conclusions Our results present a catalog of transport systems in eleven Streptomyces species. These expansive transport systems are important mediators of the complex processes including nutrient uptake, concentration balance of elements, efflux of drugs and toxins, and the timely and orderly secretion of proteins. A better understanding of transport systems will allow enhanced optimization of production processes for both pharmaceutical and industrial applications of Streptomyces, which are widely used in antibiotic production and heterologous expression of recombinant proteins. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2899-4) contains supplementary material, which is available to authorized users. Keywords StreptomycesTransport proteinsComparative genomicsDrug effluxProtein translocationThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Streptomyces is a group of soil-dwelling Gram-positive bacteria, which are well known for their ability to produce a broad array of secondary metabolites including antibiotics, antifungals, antiparasitic drugs, anticancer agents, immunosuppressants, and herbicides [1, 2]. They are also ideal systems in biotechnology for heterologous expression of recombinant proteins with simple downstream processing and high yields [3, 4]. In order to survive in the complex soil environment, streptomycetes have evolved exceptionally powerful transport systems [5, 6]. For example, in Streptomyces coelicolor, there are more than 600 predicted transport proteins with a large proportion being the ATP-binding Cassette (ABC) and Major Facilitator Superfamily (MFS) transporters, which have been implicated in the transport of secondary metabolites including antibiotics [7]. In addition to secondary metabolites, streptomycetes also secret to the environment a mass of proteins through the general secretory (Sec) pathway and the twin-arginine translocation (Tat) pathway [8–10]. These secretory systems are known to facilitate nutrient acquisition. For example, secreted cellulases and chitinases can degrade otherwise insoluble nutrient sources. Transporters are of critical importance to all living organisms in facilitating metabolism, intercellular communication, biological synthesis and reproduction. They are involved in the uptake of nutrients from the environment, the secretion of metabolites, the efflux of drugs and toxins, the maintenance of ion concentration gradient across membranes, the secretion of macromolecules, such as sugars, lipids, proteins and nucleic acids, signaling molecules, the translocation of membrane proteins, and so on [11]. A Transporter Classification (TC) system has been developed by the Saier group [11, 12]. To date, more than 10,000 non-redundant transport proteins comprising about 750 families are collected in their Transporter Classification Database (TCDB) [13]. These families are divided among seven major classes: Channels/Pores (Class 1), Electrochemical Potential-driven Transporters (Class 2), Primary Active Transporters (Class 3), Group Translocators (Class 4), Transmembrane Electron Carriers (class 5), Accessory Factors Involved in Transport (Class 8), and Incompletely Characterized Transport Systems (Class 9). This classification system has been applied to in-depth studies of transporters in a number of microbial genomes [14–17], and is being adopted in this study for Streptomyces. The availability of genomes from closely related Streptomyces species enables comprehensive analysis of the transport protein families in Streptomyces. In this study, we report a catalog and comparative genomic analysis of transporters in eleven Streptomyces species with complete genome sequences and annotations, including S. coelicolor (SCO), S. avermitilis (SAV), S. bingchenggensis (SBI), S. cattleya (SCAT), S. flavogriseus (SFLA), S. griseus (SGR), S. hygroscopicus (SHJG), S. scabiei (SCAB), S. sp. SirexAA-E (SACTE), S. venezuelae (SVEN) and S. violaceusniger (STRVI) [7, 18–24]. We identified and classified these Streptomyces transporters, using the nomenclature in the TCDB. The class, transmembrane topology and substrate specificity of these transporters are investigated in detail. An improved understanding of Streptomyces transporters will bring new insights into the mechanisms underlying the unique and powerful secretion systems of secondary metabolites and proteins in this group of bacteria of enormous economic and biomedical significance. Results and discussion Abundant transporters are present in eleven Streptomyces genomes Strong material intake and secretion capacity powered by transport systems is an adaptive attribute of soil-dwelling bacteria [1]. We used the coding sequences from eleven Streptomyces genomes to query the TCDB [13, 25] using BLASTP and identified 761-1258 transporters in these eleven genomes, which accounted for 10.2 to 13.7 % of each respective proteome (Table 1 and Additional file 1). S. bingchenggensis, which has the largest genome, and the largest number of protein-coding genes, has the largest number of transporters, whereas S. cattleya contains only 761 transporters, the lowest number and proportion of transporters among the eleven Streptomyces species.Table 1 Distribution of transporters in eleven Streptomyces genomes Organisms Accession ID Genome size (Mbp) # ORFs # Transporters % Transporters S. avermitilis NC_003155 (chr) 9.1 7676 989 12.9 % NC_004719 (pSAP1) S. bingchenggensis NC_016582 (chr) 11.9 10022 1258 12.6 % S. cattleya NC_016111(chr) 8.1 7475 761 10.2 % NC_016113(pSCAT) S. coelicolor NC_003888(chr) 9.1 8153 990 12.1 % NC_003903 (pSCP1) NC_003904 (pSCP2) S. flavogriseus NC_016114 (chr) 7.7 6572 888 13.5 % NC_016110 (pSFLA01) NC_016115 (pSFLA02) S. griseus NC_010572 (chr) 8.5 7136 975 13.7 % S. hygroscopicus NC_017765 (chr) 10.4 9108 999 11.0 % NC_017766 (pSHJG1) NC_016972 (pSHJG2) S. scabiei NC_013929 (chr) 10.1 8746 1021 11.7 % S. sp. SirexAA-E NC_015953 (chr) 7.4 6357 869 13.7 % S. venezuelae NC_018750 (chr) 8.2 7453 935 12.5 % S. violaceusniger NC_015957 (chr) 11.0 8985 989 11.0 % NC_015951(pSTRVI01) NC_015952(pSTRVI02) Streptomyces transporters show diverse transmembrane topology The capacity of a transporter is often associated with the complexity and topology of its transmembrane region(s) where the major events of substrate uptake or output across the cell membranes take place. Using the TMHMM (TransMembrane prediction using Hidden Markov Models) algorithm [26], we performed the transmembrane topology analysis for Streptomyces transporters to identify the transmembrane segments (TMSs). The number of TMSs ranges from 0 to 24. The largest number of TMSs observed in a transporter in the eleven Streptomyces genomes varies from 16 to 24 (Table 2). Except for intra-/extra-cellular transporters which have no TMS, transporters with 6 and 12 TMSs are predominant. Most transporters with 6 TMSs are ABC transporters (TC 3.A.1), and transporters with 12 TMSs are mainly members of the Major Facilitator Superfamily (MFS) (TC 2.A.1), the Amino Acid-Polyamine-Organocation (APC) superfamily (TC 2.A.3), the Resistance-Nodulation-Cell Division (RND) superfamily (TC 2.A.6) and the ABC superfamily (TC 3.A.1). It is possible that these 12-TMS transporters have arisen from the primordial 6-TMS form via intragenic duplication [27]. Among the transporters with more than 6 TMSs, the transporters with an even number of TMSs are more abundant than those with an odd number of TMSs (Fig. 1). The distribution of TMSs in S. griseus transporters is unique: this bacterium has 53 transporters with 9 TMSs, mostly ABC transporters, accounting for 5.4 % of the total transporters. This proportion is significantly higher than that of the other ten sibling species. On the other hand, S. griseus has the lowest proportion of 12-TMS transporters (7.3 %), most of which are also ABC transporters. These topology patterns suggest that during the evolution of transporters in S. griseus, the “6 + 3” events may be more frequent than the typical “6 + 6” events observed in ten other Streptomyces species [27, 28].Table 2 Distribution of topological types of transporters in eleven Streptomyces genomes TMS SACTE SAV SBI SCAB SCAT SCO SFLA SGR SHJG STRVI SVEN 0 322 382 482 424 280 344 332 372 392 371 350 1 41 41 55 33 51 47 45 33 28 37 44 2 14 17 18 21 19 19 21 15 16 15 16 3 26 28 32 20 21 26 28 30 22 26 13 4 27 29 31 36 23 32 26 29 36 28 30 5 49 55 62 52 40 58 42 58 56 58 55 6 119 130 201 141 72 135 124 122 116 143 113 7 22 29 23 15 12 24 23 22 20 19 24 8 26 32 35 30 26 34 27 25 35 31 22 9 33 25 36 28 16 35 34 53 30 21 30 10 41 55 62 46 44 54 41 43 48 41 55 11 20 28 42 32 27 30 24 26 40 35 29 12 66 72 109 89 68 89 70 71 97 106 85 13 18 26 24 19 15 20 14 21 25 23 21 14 40 37 43 31 45 38 32 46 35 32 44 15 1 1 1 1 1 1 0 2 0 0 1 16 2 1 1 1 1 1 1 3 2 2 1 17 0 0 1 2 0 2 2 1 1 1 2 18 1 1 0 0 0 0 1 1 0 0 0 19 0 0 0 0 0 0 0 1 0 0 0 24 1 0 0 0 0 1 1 1 0 0 0 Total 869 989 1258 1021 761 990 888 975 999 989 935 Note: SACTE (S. sp. SirexAA-E), SAV (S. avermitilis), SBI (S. bingchenggensis), SCAB (S. scabiei), SCAT (S. cattleya), SCO (S. coelicolor), SFLA (S. flavogriseus), SGR (S. griseus), SHJG (S. hygroscopicus), STRVI (S. violaceusniger), SVEN (S. venezuelae) Fig. 1 Distribution of transporter topologies in eleven Streptomyces genomes. The abbreviations for species are: S. sp. SirexAA-E (SACTE), S. avermitilis (SAV), S. bingchenggensis (SBI), S. scabiei (SCAB), S. cattleya (SCAT), S. coelicolor (SCO), S. flavogriseus (SFLA), S. griseus (SGR), S. hygroscopicus (SHJG), S. violaceusniger (STRVI), and S. venezuelae (SVEN) Transporters in eleven Streptomyces genomes can be divided into seven classes and 171 families The Streptomyces transporters fall into seven classes and 171 transporter families according to the TCDB system (Table 3 and Additional file 2). The distribution of transporters in each species is depicted in Fig. 2.Table 3 Distribution of Streptomyces transporters in each TC class and subclass Class Subclass SACTE SAV SBI SCAB SCAT SCO SFLA SGR SHJG STRVI SVEN 1: Channels/Proes 22 31 29 34 22 29 26 28 26 31 30 1.A: α-Type Channels 18 24 20 25 15 22 21 20 21 24 21 1.B: β-Barrel Porins 3 6 7 8 6 6 4 5 4 6 6 1.C: Pore-Forming Toxins (Proteins and Peptides) 1 1 1 1 1 1 1 3 1 1 3 1.I: Membrane-bounded channels 0 0 1 0 0 0 0 0 0 0 0 2: Electrochemical Potential-driven Transporters 212 266 330 251 239 274 217 242 305 271 269 2.A: Porters (uniporters, symporters, antiporters) 212 266 328 251 239 274 217 242 305 271 269 2.C: Ion-gradient-driven energizers 0 0 2 0 0 0 0 0 0 0 0 3: Primary Active Transporters 500 544 705 553 365 552 498 555 489 528 494 3.A: P-P-bond-hydrolysis-driven transporters 455 492 656 505 304 497 451 508 433 476 449 3.B: Decarboxylation-driven transporters 6 6 5 6 10 7 6 6 6 4 6 3.D: Oxidoreduction-driven transporters 39 46 43 42 51 48 41 41 50 47 39 3.E: Light absorption-driven transporters 0 0 1 0 0 0 0 0 0 1 0 4: Group Translocators 27 46 62 54 35 30 36 40 46 37 43 4.A: Phosphotransfer-driven group translocators 5 7 4 5 2 8 8 6 5 7 6 4.B: Nicotinamide ribonucleoside uptake transporters 1 1 0 1 1 1 3 3 1 0 3 4.C: Acyl CoA ligase-coupled transporters 21 38 58 48 32 21 25 31 40 30 34 5: Transmembrane Electron Carriers 12 13 21 19 18 26 15 13 20 16 16 5.A: Transmembrane 2-electron transfer carriers 12 12 21 18 17 26 14 13 19 15 16 5.B: Transmembrane 1-electron transfer carriers 0 1 0 1 1 0 1 0 1 1 0 8: Accessory Factors Involved in Transport 4 4 5 6 5 5 5 6 6 4 4 8.A: Auxiliary transport proteins 4 4 5 6 5 5 5 6 6 4 4 9: Incompletely Characterized Transport Systems 60 63 74 67 75 67 68 66 63 69 55 9.A: Recognized transporters of unknown biochemical mechanism 27 25 44 27 33 31 32 35 27 33 25 9.B: Putative transport proteins 33 38 30 40 42 36 35 31 36 36 30 9.C: Functionally characterized transporters lacking identified sequences 0 0 0 0 0 0 1 0 0 0 0 N/A 32 22 32 37 2 7 23 25 44 33 24 Total 869 989 1258 1021 761 990 888 975 999 989 935 Fig. 2 Distribution of transporter types according to the TC system in eleven Streptomyces genomes. Class 1: Channels/Proes; Class 2: Electrochemical Potential-driven Transporters; Class 3: Primary Active Transporters; Class 4: Group Translocators; Class 5: Transmembrane Electron Carriers; Class 8: Accessory Factors Involved in Transport; Class 9: Incompletely Characterized Transport Systems; N/A: Not assigned The Primary Active Transporters (Class 3) is the most abundant class of transporters in Streptomyces, which includes 365-705 transporters (representing about 48.0-57.5 % of the total transport machinery). This class of transporters plays important roles in various aspects of bacterial life cycle, especially in the import and export of secondary metabolites, and cation transportation. Class 2 transporters, the electrochemical potential-driven transporters, are also widely found in Streptomyces. 212-330 transporters in eleven Streptomyces genomes belong to this class, which account for 24.4 %-31.4 % of all the transporters. The porters in this class include uniporters, symporters and antiporters. The most abundant family, MFS, in Class 2 transporters has been implicated in drug efflux. Lineage-specificity is also observed in this class of transporters. For example, S. bingchenggensis possesses two Ion-gradient-driven Energizers (TC 2.C), while the other ten Streptomyces species only have Porters (uniporters, symporters, antiporters) (TC 2.A). Class 1 transporters are not abundant, but are functionally important for Streptomyces. 22-34 channel/pore transporters are present in these eleven genomes, accounting for 2.3 %-3.2 % of all the transporters. The majority of these channel-type proteins are alpha-type channels (TC 1.A), which have been implicated in stress responses of Gram-positive bacteria, especially responses to osmotic pressure [27]. A small number of proteins belong to β-type porins and a fewer are putative Channel-Forming Toxins (TC 1.C). The membrane-bounded channel (TC 1.I) subclass is rare in Streptomyces; only S. bingchenggensis has a transport protein from this subclass. Classes 4, 5, and 8 are relatively less abundant. About 3.0 %-5.3 % of all the transport proteins are Class 4 transporters. Two major subclasses observed in Class 4 are the PTS Glucose-Glucoside (Glc) family (4.A.1) and the Fatty Acid Transporter (FAT) family (4.C.1), which are responsible for the transport of glucoses-glucosides and fatty acids, respectively. Notably, S. cattleya, which has the smallest repertoire of transporters among the eleven Streptomyces, does not seem to contain any Glc transporters; it remains unknown if it uses an alternative system. Only 12-21 members of the Class 4 transporters, the Transmembrane Electron Carriers, are found in Streptomyces. Two subclasses are present, including the Prokaryotic Molybdopterin-containing Oxidoreductase (PMO) family (TC 5.A.3) and the Prokaryotic Succinate Dehydrogenase (SDH) family (TC 5.A.4), which transfer electrons mainly by redox reactions. Class 8, the Accessory Factors Involved in Transport, is the least abundant transporter class (0.4 %-0.7 %) in Streptomyces. A significant number (60-75) of transporters in Streptomyces can be grouped into Class 9, an incompletely characterized class. While their exact physiological roles are yet to be elucidated, they might be involved in the transport of ions, implicated by their sequence similarities with the members of the HlyC/CorC (HCC) family (TC 9.A.40), and the Tripartite Zn2+ Transporter (TZT) family (TC 9.B.10). Examples of important transporter families Many of the 171 transporter families are involved in the transfer of ions, saccharides, amino acids, polypeptides, proteins, drugs, toxins and other compounds. The two most abundant and perhaps also the most important families are in the ABC (TC 3.A.1) and MFS (TC 2.A.1) superfamilies. They are responsible for the secretion of a wide array of antibiotics in Streptomyces [29, 30]. The ABC transporters 32.7 %-47.5 % (249-597) of all the transport proteins in the eleven Streptomyces genomes are members of ABC superfamily. ABC transporters are characterized by a conserved ATP hydrolyzing domain for energy provision, pore-forming membrane-integrated domain(s), and a substrate-binding domain [31, 32]. The ABC transport system is composed of the intake system and the efflux system. The 30 intake families (TC 3.A.1-3.A.33) that we identified in the Streptomyces genomes are specialized in the uptake of diverse nutrient substances. This intake system includes families of Carbohydrate Uptake Transporters (TC 3.A.1.1, 3.A.1.2) that transport saccharides, Polar Amino Acid Uptake Transporters and Hydrophobic Amino Acid Uptake Transporters (TC 3.A.1.3, 3.A.1.4) that transfer amino acids, Polyamine/Opine/ Phosphonate Uptake Transporters and Quaternary Amine Uptake Transporters (TC 3.A.1.11, 3.A.1.12) that transfer amine substances, Iron Chelate Uptake Transporters and Manganese/Zinc/Iron Chelate Uptake Transporters (TC 3.A.1.14, 3.A.1.15) that transfer metal ions. Unlike the intake system, the 35 Streptomyces efflux families are involved in the transport of macromolecular substances. These transporters are believed to be essential for Streptomyces due to their roles in drug efflux and protein secretion. The drug efflux system regulates various aspects of the response to drug compounds mediated by Drug Exporters (TC 3.A.1.105, 3.A.1.117, 3.A.1.119, 3.A.1.135), Drug Resistance ATPases (TC 3.A.1.120, 3.A.1.121), Macrolide Exporters (TC 3.A.1.122), β-Exotoxin I Exporters (TC 3.A.1.126), Multidrug Resistance Exporters (TC 3.A.201) and Pleiotropic Drug Resistance transporters (TC 3.A.1.205). Potent protein transport in Streptomyces is regulated by Protein/Peptide Exporters (TC 3.A.1.109, 110, 111, 112, 123, 124, 134), Lipoprotein Translocases (TC 3.A.1.125), AmfS Peptide Exporters (TC 3.A.1.127), and SkfA Peptide Exporters (TC 3.A.1.128). The MFS transporters Unlike the ABC transporters, the MFS transporters are driven by an electrochemical potential formed by ion concentration gradients across the cytomembrane [30]. There are 90-169 (10.1 %- 15.0 %) MFS transporters in eleven Streptomyces genomes. Streptomyces possesses 39 subfamilies of MFS transporters, including 20 intake systems, 13 efflux systems and 6 systems whose transport direction is unknown. The substances transported by the intake systems are mainly saccharides and organic acids. One of the most important roles of the MFS transporters is drug efflux [30]. Diverse subfamilies of drug efflux MFS transporters are present in Streptomyces, with varying mechanisms of action, including Drug:H+ Antiporters (TC 2.A.1.2, 2.A.1.3, 2.A.1.21), Aromatic Compound/Drug Exporters (TC 2.A.1.32), Fosmidomycin Resistance transporters (TC 2.A.1.35), Acriflavin-sensitivity transporters (TC 2.A.1.36), and Microcin C51 Immunity Proteins (TC 2.A.1.61), to name a few. The wide distribution of substrates for Streptomyces transporters The capacity of the complex and powerful transporter system in Streptomyces is evidenced by the broad scope of the substrates being transported. Figure 3a shows the distribution of transporters that transport different type of substrates in Streptomyces, including carbon sources, drugs, toxicants, electrons, inorganic molecules, macromolecules, amino acids and derivatives, nucleotides and derivatives, vitamins, and accessory factors. The carbon source transporters are the most abundant, with their proportion of all the transport proteins ranging from 21.7 to 31.6 % in eleven genomes. Notably, the substrates of an average of 6.4 % of the transporters in Streptomyces genomes examined cannot be determined based on genomic analysis, and await advanced structural and biochemical characterization.Fig. 3 a Distribution of substrate types and (b) predicted polar characteristics: bidirectional transport, uptake or export in eleven Streptomyces genomes Streptomyces transporters can be divided into three classes, uptake, efflux and bidirectional, according to the direction of the substrates transported (Fig. 3b). Among the transporters of the eleven Streptomyces genomes, on average 46.5 % are involved in the uptake of substrates, 35.8 % are involved in the efflux of substrates, and 11.0 % are in charge of the bidirectional transport of substrates. The direction of 6.7 % of these proteins remains undetermined. Streptomyces have lineage-specific protein secretion systems Streptomyces have two major lineage-specific protein transport systems, the Tat system (TC 2.A.64) and the Sec system (TC 3.A.5) [8, 9]. The Tat system was shown to be related to the pathogenicity of pathogenic bacteria [33]. In S. scabies, the transporters in the Tat pathway secrete several toxicity-associated proteins [34]. While the key component proteins of the Tat system, TatA, TatB and TatC, are present in all eleven Streptomyces genomes we looked at, lineage-specificity is clearly shown with respect to the copy number variation of these genes (Table 4). Only one copy of the tatB and tatC genes is present in nine Streptomyces genomes; S. flavogriseus has two copies of the tatB genes and S. hygroscopicus has two copies of the tatC genes. The copy number of the tatA gene ranges from one to three in eleven genomes (Table 4). Phylogenetic analysis shows that the multiple copies of the tatA genes may have different evolutionary origins and can be divided into three independent clades, namely tatA1, tatA2 and tatA3 (Fig. 4a). The tatA paralogous genes in the majority of the Streptomyces genomes belong to different clades. Notably, all the three tatA paralogous genes in S. cattleya are clustered into the tatA3 clade, indicative of recent gene duplication events.Table 4 The Tat translocation system in Streptomyces (TC 2.A.64) Species tatA1 tatA2 tatA3 tatB1 tatB2 tatC1 tatC2 SACTE SACTE_1063 SACTE_6092 SACTE_3032 SACTE_4381 SACTE_1062 SAV SAV_6692 SAV_3114 SAV_6693 SBI SBI_08493 SBI_04079 SBI_08494 SCAB SCAB_73591 SCAB_31121 SCAB_73601 SCAT SCAT_3206 SCAT_2668 SCAT_4914 SCAT_4007 SCAT_5184 SCO SCO1633 SCO3768 SCO5150 SCO1632 SFLA Sfla_5203 Sfla_0514 Sfla_5510 Sfla_5507 Sfla_2146 Sfla_5204 SGR SGR_5870 SGR_6484 SGR_340 SGR_2375 SGR_5871 SHJG SHJG_2368 SHJG_3070 SHJG_0499 SHJG_6250 SHJG_2367 SHJG_3069 STRVI Strvi_6639 Strvi_3352 Strvi_1468 Strvi_6638 SVEN SVEN_1225 SVEN_4796 SVEN_1224 Fig. 4 a Phylogenetic tree of the TatA system. b Phylogenetic tree of the SecD/SecF (b) system in eleven Streptomyces genomes. The trees were constructed using the neighbor-joining method by MEGA6 [43]. The Maximum Parsimony and Maximum Likelihood methods gave virtually the same topology (data not shown) Similarly, the Sec system is also species-specific. This system includes SecA, SecY, SecE, SecG, SecD, SecF, YajC, FtsY, etc. [35], all of which are highly conserved in Streptomyces (Table 5). There is only one copy of the secE, secG, secD, secF, yajC and ftsY genes in each of the eleven Streptomyces genomes. Interestingly, there is a second set of secA2/secY2 genes in several species, which may be involved in the secretion of proteins with specific functions, for example, the secretion of toxic proteins [36]. In S. avermitilis, for instance, there are two copies of the secA genes, and S. venezuelae has two copies of the secY genes.Table 5 The Sec translocation system in Streptomyces (TC 3.A.5) Species secA1 secA2 secY secY2 secE secG SACTE SACTE_2472 SACTE_3988 SACTE_3949 SACTE_1366 SAV SAV_5071 SAV_2565 SAV_4312 SAV_4908 SAV_6299 SBI SBI_06502 SBI_06209 SBI_06158 SBI_08032 SCAB SCAB_55371 SCAB_36741 SCAB_37261 SCAB_69731 SCAT SCAT_2009 SCAT_3612 SCAT_3559 SCAT_1102 SCO SCO3005 SCO4722 SCO4646 SCO1944 SFLA Sfla_3902 Sfla_2503 Sfla_2541 Sfla_4882 SGR SGR_4531 SGR_2814 SGR_2876 SGR_5576 SHJG SHJG_4468 SHJG_5817 SHJG_5775 SHJG_3400 STRVI Strvi_8396 Strvi_0893 Strvi_0854 Strvi_7031 SVEN SVEN_2748 SVEN_4399 SVEN_0354 SVEN_4338 SVEN_1573 Species secD secF secDF yajC ftsY SACTE SACTE_0919 SACTE_0918 SACTE_5723 SACTE_0920 SACTE_4801 SAV SAV_6837 SAV_6838 SAV_6836 SAV_2654 SBI SBI_02394 SBI_02393 SBI_02395 SBI_03477 SCAB SCAB_74911 SCAB_74921 SCAB_6041 SCAB_74901 SCAB_26291 SCAT SCAT_5307 SCAT_5308 SCAT_5306 SCAT_4417 SCO SCO1516 SCO1515 SCO6160 SCO1517 SCO5580 SFLA Sfla_5348 Sfla_5349 Sfla_0862 Sfla_5347 Sfla_1718 SGR SGR_6019 SGR_6020 SGR_1134 SGR_6018 SGR_1898 SHJG SHJG_2940 SHJG_2939 SHJG_8531 SHJG_2941 SHJG_6701 STRVI Strvi_3032 Strvi_3033 Strvi_3031 Strvi_1937 SVEN SVEN_1116 SVEN_1115 SVEN_0190 SVEN_1117 SVEN_5276 The evolutionary pattern in the secD and the secF genes is particularly interesting (Fig. 4b). In bacteria, these genes encode accessory factors in the Sec pathway that can accelerate the translocation of protein substrates. There are two forms of the secD and secF genes: in the first form, these two genes are adjacent but separate, while in the second form, the two genes are fused into a single secDF gene. The fused secDF is present in seven Streptomyces genomes. Unlike most bacteria that have one of the two forms, the majority of Streptomyces species have both the separated form and the fused form [37]. The acquisition of a second copy may confer a selective advantage to Streptomyces by enhancing the capacity and the effectiveness of protein transport. Conclusions Comparative genomic analyses of eleven Streptomyces genomes revealed an abundant repertoire of 761-1258 transporters, belonging to seven transporter classes and 171 transporter families. The powerful transport systems in Streptomyces play critical roles in drug efflux, protein secretion and stress response. A better understanding of transport systems will allow enhanced optimization of production processes for both pharmaceutical and industrial applications of Streptomyces. Methods Data The completed whole genome data of the eleven Streptomyces species (Table 1), including amino acid sequences and functional annotations of all the proteins were downloaded from the NCBI database (http://www.ncbi.nlm.nih.gov/genome/browse/). The transporter classification and amino acid sequences of all classified transporters were downloaded from the TCDB database (http://www.tcdb.org/) [13]. We also collected data from the TransporterDB database [38] (http://www.membranetransport.org/) which included the transporter classification data of S. coelicolor and S. avermitilis, and from the Transporter Inference Parser database [39] (http://biocyc.org/), which identified transporter according to their function annotation and included the relevant data of S. coelicolor, S. avermitilis, S. griseus and S. scabies. Identification and classification of transporters The BLASTP search of all the proteins in eleven Streptomyces species versus all the transport proteins in TCDB database was conducted to identify transporters in Streptomyces that are homologs to known and predicted transporters in the TCDB [13, 25]. The threshold for homologous genes was set as follow: E-value ≤ 10-5, similarity ≥ 50 %, and the sequence coverage ≥ 30 %. We classified a Streptomyces transporter based on its homologous gene with known function in the TCDB that had the lowest expected value, the highest similarity score and the highest coverage. The classification of Streptomyces transporters in the TransporterDB and the Transporter Inference Parser, the annotations and the conserved domain information helped to filter false negative and false positive predictions. The Pfam search program based on the Hidden Markov Models (HMMs) (http://pfam.xfam.org/) [40] was used to identify conserved structure domains of Streptomyces transporters, with Pfam GA as the threshold. TMHMM (http://www.cbs.dtu.dk/services/TMHMM/) [26] was used to analyze the transmembrane structures and the number of putative TMSs of Streptomyces transporters. On the basis of the degree of similarities with known or predicted transporters in the TCDB, as well as the conserved domains and the number and location of TMSs, we further classified the Streptomyces transporters into families and subfamilies of homologous transporters according to the TC system [13]. The TC number generally has five components: V.W.X.Y.Z, representing the transporter class, subclass, family, subfamily and the substrate or range of substrates transported [11, 12]. Most Streptomyces transporters were classified at the transporter family level. The transporters in superfamilies such as ABC and MFS were classified at the subfamily level. The substrate and transport direction of each Streptomyces transporter was predicted based on homology to functionally characterized transporters in the TCDB. Classification of a putative transporter into a family or subfamily according to the TC system allows for the prediction of substrate types and transport direction with confidence [13, 17, 41]. Phylogenetic analysis of transport protein families Multiple sequence alignments were obtained using Clustal X 2.1 [42]. Phylogenetic trees were reconstructed using MEGA6 with neighbor-joining (NJ), maximum parsimony (MP) and maximum likelihood (ML) methods [43]. Additional files Additional file 1: A detailed description of transporters in eleven Streptomyces genomes. The file includs protein IDs, names, annotations, protein lengths, Pfam domains, number of TMSs, and their homologs in TCDB with the BLASTP E-value. (XLSX 918 kb) Additional file 2: The classification of Streptomyces transporters. (XLSX 76 kb) Acknowledgements We thank the Computational Biology Initiative at UTSA for providing computational support. This work was supported by grants from the National Natural Science Foundation of China (31501021) and the Zhejiang Provincial Natural Sciences Foundation of China (LY15C060001) to ZZ, grants from the National Basic Research Program of China (973 Program, 2012CB721005) and the National Natural Science Foundation of China (30870033) to YQL, grants from the National Institutes of Health (GM100806, AI080579, and GM081068) to YW. ZZ was also supported by a government scholarship from the China Scholarship Council. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Declarations Publication charges for this article have been funded by the National Natural Science Foundation of China (31501021) to ZZ. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The datasets supporting the conclusions of this article are included within the article and its additional files. Authors’ contributions YW, YQL and ZZ conceived and designed the study. ZZ, NS, SW and YW performed data analysis. YW and ZZ drafted the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Hopwood DA Streptomyces in Nature and Medicine: The Antibiotic Makers 2007 New York Oxford University Press 2. Garrity GM, Lilburn TG, Cole JR, Harrison SH, Euzéby J., B.J. T. Part 10 - The Bacteria: Phylum “Actinobacteria”: Class Actinobacteria. In: Taxonomic Outline of the Bacteria and Archaea. 2007: Release 7.7.: 399-539. 3. Anne J Maldonado B Van Impe J Van Mellaert L Bernaerts K Recombinant protein production and streptomycetes J Biotechnol 2012 158 4 159 167 10.1016/j.jbiotec.2011.06.028 21777629 4. Li YD Zhou Z Lv LX Hou XP Li YQ New approach to achieve high-level secretory expression of heterologous proteins by using Tat signal peptide Protein Pept Lett 2009 16 6 706 710 10.2174/092986609788490096 19519532 5. Zhou Z Gu J Li YQ Wang Y Genome plasticity and systems evolution in Streptomyces BMC Bioinformatics 2012 13 Suppl 10 S8 10.1186/1471-2105-13-S10-S8 22759432 6. Zhou Z Gu J Du YL Li YQ Wang Y The -omics era- toward a systems-level understanding of streptomyces Curr Genomics 2011 12 6 404 416 10.2174/138920211797248556 22379394 7. Bentley SD Chater KF Cerdeno-Tarraga AM Challis GL Thomson NR James KD Harris DE Quail MA Kieser H Harper D Complete genome sequence of the model actinomycete Streptomyces coelicolor A3(2) Nature 2002 417 6885 141 147 10.1038/417141a 12000953 8. Chater KF Biro S Lee KJ Palmer T Schrempf H The complex extracellular biology of Streptomyces FEMS Microbiol Rev 2010 34 2 171 198 10.1111/j.1574-6976.2009.00206.x 20088961 9. Widdick DA Dilks K Chandra G Bottrill A Naldrett M Pohlschroder M Palmer T The twin-arginine translocation pathway is a major route of protein export in Streptomyces coelicolor Proc Natl Acad Sci U S A 2006 103 47 17927 17932 10.1073/pnas.0607025103 17093047 10. Morosoli R Shareck F Kluepfel D Protein secretion in streptomycetes FEMS Microbiol Lett 1997 146 2 167 174 10.1111/j.1574-6968.1997.tb10188.x 9011036 11. Busch W Saier MH Jr The transporter classification (TC) system, 2002 Crit Rev Biochem Mol Biol 2002 37 5 287 337 10.1080/10409230290771528 12449427 12. Busch W Saier MH Jr The IUBMB-endorsed transporter classification system Mol Biotechnol 2004 27 3 253 262 10.1385/MB:27:3:253 15247498 13. Saier MH Jr Reddy VS Tamang DG Vastermark A The transporter classification database Nucleic Acids Res 2014 42 Database issue D251 258 10.1093/nar/gkt1097 24225317 14. Paulsen IT Nguyen L Sliwinski MK Rabus R Saier MH Jr Microbial genome analyses: comparative transport capabilities in eighteen prokaryotes J Mol Biol 2000 301 1 75 100 10.1006/jmbi.2000.3961 10926494 15. Kumar U Saier MH Jr Comparative genomic analysis of integral membrane transport proteins in ciliates J Eukaryot Microbiol 2015 62 2 167 187 10.1111/jeu.12156 25099884 16. Paparoditis P Vastermark A Le AJ Fuerst JA Saier MH Jr Bioinformatic analyses of integral membrane transport proteins encoded within the genome of the planctomycetes species, Rhodopirellula baltica Biochim Biophys Acta 2014 1838 1 Pt B 193 215 10.1016/j.bbamem.2013.08.007 23969110 17. Getsin I Nalbandian GH Yee DC Vastermark A Paparoditis PC Reddy VS Saier MH Jr Comparative genomics of transport proteins in developmental bacteria: Myxococcus xanthus and Streptomyces coelicolor BMC Microbiol 2013 13 279 10.1186/1471-2180-13-279 24304716 18. Barbe V Bouzon M Mangenot S Badet B Poulain J Segurens B Vallenet D Marliere P Weissenbach J Complete genome sequence of Streptomyces cattleya NRRL 8057, a producer of antibiotics and fluorometabolites J Bacteriol 2011 193 18 5055 5056 10.1128/JB.05583-11 21868806 19. Wang XJ Yan YJ Zhang B An J Wang JJ Tian J Jiang L Chen YH Huang SX Yin M Genome sequence of the milbemycin-producing bacterium streptomyces bingchenggensis J Bacteriol 2010 192 17 4526 4527 10.1128/JB.00596-10 20581206 20. Ohnishi Y Ishikawa J Hara H Suzuki H Ikenoya M Ikeda H Yamashita A Hattori M Horinouchi S Genome sequence of the streptomycin-producing microorganism streptomyces griseus IFO 13350 J Bacteriol 2008 190 11 4050 4060 10.1128/JB.00204-08 18375553 21. Ikeda H Ishikawa J Hanamoto A Shinose M Kikuchi H Shiba T Sakaki Y Hattori M Ōmura S Complete genome sequence and comparative analysis of the industrial microorganism Streptomyces avermitilis Nat Biotechnol 2003 21 5 526 531 10.1038/nbt820 12692562 22. Bignell DRD Seipke RF Huguet-Tapia JC Chambers AH Parry RJ Loria R Streptomyces scabies 87-22 contains a coronafacic acid-like biosynthetic cluster that contributes to plant-microbe interactions Mol Plant Microbe Interact 2010 23 2 161 175 10.1094/MPMI-23-2-0161 20064060 23. Wu H Qu S Lu CY Zheng HJ Zhou XF Bai LQ Deng ZX Genomic and transcriptomic insights into the thermo-regulated biosynthesis of validamycin in Streptomyces hygroscopicus 5008 BMC Genomics 2012 13 337 10.1186/1471-2164-13-337 22827618 24. Yaxley AM. Study of the complete genome sequence of Streptomyces scabies (or scabiei) 87.22. University of Warwick, Coventry, UK; 2009. 25. Saier MH Jr Tran CV Barabote RD TCDB: the Transporter Classification Database for membrane transport protein analyses and information Nucleic Acids Res 2006 34 Database issue D181 186 10.1093/nar/gkj001 16381841 26. Krogh A Larsson B von Heijne G Sonnhammer EL Predicting transmembrane protein topology with a hidden Markov model: application to complete genomes J Mol Biol 2001 305 3 567 580 10.1006/jmbi.2000.4315 11152613 27. Saier MH Jr Tracing pathways of transport protein evolution Mol Microbiol 2003 48 5 1145 1156 10.1046/j.1365-2958.2003.03499.x 12787345 28. Lam VH Lee JH Silverio A Chan H Gomolplitinant KM Povolotsky TL Orlova E Sun EI Welliver CH Saier MH Jr Pathways of transport protein evolution: recent advances Biol Chem 2011 392 1-2 5 12 10.1515/bc.2011.018 21194372 29. Mendez C Salas JA The role of ABC transporters in antibiotic-producing organisms: drug secretion and resistance mechanisms Res Microbiol 2001 152 3-4 341 350 10.1016/S0923-2508(01)01205-0 11421281 30. Saidijam M Benedetti G Ren Q Xu Z Hoyle CJ Palmer SL Ward A Bettaney KE Szakonyi G Meuller J Microbial drug efflux proteins of the major facilitator superfamily Curr Drug Targets 2006 7 7 793 811 10.2174/138945006777709575 16842212 31. Tomii K Kanehisa M A comparative analysis of ABC transporters in complete microbial genomes Genome Res 1998 8 10 1048 1059 9799792 32. Wang B Dukarevich M Sun EI Yen MR Saier MH Jr Membrane porters of ATP-binding cassette transport systems are polyphyletic J Membr Biol 2009 231 1 1 10 10.1007/s00232-009-9200-6 19806386 33. De Buck E Lammertyn E Anne J The importance of the twin-arginine translocation pathway for bacterial virulence Trends Microbiol 2008 16 9 442 453 10.1016/j.tim.2008.06.004 18715784 34. Joshi MV Mann SG Antelmann H Widdick DA Fyans JK Chandra G Hutchings MI Toth I Hecker M Loria R The twin arginine protein transport pathway exports multiple virulence proteins in the plant pathogen Streptomyces scabies Mol Microbiol 2010 77 1 252 271 10.1111/j.1365-2958.2010.07206.x 20487278 35. Driessen AJ Fekkes P van der Wolk JP The Sec system Curr Opin Microbiol 1998 1 2 216 222 10.1016/S1369-5274(98)80014-3 10066476 36. Rigel NW Braunstein M A new twist on an old pathway--accessory secretion systems Mol Microbiol 2008 69 2 291 302 10.1111/j.1365-2958.2008.06294.x 18544071 37. Zhou Z Li Y Sun N Sun Z Lv L Wang Y Shen L Li YQ Function and evolution of two forms of SecDF homologs in Streptomyces coelicolor PLoS One 2014 9 8 e105237 10.1371/journal.pone.0105237 25140821 38. Ren Q Chen K Paulsen IT TransportDB: a comprehensive database resource for cytoplasmic membrane transport systems and outer membrane channels Nucleic Acids Res 2007 35 Database issue D274 279 10.1093/nar/gkl925 17135193 39. Lee TJ Paulsen I Karp P Annotation-based inference of transporter function Bioinformatics 2008 24 13 i259 267 10.1093/bioinformatics/btn180 18586723 40. Finn RD Bateman A Clements J Coggill P Eberhardt RY Eddy SR Heger A Hetherington K Holm L Mistry J Pfam: the protein families database Nucleic Acids Res 2014 42 Database issue D222 230 10.1093/nar/gkt1223 24288371 41. Saier MH Jr A functional-phylogenetic classification system for transmembrane solute transporters Microbiol Mol Biol Rev 2000 64 2 354 411 10.1128/MMBR.64.2.354-411.2000 10839820 42. Larkin MA Blackshields G Brown NP Chenna R McGettigan PA McWilliam H Valentin F Wallace IM Wilm A Lopez R Clustal W and Clustal X version 2.0 Bioinformatics 2007 23 21 2947 2948 10.1093/bioinformatics/btm404 17846036 43. Tamura K Stecher G Peterson D Filipski A Kumar S MEGA6: Molecular evolutionary genetics analysis version 6.0. Mol Biol Evol 2013 30 12 2725 2729 10.1093/molbev/mst197 24132122

PMC005xxxxxx/PMC5001238.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557118290810.1186/s12864-016-2908-7ResearchTowards precision medicine-based therapies for glioblastoma: interrogating human disease genomics and mouse phenotypes Chen Yang yxc233@case.edu 1Gao Zhen zxg119@case.edu 1Wang Bingcheng bxw14@case.edu 2Xu Rong rxx@case.edu 11 Department of Epidemiology and Biostatistics, Case Western Reserve University, Cleveland, Ohio USA 2 Department of Pharmacology, Case Western Reserve University, Cleveland, Ohio USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.516© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Glioblastoma (GBM) is the most common and aggressive brain tumors. It has poor prognosis even with optimal radio- and chemo-therapies. Since GBM is highly heterogeneous, drugs that target on specific molecular profiles of individual tumors may achieve maximized efficacy. Currently, the Cancer Genome Atlas (TCGA) projects have identified hundreds of GBM-associated genes. We develop a drug repositioning approach combining disease genomics and mouse phenotype data towards predicting targeted therapies for GBM. Methods We first identified disease specific mouse phenotypes using the most recently discovered GBM genes. Then we systematically searched all FDA-approved drugs for candidates that share similar mouse phenotype profiles with GBM. We evaluated the ranks for approved and novel GBM drugs, and compared with an existing approach, which also use the mouse phenotype data but not the disease genomics data. Results We achieved significantly higher ranks for the approved and novel GBM drugs than the earlier approach. For all positive examples of GBM drugs, we achieved a median rank of 9.2 45.6 of the top predictions have been demonstrated effective in inhibiting the growth of human GBM cells. Conclusion We developed a computational drug repositioning approach based on both genomic and phenotypic data. Our approach prioritized existing GBM drugs and outperformed a recent approach. Overall, our approach shows potential in discovering new targeted therapies for GBM. Keywords GlioblastomaDrug repositioningCancer genomicsMouse phenotypeThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Glioblastoma (GBM) is one of the leading causes of cancer-related deaths in both the pediatric and adult populations [1]. The standard treatment includes radiation plus chemotherapy following maximal safe resection of cancer mass [2]. However, the prognosis of GBM patients remains poor even with optimal radio- and chemo-therapies: the mean survival is 15 months and most patients die within two years [2, 3]. In addition, GBM is not a priority for new drug development because of socioeconomic problems and medical difficulties [3]. Both the grim prognosis and urgent clinical needs have motivated us to develop an “in silico” drug repositioning approach and pursue FDA-approved agents that has the potential to treat GBM but not previously identified as GBM therapeutics. Since GBMs are highly heterogeneous at the genomic, histological and differentiation level, the lack of specific therapies contributes to the treatment failures. Cancer therapies that target on specific molecular profiles of individual tumors have the potential to maximize the efficacy [4]. For example, Imatinib has been used to successfully treat a subtype of leukemia with mutations in the BCR-ABL fusion protein and has achieved a median survival of five years [5]. Over the past two decades, extensive researches have identified hundreds of genetic mutations that likely drive the GBM formation [6, 7]. More recently, systemic multi-platform analysis of glioma and bioinformatic mining by The Cancer Genome Atlas (TCGA) has led to the classification of GBM into distinct molecular subtypes according to the genes altered during gliomagenesis [8, 9]. Here, we use the accumulated genomic data for GBM to guide the drug repositioning approach towards discovering precise targeted drugs for GBM. Disease genetic and genomic profiles have been demonstrated useful in computational drug discovery approaches [10–15]. These approaches estimate the association between a drug and a disease through calculating their genomic profile similarities. They show increased ability in discovering new drug-disease pairs comparing with drug-based and disease-based repositioning strategies (Fig. 1), which depend on existing drug-indication knowledge to infer new drug-disease associations. However, the profile-based approach (Fig. 1(c)) has an inherent challenge: the lower-level genomics profile similarities between drugs and diseases do not necessarily translate into higher-level drug treatment efficacy in diseases. Previous studies have demonstrated that phenotypic data are critical in computational drug discovery approaches [16–19]. Recently, the Mouse Genome Informatics (MGI) database [20] has provided large amounts of phenotypic descriptions for mouse genetic mutations based on systematic gene knockouts, which are impossible on human. These causal gene-phenotype associations in mice have been demonstrated useful in discovering of new disease-associated genes [21] and drug targets [22], and also have the potential to overcome the challenge in genomics-based drug repositioning approaches. Fig. 1 Computational drug repositioning strategies: a Disease-based methods (Similar diseases may be treated with the same drug), b Drug-based methods (similar drugs may treat the same disease), and c Profile-based methods (the association between a drug and a disease is estimated by their profile similarity) In this study, we develop a novel GBM drug repositioning strategy leveraging both lower-level disease and drug genomics and higher-level mouse phenotypes. We first identify GBM-specific mouse phenotypes using a compiled list of GBM-associated genes identified by multiple TCGA studies [6, 8]. Then we screen all the FDA-approved drugs for candidates that share similar mouse phenotype profiles with GBM. We validate the approach using approved GBM drugs, and approximate the performance in detecting novel GBM drugs using two evaluation sets: a set of potential GBM therapies tested in clinical trials and a set of off-label GBM drugs in the post-marketing surveillance system. Finally, we investigate the top 10 % drug predictions. Overall, we combine the genomic and phenotypic data for diseases and drugs towards identifying novel targeted therapies for GBM. Methods Our approach ranks 1348 approved drugs by the mouse phenotype profile similarities between GBM and each drug. Figure 2 shows two steps in the algorithm: (1) identifying the phenotypes in mice for GBM and each approved drug, using the well-studied disease-associated genes and drug target genes, respectively; and (2) calculating the semantic similarities of the mouse phenotype profiles between the disease and drugs. The rank of drugs based on the phenotype similarities suggests how likely the drug can be used to treat GBM. The following parts describe each step as well as the evaluation methods in detail. Fig. 2 Our method contains two parts: a Identify mouse phenotype profiles for GBM and all approved drugs, and b Rank candidate drugs by mouse phenotype similarities with GBM Identify mouse phenotype profiles for GBM and drugs using disease genetics and drug target genes TCGA Research Network provides a comprehensive catalog of genomic abnormalities driving tumorgenesis. We compiled a list of 102 genetic mutations that significantly differentiate GBM tumors and healthy tissue from several recent TCGA studies [6, 8]. The list primarily contains the genes in the core GBM-associated pathways, including p53, Rb, and receptor tyrosine kinase (RTK)/Ras/phosphoinositide 3-kinase (PI3K) signaling. We extracted the mouse phenotypes that are linked to 102 GBM-associated genes from the MGI database. Each phenotype was weighted and ranked by the number of genes it is linked with. The mouse phenotype terms were removed from the list if their weights are smaller than the median of all weights. At last, we identified a list of 945 GBM-specific mouse phenotypes. We mapped the phenotype terms into 26 categories by tracing the isa relationship in the mammalian phenotype ontology. A score was calculated for each category as the sum of weights of all phenotypes in it. We ranked the phenotype categories by their scores and investigated the top five categories. Then we identified the mouse phenotype profile for each of the 1348 FDA-approved drug. The drug target genes were first extracted from the STITCH database, and each drug-target link has a confidence score. Then we extracted the mouse phenotypes that are linked with the target genes for each drug. The phenotype terms are weighted by the sum of confidence scores of the corresponding target genes. Finally, we obtained a vector of weighted mouse phenotype features for each candidate drug. Rank candidate drugs for GBM using mouse phenotype similarities between GBM and drugs We calculated the phenotypic similarity between GBM and the drugs in order to rank the candidate drugs by their similarity to GBM. Phenotype terms associated with both GBM and the drugs were normalized by concepts in the ontology, which provides semantic relationships between concepts and has been widely used in biomedical applications [17, 21, 23, 24]. We calculated the semantic distances between the mouse phenotype vectors for GBM and the candidate drugs in the context of the mouse phenotype ontology. We first quantified the information content for each phenotype term t as −logp(t), in which p(t) represents the frequency among phenotype annotations to all the 7568 mouse genes. In calculating the information content, if a gene is annotated by one phenotype term, we assumed that it is also annotated by the ancestors of this term in the hierarchy of mammalian phenotype ontology. Hence, a phenotype term has higher information content than its ancestors, which lie on higher levels in the ontology. Then we defined the semantic distance sim(t1,t2) between phenotype terms t1 and t2 as: 1 sim(t1,t2)=maxa∈A(t1,t2)−logp(a), where A(t1,t2) is the set of common ancestors for t1 and t2 in the ontology. To calculate the distance from the phenotype vector p1 to p2, we matched each phenotype feature in p1 to the most similar feature in p2 and took the average: 2 sim(p1→p2)=avg∑t1∈p1maxt2∈p2sim(t1,t2) To calculate the distance between p1 and p2, we averaged the semantic distances in both directions: 3 sim(p1,p2)=12sim(p1→p2)+12sim(p2→p1) A similar definition of distance between a pair of concepts in an ontology was also used before [24]. Validate our approach through de novo prediction of approved, potential, as well as off-label GBM drugs We tested whether our approach can prioritize the existing and novel GBM drug therapies in the top among 1348 candidates. We compiled three evaluation drug sets based on previous studies [25, 26]: the approved GBM drugs, potential GBM drugs that have been tested in clinical trials, and off-label GBM drugs identified from a post-marketing drug surveillance system. The approved GBM drug set contains temozolomide and carmustine, which are cytotoxic (non-targeted) chemical drugs, and bevacizumab, which is the first targeted drug approved for brain tumor. The potential GBM drug set contains 52 drugs collected from the clinical trials. In addition, the FDA drug surveillance system contains large-scale drug-disease data collected from hospitals, patients, and pharmaceutical companies. A total of 36 off-label uses for GBM were extracted from this system (containing zero overlap with the 52 potential GBM drugs). We have removed the approved GBM drugs from both the potential and off-label GBM drug sets, and evaluated the ranks for these two sets to approximate the performance of the proposed approach in predicting novel GBM drugs. We compared the performance of our approach with a recent drug repositioning approach proposed by Hoehndorf [27] in ranking the above evaluation sets. The Hoehndorf’s method also used the mouse phenotype data, but did not incorporate the human disease genomics data. They matched the human phenotype ontology [24] and the mammalian phenotype ontology [28] to predict genes for a human disease using the gene-phenotype relationships in animal models. After that, they linked the predicted disease genes with the drug target genes to suggest candidate drugs for the given disease. We first evaluated the ranks for approved GBM drugs, and the median ranks for the potential and off-label GBM drug sets. We tested the median rank instead of the average, because the median is not affected by individual large ranks. For example, if the rank for a GBM drug is 1348 using method A and 1000 using method B, both methods fail in detecting this positive example. But method B may achieve much higher average ranks than method A, affected by the large values of these two ranks. We also extracted the overlapping drugs between the ranked drug lists generated by the two methods, and performed the paired student’s t-test to evaluate the significance of their ranking difference. Then we combined the three evaluation sets, assumed all the drugs as the positive examples, and compared the precision-recall curve as well as the mean average precision between methods. Result Identified mouse phenotypes are associated with GBM pathogenesis We classified the GBM-specific mouse phenotypes detected through GBM-associated genes, and ranked the phenotype categories. Table 1 shows that the top-ranked phenotype categories are “tumorigenesis” and “nervous system phenotype” as expected. Besides, the result shows that GBM interacts with the immune system and hematopoietic system, which is consistent with a series of previous researches. A recent mouse model study [29] reveals that the GBM cells are able to migrate along the cerebral blood vessels and extract nutrients from the blood for themselves. They also replace the specialized brain cell named astrocytes to create a breakdown in the blood-brain barrier (BBB), which tightly controls the lymphocyte traffic into the central nervous system (CNS) in healthy people. Then the GBM cells evade the immune responses through inhibiting the T cell proliferation [30], inducing immunosuppressive microglia [31] and other channels. Studies on the pathways involving these immune evasion strategies have led to several recent advances in developing targeted immunotherapies for GBM [32, 33]. Table 1 The top-ranked categories of GBM-specific mouse phenotypes detected through disease genetics Rank Phenotype category Example phenotype 1 Tumorigenesis Increased glioblastoma incidence 2 Nervous system phenotype Abnormal astrocyte morphology 3 Hematopoietic system Abnormal hematopoiesis Phenotype 4 Mortality/aging Decreased survivor rate 5 Immune system phenotype Decreased leukocyte cell number Our approach outperforms an existing drug repositioning approach in prioritizing approved, potential and off-label GBM drugs Using the phenotype profiles detected through GBM and drug genomics, our approach prioritized the approved GBM drug bevacizumab in top 24.4 % among a total of 1348 chemicals, which is a much higher rank than Hoehndorf’s rank in 67.9 %. In addition, we identified the other two GBM approved drugs temozolomide and carmustine and ranked them within top 6.7 % and 10.7 %, respectively, while Hoehndorf’s ranking list does not contain these two drugs. Table 2 shows that our median rank for the potential GBM drugs in clinical trials is 7.8 %, which is 4.8-fold higher than Hoehndorf’s median rank. For the off-label GBM drugs from the post-marketing surveillance system, our median rank is 15.3 %, which is 1.8-fold higher than Hoehndorf’s approach. We generated significantly better ranks for both the potential drug set (p=0.003) and the off-label drug set (p=0.02) than Hoehndorf’s approach based on result of the paired t-test. Besides, we ranked the potential drugs higher than the off-label drugs, possibly because of the mixed and noisy data sources of the post-marketing drug surveillance system. Table 2 The ranks for GBM drugs in the three evaluation sets generated by our approach and Hoehndorf’s approach Evaluation drug (set) Our approach Hoehndorf’s approach p-value Approved drugs Temozolomide 6.7 % NA NA Carmustine 10.7 % NA NA Bevacizumab 24.4 % 67.9 % NA Potential drugs (clinical trials) 7.8 % 45.4 % p=0.003 Off-label drugs (post-marketing surveillance) 15.3 % 44.2 % p=0.02 Combination 9.2 % 45.6 % p=0.0003 We combined the three evaluation sets as a positive sample set and found their median rank is within top 9.2 % (Table 2). Comparing with Hoehndorf’s approach, we achieved significantly better performance in ranking these positive drugs (p=3e−4). Figure 3 shows the precision-recall curve for the two methods. The mean average precision calculated based on the curve is 0.29 for our approach, comparing to 0.20 for Hoehndorf’s approach. Fig. 3 Precision-recall curve in ranking the positive examples of GBM drugs for our approach and Hoehndorf’s approach We classified the drugs in all evaluation sets into three types, namely non-targeted cancer drugs, targeted cancer drugs and non-cancer drugs (Table 3). Our approach achieved the best performance in ranking the targeted cancer drugs, which has a median rank of 7.3 %. On the other hand, Hoehndorf’s approach performed best when predicting the non-targeted cancer therapies. This may be due to the different input data for the two methods: we incorporated the disease genomics data, while Hoehndorf’s approach directly analyzed the disease phenotypes. Overall, our approach works better than the baseline approach in ranking the evaluation drugs, which are more likely to be able to treat GBM than random drugs. The most significant difference between our approach and the baseline approach lies in ranking the non-cancer drugs that have been tested or in off-label use for GBM, and the paired t-test yielded a p-value of 6e−4. Table 3 Median ranks for different types of drugs in the combined evaluation set Drug type Our approach Hoehndorf’s approach p-value Non-targeted cancer therapies (chemotherapies) 9.5 % 25.8 % p=0.023 Targeted cancer drugs 7.3 % 56.4 % p=0.015 Non-cancer drugs 13.3 % 67.4 % p=0.0006 Together, the result suggests that our approach performed significantly better than an existing method that also utilizes the mouse phenotype data in prioritizing all approved and novel GBM drugs, and specially in identifying potential targeted GBM drugs. One possible reason is that we used the most recent discoveries of GBM associated gene mutations and a more comprehensive drug-target database, which provides opportunities for discovering targeted therapies for GBM. Table 4 lists five examples in our top 5 % predictions and their traditionally approved indications. Among them, rosiglitazone is a PPAR γ agonist that shows the ability to inhibit proliferation of human GBM cell lines [34]. Bortezomib may overcome MGMT-related resistance of GBM cell lines to temozolomide [35]. Estradiol is a form of estrogen and induces JNK-dependent apoptosis in human GBM and rat glioma cells [36]. Simvastatin was identified by a recent drug screening study using human cell lines [37]. Decitabine can efficiently induce the differentiation and growth inhibition in IDH1 mutant glioma cells [38]. Table 4 Examples in our top 5 % drug predictions for GBM Drug Traditional indication Rosiglitazone Type 2 diabetes Bortezomib Multiple myeloma Estradiol Symptoms of menopause Simvastatin High cholesterol and triglyceride Decitabine Myelodysplastic syndrome Discussion In this study, we predict candidate targeted drugs for GBM through combining discoveries on disease genomics and large-scale mouse phenotype data. We currently have not considered the blood-brain barrier (BBB) permeability of the candidate drugs, which is a major challenge for drug discovery for CNS diseases. No readily available BBB permeable drug database can be publicly accessed to enable simple filtering among the candidate GBM drugs. Computational approaches based on decision tree have been developed to identify BBB permeable drugs [39]. It is also possible to modify the drug chemically or pharmaceutically to increase its permeability [40]. In summary, our future work contains further selecting the candidate GBM drugs that can be delivered into the brain. TCGA recently classified GBM into four types: Proneural, Neural, Classical and Mesenchymal [6, 8]. Each class has distinct genomic profiles. The Classical GBM has increased EGFR expression and lacks TP53 mutations. The Proneural subtype shows alterations of PDGFRA and point mutations in IDH1. The Neural subtype is characterized by expressions of neuron markers. And the Mesenchymal GBM shows deletions of NF1, expression of mesenchymal markers, and high expressions of the TNF super family pathway and NF- κB pathway [8]. Patients of the four types also respond differently to chemo- and/or radiotherapy [8]. In the future, We will predict drugs for each of the four types targeting on their distinct genetic and genomic features towards achieving precision medicine for GBM. We expect specific and different drug predictions across the GBM subtypes. In addition, human disease phenotypes, disease phenotypic similarities and drug similarities may also contribute to GBM drug repositioning. For the drug-gene interaction database, we currently use the STITCH database, but other sources like Cancer Cell Line Encyclopedia (CCLE) [41] may contain different knowledge. In the future, we will develop algorithms to seamlessly integrate more comprehensive data to further filter strong candidate GBM drugs. We will also test the candidate drugs in biomedical experiments and clinical studies. Conclusions We screened 1348 approved drugs and predicted targeted drugs for GBM through combining disease genomic and mouse phenotype data. Our approach prioritized the approved GBM drugs, and outperformed a recent drug repositioning method in identifying novel GBM drugs. For all positive examples of GBM drugs, we achieved a median rank of 9.2 %, comparing to 45.6 % generated by the earlier approach. In the paired t-test, our approach generated significantly higher ranks for the evaluation drugs than the baseline approach (p=3e−4). We found that many of our top-ranked predictions have been demonstrated effective in inhibiting the growth of human GBM cells. Overall, the results show that our drug repositioning approach has the potential in finding new targeted therapies for GBM. Abbreviations GBM, glioblastoma; TCGA, the cancer genome atlas; FDA, food and drug administration; MGI, mouse genome informatics; RTK, receptor tyrosine kinase; PI3K, phosphoinositide 3-kinase; STITCH, search tool for interactions of chemicals; BBB, blood-brain barrier; CNS, central nervous system; ccle, cancer cell line encyclopedia; MAP, mean average precisio From The International Conference on Intelligent Biology and Medicine(ICIBM) 2015 Indianapolis, IN, USA.13-15 November 2015 Acknowledgements YC, ZG and RX are supported by the Eunice Kennedy Shriver National Institute of Child Health & Human Development of the National Institutes of Health under the NIH Director’s New Innovator Award DP2HD084068. BW is supported by NCI award R01CA155676 and R01CA152371. Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials Data is available by contacting Rong Xu at rxx@case.edu. Authors’ contributions RX conceived the study. YC designed the methods, performed the experiments and wrote the manuscript. All authors have participated study discussion and manuscript preparation. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1 Wen PY Kesari S Malignant gliomas in adults New Engl J Med 2008 359 5 492 507 10.1056/NEJMra0708126 18669428 2 Stupp R Tonn JC Brada M Pentheroudakis G Group EGW High-grade malignant glioma: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up Ann Oncol 2010 21 suppl 5 v190—v193 20555079 3 Omuro A DeAngelis LM Glioblastoma and other malignant gliomas: a clinical review Jama 2013 310 17 1842 1850 10.1001/jama.2013.280319 24193082 4 Bai RY Staedtke V Riggins GJ Molecular targeting of glioblastoma: drug discovery and therapies Trends Mol Med 2011 17 6 301 312 10.1016/j.molmed.2011.01.011 21411370 5 Fausel C Targeted chronic myeloid leukemia therapy: seeking a cure Am J Health Syst Pharm 2007 64 24 Supplement 15 S9—S15 18056932 6 Brennan CW Verhaak RG McKenna A Campos B Noushmehr H Salama SR The somatic genomic landscape of glioblastoma Cell 2013 155 2 462 477 10.1016/j.cell.2013.09.034 24120142 7 Ostrom QT Bauchet L Davis FG Deltour I Fisher JL Langer CE The epidemiology of glioma in adults: “state of the science” review Neuro Oncol 2014 16 7 896 913 10.1093/neuonc/nou087 24842956 8 Verhaak RG Hoadley KA Purdom E Wang V Qi Y Wilkerson MD Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, EGFR, and NF1 Cancer cell 2010 17 1 98 110 10.1016/j.ccr.2009.12.020 20129251 9 Network TCGAR Comprehensive, integrative genomic analysis of diffuse lower-grade gliomas N Engl J Med 2015 372 2481 2498 10.1056/NEJMoa1402121 10 Sanseau P Agarwal P Barnes MR Pastinen T Richards JB Cardon LR Use of genome-wide association studies for drug repositioning Nat Biotechnol 2012 30 4 317 320 10.1038/nbt.2151 22491277 11 Wang ZY Zhang HY Rational drug repositioning by medical genetics Nat Biotechnol 2013 31 12 1080 1082 10.1038/nbt.2758 24316641 12 Okada Y Wu D Trynka G Raj T Terao C Ikari K Genetics of rheumatoid arthritis contributes to biology and drug discovery Nature 2014 506 7488 376 381 10.1038/nature12873 24390342 13 Chen Y Xu R Network-based gene prediction for Plasmodium falciparum malaria towards genetics-based drug discovery BMC Genomics 2015 16 Suppl 7 S9 10.1186/1471-2164-16-S7-S9 26099491 14 Dudley JT Sirota M Shenoy M Pai RK Roedder S Chiang AP Computational repositioning of the anticonvulsant topiramate for inflammatory bowel disease Sci Transl Med 2011 3 96 96ra76 96ra76 10.1126/scitranslmed.3002648 21849664 15 Jahchan NS Dudley JT Mazur PK Flores N Yang D Palmerton A A drug repositioning approach identifies tricyclic antidepressants as inhibitors of small cell lung cancer and other neuroendocrine tumors Cancer discovery 2013 3 12 1364 1377 10.1158/2159-8290.CD-13-0183 24078773 16 Chen Y Li L Zhang GQ Xu R Phenome-driven disease genetics prediction toward drug discovery Bioinformatics 2015 31 12 i276—i283 10.1093/bioinformatics/btv094 26072493 17 Chen Y Zhang X Zhang Gq Xu R Comparative analysis of a novel disease phenotype network based on clinical manifestations J Biomed Inform 2015 53 113 120 10.1016/j.jbi.2014.09.007 25277758 18 Chen Y Li L Xu R Disease Comorbidity Network Guides the Detection of Molecular Evidence for the Link Between Colorectal Cancer and Obesity AMIA Summits Transl Sci Proc 2015 2015 201 26306270 19 Xu R Wang Q PhenoPredict: A disease phenome-wide drug repositioning approach towards schizophrenia drug discovery J biomed inform 2015 56 348 355 10.1016/j.jbi.2015.06.027 26151312 20 Eppig JT Blake JA Bult CJ Kadin JA Richardson JE Group MGD The Mouse Genome Database (MGD): facilitating mouse as a model for human biology and disease Nucleic Acids Res 2015 43 D1 D726—D736 10.1093/nar/gku967 25348401 21 Hoehndorf R Schofield PN Gkoutos GV PhenomeNET: a whole-phenome approach to disease gene discovery Nucleic Acids Res 2011 39 18 e119—e119 10.1093/nar/gkr538 21737429 22 Hoehndorf R Hiebert T Hardy NW Schofield PN Gkoutos GV Dumontier M Mouse model phenotypes provide information about human drug targets Bioinformatics 2014 30 5 719 725 10.1093/bioinformatics/btt613 24158600 23 Chen Y Ren X Zhang GQ Xu R Ontology-guided organ detection to retrieve web images of disease manifestation: towards the construction of a consumer-based health image library J Am Med Inform Assoc 2013 20 6 1076 1081 10.1136/amiajnl-2012-001380 23792805 24 Robinson PN Köhler S Bauer S Seelow D Horn D Mundlos S The Human Phenotype Ontology: a tool for annotating and analyzing human hereditary disease Am J Human Genet 2008 83 5 610 615 10.1016/j.ajhg.2008.09.017 18950739 25 Xu R Wang Q Large-scale combining signals from both biomedical literature and the FDA Adverse Event Reporting System (FAERS) to improve post-marketing drug safety signal detection BMC bioinformatics 2014 15 1 17 10.1186/1471-2105-15-17 24428898 26 Xu R Wang Q Large-scale extraction of accurate drug-disease treatment pairs from biomedical literature for drug repurposing BMC bioinformatics 2013 14 1 181 10.1186/1471-2105-14-181 23742147 27 Hoehndorf R, Oellrich A, Rebholz-Schuhmann D, Schofield PN, Gkoutos GV. Linking pharmgkb to phenotype studies and animal models of disease for drug repurposing. In: Pac Symp Biocomput. Pacific Symposium on Biocomputing: 2012. p. 388–399. 28 Smith CL Eppig JT The mammalian phenotype ontology: enabling robust annotation and comparative analysis Wiley Interdiscip Rev Syst Biol Med 2009 1 3 390 399 10.1002/wsbm.44 20052305 29 Watkins S, Robel S, Kimbrough IF, Robert SM, Ellis-Davies G, Sontheimer H. Disruption of astrocyte–vascular coupling and the blood–brain barrier by invading glioma cells. Nature communications. 2014:5. doi:10.1038/ncomms5196. 30 Wei J Barr J Kong LY Wang Y Wu A Sharma AK Glioblastoma cancer-initiating cells inhibit T-cell proliferation and effector responses by the signal transducers and activators of transcription 3 pathway Mol Cancer Ther 2010 9 1 67 78 10.1158/1535-7163.MCT-09-0734 20053772 31 Wu A Wei J Kong LY Wang Y Priebe W Qiao W Glioma cancer stem cells induce immunosuppressive macrophages/microglia Neuro-oncology 2010 12 11 1113 1125 10.1093/neuonc/noq082 20667896 32 Dotti G Gottschalk S Savoldo B Brenner MK Design and development of therapies using chimeric antigen receptor-expressing T cells Immunological reviews 2014 257 1 107 126 10.1111/imr.12131 24329793 33 Pyonteck SM Akkari L Schuhmacher AJ Bowman RL Sevenich L Quail DF CSF-1R inhibition alters macrophage polarization and blocks glioma progression Nature medicine 2013 19 10 1264 1272 10.1038/nm.3337 24056773 34 Morosetti R Servidei T Mirabella M Rutella S Mangiola A Maira G The PPAR γ ligands PGJ2 and rosiglitazone show a differential ability to inhibit proliferation and to induce apoptosis and differentiation of human glioblastoma cell lines Int J Oncol 2004 25 2 493 502 15254749 35 Vlachostergios PJ Hatzidaki E Befani CD Liakos P Papandreou CN Bortezomib overcomes MGMT-related resistance of glioblastoma cell lines to temozolomide in a schedule-dependent manner Investig New Drugs 2013 31 5 1169 1181 10.1007/s10637-013-9968-1 23645448 36 Altiok N Ersoz M Koyuturk M Estradiol induces JNK-dependent apoptosis in glioblastoma cells Oncology letters 2011 2 6 1281 1285 22848302 37 Jiang P Mukthavavam R Chao Y Bharati IS Fogal V Pastorino S Novel anti-glioblastoma agents and therapeutic combinations identified from a collection of FDA approved drugs Journal of translational medicine 2014 12 1 1 10.1186/1479-5876-12-1 24387290 38 Turcan S Fabius AW Borodovsky A Pedraza A Brennan C Huse J Efficient induction of differentiation and growth inhibition in IDH1 mutant glioma cells by the DNMT Inhibitor Decitabine Oncotarget 2013 4 10 1729 1736 10.18632/oncotarget.1412 24077826 39 Suenderhauf C Hammann F Huwyler J Computational prediction of blood-brain barrier permeability using decision tree induction Molecules 2012 17 9 10429 10445 10.3390/molecules170910429 22941223 40 Pardridge WM Drug transport across the blood–brain barrier J Cereb Blood Flow Metab 2012 32 11 1959 1972 10.1038/jcbfm.2012.126 22929442 41 Barretina J Caponigro G Stransky N Venkatesan K Margolin AA Kim S The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity Nature 2012 483 7391 603 607 10.1038/nature11003 22460905

PMC005xxxxxx/PMC5001239.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556922290610.1186/s12864-016-2906-9ResearchAn integrative genomics approach for identifying novel functional consequences of PBRM1 truncated mutations in clear cell renal cell carcinoma (ccRCC) Wang Yuanyuan yuanyuan.wang@vanderbilt.edu 1Guo Xingyi xingyi.guo@vanderbilt.edu 12Bray Michael J. michael.j.bray@vanderbilt.edu 3Ding Zhiyong zding@mdanderson.org 4Zhao Zhongming zhongming.zhao@uth.tmc.edu 15671 Department of Biomedical Informatics, Vanderbilt University School of Medicine, Nashville, TN 37203 USA 2 Division of Epidemiology, Department of Medicine, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 3 Vanderbilt Genetics Institute, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 4 Department of Systems Biology, University of Texas MD Anderson Cancer Center, Houston, TX 77030 USA 5 Department of Cancer Biology, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 6 Department of Psychiatry, Vanderbilt University School of Medicine, Nashville, TN 37212 USA 7 Center for Precision Health, School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, TX 77030 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.515© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Clear cell renal cell carcinoma (ccRCC) is the most common type of kidney cancer. Recent large-scale next-generation sequencing analyses reveal that PBRM1 is the second most frequently mutated gene harboring many truncated mutations and has a suspected tumor suppressor role in ccRCC. However, the biological consequences of PBRM1 somatic mutations (e.g., truncated mutations) that drive tumor progression in ccRCC remain unclear. Methods In this study, we proposed an integrative genomics approach to explore the functional consequences of PBRM1 truncated mutations in ccRCC by incorporating somatic mutations, mRNA expression, DNA methylation, and microRNA (miRNA) expression profiles from The Cancer Genome Atlas (TCGA). We performed a systematic analysis to detect the differential molecular features in a total of 11 ccRCC samples harboring PBRM1 truncated mutations from the 33 “pan-negative” ccRCC samples. We excluded the samples that had any of the five high-confidence driver genes (VHL, BAP1, SETD2, PTEN and KDM5C) reported in ccRCC to avoid their possible influence in our results. Results We identified 613 differentially expressed genes (128 up-regulated and 485 down-regulated genes using cutoff |log2FC| < 1 and p < 0.05) in PBRM1 mutated group versus “pan-negative” group. The gene function enrichment analysis revealed that down-regulated genes were significantly enriched in extracellular matrix organization (adjusted p = 2.05 × 10−7), cell adhesion (adjusted p = 2.85 × 10−7), and ion transport (adjusted p = 9.97 × 10−6). Surprisingly, 26 transcriptional factors (TFs) genes including HOXB9, PAX6 and FOXC1 were found to be significantly differentially expressed (23 over expressed TFs and three lower expressed TFs) in PBRM1 mutated group compared with “pan-negative” group. In addition, we identified 1405 differentially methylated CpG sites (targeting 1308 genes, |log2FC| < 1, p < 0.01) and 185 significantly altered microRNAs (|log2FC| < 1, p < 0.05) associated with truncated PBRM1 mutations. Our integrative analysis suggested that methylation and miRNA alterations were likely the downstream events associated with PBRM1 truncation mutations. Conclusions In summary, this study provided some important insights into the understanding of tumorigenesis driven by PBRM1 truncated mutations in ccRCC. The approach may be applied to many driver genes in various cancers. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2906-9) contains supplementary material, which is available to authorized users. Keywords Clear cell renal cell carcinoma (ccRCC)Driver genePBRM1ExpressionMethylationmicroRNAThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Renal cell carcinoma (RCC) is the most common type of kidney cancer (>85 %), which causes ~3 % deaths in men in the United States every year [1, 2]. RCC can be classified into four clinical subtypes including clear cell renal cell carcinoma (ccRCC), papillary RCC (pRCC), chromophobe RCC (chRCC), and renal oncocytoma (RO). Among them, ccRCC is the most common type representing 75–85 % of all RCC cases [2, 3]. Unlike other cancer types that are found to have recurrent mutations in oncogenes [4–7], ccRCC tumors are mainly associated with somatic mutations in tumor suppressor genes such as VHL, PBRM1, BAP1 and SETD2 [8–10]. PBRM1 (Polybromo-1, pb1, encoding BAF180 protein), which maps to 3p21, plays an ATP-dependent chromatin-remodeling role as a subunit of the SWI/SNF (SWItch/Sucrose Non-Fermentable) complex [11–13]. PBRM1 is found to mediate gene regulation of cell growth, migration, proliferation and differentiation in multiple cancer types including kidney, bladder, and breast. Among these cancer types, PBRM1 is one of the most frequently mutated and studied genes in ccRCC than any other cancer types [11, 12, 14–18]. In ccRCC, PBRM1 is the second most frequently mutated gene; it is observed in ~40 % of tumor cases and functions as a driver tumor suppressor gene [3, 9, 10, 13, 18–20]. PBRM1 mutations in ccRCC samples may lead to a dysregulation of several critical cell signaling pathways including actin-based motility by rho, tight junction signaling, axonal guidance signaling and germ cell-sertoli cell junction signaling [21]. Furthermore, mutations in PBRM1 are identified as the root of tumor evolution in a subgroup of ccRCC [22]. While previous studies have focused on the exploration of particular downstream genes and pathways directly regulated by PBRM1 gene, an in-depth integrative analysis on the biological consequences of PBRM1 truncated mutations has not been done yet. Such an analysis is important because tumor suppressor genes play function largely through truncated mutations [23]. Here, we performed an integrative genomics analysis to investigate the biological consequences of truncated PBRM1 mutations in ccRCC. We downloaded multiple -omics data including RNA-Seq, DNA methylation, and microRNA-Seq data of ccRCC samples from The Cancer Genome Atlas (TCGA). We systemically compared molecular features in a total of 11 mutated PBRM1 samples with those in 33 “pan-negative” samples; and those samples were all exclusive of any of the five known ccRCC driver genes (VHL, BAP1, SETD2, PTEN and KDM5C) [13, 15]. The approach allowed us to maximally reduce the noise from the observed molecular signals. We identified a substantial proportion of molecular alterations including changes in gene expression, DNA methylation, and dysregulation of microRNAs (miRNAs) that were significantly associated with truncated PBRM1 mutations, as well as the follow up pathway, co-expression network, and hypothesized mechanism analysis. Results Workflow for defining PBRM1-mutated and “pan-negative” sample groups Somatic mutation profiles for 548 tumor samples in ccRCC, or kidney renal clear cell carcinoma (KIRC), were downloaded from TCGA (data accessed on January 20, 2015). After examining PBRM1 mutations, we separated samples into two groups including 177 mutated PBRM1 samples and 371 non-mutated PBRM1 samples, respectively (Additional file 1) [13]. We further excluded a total of 146 and 262 samples for downstream analysis because they carried mutations in five high-confidence ccRCC driver genes (VHL, BAP1, SETD2, PTEN, and KDM5C) [13, 15]. This process resulted in 31 PBRM1 mutated samples and 109 “pan-negative” samples, respectively (Fig. 1a, Additional file 1). In the next step, we identified the samples with matched RNA-Seq, DNA methylation, and microRNA-Seq data; this resulted in a total of 11 mutated PBRM1 samples and 33 “pan-negative” samples. They were used for the analyses for downstream pre-transcriptional and transcriptional events (Fig. 1a, and b, Additional file 1). Importantly, those 11 samples carried “loss of function” mutations in PBRM1 gene, including five nonsense mutations, three splice sites mutations and three frame shift deletions (Fig. 1b, Additional file 2: Table S1).Fig. 1 Sample filtering workflow used for integrative genomic analyses and differential expression results by comparing 11 PBRM1 mutated and 33 “pan-negative” ccRCC samples. a A sample filtering workflow was used for integrative genomic analyses. First, 548 ccRCC samples were split into PBRM1 mutated group (177 samples) and PBRM1 non-mutated group (371 samples). Five high-confidence ccRCC driver genes (VHL, BAP1, SETD2, PTEN and KDM5C) were excluded in both groups, resulting in 31 PBRM1 mutated samples and 109 “pan-negative” samples. After that, samples that have all DNA methylation, RNA-Seq, and miRNA-Seq data were extracted; resulting in 11 PBRM1 mutated samples and 33 “pan-negative” samples for further in-depth integrative analysis. b Cartoon representation of mutation types and locations in 11 PBRM1 truncation mutated samples. Five nonsense mutations (red diamond), three splice sites mutations (green round), and three frame shift deletions (purple square) were observed in 11 PBRM1 truncated mutation samples. c Volcano plot of significance of gene expression difference between PBRM1 mutated group and “pan-negative” group at gene expression levels. Each dot represents one gene. The x axis shows the gene expression difference by a log transformed fold change while the y axis shows significance by –log10 transformed p-value value obtained from edgeR. A gene is called significantly and differentially expressed if its |log(FC)| > 2 and p-value < 0.05. Red dashed line shows |log(FC)| =2 or p-value = 0.05. d Bar plot of log transfer of fold change in differentially expressed transcriptional factors. 23 transcriptional factors were found to be down-regulated in PBRM1 mutated group while three transcriptional factors were found up-regulated Identification of transcriptional factors from differentially expressed genes associated with PBRM1 truncated mutations We performed a comparative analysis on gene expression profiles to identify the differential expressed genes (DEGs) between the PBRM1 mutated group and “pan-negative” group using edgeR [24]. At a significance threshold of absolute log2 transferred fold change (|log2FC|) > 1 and p < 0.05, a total of 613 DEGs were identified including 128 genes having over expression and 485 genes showing lower expression in PBRM1 mutated samples compared with the “pan-negative” group (Fig. 1c, Additional file 1 and Additional file 3). Of those DEGs, 26 transcription factors were observed, 23 were down-regulated but only three were up-regulated (Fig. 1d). Interestingly, four Antp homeobox family and two forkhead family transcriptional factors (HOXA1, HOXB5, HOXB8, HOXB9, FOXP1, and FOXC1) that are involved in cell development and proliferation [25] were found to be down-regulated in the PBRM1 mutated group versus “pan-negative” group. Additionally, GATA3, a transcription factor that was observed to be down-regulated in PBRM1 mutated group in our study, was previously found to be an important early event and potential regulator that associated with loss of TGFβ receptor expression in ccRCC [26, 27] (Fig. 1d). Gene function enrichment analysis showed that down-regulated genes were significantly enriched in extracellular matrix organization (adjust p = 2.05 × 10−7), cell adhesion (adjust p = 2.82 × 10−7) and ion transport (adjust p = 1.61 × 10−5), while up-regulated genes were significantly enriched in pathway-restricted SMAD protein phosphorylation (adjust p = 3.59 × 10−3) (Fig. 2a and b, Additional file 2: Tables S2 and S3, Additional file 3). We further examined gene expression and methylation, as hypo-methylation is often related to active transcription and gene expression. Our examination the relationship between lower expressed genes and hyper-methylated genes showed that 33 down-regulated genes were hyper-methylated (we abbreviated as hyper-down genes), including BCAT1 associated with cell growth, HOXB9 encoding a cell cycle regulation transcription factor, and PAX6 encoding a cellular development associated transcription factor (Additional file 1) [25].Fig. 2 Functional enrichment results of differentially expressed genes from RNA-Seq of PBRM1 mutation ccRCC samples. a Clustered function and pathway enrichment results of up-regulated genes in PBRM1 mutated group compared with “pan-negative” group, with p-value < 0.01 results shown. Different clusters were shown in different colors. b Clustered function and pathway enrichment results of down-regulated genes in PBRM1 mutated group compared with “pan-negative” group, with p-value < 0.001 results shown. Different clusters were shown in different colors Widespread epigenetic silencing associated with PBRM1 truncated mutations We analyzed genome-scale DNA methylation profiles by comparing β-value changes (measured as β-differences) between mutated PBRM1 group and “pan-negative” group (see Methods). A total of 1308 differentially methylated genes covering 1405 differentially methylated CpG sites were identified using |β-difference| > 0.15 and p < 0.05 cutoff (Fig. 3a and b, Additional file 1, Additional file 4: Figure S1). Among those genes, 1229 hyper-methylated (94 %) and 79 hypo-methylated genes (6 %) were observed in PBRM1 mutated samples compared to the “pan-negative” samples, suggesting that an global gene inactivation may be associated with PBRM1 truncated mutations (Additional file 2: Table S4, Additional file 4: Figure S2). This observation is consistent with the differential gene expression results above (more down-regulated genes than up-regulated genes in PBRM1 group); however, these genes may not be immediately regulated by PBRM1 because truncated mutations in a tumor suppressor gene are expected to result in up-regulation of its immediately regulated gene according to the “loss of function” model. Functional enrichment analyses indicated that those hyper-methylated genes were significantly enriched in multiple processes including generation of neurons (q = 1.20 × 10−5), cell differentiation (q = 1.22 × 10−5), and regulation of catabolic process (q = 4.02 × 10−5), while glomerulus development was observed to be most significant in hypo-methylated genes (q = 3.21 × 10−3) (Additional file 2: Tables S5 and S6, Additional file 4: Figure S2). Interestingly, we found that hyper-methylated CpG sites exhibited a significantly higher proportion residing in several gene regions including promoters and gene bodies than hypo-methylated genes (Additional file 4: Figure S3) [28].Fig. 3 Methylation pattern and miRNA expression pattern in PBRM1 mutated ccRCC a Volcano plot of significance of DNA methylation pattern difference (β-difference) between PBRM1 mutated group and “pan-negative” group. Each dot represents one methylation probes. The x axis shows the difference in β-value (β-difference) while the y axis shows the significance by –log transformed p-value obtained from Samr. A probe is called significantly and differentially expressed if its |β-difference| > 0.15 and p-value < 0.01. The red dashed line shows |β-difference| =0.15 or p-value = 0.01. b Heat map of differential expressed methylation probes between PBRM1 mutated group and “pan-negative” group. c Volcano plot of significance of miRNA expression differences between PBRM1 mutated group and “pan-negative” group. Each dot represents one miRNA. The x axis shows log transformed fold changes of miRNA expression while the y axis shows significance by –log10 transformed p-value obtained from edgeR. A probe is called significantly and differentially expressed if its |log(FC)| > 1 and p-value < 0.05. Red dashed line |log(FC)| =1 or p-value = 0.05. d Bar plot of top ten up-regulated miRNAs and down-regulated miRNAs that revealed in PBRM1 mutated samples compared with “pan-negative” ccRCC samples miRNA dysregulation associated with PBRM1 truncation mutations A total of 185 differentially expressed miRNAs were identified to be associated with PBRM1 truncation mutations using the cutoffs: absolute log2 transferred fold change (|log2FC|) > 1 and p < 0.05. Among them, 87 miRNAs exhibited up-regulation pattern in PBRM1 mutated samples while the remaining 98 miRNAs exhibited down-regulation pattern (Fig. 3c, Additional file 1). The 10 most differentially expressed miRNAs were shown in Fig. 3d. Interestingly, three identified miRNAs (miR-221, miR-222 and miR-16) exhibiting down-regulation patterns in PBRM1 mutated group were consistent with the previous reports by experimental studies [13]. Next, we performed the analysis of those predicted targets genes that may be regulated by these differentially expressed miRNAs. Among the differentially expressed miRNAs, 64 up-regulated miRNAs and 56 down-regulated miRNAs had targets in TarBase [29] or miRTarBase [30] database. We observed 3093 and 3945 target genes for up-regulated miRNAs and down-regulated miRNAs, respectively. Comparisons between miRNA targets and DEGs revealed that 14 miRNA target genes were up-regulated while 129 were down-regulated, in which nine miRNA target genes were hyper-methylated and also down-regulated in PBRM1 mutated group (Fig. 4a, Additional file 1). Functional enrichment analysis revealed that 24 functional terms and pathways, including extracellular matrix organization and extracellular structure organization pathways, were observed in more than one gene set; and these gene sets are differentially expressed genes, differentially methylated genes, and differential expressed miRNA targets genes (Fig. 4b).Fig. 4 Integrative analysis results of function terms and pathway enrichment. a Venn representation of the overlaps among up-regulated genes (DEG-up), down-regulated genes (DEG-down), target genes of up-regulated miRNAs (Up miRNA targets) and target genes of down-regulated miRNAs (Down miRNA targets). b Venn representation of overlaps among function and pathway enrichment results from differential methylated genes (methylation), differential expressed genes (RNA-Seq) and targets genes of differential expressed miRNA (miRNA-Seq) Integrated analysis for PBRM1 truncated mutations in ccRCC To further explore the regulatory mechanisms of the identified genes and miRNAs associated with PBRM1 truncated mutations in ccRCC, we constructed co-expression networks using R software based on mRNA expression results (Fig. 5a and b, detailed information is in Methods). To identify miRNAs that involved in gene co-expression networks, miRNAs target genes that were found co-expressed with other genes and corresponding miRNAs were also included in co-expression networks. Six miRNAs (miR-17-5p, miR-9-5p, miR-16-5p, miR-615-3p, miR-124-3p, and miR-93-5p) were observed in both up-regulated and down-regulated co-expression networks, in which different possible targets were involved. The miRNA target genes including SLC39A14 and EGR2 that are related to ion transport and cell growth were observed in the PBRM1-specific up-regulated co-expression network, suggesting that miRNAs may be involved in ion transport and a cell growth process in PBRM1-driven dysregulation. In the PBRM1 specific down-regulated co-expression network, two down-regulated DNA-binding transcription factors HOXB9 and PAX6 were observed as positively co-expressed with several genes and regulated by miRNAs, suggesting their essential role in PBRM1-related down-regulation (Additional file 1). Similarly, SDCBP2 and PAX6 were found to be positively co-expressed with many genes in the down-regulated co-expression network (Additional file 1), which further verified the association of compound metabolisms and development with PBRM1 truncation mutations [25].Fig. 5 PBRM1 mutation specific, up-regulated and down-regulated co-expression network. Highly co-expressed genes in PBRM1 mutated groups were mapped into a protein-protein interaction network from PINA2, as reference network. 128 up-regulated genes and 33 hyper-down (hyper-methylated and down-regulated) genes were mapped into the reference network, as up-regulated co-expression network (a) and down-regulated core co-expression network. b In down-regulated core co-expression network, only first neighbors of 33 hyper-down genes in down-regulated genes were kept in network. In both networks, only genes with degree above there were kept for better version Collectively, PBRM1 truncated mutations may lead to the pre-transcriptional deregulation at DNA methylation level and the post-transcriptional deregulation at the miRNA expression level. Accordingly, this resulted in widespread hyper-methylation and miRNA expression alteration in ccRCC tumor genomes (Fig. 5). Based on our integrative genomic analysis results, we proposed the possible regulations linked to the PBRM1 truncated mutations in the tumorigenesis of ccRCC (Fig. 6). These functional alterations include both up-regulation and down-regulation of molecules and pathways that are associated with the miRNA and methylation changes in PBRM1-truncated mutation tumor cells.Fig. 6 Hypothesized mechanisms of PBRM1 truncated mutation functions in the tumor genetics of ccRCC. Hyper-methylation and altered miRNAs expression were found associated with PBRM1 truncated mutation in ccRCC. Up-regulated genes and pathways were shown in red (left) while down-regulated genes and pathways were shown in blue (right) Discussion This study highlights the association between PBRM1 truncated mutations and decreased extracellular matrix organization, cell adhesion, ion transport and tissue development. This suggests that PBRM1 plays an important regulatory role in cell-cell crosstalk in the tumorigenesis of ccRCC. In this study, there are more differentially methylated genes (1308 genes) than differential expressed genes (613 genes) in PBRM1 mutated group, suggesting a complicated pre-transcriptional level regulation with DNA methylation involved in PBRM1 mutations. Studying the downstream events of a driver gene has become important now because the scientific community has witnessed large amount of genomic data allowing the sample stratification by driver mutation and also because a driver gene may lead to many critical biological events linking to tumorigenesis or drug treatment [31, 32]. We recently develop approaches to study the downstream events of a specific mutation in a driver gene (BRAFV600E and NRASQ61) in melanoma [4, 5]. To our knowledge, this is the first study to integrate pre-transcriptional and post-transcriptional level data to investigate the main effects of a driver gene (PBRM1) through its truncated mutations in a cancer (ccRCC). Observations in this study are based on 11 PBRM1 mutated and 33 “pan-negative” ccRCC samples, which may have some bias because of the small sample size. However, by an integrative analysis of multiple -omics data, we could still achieve reliable results for further validation. As we did similarly in melanoma [4, 5], the stratification of samples by driver mutation only (cases) and “pan-negative” samples (controls) would likely increase the power because it effectively removed the noise from similar samples with other driver mutations. This is especially important in cancer genomics studies because driver mutations may affect the same or similar signaling pathways (e.g., Ras pathway). Our results suggest that PBRM1 mutations are an important event in the early stage of ccRCC tumor genetics, which paves the way for further PBRM1-related research in ccRCC. To excluded the influences of other driver genes and highlight the effects of PBRM1 in ccRCC, we defined the “pan-negative” ccRCC sample set by excluding samples that contained somatic mutations in any of the five well-known driver genes in ccRCC. Future validation may apply the similar strategies. Our integrative analysis using methylation, gene expression, and miRNA expression is the first to study the PBRM1 truncation mutation specific dysfunction in co-expressed networks. All mutations in 11 PBRM1 mutated samples are truncation mutations, which signify dysfunction state of PBRM1 as a tumor suppressor gene in ccRCC. There are several limitations in this study. First, how our results are related to the influence of PBRM1 on tumor prognosis needs further investigation because previous studies suggest the association between PBRM1 mutations and prognosis of ccRCC is still unclear [13, 22, 33, 34]. In addition, copy number variants of PBRM1 are not considered either since we only focus on the downstream consequences that associate with early somatic mutation events in PBRM1. No validation cohorts of PBRM1 have involved in this study yet because of the limited results available related to PBRM1 at the current stage. We hope more reports will become available from other groups in the near future so that our results may be experimentally validated. Our analysis focuses on the gene level changes that associated with PBRM1 truncated mutation, in which protein level changes were not considered because of the complicated regulation from gene expression level to protein level. PBRM1 is found to be highly mutated in several cancer types. It is most frequently mutated in ccRCC. Loss of function and expression of PBRM1 was less common in non-ccRCC than in ccRCC, suggesting a specific regulatory role of PBRM1 truncation mutations in ccRCC [35]. In breast cancer, PBRM1 is shown to be a core regulator of p21 [14]; however, we could not find a similar pattern in ccRCC. The result suggests that PBRM1 may act differently through its regulation mechanisms in different cancer types. Future studies to dissect the role of PBRM1 in different cancer types would be helpful to better understand the mechanisms of PBRM1 truncation mutations and tumorigenesis. More cancer genomic data is expected from large consortia like the International Cancer Genome Consortium (ICGC). So, a follow up study is needed in future. Conclusion Our study investigated molecular alterations including gene expression, methylation, and miRNA expression that associated with PBRM1 truncation mutations in clear cell renal cell carcinoma. Our analysis results identified 613 differentially expressed genes, 1308 differentially methylated genes and 185 differentially expressed miRNAs between PBRM1 mutated group and “pan-negative” group. Hypothesized mechanisms of PBRM1 mutations in ccRCC were explored based on the integrative analysis results. Our results provide some important insights into the PBRM1 regulation in the tumor development of ccRCC. Methods Summary of ccRCC samples A total of 548 ccRCC (KIRC) samples were downloaded from TCGA. Level 2 results from both BI Mutation Calling and BCM Mutation Calling were utilized to find somatic mutations in all samples. 177 of 548 ccRCC samples (32.3 %) were identified to have PBRM1 mutations and 371 samples (67.7 %) were identified as PBRM1 non-mutated or control samples. To eliminate the influence of other driver genes, five well-known mutation genes (VHL, BAP1, SETD2, PTEN and KDM5C) were suggested as highly potential driver genes of ccRCC based on the somatic mutation results and earlier researches [13]. Samples with somatic mutations of those five genes were excluded from both mutated and non-mutated PBRM1 samples, resulting in 31 PBRM1 mutated samples and 109 “pan-negative” samples (Fig. 1a). Finally, 11 PBRM1 mutated samples and 33 “pan-negative” samples that had DNA methylation, gene expression, and miRNA expression data were utilized for all the analyses in this study. RNA-Seq and miRNA-Seq data pre-processing and differential expression analysis RNA-Seq and miRNA-Seq data were downloaded from IlluminaHiSeq_RNASeqV2 and BCGSC IlluminaHiSeq_miRNASeq platform in TCGA database, respectively. Level 3 data were utilized to find RNA expression and miRNA expression. In each group, genes/miRNAs with no expression were removed, while only genes/miRNAs with counts per million (cpm) >1 in at least two samples were kept for further analysis. edgeR package [24] in R software was used in differential RNA-Seq and miRNA expression analysis. We defined significantly DEGs or differentially expressed miRNAs if they had |log2FC| > 1 and p < 0.05. MiRNA target genes were retrieved from databases TarBase [29] and miRTarBase [30]. Methylation analysis Illumina HumanMethylation450K BeadChip Kit containing 486,428 CpG sites was used to explore DNA methylation profile on the genome scale. Probes targeting the X and Y chromosome, probes containing a single-nucleotide polymorphism (SNP) within five base pairs of CpG site, and probes that had no reference gene location were also removed. In total, 312,777 probes were kept for further analysis. β-values that ranged between 0 and 1 were used to represent the relative methylation level, which was measured as logistic transformation of the ratio of the methylated probe intensity over all methylation probe intensities [36]. β-difference value (differences between β-values) was used to characterize different methylation levels between PBRM1 mutated group and non-mutated “pan-negative” group. All methylation analysis was performed in R/Bioconductor packages [37]. Samr package in R software [37] was used to calculate the significance of each CpG site. Probes with |β-difference| > 0.15 and p < 0.01 were selected as differentially methylated probes, and the gplots package in R software was used to obtain a heatmap of differentially methylated probes. Gene function and pathway enrichment analysis The ClueGO plugin [38] in Cytoscape software [39] was used for gene function and pathway enrichment analysis. Catalogues in GO Biological Process, KEGG, REACTOME and WikiPathways databases that catalogued in ClueGo were applied for the functional enrichment analysis. The Benjamin-Hochberg method [40] was used in the adjustment of p (false discovery rate), and other parameters were retained as default in GlueGO. Gene sets or pathways with adjusted p < 0.05 were retained for further analysis. Transcription factors were annotated based on the TRANSFAC database (downloaded on April 1, 2015) [41]. PBRM1 mutation specific, differentially regulated co-expression network The Pearson correlation coefficient in R software was used to calculate the correlation of each pair on all the 14270 genes that were extracted from RNA-Seq results after excluding low expression genes in PBRM1 mutated group. The top 5 % co-expressed gene pairs were kept as co-expressed and protein-protein interactions from PINA2 [42] were used to find out the relationships between co-expressed genes, which resulted in a PBRM1 mutation specific background network that contains 335,726 gene interaction pairs. 128 up-regulated genes and miRNAs with targets in up-regulated genes were mapped into the reference network, resulting in a PBRM1 mutation specific, up-regulated co-expression network. 485 down-regulated genes and miRNAs with targets in down-regulated genes were mapped into the reference network, resulting in a PBRM1 mutation specific, down-regulated co-expression network. To explore the essential genes associated with PBRM1 mutations, only 33 hyper-down genes and their first neighbors were kept, resulting in PBRM1 mutation specific, down-regulated core co-expression network. The Cytoscape software was used to make the network visualization, with genes that have three or more degrees being shown in Fig. 5a and b. Abbreviations ccRCC, clear-cell renal cell carcinoma; KIRC, kidney renal clear cell carcinoma; TCGA, the cancer genome Atlas. Additional files Additional file 1: Related sample IDs, differential methylated gene IDs, differential expressed gene IDs and altered miRNAs and their target gene IDs. (XLS 493 kb) Additional file 2: Table S1. Detailed information of somatic mutations in 11 PBRM1 mutated ccRCC samples. Table S2. Functional and pathway enrichment results of up-regulated genes. Table S3. Functional and pathway enrichment results of down-regulated genes. Table S4. Alterations of β-value distributions in PBRM1 mutated group and “pan-negative” group. Table S5. Top 20 GO Terms in functional enrichment results of hyper-methylated genes. Table S6. Functional enrichment results of hypo-methylated genes. (DOCX 31 kb) Additional file 3: Detailed information and functional enrichment results of 128 up-regulated genes and 485 down-regulated genes. (XLS 399 kb) Additional file 4: Figure S1. Global methylation density in PBRM1 mutated group and “pan-negative” group. Figure S2. Statics results of altered methylated genes numbers and functions. Figure S3. Percentage of different methylated CpG island region (promoter, 5’UTR, first exon, gene body and 3’UTR) in hyper-methylated and hypo-methylated genes. (DOCX 1378 kb) Acknowledgements The authors would like to thank Drs. Wei Jiang, Feixiong Cheng, Junfei Zhao, Peilin Jia and Ramkrishna Mitra for valuable suggestion and discussion on the data analysis. We thank Vanderbilt Advanced Computing Center for Research & Education (ACCRE) for providing computing resources and support. Declarations Publication of this article was charged from the faculty retention funds to Dr. Zhao from Vanderbilt University. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Funding This work was partially supported by National Institutes of Health (NIH) grants (R01LM011177 and R21CA196508) and Ingram Professorship Funds (to Z.Z.). The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. Availability of data and materials All the data used in this study is from the public sources with the links being included in the publication. Also, additional files, which may be needed to reproduce the results presented in the manuscript, are made available as supplementary material. Authors’ contributions ZZ, XG and YW designed the project, YW and MJB collected the data, YW and XG performed the experiments and analyzed the data, YW, XG, MJB and ZZ drafted the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Siegel RL Miller KD Jemal A Cancer statistics, 2015 CA Cancer J Clin 2015 65 1 5 29 10.3322/caac.21254 25559415 2. Randall JM Millard F Kurzrock R Molecular aberrations, targeted therapy, and renal cell carcinoma: current state-of-the-art Cancer Metastasis Rev 2014 33 4 1109 1124 10.1007/s10555-014-9533-1 25365943 3. Sato Y Yoshizato T Shiraishi Y Maekawa S Okuno Y Kamura T Shimamura T Sato-Otsubo A Nagae G Suzuki H Integrated molecular analysis of clear-cell renal cell carcinoma Nat Genet 2013 45 8 860 867 10.1038/ng.2699 23797736 4. Guo X Xu Y Zhao Z In-depth genomic data analyses revealed complex transcriptional and epigenetic dysregulations of BRAFV600E in melanoma Mol Cancer 2015 14 60 10.1186/s12943-015-0328-y 25890285 5. Jiang W Jia P Hutchinson KE Johnson DB Sosman JA Zhao Z Clinically relevant genes and regulatory pathways associated with NRASQ61 mutations in melanoma through an integrative genomics approach Oncotarget 2015 6 4 2496 2508 10.18632/oncotarget.2954 25537510 6. Eser S Schnieke A Schneider G Saur D Oncogenic KRAS signalling in pancreatic cancer Br J Cancer 2014 111 5 817 822 10.1038/bjc.2014.215 24755884 7. Croce CM Oncogenes and cancer N Engl J Med 2008 358 5 502 511 10.1056/NEJMra072367 18234754 8. Gossage L Eisen T Maher ER VHL, the story of a tumour suppressor gene Nat Rev Cancer 2015 15 1 55 64 10.1038/nrc3844 25533676 9. Brugarolas J Molecular genetics of clear-cell renal cell carcinoma J Clin Oncol Off J Am Soc Clin Oncol 2014 32 18 1968 1976 10.1200/JCO.2012.45.2003 10. Duns G Hofstra RM Sietzema JG Hollema H van Duivenbode I Kuik A Giezen C Jan O Bergsma JJ Bijnen H Targeted exome sequencing in clear cell renal cell carcinoma tumors suggests aberrant chromatin regulation as a crucial step in ccRCC development Hum Mutat 2012 33 7 1059 1062 10.1002/humu.22090 22461374 11. da Costa WH Rezende M Carneiro FC Rocha RM da Cunha IW Carraro DM Guimaraes GC de Cassio ZS Polybromo-1 (PBRM1), a SWI/SNF complex subunit is a prognostic marker in clear cell renal cell carcinoma BJU Int 2014 113 5b E157 E163 10.1111/bju.12426 24053427 12. Wilson BG Roberts CW SWI/SNF nucleosome remodellers and cancer Nat Rev Cancer 2011 11 7 481 492 10.1038/nrc3068 21654818 13. Cancer Genome Atlas Research N Comprehensive molecular characterization of clear cell renal cell carcinoma Nature 2013 499 7456 43 49 10.1038/nature12222 23792563 14. Xia W Nagase S Montia AG Kalachikov SM Keniry M Su T Memeo L Hibshoosh H Parsons R BAF180 is a critical regulator of p21 induction and a tumor suppressor mutated in breast cancer Cancer Res 2008 68 6 1667 1674 10.1158/0008-5472.CAN-07-5276 18339845 15. Huang L Peng Y Zhong G Xie W Dong W Wang B Chen X Gu P He W Wu S PBRM1 suppresses bladder cancer by cyclin B1 induced cell cycle arrest Oncotarget 2015 6 18 16366 16378 10.18632/oncotarget.3879 25978027 16. Ryan RJ Bernstein BE Molecular biology. Genetic events that shape the cancer epigenome Science 2012 336 6088 1513 1514 10.1126/science.1223730 22723401 17. Wang Z Zhai W Richardson JA Olson EN Meneses JJ Firpo MT Kang C Skarnes WC Tjian R Polybromo protein BAF180 functions in mammalian cardiac chamber maturation Genes Dev 2004 18 24 3106 3116 10.1101/gad.1238104 15601824 18. Jiao Y Pawlik TM Anders RA Selaru FM Streppel MM Lucas DJ Niknafs N Guthrie VB Maitra A Argani P Exome sequencing identifies frequent inactivating mutations in BAP1, ARID1A and PBRM1 in intrahepatic cholangiocarcinomas Nat Genet 2013 45 12 1470 1473 10.1038/ng.2813 24185509 19. Varela I Tarpey P Raine K Huang D Ong CK Stephens P Davies H Jones D Lin ML Teague J Exome sequencing identifies frequent mutation of the SWI/SNF complex gene PBRM1 in renal carcinoma Nature 2011 469 7331 539 542 10.1038/nature09639 21248752 20. Benusiglio PR Couve S Gilbert-Dussardier B Deveaux S Le Jeune H Da Costa M Fromont G Memeteau F Yacoub M Coupier I A germline mutation in PBRM1 predisposes to renal cell carcinoma J Med Genet 2015 52 6 426 430 10.1136/jmedgenet-2014-102912 25911086 21. Kapur P Pena-Llopis S Christie A Zhrebker L Pavia-Jimenez A Rathmell WK Xie XJ Brugarolas J Effects on survival of BAP1 and PBRM1 mutations in sporadic clear-cell renal-cell carcinoma: a retrospective analysis with independent validation Lancet Oncol 2013 14 2 159 167 10.1016/S1470-2045(12)70584-3 23333114 22. Gerlinger M Horswell S Larkin J Rowan AJ Salm MP Varela I Fisher R McGranahan N Matthews N Santos CR Genomic architecture and evolution of clear cell renal cell carcinomas defined by multiregion sequencing Nat Genet 2014 46 3 225 233 10.1038/ng.2891 24487277 23. Zhao M Sun J Zhao Z TSGene: a web resource for tumor suppressor genes Nucleic Acids Res 2013 41 Database issue D970 D976 10.1093/nar/gks937 23066107 24. Robinson MD McCarthy DJ Smyth GK edgeR: a Bioconductor package for differential expression analysis of digital gene expression data Bioinformatics 2010 26 1 139 140 10.1093/bioinformatics/btp616 19910308 25. Safran M Dalah I Alexander J Rosen N Iny Stein T Shmoish M Nativ N Bahir I Doniger T Krug H GeneCards Version 3: the human gene integrator Database J Biol Databases Acuration 2010 2010 baq020 26. Cooper SJ Zou H Legrand SN Marlow LA von Roemeling CA Radisky DC Wu KJ Hempel N Margulis V Tun HW Loss of type III transforming growth factor-beta receptor expression is due to methylation silencing of the transcription factor GATA3 in renal cell carcinoma Oncogene 2010 29 20 2905 2915 10.1038/onc.2010.64 20208565 27. Tun HW Marlow LA von Roemeling CA Cooper SJ Kreinest P Wu K Luxon BA Sinha M Anastasiadis PZ Copland JA Pathway signature and cellular differentiation in clear cell renal cell carcinoma PLoS One 2010 5 5 e10696 10.1371/journal.pone.0010696 20502531 28. Esteller M CpG island hypermethylation and tumor suppressor genes: a booming present, a brighter future Oncogene 2002 21 35 5427 5440 10.1038/sj.onc.1205600 12154405 29. Sethupathy P Corda B Hatzigeorgiou AG TarBase: A comprehensive database of experimentally supported animal microRNA targets RNA 2006 12 2 192 197 10.1261/rna.2239606 16373484 30. Hsu SD Lin FM Wu WY Liang C Huang WC Chan WL Tsai WT Chen GZ Lee CJ Chiu CM miRTarBase: a database curates experimentally validated microRNA-target interactions Nucleic Acids Res 2011 39 Database issue D163 D169 10.1093/nar/gkq1107 21071411 31. Xu Y Guo X Sun J Zhao Z Snowball: resampling combined with distance-based regression to discover transcriptional consequences of a driver mutation Bioinformatics 2015 31 1 84 93 10.1093/bioinformatics/btu603 25192743 32. Xia J Jia P Hutchinson KE Dahlman KB Johnson D Sosman J Pao W Zhao Z A meta-analysis of somatic mutations from next generation sequencing of 241 melanomas: a road map for the study of genes with potential clinical relevance Mol Cancer Ther 2014 13 7 1918 1928 10.1158/1535-7163.MCT-13-0804 24755198 33. Nam SJ Lee C Park JH Moon KC Decreased PBRM1 expression predicts unfavorable prognosis in patients with clear cell renal cell carcinoma Urol Oncol 2015 33 8 340 e349 340 e316 10.1016/j.urolonc.2015.01.010 26003625 34. Hakimi AA Ostrovnaya I Reva B Schultz N Chen YB Gonen M Liu H Takeda S Voss MH Tickoo SK Adverse outcomes in clear cell renal cell carcinoma with mutations of 3p21 epigenetic regulators BAP1 and SETD2: a report by MSKCC and the KIRC TCGA research network Clin Cancer Res Off J Am Assoc Cancer Res 2013 19 12 3259 3267 10.1158/1078-0432.CCR-12-3886 35. Ho TH Kapur P Joseph RW Serie DJ Eckel Passow JE Parasramka M Cheville JC Wu KJ Frenkel E Rakheja D Loss of PBRM1 and BAP1 expression is less common in non-clear cell renal cell carcinoma than in clear cell renal cell carcinoma Urol Oncol 2015 33 1 23 e29 14 10.1016/j.urolonc.2014.10.014 25465300 36. Du P Zhang X Huang CC Jafari N Kibbe WA Hou L Lin SM Comparison of Beta-value and M-value methods for quantifying methylation levels by microarray analysis BMC Bioinformatics 2010 11 587 10.1186/1471-2105-11-587 21118553 37. Gentleman RC Carey VJ Bates DM Bolstad B Dettling M Dudoit S Ellis B Gautier L Ge Y Gentry J Bioconductor: open software development for computational biology and bioinformatics Genome Biol 2004 5 10 R80 10.1186/gb-2004-5-10-r80 15461798 38. Bindea G Mlecnik B Hackl H Charoentong P Tosolini M Kirilovsky A Fridman WH Pages F Trajanoski Z Galon J ClueGO: a Cytoscape plug-in to decipher functionally grouped gene ontology and pathway annotation networks Bioinformatics 2009 25 8 1091 1093 10.1093/bioinformatics/btp101 19237447 39. Shannon P Markiel A Ozier O Baliga NS Wang JT Ramage D Amin N Schwikowski B Ideker T Cytoscape: a software environment for integrated models of biomolecular interaction networks Genome Res 2003 13 11 2498 2504 10.1101/gr.1239303 14597658 40. Benjamini Y Hochberg Y Controlling the False Discovery Rate - a Practical and Powerful Approach to Multiple Testing J Roy Stat Soc B Met 1995 57 1 289 300 41. Matys V Fricke E Geffers R Gossling E Haubrock M Hehl R Hornischer K Karas D Kel AE Kel-Margoulis OV TRANSFAC: transcriptional regulation, from patterns to profiles Nucleic Acids Res 2003 31 1 374 378 10.1093/nar/gkg108 12520026 42. Cowley MJ Pinese M Kassahn KS Waddell N Pearson JV Grimmond SM Biankin AV Hautaniemi S Wu J PINA v2.0: mining interactome modules Nucleic Acids Res 2012 40 Database issue D862 D865 10.1093/nar/gkr967 22067443

PMC005xxxxxx/PMC5001240.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556418279110.1186/s12864-016-2791-2ResearchReconstructing directed gene regulatory network by only gene expression data Zhang Lu 1Feng Xi Kang 1Ng Yen Kaow 2Li Shuai Cheng shuaicli@gmail.com 11 Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong 2 Faculty of Information and Communication Technology, University Tunku Abdul Rahman, Kampar, Perak Malaysia 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.430© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Accurately identifying gene regulatory network is an important task in understanding in vivo biological activities. The inference of such networks is often accomplished through the use of gene expression data. Many methods have been developed to evaluate gene expression dependencies between transcription factor and its target genes, and some methods also eliminate transitive interactions. The regulatory (or edge) direction is undetermined if the target gene is also a transcription factor. Some methods predict the regulatory directions in the gene regulatory networks by locating the eQTL single nucleotide polymorphism, or by observing the gene expression changes when knocking out/down the candidate transcript factors; regrettably, these additional data are usually unavailable, especially for the samples deriving from human tissues. Results In this study, we propose the Context Based Dependency Network (CBDN), a method that is able to infer gene regulatory networks with the regulatory directions from gene expression data only. To determine the regulatory direction, CBDN computes the influence of source to target by evaluating the magnitude changes of expression dependencies between the target gene and the others with conditioning on the source gene. CBDN extends the data processing inequality by involving the dependency direction to distinguish between direct and transitive relationship between genes. We also define two types of important regulators which can influence a majority of the genes in the network directly or indirectly. CBDN can detect both of these two types of important regulators by averaging the influence functions of candidate regulator to the other genes. In our experiments with simulated and real data, even with the regulatory direction taken into account, CBDN outperforms the state-of-the-art approaches for inferring gene regulatory network. CBDN identifies the important regulators in the predicted network: 1. TYROBP influences a batch of genes that are related to Alzheimer’s disease; 2. ZNF329 and RB1 significantly regulate those ‘mesenchymal’ gene expression signature genes for brain tumors. Conclusion By merely leveraging gene expression data, CBDN can efficiently infer the existence of gene-gene interactions as well as their regulatory directions. The constructed networks are helpful in the identification of important regulators for complex diseases. Keywords Gene regulatory networkRegulatory directionImportant regulatorsGene expressionIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Understanding of regulatory mechanisms can help us bridge the gap from genotype to phenotype and enlighten us with more insights on the synthesizing effects of different elements in cells. The advent of high-throughput technology provides us an unprecedent opportunity to construct an atlas of these regulatory mechanisms—the gene regulatory network (GRN)—from which one can study important dynamics such as cell proliferation, differentiation, metabolism, and apoptosis. GRN is often inferred from gene expression data, which is available in abundance from high-throughput microarray and RNA-Seq. Many computational approaches have been developed to infer the dependencies between transcription factor (TF) and its target genes from expression data. The intuitive method is to consider a regulatory dependency as the correlation of the expressions of the TF-target pair, computed through a measure such as mutual information (MI), Pearson correlation, etc. However, the correlations captured within the expression data include the effects of intermediary factors; unless taken into account, they will result in the inclusion of transitive edges in the GRN inferred. To overcome this phenomenon, ARACNE [1], an MI-based method, distinguishes between direct and indirect dependencies by applying data processing inequality. It considers the lowest MI value among any triplet of genes as a transitive edge. CLR (context likelihood of relatedness) [2] presents a framework to consider background noise, which naturally accounts for the transitive effects. The method works on the fact that each gene’s MIs or Pearson correlations with other genes follow the Gaussian distribution. This allows the gene-gene correlations to be expressed as Z-scores, thus allowing the comparison of their strengths. Methods based on regression have also been proposed. They incorporate all the genes in a regression model; one as response variable and the others as regressors. Regression-based methods face two difficulties: 1. most of the regressors are not actually independent, hence potentially resulting in erratic regression coefficients for these variables; 2. The model suffers from severe overfitting which necessitates the use of variable selection strategies. A few successful methods have been reported. TIGRESS [3] treats GRN inference as a sparse regression problem and introduce least angle regression in conjunction with stability selection to choose target genes for each TF. GENIE3 [4] performs variables selection based on an ensemble of regression trees (Random Forests or Extra-Trees). Another kinds of methods are proposed to improve the predicted GRNs by introducing additional information. Considering the heterogeneity of gene expression across different conditions, cMonkey [5] is designed as a bi-clustering algorithm to group genes by assessing their co-expressions and the co-occurrence of their putative cis-acting regulatory motifs. The genes grouped in the same cluster are implied to be regulated by the same regulator. Inferelator [6] is developed to infer the GRN for each gene cluster from cMonkey by regression and L1-norm regularization on gene expression or protein abundance. Recently, Chen et al. [7] demonstrated that involving three dimensional chromatin structure with gene expression can improve the GRN reconstruction. While these methods have relatively good performance in reconstructing GRNs, they are unable to infer regulatory directions. There have been many attempts at the inference of regulatory directions by introducing external data. The regulatory direction may be determined from cis expression single nucleotide polymorphism data, called cis-eSNP. The cis-eSNPs are thought of as regulatory anchors by influencing the expression of nearby genes. Zhu et al. [8] developed a method called RIMBANET which reconstructs the GRN through a Bayesian network that integrates both gene expression and cis-eSNPs. The cis-eSNPs determine the regulatory direction with these rules: 1. The genes with cis-eSNPs can be the parent of the genes without cis-eSNPs; 2. The genes without cis-eSNPs cannot be the parent of the genes with cis-eSNPs. These strategies have been very successful [9–11]. However, their applicability is limited by the availability of both SNP and gene expression data. The inference of interaction networks is also actively studied in other fields. Recently, Dror et al. [12] proposed the use of a partial correlation network (PCN) to model the interaction network of a stock market. PCN computes the influence function of stock A to B, by averaging the influence of A in the connectivity between B and other stocks. The influence function is asymmetric, so the node with larger influence to the other one is assigned as parent. Their framework has been extended to other fields such as immune system [13] and semantic networks [14]. Nevertheless, there is an obvious drawback in using PCNs for the inference of GRNs: PCNs only determine whether one node is at a higher level than the other. They do not distinguish between the direct and transitive interactions. Another primary goal of GRN analysis is to identify the important regulator in a network. An important regulator is a gene that influences most of the gene expression signature (GES) genes (e.g. differentially expressed genes) in the network. Carro et al. [15] identified C/EBPβ and STAT3 as important regulators for brain tumor by calculating the overlap between the TF’s targets and ‘mesenchymal’ GES genes based on Fisher’s exact test. TFs were ranked by the number of overlap genes to avoid the influence of the different size of their targets. However, this study only considers the direct influence (Fig. 1(a))of transcription factors to their target genes, the indirect influence (Fig. 1(b)), through transitive genes, are neglected. Zhang et al. [16] developed a method called KDA (key driver analysis) to calculate whether the GES genes are enriched in the targets of regulators by searching h-layer neighborhood dynamically or statically with respect to the given directed network. KDA has been extended to search indirect nodes that are influenced by those regulators, but the influence function is qualitative. It ignores the regulatory strength between regulators and their target genes. On the other hand, because the directed network is quantitatively predicted from gene expression data, we cannot regard the interactions as having the same weight. Fig. 1 Two types of important regulators with directed influence (a) and indirect influence (b) to the other genes in the network In this study, we propose a new method, Context Based Dependency Network (CBDN), which introduces the use of an influence function to decide the edge direction. In addition, we show a directed data processing inequality (DDPI), a property of the influence function, which is used to remove transitive interactions in the partial correlation network. Thus each edge predicted by CBDN is both causal and directed, which can be further applied to infer the important regulators quantitatively. The performance of CBDN is compared to a few well-known algorithms, namely ARACNE, CLR, TIGRESS and GENIE3. In the simulation study, CBDN’s result is comparable to the best result of these methods in each situation and proves its outstanding ability to predict regulatory direction. For a realistic test, we point out the TYROBP-oriented network which is related to Alzheimer’s disease [17]. In this test, CBDN is superior to other methods in inferring both network structure and regulatory direction. CBDN also successfully infers TYROBP as the important regulator by quantitatively considering TYROBP’s influences on the other genes. For another real expression data from the patients affected by human brain tumors, CBDN predicts two potential important regulators ZNF329 and RB1 whose function are associated with brain tumors. All of these results demonstrate the strength of CBDN in the inference of directed GRNs from gene expression data as well as its potential in predicting important regulators. Result CBDN is designed to construct directed regulatory network by only gene expression data. The computation of CBDN consists of three stages: In the first stage, the influence of each gene to the others is calculated to determine the edge direction. This is done through a partial correlation network constructed from the gene expression data; In the second stage, the transitive interactions are removed by DDPI; In the third stage, the important regulators are inferred by ranking the regulators based on their total influences to the GES genes. Determine the edge direction CBDN infers the regulatory interaction through the influence function. The influence function of gene A to gene B (denoted as D(A→B)) is calculated by averaging the Pearson correlation changes between gene B and the other genes in the network, with or without gene A. Notice that the influence function is asymmetric that means D(A→B)≠D(B→A), this phenomenon is adopted to determine the direction of regulatory edge by selecting the genes with larger influence function as the parents. The influence function is derived from partial correlation network, the detailed description can be found in “Methods”. Here we give a schematic example based on the simulated GRN structure in Fig. 2(a) to interpret how CBDN determines the edge directionality. Fig. 2 The simulated gene regulatory network structures and edge directions with 10 (a), 20 (b), 50 (c) and 100 (d) nodes Here, we denote the variable of node i as Xi. For instance, the direction between X1 and X4 is determined by comparing D(X1→X4) and D(X4→X1). X4 merely affects the correlation between X1 and X10 (see Methods), 1 D(X4→X1)=|Corr(X1,X10)|9 Corr(Xi,Xj) denotes the Pearson correlation between the two variables Xi and Xj. the correlation between X1 and other variables are not influenced given X4. When conditioning on X1, the influences are extended to seven variables (X2,X3,X5,X6,X7,X8 and X9), 2 D(X1→X4)=Σi2,3,5,6,7,8,9|Corr(X1,Xi)|9 The upper bound of D(X4→X1) (D(X4→X1)≤1) is smaller than D(X1→X4) (D(X1→X4)≤7) in general, so CBDN concludes that D(X4→X1)≤D(X1→X4). The edge direction is from X1 to X4. Directed data processing inequality The influence function described above only determines whether one gene is the parent or child of another gene; it does not provide the regulatory relationship. As an example, the partial correlation network in Fig. 3 identifies Xi as the parent of Xk, but does not distinguish whether a transitive relation (Xi→Xj→Xk) exists or not (Xi→Xk). Data processing inequality (DPI) can be used to remove transitive interactions by assuming the post-processing cannot increase the mutual information. If Xi, Xj and Xk form a Markov chain, denoted as Xi→Xj→Xk 3 MI(Xi;Xk)≤MI(Xi;Xj) Fig. 3 The diagram for how to remove transitive interactions according to DDPI. We assume X i regulates X j, DDPI is calculated to determine whether X i directly regulate X k (red dashed arrow) or through X j (blue solid arrows) which shows that the mutual information between the genes with transitive interaction cannot be greater than direct interaction. This observation has been used in ARACNE to remove transitive interactions for every triplet of genes. Considering the edge direction and the nature of influence function, we propose a directed data processing inequality to show that the influence of a gene which interacts transitively with its target genes cannot be greater than that of a gene which interacts directly, that is 4 D(Xi→Xk)≤D(Xj→Xk) The mathematical proof is straightforward and presented in Methods. We give an example to show how DDPI distinguishes direct (X2 to X6) and transitive (X1 to X6) interactions in Fig. 2(a). Given X6, all the other variables are divided into two categories: non-descendent of X2 and descendent of X2. The set U denotes non-descendent of X2, including X1,X2,X3,X4,X8,X9,X10. The descendents of X2, presented as V, consists of X5 and X7. For all the variables in U, the influence functions for X1 (D1(X1→X6)) and X2 (D1(X2→X6)) are 5 D1(X1→X6)=Σi3,4,8,9,10|Corr(Xi,X6)|6D1(X2→X6)=Σi1,3,4,8,9,10|Corr(Xi,X6)|6 For all the variables in V, the influence functions for X1 (D2(X1→X6)) and X2 (D2(X2→X6)) are 6 D2(X1→X6)=0D2(X2→X6)=Σi5,7|Corr(Xi,X6)|2 Then we have 7 D1(X2→X6)>D1(X1→X6)D2(X2→X6)>D2(X1→X6)D(X2→X6)=D1(X2→X6)+D2(X2→X6)>D1(X1→X6)+D2(X1→X6)=D(X1→X6) X2 is prefer to be the direct parent of X6 instead of X1 according to Eq. 7. Thus the regulatory structure in Fig. 2(a) should be X2→X6 rather than X1→X6. To account for the influence of noise, we introduce a tolerance parameter τ. A transitive relationship Xj→Xk is removed when D(Xi→Xk)−D(Xj→Xk)>τ. Otherwise, Xi→Xk is removed. Large τ implies much more noise exists in the expression data to influence D(Xi→Xk) and D(Xj→Xk). Determine the important regulators The important regulator identified by CBDN is not required to regulate most of the GES genes. Instead, it should have large influence on them, which guarantees such regulator is always on the top level. In this example, X1 has the largest influence on the other genes in the network and is located on the top level (Methods). Simulation Tree structure simulation In order to explicitly reflect the nature of directed interactions in the gene regulatory network, we simulate a tree structure in which each node has only one parent (except the root) and is merely regulated by its parent (only one arrow from its parent, shown in Fig. 2). In other words, the expression profiles of the descendents are only determined by their parents. The expression profiles for each node were sampled from Gaussian distribution. The joint distribution of the parent and one of its descendent follows bivariate Gaussian distribution with specified covariance and noise. In addition, we mix uniform distributed noise weighted by ωκ to the simulated expression profiles, where “ ω" presents the amount of noise and “ κ” denotes the noise level. We set “ ω" to a constant (ω=3) and change “ κ” from 0 to 2 in the simulations. The expression profiles of 10, 20, 50, 100 nodes are simulated, each of them derived from 1000 samples. The network structure and edge direction are shown in Fig. 2. Infer edge direction Based on the partial correlation network, CBDN can predict the interaction edge direction by only gene expression data. In the simulation, we calculate the proportion of edges that are assigned the directions correctly to evaluate the CBDN’s performance. Our simulation results demonstrate excellent performance of CBDN in predicting edge direction (Fig. 4). There are 83.3 % of the simulations (66/72) where at least 60 % of the edges are correctly assigned directions. As the covariance between these nodes increased, the predicted accuracy increases, and reaches optimality when the covariance is above 0.4. The influence of noise is more severe for the networks with small number of nodes (Fig. 4(a), (b) and (f)). The low covariance makes the performance in large networks declined dramatically (Fig. 4(a) and (b)). Fig. 4 The performance of predicting edge direction by PCN. The increasing covariance spectrum is assigned from 0.1-0.9 in (a)-(f). Different situations such as the amount of mixed noise and the number of nodes are also evaluated in each subfigure Compare CBDN with other methods We evaluate the overall performance of CBDN (including predicted edges and their directions) by comparing it with other famous methods based on a variety of simulated datasets. The true positive rate (TPR) and false positive rate (FPR) are used to plot the receiver operating characteristics (ROC) curve, where TPR=TPTP+FN, FPR=FPFP+FN (TP:true positive, FN:false negative, FP:false positive). The area under ROC curve (AUC) was applied to evaluate the performance of CBDN. We apply the same tests on four state-of-the-art approaches (ARACNE, CLR, GENIE3 and TIGRESS) for comparison. In Table 1, CBDN’s result is the best when no noise exists. Even with small covariance, CBDN correctly revealed the structure and regulatory orientations (Table 1(a)). When noise is introduced, CBDN’s result remains comparable with the best result in each situation. CBDN worked well in general under medium covariance; large or small covariance make it difficult to distinguish direct and transitive interactions, especially when a large amount of noise is introduced (Table 1). However, our comparison is very conservative here, since the performance of CBDN is evaluated by considering both structure and direction, while the other four methods are evaluated only on the inferred structures. Nevertheless, CBDN achieves sufficiently good performance in reconstructing the directed GRNs. We also simulate tree structures with 20, 50,100 nodes, in which CBDN achieves very similar results as the network with 10 nodes simulation (See Tables 2, 3 and 4). Table 1 Simulation result for 10 nodes tree by comparing CBDN with other methods by AUC Covariance ARACNE CLR GENIE3 TIGRESS CBDN (a) Simulation without any noise 0.1 0.8367 0.8009 0.8765 0.8157 0.8750 0.2 1 1 1 0.8410 1 0.4 1 1 1 0.8502 1 0.6 1 1 1 0.8272 1 0.8 1 1 1 1 1 b) Simulation with 1/3 random noise 0.1 0.6304 0.6358 0.5879 0.8107 0.8571 0.2 0.9192 0.9846 0.9884 0.8162 1 0.4 1 1 1 0.8327 1 0.6 1 1 1 0.8557 1 0.8 1 1 0.9985 0.8338 1 (c) Simulation with 2/3 random noise 0.1 0.6904 0.6172 0.6813 0.6241 0.8571 0.2 0.6889 0.8086 0.8480 0.8309 1 0.4 0.9531 0.9599 0.9437 0.8428 1 0.6 1 1 0.9931 0.8424 0.8750 0.8 0.9333 0.9907 0.9807 0.8058 0.8750 Table 2 Simulation result for 20 nodes tree by comparing CBDN with other methods by AUC Covariance ARACNE CLR GENIE3 TIGRESS CBDN (a) Simulation without any noise 0.1 0.8775 0.9332 0.9747 0.7916 0.9306 0.2 0.9961 0.9963 0.9985 0.8034 1 0.4 1 1 1 0.8245 1 0.6 1 1 1 0.7975 1 0.8 1 1 1 0.8015 1 (b) Simulation with 1/3 random noise 0.1 0.7261 0.8864 0.8369 0.7812 0.8269 0.2 0.9166 0.9836 0.9877 0.7940 0.9286 0.4 1 1 1 0.8249 1 0.6 1 1 1 0.7845 1 0.8 1 1 0.9996 0.8387 1 (c) Simulation with 2/3 random noise 0.1 0.6364 0.5499 0.5748 0.5848 0.7500 0.2 0.7797 0.8680 0.9146 0.7735 0.8462 0.4 0.9825 0.9905 0.9988 0.8126 1 0.6 0.9977 1 0.9994 0.8465 0.9000 0.8 0.8804 0.9920 0.9911 0.8146 1 Table 3 Simulation result for 50 nodes tree by comparing CBDN with other methods by AUC Covariance ARACNE CLR GENIE3 TIGRESS CBDN (a) Simulation without any noise 0.1 0.7643 0.8991 0.9225 0.8562 0.8646 0.2 0.9988 0.9997 0.9999 0.8352 0.9762 0.4 1 1 1 0.8448 0.9286 0.6 1 1 1 0.8483 0.9902 0.8 1 1 1 0.8470 1 (b) Simulation with 1/3 random noise 0.1 0.7018 0.7831 0.8208 0.8151 0.7561 0.2 0.9617 0.9936 0.9985 0.8409 0.9748 0.4 1 0.9999 1 0.8738 0.9688 0.6 1 1 1 0.9032 1 0.8 1 0.9994 0.9998 0.9300 1 (c) Simulation with 2/3 random noise 0.1 0.6266 0.5486 0.6385 0.6712 0.7561 0.2 0.6196 0.7746 0.8675 0.8139 0.9625 0.4 0.9893 0.9967 0.9991 0.8673 0.8600 0.6 0.9948 0.9982 0.9982 0.8828 0.9697 0.8 0.9286 0.9943 0.9942 0.9043 1 Table 4 Simulation result for 100 nodes tree by comparing CBDN with other methods by AUC Covariance ARACNE CLR GENIE3 TIGRESS CBDN (a) Simulation without any noise 0.1 0.7445 0.8674 0.9388 0.8394 0.9804 0.2 0.9976 0.9995 1 0.8632 0.9231 0.4 1 1 1 0.8676 0.9792 0.6 1 1 1 0.8872 1 0.8 1 1 0.8426 0.9018 1 (b) Simulation with 1/3 random noise 0.1 0.6929 0.7572 0.8303 0.7765 0.8333 0.2 0.9561 0.9915 0.9992 0.8615 0.9894 0.4 1 1 1 0.8745 0.9875 0.6 1 1 1 0.9071 0.9905 0.8 1 0.9992 1 0.9511 0.9965 (c) Simulation with 2/3 random noise 0.1 0.4874 0.6362 0.6480 0.6547 0.9756 0.2 0.7527 0.8294 0.8867 0.8169 0.9794 0.4 0.9737 0.9871 0.9976 0.8843 0.9938 0.6 0.9990 0.9996 0.9998 0.9237 0.9907 0.8 0.9520 0.9973 0.9979 0.9123 0.9965 Infer important regulators From the network structure for simulation (Fig. 2), the confirmed important regulator is node 1, which is the parent of all the other nodes in the network. Here, we calculate the proportion of those nodes in the network, whose total influence value TIV (Methods) is smaller than the TIV for node 1, to evaluate the inference ability of CBDN. From Fig. 5(a) and (b), we see that smaller networks are in general inferred more accurately, while the effects of noise is unpredictable. For example, for the 50 nodes network in Fig. 5(a), the case with 2/3 noise applied is better predicted than the cases with smaller noise. The important regulator prediction is unstable and unbelievable in the network with weak correlation. The proportion tends to one when the covariance is larger than 0.6 and the nodes in the network are larger than 20 (Fig. 5(d), (e) and (f)), which suggest that the inference is quite reliable for above medium covariance. Fig. 5 The performance of predicting important regulator by DDPI. The increasing covariance spectrum is assigned from 0.1-0.9 in (a)-(f). Different situations such as the amount of mixed noise and the number of nodes are also evaluated in each subfigure Real data For this test, we download the processed expression data from GEO [18] (GSE44770), which is from dorsolateral prefrontal cortex of human brains. The expression data include 230 tissues from the individuals with or without Alzheimer’s disease. The negative expression values are considered missing values because of their low intensities compared to background noise. We impute these missing values with the average positive expression values across all the samples of the same gene. Using gene expression and cis-eSNPs data, Zhang et al. [17] had earlier found the disease-related network to be regulated by TYROBP. In addition, loss-of-function-mutations were recognized in TYROBP in Finnish and Japanese patients affected by presenile dementia with bone cysts [19]. Zhang et al. also overexpressed either full-length or a truncated version of TYROBP in microglia cells from mouse embryonic stem cells to confirm the structure and direction of the regulatory network (Fig. 6). From the TYROBP regulatory network, we choose 47 GES genes, the expressions of which are altered when TYROBP is overexpressed and captured by microarray data, multiple probes designed for the same gene are combined by averaging their expression values. Fig. 6 The network structure for the TYROBP oriented regulatory network for Alzheimer’s disease This dataset is then used as the input for ARACNE, CLR, GENIE3, TIGRESS, and CBDN. The results are compared with the true network structure and edge directions from mouse embryonic stem cells experiment. Figure 7 demonstrates the AUC scores for the five methods. CBDN achieves the best performance, which is 2 % higher than the second best result from GENIE3. To evaluate the capability of CBDN in predicting the regulatory direction and important regulator, we assume all the genes to be potential regulators and ranked them based on TIV. If one gene is assessed as a regulators, other genes are assumed to be GES genes. Figure 8 lists the top 10 genes with the largest TIV, only the values of TYROBP and SLC7A7 are above 8, the validate important regulator TYROBP is ranked at the top. SLC7A7 regulates eleven GES genes (HCLS1, IL10RA, RNASE6, GIMAP2, RGS1, TNFRSF1B, IL18, SFT2D2, KCNE3, LHFPL2 and MAF) and may be another candidate regulator and required to be validated in the future. Fig. 7 The performance of different methods for predicting TYROBP oriented regulatory network Fig. 8 The top ten genes with the largest TIV values for Alzheimer’s disease For another experiment, we download the expression data for brain tumors (GSE19114) and pre-process them as for Alzheimer’s disease. Eventually, we choose 132 ’mesenchymal’ gene expression signature (MGES) genes and 883 TFs from Supplementary Tables 1 and 2 from the original paper [15]. Both MGES genes and TFs are combined together to calculate TIV for each TFs, because we are also required to consider the regulatory relationships between TFs. We are unable to identify the two key regulators (STAT3 and C/EBPβ) described in the original papers from the top TIV ranked TFs (Fig. 9), because we adopt different definitions and inherent characteristics of important regulators. The top two TFs, ZNF329 and RB1 with TIVs exceed 120, are selected as new candidate important regulators. The relationship between ZNF329 and brain tumors is still unclear, but zinc finger protein family has been proved to be associated with brain tumor. Zhao et al. [20] identified ZNF325 as a transcription repressor in MAPK/ERK signaling pathway. Recently, Das et al. [21] made a comprehensive review to clarify the relationship between MAPK/ERK signaling pathway and brain tumors and how can one inhibit this pathway to treat paediatric brain tumors. RB1 gene is the most important cell cycle regulatory genes and the first reported human tumor suppressor gene. It has been identified to be related with a variety of human cancers including brain tumors [22]. Mathivanan et al. found loss of heterozygosity and deregulated expression of RB1 in human brain tumors [23]. Fig. 9 The top ten genes with the largest TIV values for brain tumors Discussion In this paper, we propose a new computational method called Context Based Dependency Network (CBDN), which constructs directed GRNs from only gene expression data. This provides us an opportunity to gain deeper insights from the readily available gene expression data that we have accumulated for years in databases such as GEO. Although gene expression data can reflect the gene-gene interactions in GRN, there are still three limitations that must be addressed. First, the transcription factors prefer to act together as a protein complex rather than individually. The protein complex may be blocked or inactivated, for reasons such as incorrect folding, being restricted in the nucleus or inactivated by the phosphorylation or other modifications, etc., even if its transcribed mRNA has high expression level. Second, the expression of TF and TF binding are time-dependent. Because the time delay exists between transcription and translation, high mRNA expression level does not imply a simultaneous high in protein abundance. Third, even when TFs are bound to their target genes, they may demonstrate different effects because of their three dimensional distances and histone modification. The probes with low florescence signals are impossible to be distinguished from background noise. CBDN treats them as missing values and imputes them by the average value of the other samples. We have further tested other gene expression imputation methods such as the impute package from Bioconductor or BPCA [24], the reconstructed GRN seems stable and consistence. In the future, some noise filtering methods should be incorporated in CBDN such as described in [25, 26]. The performances of CBDN are underestimated for both simulated and real expression data. Except CBDN, the true positive results are defined as the interactions exist in both predictions and ground truth, which neglect the edge direction. For CBDN, both of the interactions and directions are taken into consideration for evaluating its performance. Even though only 2 % of AUC is improved in TYROBPoriented GRN inference, the result is more powerful and useful since they incorporate edge directions. The performance of CBDN is significantly better than other methods in some situations such as Table 1(c) with covariance =0.1, but most of the time CBDN is only slightly better or comparable with other methods. We believe that CBDN will be invaluable to biomedical studies by transcriptome sequencing, where there is a need for the identification of important regulators. Such studies used to be limited by the availability of SNP data to anchor regulatory directions. However, CBDN may be able to infer such important regulators from gene expression data alone, as it identifies the important regulator TYROBP in Alzheimer’s disease. Because CBDN uses new concept of important regulators, it can also help us get new findings which may be neglected by the previous approaches. This paper also contributes to mathematics in the form of an inequality for directed data processing (DDPI) which naturally extends the data processing inequality for mutual information. DDPI is applied to remove transitive interactions in CBDN. In the future CBDN should be extended to predict bi-directed interactions which are quite common in nature. By incorporating external data, we hope to use it to tackle the situations where more than one TFs co-regulate a gene simultaneously. Conclusion The reconstruction of gene regulatory network has been actively researched in the past decade, many methods have been designed to achieve this using only high-throughput gene expression data. However, the edge direction is usually unknown and seems hard to be determined by only gene expression data. Even when the directions can be affirmed, the available approaches is unable to remove transitive interactions from directed network. Here, we propose a novel method CBDN, which can reconstruct direct gene regulatory network by only gene expression data. CBDN first constructs an asymmetric partial correlation network to determine the two influence functions for each pair of genes and determine the edge direction between them. DDPI extends data processing inequality applied in directed network to remove transitive interactions. By aggregating the influence function to all the nodes in the network, the total influence value is calculated to assess whether the node is an important regulator. For both simulation and real data test, CBDN demonstrated superior performance compared to other available methods in reconstructing directed gene regulatory network. It also successfully identified the important regulators for Alzheimer’s disease and brain tumors. Methods Partial correlation network In CBDN, a partial correlation network is first constructed to compute the influence of each node to the others. Interaction directions are resolved by choosing the node with a larger influence as the parent. The influence of gene A to gene B is calculated by averaging the difference between the shortest topological paths of gene B to other genes with or without gene A. We assume the input data is an m×n matrix, E=(ei,j)m×n, where each row i (denoted Ei,∙) represents a sample; that is, one expression value per gene; and each column j (denoted E∙,j) represents the expression values of a gene across all the samples. The partial correlation between Xi and Xk given Xj is calculated as 8 PC(Xi,Xk|Xj)=Corr(Xi,Xk)−Corr(Xi,Xj)Corr(Xk,Xj)[1−Corr(Xi,Xj)2][1−Corr(Xk,Xj)2] Where Corr(Xi,Xj) is the Pearson correlation between two genes Xi and Xj. The influence of gene Xj for the correlation between Xi and Xk (k≠j) is defined as the difference between Corr(Xi,Xj) and PC(Xi,Xk|Xj), 9 d(Xi,Xk|Xj)=Corr(Xi,Xk)−PC(Xi,Xk|Xj) The influence of gene Xj to Xi, D(Xj→Xi) is the average d(Xi,Xk|Xj) across all the gene Xk, 10 D(Xj→Xi)=1n−1Σk≠jn−1|d(Xi,Xk|Xj)| CBDN assumes no two-gene cyclic regulation in the network, so we remove the interaction Xi→Xj if D(Xi→Xj)<D(Xj→Xi), and vice versa. Proof for directed data processing inequality In the directed GRN, we assume three genes (Xi, Xj and Xk) form a Markov chain (Xi→Xj→Xk), the other genes are separated into two categories: non-descendents of Xi (U={Xm⋯Xn}) and descendents of Xi (V={Xp⋯Xa}). For the elements in U, 11 D1(Xi→Xk)=1|U|Σt≠i|U||d(Xk,Xt|Xi)| 12 D1(Xj→Xk)=1|U|Σt≠j|U||d(Xk,Xt|Xj)| Based on Eq. 9, Xk is conditionally independent with the elements in U given Xi or Xj, thus we have PC(Xk,Xt|Xj)=PC(Xk,Xt|Xi)=0, |d(Xk,Xt|Xi)|=|d(Xk,Xt|Xj)|=|Corr(Xk,Xt)|,∀t∈U. For the genes in U, Xi and Xj have the same influence to Xk, D1(Xi→Xk)=D1(Xj→Xk). For the elements in V 13 D2(Xi→Xk)=1|V|Σt≠i|V||d(Xk,Xt|Xi)| 14 D2(Xj→Xk)=1|V|Σt≠j|V||d(Xk,Xt|Xj)| Because Xk is the direct descendent of Xj, Xk is independent with other genes in V given Xj (PC(Xk,Xt|Xj)=0,d(Xk,Xt|Xj)=|Corr(Xk,Xt)|≥0,∀t∈V). The correlations between Xk and the other genes in V do not change when given Xi, so |d(Xk,Xt|Xi)|=0,∀t∈V. We conclude that D2(Xi→Xk)=0 and D2(Xj→Xk)≥0 15 D(Xi→Xk)=D1(Xi→Xk)+D2(Xi→Xk)≤D1(Xj→Xk)+D2(Xj→Xk)=D(Xj→Xk) Determine the important regulators We propose a new method to identify the important regulators in a quantitative way. Assume the genes with gene expression signature (GES) (eg. differentially expressed genes) are Xs1,Xs2,…,Xsn, the total influence value (TIV) of gene Xi is TIV(Xi)=Σt=1nD(Xi→Xst). Regulators are ranked by their TIVs. From IEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA. 9-12 November 2015 Declarations Publication of this article was funded by GRF Grant NO. 9041901 (CityU 118413). This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4. Authors’ contributions SL supervised the work and together with LZ, developed CBDN and procedure of experiment. LZ implemented the CBDN method in matlab. LZ, XK did the experiments on simulation and real data. LZ, SL and YK wrote the manuscript. All authors have read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. ==== Refs References 1 Margolin AA Nemenman I Basso K Wiggins C Stolovitzky G Dalla Favera R Califano A ARACNE: an algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context BMC Bioinformatics 2006 7 Suppl 1 7 10.1186/1471-2105-7-S1-S7 16401345 2 Faith JJ Hayete B Thaden JT Mogno I Wierzbowski J Cottarel G Kasif S Collins JJ Gardner TS Large-scale mapping and validation of Escherichia coli transcriptional regulation from a compendium of expression profiles PLoS Biol 2007 5 1 8 10.1371/journal.pbio.0050008 3 Haury AC Mordelet F Vera-Licona P Vert JP TIGRESS: trustful inference of gene regulation using stability selection BMC Syst Biol 2012 6 145 10.1186/1752-0509-6-145 23173819 4 Huynh-Thu VA Irrthum A Wehenkel L Geurts P Inferring regulatory networks from expression data using tree-based methods PLoS ONE 2010 5 e12776 1 10 5 Reiss DJ Baliga NS Bonneau R Integrated biclustering of heterogeneous genome-wide datasets for the inference of global regulatory networks BMC Bioinformatics 2006 7 280 10.1186/1471-2105-7-280 16749936 6 Bonneau R Reiss DJ Shannon P Facciotti M Hood L Baliga NS Thorsson V The Inferelator: an algorithm for learning parsimonious regulatory networks from systems-biology data sets de novo Genome Biol 2006 7 5 36 10.1186/gb-2006-7-5-r36 7 Chen H Chen J Muir LA Ronquist S Meixner W Ljungman M Ried T Smale S Rajapakse I Functional organization of the human 4D Nucleome Proc Natl Acad Sci U S A 2015 112 26 8002 7 10.1073/pnas.1505822112 26080430 8 Zhu J Lum PY Lamb J GuhaThakurta D Edwards SW Thieringer R Berger JP Wu MS Thompson J Sachs AB Schadt EE An integrative genomics approach to the reconstruction of gene networks in segregating populations Cytogenet Genome Res 2004 105 2-4 363 74 10.1159/000078209 15237224 9 Zhu J Zhang B Smith EN Drees B Brem RB Kruglyak L Bumgarner RE Schadt EE Integrating large-scale functional genomic data to dissect the complexity of yeast regulatory networks Nat Genet 2008 40 7 854 61 10.1038/ng.167 18552845 10 Yang X Zhang B Molony C Chudin E Hao K Zhu J Gaedigk A Suver C Zhong H Leeder JS Guengerich FP Strom SC Schuetz E Rushmore TH Ulrich RG Slatter JG Schadt EE Kasarskis A Lum PY Systematic genetic and genomic analysis of cytochrome P450 enzyme activities in human liver Genome Res 2010 20 8 1020 36 10.1101/gr.103341.109 20538623 11 Wang IM Zhang B Yang X Zhu J Stepaniants S Zhang C Meng Q Peters M He Y Ni C Slipetz D Crackower MA Houshyar H Tan CM Asante-Appiah E O’Neill G Luo MJ Thieringer R Yuan J Chiu CS Lum PY Lamb J Boie Y Wilkinson HA Schadt EE Dai H Roberts C Systems analysis of eleven rodent disease models reveals an inflammatome signature and key drivers Mol Syst Biol 2012 8 594 10.1038/msb.2012.24 22806142 12 Kenett DY Tumminello M Madi A Gur-Gershgoren G Mantegna RN Ben-Jacob E Dominating clasp of the financial sector revealed by partial correlation analysis of the stock market PLoS ONE 2010 5 12 15032 10.1371/journal.pone.0015032 13 Madi A Kenett DY Bransburg-Zabary S Merbl Y Quintana FJ Boccaletti S Tauber AI Cohen IR Ben-Jacob E Analyses of antigen dependency networks unveil immune system reorganization between birth and adulthood Chaos 2011 21 1 016109 10.1063/1.3543800 21456851 14 Kenett YN Kenett DY Ben-Jacob E Faust M Global and local features of semantic networks: evidence from the Hebrew mental lexicon PLoS ONE 2011 6 8 23912 10.1371/journal.pone.0023912 15 Carro MS Lim WK Alvarez MJ Bollo RJ Zhao X Snyder EY Sulman EP Anne SL Doetsch F Colman H Lasorella A Aldape K Califano A Iavarone A The transcriptional network for mesenchymal transformation of brain tumours Nature 2010 463 7279 318 25 10.1038/nature08712 20032975 16 Zhang B, Zhu J. Identification of key causal regulators in gene networks. In: Proceedings of the World Congress on Engineering 2013. London, U.K: Newswood Limited. 2013 (2). 17 Zhang B Gaiteri C Bodea LG Wang Z McElwee J Podtelezhnikov AA Zhang C Xie T Tran L Dobrin R Fluder E Clurman B Melquist S Narayanan M Suver C Shah H Mahajan M Gillis T Mysore J MacDonald ME Lamb JR Bennett DA Molony C Stone DJ Gudnason V Myers AJ Schadt EE Neumann H Zhu J Emilsson V Integrated systems approach identifies genetic nodes and networks in late-onset Alzheimer’s disease Cell 2013 153 3 707 20 10.1016/j.cell.2013.03.030 23622250 18 Barrett T Wilhite SE Ledoux P Evangelista C Kim IF Tomashevsky M Marshall KA Phillippy KH Sherman PM Holko M Yefanov A Lee H Zhang N Robertson CL Serova N Davis S Soboleva A NCBI GEO: archive for functional genomics data sets–update Nucleic Acids Res 2013 41 Database issue 991 5 10.1093/nar/gks1193 19 Paloneva J Kestila M Wu J Salminen A Bohling T Ruotsalainen V Hakola P Bakker AB Phillips JH Pekkarinen P Lanier LL Timonen T Peltonen L Loss-of-function mutations in TYROBP (DAP12) result in a presenile dementia with bone cysts Nat Genet 2000 25 3 357 61 10.1038/77153 10888890 20 Zhao Y Zhou L Liu B Deng Y Wang Y Wang Y Huang W Yuan W Wang Z Zhu C Liu M Wu X Li Y ZNF325, a novel human zinc finger protein with a RBaK-like RB-binding domain, inhibits AP-1- and SRE-mediated transcriptional activity Biochem Biophys Res Commun 2006 346 4 1191 1199 10.1016/j.bbrc.2006.06.031 16806083 21 Paramita Das DJG Treating Pediatric Brain Tumors by Inhibiting the RAS-ERK Signaling Pathway: A Review J Pediatric Oncol 2015 3 1 1 7 10.14205/2309-3021.2015.03.01.1 22 Mathivanan J K R Gope ML Gope R Possible role of the tumor suppressor gene retinoblastoma (rb1) in human brain tumor development Ann Neurosci 2007 14 3 72 82 10.5214/ans.0972.7531.2007.140304 23 Mathivanan J Rohini K Gope ML Anandh B Gope R Altered structure and deregulated expression of the tumor suppressor gene retinoblastoma (RB1) in human brain tumors Mol Cell Biochem 2007 302 1-2 67 77 10.1007/s11010-007-9428-3 17318405 24 Oba S Sato MA Takemasa I Monden M Matsubara K Ishii S A Bayesian missing value estimation method for gene expression profile data Bioinformatics 2003 19 16 2088 96 10.1093/bioinformatics/btg287 14594714 25 Aris VM Cody MJ Cheng J Dermody JJ Soteropoulos P Recce M Tolias PP Noise filtering and nonparametric analysis of microarray data underscores discriminating markers of oral, prostate, lung, ovarian and breast cancer BMC Bioinformatics 2004 5 185 10.1186/1471-2105-5-185 15569388 26 Zeisel A Amir A Kostler WJ Domany E Intensity dependent estimation of noise in microarrays improves detection of differentially expressed genes BMC Bioinformatics 2010 11 400 10.1186/1471-2105-11-400 20663218

PMC005xxxxxx/PMC5001241.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27535360279710.1186/s12864-016-2797-9ResearchThe biorepository portal toolkit: an honest brokered, modular service oriented software tool set for biospecimen-driven translational research Felmeister Alex S. felmeistera@email.chop.edu 12Masino Aaron J. 1Rivera Tyler J. 1Resnick Adam C. 13Pennington Jeffrey W. 11 Department of Biomedical and Health Informatics, The Children’s Hospital of Philadelphia, 3401 Civic Center Blvd, Philadelphia, PA USA 2 College of Computing and Informatics, Drexel University, 3141 Chestnut Street, Philadelphia, PA USA 3 Department of Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, 3400 Civic Center Boulevard, Building 421, Philadelphia, PA USA 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.434© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background High throughput molecular sequencing and increased biospecimen variety have introduced significant informatics challenges for research biorepository infrastructures. We applied a modular system integration approach to develop an operational biorepository management system. This method enables aggregation of the clinical, specimen and genomic data collected for biorepository resources. Methods We introduce an electronic Honest Broker (eHB) and Biorepository Portal (BRP) open source project that, in tandem, allow for data integration while protecting patient privacy. This modular approach allows data and specimens to be associated with a biorepository subject at any time point asynchronously. This lowers the bar to develop new research projects based on scientific merit without institutional review for a proposal. Results By facilitating the automated de-identification of specimen and associated clinical and genomic data we create a future proofed specimen set that can withstand new workflows and be connected to new associated information over time. Thus facilitating collaborative advanced genomic and tissue research. Conclusions As of Janurary of 2016 there are 23 unique protocols/patient cohorts being managed in the Biorepository Portal (BRP). There are over 4000 unique subject records in the electronic honest broker (eHB), over 30,000 specimens accessioned and 8 institutions participating in various biobanking activities using this tool kit. We specifically set out to build rich annotation of biospecimens with longitudinal clinical data; BRP/REDCap integration for multi-institutional repositories; EMR integration; further annotated specimens with genomic data specific to a domain; build application hooks for experiments at the specimen level integrated with analytic software; while protecting privacy per the Office of Civil Rights (OCR) and HIPAA. Keywords Biorepository researchTranslational bioinformaticsPrecision medicineHonest brokerCancer genomicsData integrationData representationOpen sourcePatient health information protectionPatient privacyIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Current research is yielding rapid advances in personalized, precision medicine through targeted therapies based on an individual’s genome, genomic biomarkers, and cell biology across adult and pediatric translational research [1, 2]. This type of research has become increasingly dependent on the collection of large cohorts of high quality human biospecimens that are paired with clinical annotations [3]. While biospecimen-driven research is widely practiced, it is often limited in scope because it requires time-consuming manual processes such as retrospective annotation, cohort identification and institutional human subjects research oversight [4]. Consequently, many academic medical centers are creating large institutional biospecimen resources that can be leveraged by numerous investigators [5]. There is a trend towards these resources becoming indispensable in academic medical centers [3, 6]. Biorepository data is typically captured in longitudinal, asynchronous workflows that pose software design and data integration challenges [7]. An optimal system must provide de-identified, granular and longitudinal data to researchers while also enabling data collection workflows that require patient identification [8]. The required data often resides in separate systems such as a Laboratory Information Management System (LIMS), Research Data Capture tools, the Electronic Health Record (EHR), genomic data stores and high performance computing clusters [9, 10]. Integrative solutions are necessary at the point of collection and at information and specimen retrieval. The data must be curated to ensure it is persisted in an understandable representation for researchers in a specific medical domain [11, 12]. As biorepository resources include more clinical information and grow in scale, there are more opportunities for protected health information (PHI) to be injected into the process [13, 14]. Therefore, a central component to this toolkit is an informatics-based approach to honest brokering [15]. We build on methods described in Dhir et al. and Boyd et al. that describe specific implementations of software to aid in the honest brokering between various types of clinical data collection and de-identified biorepositories [14, 16]. We take a slightly different approach by creating non-user facing software service similar to Boyd, et al. for the honest broker as one of many components of a toolkit of connected operational biorespository informatics resources. We remove the human component completely from the de-identification and re-identification of research records in connected research systems. In this paper, we address the creation of a robust biorepository management platform that enables association of a physical biospecimens, clinical diagnoses, and patient, genomic and research. This platform utilizes a modular software architecture developed at the Children’s Hospital of Philadelphia (CHOP) in partnership with the Children’s Brain Tumor Tissue Consortium (CBTTC) [17]. The platform was developed in the specific context of distributed biorepository and biobank studies in biological tissue and genomic research. In this manuscript, we describe the requirements, challenges in architectural design and implementation to create integrated data resources in biospecimen-driven translational research. We designed and utilized an open source, modular software toolkit that supports biorepository operations and de-identified secondary usage. We created an operational and scientific resource that protects subject privacy, allows for variable specimen and data management workflows and flexible resource queries. When new systems and workflows are introduced, the toolkit allows for flexible introduction of new data types, systems and operational workflows spanning specimen, clinical, imaging and genomic data. Our platform allows for extensible software and data resources for biospecimen-driven translational research. Cancer focus Cancer is a main focus of current precision medicine initiatives. This is reflected in politics, the media, public funding and medical research community priorities [18]. We are in an age of increased use of web technologies that allow us to reach new levels of productivity and connectivity in business, finance, government and entertainment [19]. The time has come for us to use these same techniques to unravel the complexities of cancer [20]. New breakthroughs are helping us use our own immune systems to target an increasing list of common cancers [21]. Unfortunetely, time is not on the side of children suffering from rare brain tumors. Recent research and government population health programs identify over 120 types of pediatric brain cancers [22]. To make matters more complex, the origins of brain tumors in kids is widely unknown [22–25]. Pediatric cancer patients are treated for cancers with adult-based therapies and there is a lack of investment from pharmaceutical companies in the specific diseases affecting children [26]. It is essential to create biologic- and data-centric resources to find pathways and molecularly describe disease seen in research similar to Bastianos et al. and Parsons et al. where developments, respectively, uncover a molecular pathway in Craniopharyngiomas and a comprehensive molecular description of the common childhood brain tumor, Medulloblastoma [27, 28]. Though molecularly based research has become common with the availability of high throughput technologies, further progress is needed in infrastructure, specifically in cancer research, that allows for complete clinical annotation of specimen and genomic data from consented subjects ([29], p. 549). Initial research of rare tumors at CHOP brought biorepository data sharing, management, and annotation issues to the forefront. A need for enhanced capabilities was particularly evident in two proposed studies targeted by our initial software system design. These tumor biorepository studies originally used a human honest broker to manage the de-identification and re-identification of records to exclude protected patient information (a/k/a patient health information) from the research LIMS. The process began with manual data intake by a data manager on hardcopy REDCap case report forms (CRFs) [30]. The CRFs contained the patient identifiers: Medical Record Number (MRN), First Name, Last Name and Date of Birth. The CRF was physically delivered to a human honest broker that would create a new electronic REDCap record with a research identifier. The hardcopy CRF was then returned to the data manager with the research identifier attached and the patient identifiers removed. The data manager then abstracted the hand written CRFs to the associated REDCap project record. Each longitudinal data collection event required manual re-identification by the human honest broker. This process became unsustainable as biospecimen and clinical data collection increased and molecular experimentation associated with records began. It was also difficult to complete the CRF in a single patient encounter due to variations in encounter length and frequency. This experience clearly illustrated the need for a scalable solution that would abide by NCI Best Practices for Biospecimen Resources [15]. The Biorepository Portal Toolkit (BRP Toolkit) project was subsequently developed to support biorepository development at institutional scale. Methods Modular approach We took a modular and entity-driven integrated systems approach to facilitate variable specimen acquisition and data collection events. The primary entity is the patient enrolled as a research subject on the study. The subject entity is created in the electronic honest broker (eHB) and assoicated with a master patient index (MPI) and the subject’s instutional origin. Each external research record associated with the subject record, in this case the data management tool, REDCap, is not limited to a one-to-one relation of subject-to-REDCap record. The subject entity can be assoicated with many projects, forms and records in a 1-to-many entity relationaship [31]. We, in tandem, built a research portal, dubbed the Biorepository Portal (BRP). The BRP can access subject records in the electronic Honest Broker (eHB) and subsequent external research records through token-based authorization from that client system. The BRP reproduces the REDCap electronic Case Report Forms (eCRFs) based on records stored in the eHB with a custom REDCap client utilizing the REDCap application programing interface (API), in real-time. This produces a complete form for that subject at time of access. It displays the subject information and identifiers at the top of the screen at all times during form data entry and while shifting from form to form. A research coordinator or data manager can enter any temporal and longitudinal research data based on their protocol subject list at anytime or in any order (i.e., asynchronously) while maintaining the continuous de-identification and re-identification of research data automatically. The CHOP Biorepository Core Facility utilizes ThermoFisher Nautilus as its LIMS. As part of our method, we also built a client to this LIMS that allows for association of an arbitrary number of specimen records in the LIMS with the corresponding subject record. In this way specimens can also be collected longitudinally over time. Data and specimen coordinators have the ability to associate sets of specimens with a subject or event and annotate that specimen on the fly in one system. For downstream integration, we use the same eHB software service to perform our Extract Transform Load (ETL) processes that are tailored to each project. The result is a regularly updated non-human subject research database that allows for seamless queries spanning research and clinical data sets. We allow collaborators to access specific sets of data via the data exploration tool, Harvest [32], customized for each project. The phenotypic data associated with specimen records can be integrated with direct clinical data from the EHR with appropriate institutional permissions. The modular approach allows us to integrate genomic visualization tools at the specimen level where applicable. For cancer genomics specifically, we utilize the CBioPortal [33, 34], an open source tool to visualize mutation and gene expression data from The Cancer Genome Atlas (TCGA). Web service oriented architecture The integration of tools is accomplished by taking advantage of modern web technologies. Our methods are rooted in web service oriented architecture (SOA). This pattern pervades the current generation of computing and web technology and is rapidly expanding through virtual resources accessed via network resources (i.e., cloud computing) [35]. We created a plug-and-play experience working with multiple tools in a web environment. We employ REST (Representational State Transfer) API architecture over HTTP protocols providing uniform channels for applications integrated into the tool kit [36]. In this section, we describe each specific technology in the stack, each with its own set of RESTful end-points that allow us to guide users through multiple tools as they interact with biorepository resources. Though SOAs are widely used in the field of biomedical informatics to build complex application tool chains, they are not transparent as to their usage and sometimes have very poor adoption because of their complexity [37]. To address this, we mask technical complexity with end-user tools that are familiar to research coordinators, specimen coordinators and data managers. There is wide recognition that there is no compliant Health Level Seven (HL7)-type interoperability standards when using SOA’s [9]. The operational applications and scientific applications we implemented utilize an honest broker software service, a well-known method of dealing with de-identification in biorepository research that maintain the HIPAA compliance in downstream systems [14, 38–41]. Figure 1 illustrates a high-level architecture description along with external integration points for key scientific and clinically relevant data points for data entry and reuse. This figure is split into three parts, the last of which describes the fully realized modular approach integrating both clinical, genomic and research specimen resources.Fig. 1 Three workflow iterations. Iterations of the introduction of the modular tool kit from human intervention, electronic intervention to fully realized modular approach Electronic Honest Broker (eHB) software service The concept of an honest broker has been implemented in other academic medical center environments to protect privacy when integrating research data ([14, 42], pp. 56–107). Central to our solution is the eHB, a web-based software service with end-to-end encryption that maintains an index of subjects linked to their associated research records. The initial studies/projects targeted with this solution began as a manual paper-based process of considerable complexity, and are now an “informatics tool” [15]. The index in the eHB uniquely identifies each patient through a combination of organizational association (e.g., The Children’s Hospital of Philadelphia) and a unique organization provided identifier (e.g., Medical Record Number). The eHB associates each subject record with trusted external systems and known record identifiers in those systems. For annotated biorepository studies, the eHB maintains associations to de-identified records in REDCap and the LIMS. Following our SOA approach, the eHB makes data and system functionality available via a REST web service. This allows the addition of new data management application clients in a way that is system- and programming language-agnostic. To control access, the eHB uses token-based authorization, and encrypts its data both at rest and in transit, relying on client-side keys to decrypt the payload received from the API. For known applications, the eHB provides subject data with few restrictions. Client applications determine the context of what information is appropriate to display to a user, thereby enabling flexibility to meet different workflow and protocol requirements. The eHB has a limited web-facing user interface that allows for the administration of access tokens and users. It can be managed through a comprehensive set of create, read, update, delete (CRUD) operations exposed by the REST API. Client applications, described in subsequent sections, determine the context of the request. The client application requests resources of the eHB service via a URL endpoint secured using transport layer security (TLS) and, with appropriate keys and credentials, can read and write data to and from the eHB service. The eHB REST service handles authorization of the application, encrypts data and formats a response. The actual database behind the eHB service stores only encrypted information and would be unreadable if accessed. This type of encryption decouples the identifiers and encrypts any and all information going into the eHB database and can be considered “privacy-by-design” by selectively sending and granting access to information based on context and only storing the minimum set of information needed to stitch together a record for data management or data query [16, 43]. The architectural design of the eHB, illustrated in Fig. 2, utilizes web request type architecture to be a completely independent component of the tool-kit. The eHB model is similar to prior research in clinical informatics and the EHR. Architecturally, the health record must have the element of being future-proof. There is an assurance of openness and portability through standards, flexibility and scalability, semantic interoperability and acceptance from the domain experts. In Blobel et al., the authors discuss the fundamentals of future-proof health systems describing an “atomic component” [44]. We apply this notion by creating an “atomic component” that must be guaranteed utilizing the eHB to associate the subject on the study (as the atomic component) with their biospecimens, phenotypic and genomic data.Fig. 2 The interaction between the eHB and client applications. This takes place over HTTP through CRUD (GET, POST, PUT, DELETE) operations. Client applications (i.e., the BRP) determine the context of the information sent and received to the eHB service. Within the eHB, data is encrypted at rest, and in transit. A query of data can only take place through the eHB software service The Biorepository Portal (BRP) The eHB service discussed in the previous section is exposed solely to external client systems through APIs and lacks a user interface. The user-facing BRP component provides an integrated view of multiple data management software applications. We implement a web-based application that uses HTTP protocols to communicate with the eHB and external systems. The BRP allows research staff to work with subject identifiers, subject research data and associate specimen records. The BRP presents the honest brokered data by integrating custom clients using external system APIs to integrate in real-time, the research data and patient identifiers and records. The BRP allows access via authentication utilizing institutional identity management systems that comply with all network access guidelines for hospital clinical systems then limits users to specific cohorts, institutions and data resources accessed through the portal. Figure 3 describes the layers of user access to protocols and data sources representing applications connected to the tool in a sudo-ER diagram.Fig. 3 Sudo ER diagram of complex relationships between users and data sources. Users are granted access to subject records based on protocols. Each protocol groups a list of subjects that only authorized users can manage and are part are assigned to. Within the walls of a protocol, there are groups of subjects that allow for users to differentiate between institutional or site-based cohorts (i.e., organizations). Each unique subject enrollment creates a new record for that subject in the eHB. Within a protocol there are specific data entry points set up as protocol data sources (e.g., a REDCap case report form project or LIMS specimen project), and each user has credentials to each data source. Data sources are set up as clients to each external system represented in the BRP. In our case, we have created data sources for two systems, REDCap and ThermoFisher Nautilus LIMS The BRP provides context to operational tools through clients to tools used in the context of biospecimen management. These clients are configured in the BRP as data sources. Display options for each client can be specified by a configuration encoded in JSON [45]. JSON formatted configuration files enable the portal to manage the display and access of REDCap forms and events in the data management processes based on workflows for capturing longitudinal data. The BRP not only relates an honest brokered subject record to external research systems, but also relates records to each other across systems. The following describes two implementations of external data sources; in this case a laboratory system and a data management system. In this case, records in these systems are linked in the BRP to provide further context around the connection of external research system (i.e., specimen collection record to case report form). Figure 4 is an example of the record listing in the BRP.Fig. 4 Screenshot example. A screenshot from the BRP displaying identifiable data entry along the de-identified entry points for client systems’ data sources Data sources REDCap client The REDCap client in the BRP makes a request for the metadata and data from a subject record stored in the eHB and recreates the form requested utilizing the REDCap API. The REDCap form client shows the patient Medical Record Number (MRN), last name, first name and date of birth along with each customized research data capture form. After the form record is saved, the BRP utilizes the eHB software service to either create a new record or modify an existing record in the REDCap project. The REDCap project record identifier is hashed and randomly generated without use of derived patient information. eHB identifiers are generated utilizing the application client key, in combination with a salted hash value which is guaranteed to be unique [46]. Creating a research identifier not derived from a direct patient identifier is required when using patient data for research [13]. Research identifiers are created by the connected research system randomly and are not derived from any patient information. Adding a layer that removes the subject entity from the REDCap projects associated with a study allows for REDCap to facilitate user access to projects, form building, data logging, and managing a study data dictionary [30]. The ability to supplement an entire REDCap project(s) or form(s) as specimen annotations is accomplished by associating another REDCap project with a portal project and, in turn, a subject. Our approach includes the ability to have variable numbers of projects and nested project records per patient. There are many variations of studies that use a variable number of REDCap projects/forms and project records depending on the need of the domain. For example, a BRP protocol can capture demographics one time in one REDCap project while collecting many diagnosis-type forms with longitudinal events in another project that allows for multiple records per subject. The eHB mediates and stores the links between the subject entity and their project records. Conversely, we allow for the tools to maintain separate cohorts of identified subjects where the data are stored in the same REDCap project. This is particularly important for studies in which multiple institutions are participating in sending data and specimens to one data and specimen-coordinating center (DCC/SCC). The link to REDCap records depends on the domain and temporal requirements of a biorepository study. LIMS client A key requirement in our choice of a LIMS was that it implements an API that exposed the ability to create new specimen records, print labels and update specimen records with tags from external systems. The LIMS assigns a unique identifier to a specimen collection event, and this identifier is associated with the subject entity in the eHB by user input via the BRP. The BRP has a custom client that allows specimen coordinators with the proper credentials to associate pre-labeled specimen accession kits with the subject entity. Specimen collection kits with proper collection tubes and labels are created prior to subject enrollment in the CHOP Biorepository Core Facility. The specimen coordinator then scans the barcoded label on the kit through a LIMS client in the BRP to associate the kit with the patient. Any downstream laboratory work, for example; receiving, processing, analysis and storage are performed directly in the LIMS. The laboratory technicians processing and receiving specimen kits do not see patient identifiers, only the LIMS assigned identifier. This facilitates the longitudinal capture of multiple specimen collection events associated with one subject. Access Extract transform load As Goble, et al., describe service oriented technology mechanisms “…[o]nce plumbed, the data have to be massaged and cleaned to make them fit together or conform to new schema” ([36], p. 689). We meet this requirement with ETL scripts written to integrate the disparate and de-identified data together for scientific use by researchers. This ETL process acts as another application with client access to the eHB. The first part of the ETL script uses application key-driven access to obtain a list of subject entities on specific protocols linked their respective research identifiers in data management applications. This linked list is used throughout the ETL process to join together and integrate data from disparate research systems and perform further de-identification where required by a study protocol. The ETL process produces data in a relational domain model suitable for researcher query. The ETL process is also where we integrate systems and data that are not part of the data entry in the BRP. If allowed by the protocol, the ETL process can query the eHB for patient records and pull clinical data from the health record and move it to the non-human subject biorepository database. Researcher query tool and non-human subject research data resource Integrating data sets through an ETL process is the starting point to fully realize the research potential of an integrated biobank. Researchers need to be able to discover available data and formulate queries for case definition and cohort creation. To enable this, we created a query tool that allows researchers to get quick answers to questions using available data without involving the Institutional Review Board (IRB) because all data is de-identified in the query tool [11]. This tool is implemented using the open-source Harvest framework [32]. Harvest gives the informatics team the ability to customize the application where necessary, but also have an out-of-box query tool for domains of relevant data. An example of a Harvest instance where users can search the multi-institutional biospecimen and annotation data of the CBTTC is shown in Fig. 5. This figure shows point-and-click access to multi-dimensional and disparate data in one place.Fig. 5 Integrated researcher view. Integrated non-human subject data and specimen query tool built on the Harvest platform. This interface allows for query across multiple systems in one place. The platform can be customized to allow for only certain data elements to be utilized that depend on domain and researcher requirements. Exposed in this example are elements captures in CRFs and the LIMS. There are also links out to genomics analysis tools Results and discussions Usage results The toolkit supports multiple biospecimen-driven research studies. In these studies, the accessioning of specimens and related data has grown and completely changed in scale and volume. Figure 6 is a graph showing specimen accessioning 2011 to January, 2016 for the CBTTC. In 2012 the toolkit was adopted and the graph illustrates the change in accession events happening after the toolkit adoption with a varied pace of specimen accession. This means data is available to researchers in near real-time as accessioning happens compared to pre 2012 where data was only available as the operations center could sort and enter data. Table 1 is a list of projects that utilize the toolkit with summary counts of subjects, specimens and scientific data points as of the end of 2015.Fig. 6 Usage. A graph of specimen accessioning for the CBTTC which has accessioned over 10,000 specimens to date (end of January 2016). The blue shaded area represents the sum of all specimens collected by the project Table 1 Project list this table is a list of select high-volume projects utilizing the modular tool-kit architecture described in this manuscript that particularly integrate multiple research resources through the toolkit Project Subjects (rounded) Specimens accessioned (rounded) Data integration points The Children’s Brain Tumor Tissue Consortium 1400 9200 -Case Report Forms -LIMS -Cancer Genomics University of Pennsylvania-CHOP Neurosurgery Tissue Collaborative 1500 14,000 -Case Report Forms -LIMS IBD Center Biorepository Studies 177 1100 -Case Report Forms -LIMS -Electronic Medical Record Center for Childhood Cancer Research 470 2200 -Case Report Forms -LIMS -Electronic Medical Record PennCHOP Microbiome Center 60 1500 -Case Report Forms -LIMS -Electronic Medical Record -Genomic Analysis Pipelines The table includes the project, a rounded number of individual subjects consented to the study, specimens and the integrated data points for the project. These numbers are rounded and are from the end of year, 2015 Operational and scientific complexity Our methods allow for a significant amount of complexity in scientific data management. With a modern, web-service oriented architecture we abstract the subject entity from multiple related project data in supporting research systems, providing increased flexibility and adaptability over comparable monolithic systems. The tools facilitate the longitudinal collection of clinical phenotypic data over an arbitrary time period. We also promote asynchronous and variable collection of specimens. Protection of subject privacy The architecture of the eHB helps to protect patient privacy by limiting exposure of patient identifiers to research staff. Identifiers are only available to study operations staff responsible for associating data and specimens with subjects in the context of specific IRB protocols. The toolkit shares implementation of security and access controls with the connected downstream research system (e.g., access and logging features of REDCap). Similarly, authorization to connect LIMS specimen records to a patient record are configured with LIMS named users, and the toolkit relies on the security protocols exposed through the LIMS API. Researchers only access secondary data resources that result from downstream ETL and are never directly connected to the honest broker component of the tool kit. The ability to query the eHB allows the team to build domain specific data repositories and a web query tool limited risk of exposing patient information. Evolving data types and workflows Downstream data management systems and workflows can change project to project or as the science changes within a project. We allow all downstream applications do what they were designed to do without any impact on the underlying architecture of the toolkit. The toolkit is resilient to change, and the service-oriented architecture provides hooks to incorporate additional systems, reflecting the certainty of changing requirements, data and workflows. Discussion Annotation of biospecimens with longitudinal clinical data was the initial impetus of creating this toolkit, but it has gone much further. The BRP and REDCap integration have enabled multiple multi-institutional biorepository studies, both international and national. As in Harris et al., we enable concurrent multi-institutional projects to be controlled by staff through a single common interface similar to other new research informatics initiatives [30]. Exploration of HL7 for Oncology and honest brokering is ongoing, but we have not been able to implement these medical informatics type interchange languages for this purpose [47]. Additionally, web services were not built to be HIPAA compliant [9], but we find that designing modular tools that have privacy built in by their specific usage to be sufficient in research. Further, we find the modern web architecture to allow for the integration of the favored web tools scientists are using for their research to be a novel approach. The Children’s Brain Tumor Tissue Consortium In Fig. 5, the data and specimen query tool of the Children’s Brain and Tumor Tissue Consortium, is some of the results of ambitious biorepository collaboration between six children’s hospitals. Informatics staff utilize this set of integrated tools to support the data and specimen-coordinating To date, the toolkit has been central to a number of grant-funded projects focused on next generation sequencing of biospecimens, and the identification of causative mutations in tumors and data integration heavy microbiologic research (See Table 1). A researcher can quickly determine study feasibility by reviewing available specimens and data in childhood brain cancer research and formulate new studies focused on finding pathways in rare cancer similar to Brastianos et al. and Parsons et al. [27, 28]. Current efforts are focused on integration of highly dimensional genomic data sets generated by such studies. EMR integration Integrating or retrieving data from the Electronic Medical Record (EMR) is a growing need for users of the toolkit. Projects originating at CHOP have been able to take advantage of the eHB to re-identify patients for the purpose of extracting clinical data from the CHOP EMR. We are able to, with appropriate IRB permission, obtain and de-identify clinical data directly from the EMR and incorporate these data as annotations on specimens and genomic data. This has proven useful in studies that require observational data such as medication at time of specimen collection, particularly in high frequency specimen collection in microbiologic research. Genomic data integration Continued success of biorepository-driven research relies on tight integration of other scientific data points. A primary example of this is the phenotypic and genomic information relation. In the case of the cancer tissue biorepository, the specific cancer genomic alterations are specifically important to annotate specimens. We found that allowing a researcher to query specific somatic mutations from comparative analysis of sequencing the germ-line and the tumor and find the physical specimens with that mutation to be a powerful feature. We accomplish the query of specific tumor mutations by integrating the CBioPortal into our toolkit. The CBioPortal an open source tool developed to view cancer-based genomic analysis data originally developed at Memorial Sloan Kettering Cancer Center (MSKCC) is an open access resource for exploration of multi-dimensional cancer genomics data sets [33]. Figure 7 shows how the integration of the CBioPortal with the clinical and specimen based query tool with a mix of traditional ETL and web based endpoints. Figure 8 is a set of screen shots of this cancer genomic data integration. This integration point opens up a capability unavailable in previous tissue focused biorepositories. Users can search for and view mutations called on specimens in the local biorepository data repository to pull a physical specimen for biological research. Conversely, the user can search for pathways of interest across published studies and discover if those pathways are shown in local specimens. The user can be moved back to the specimen request tool with the context of specimens from a genomic mutation-based query. Similarly as we continue to add scientific platforms to the scientific side of the toolkit pertinent to domains we are integrating with multiple commercial data management service providers. We will extend this concept by pushing the platform further to incorporate workflows dedicated to the classification of pathology diagnoses and collaborative discussion about pathology calls derived from high-resolution pathology slide scans. All of this would not be possible without a modular approach.Fig. 7 CbioPortal – harvest integration architecture. The integrated query tool allows for a back and forth search between genes of interest and visualization in CBioPortal, and the tool for phenotype and specimen requests. We perform this integration in a similar fashion to the other tools in the tool-kit. We utilize a combination of constructing web endpoints and traditional ETL. The cancer genomics integration starts with a scripted pull of mutation data via a secure database connection utilizing elements of CBioPortal’s relational data structure to store this large set of data. Specimens known to the repository are loaded into the CBioPortal by the bioinformatics team with known specimen identifiers from the LIMS. This creates a natural link between any granular genomic data, the specimen and ultimately the subject. URLs are constructed in the query platform that allow for researchers to move from clinical and specimen driven queries directly to CBioPortal to visualize mutation data of interest Fig. 8 Specimen to cancer genomics. Screen shot of integration of the Harvest-based data and specimen query tool with a the CBioPortal for cancer genomic visualization of a specific case Conclusions As of January of 2016 there are 23 unique protocols/patient cohorts being managed in the Biorepository Portal (BRP). There are over 4000 unique subject records in the electronic honest broker (eHB), over 30,000 specimens accessioned and 8 institutions participating in various biobanking activities using this tool kit. We specifically set out to build rich annotation of biospecimens with longitudinal clinical data; BRP/REDCap integration for multi-institutional repositories; EMR integration; further annotated specimens with genomic data specific to a domain; build application hooks for experiments at the specimen level integrated with analytic software; while protecting privacy per the Office of Civil Rights (OCR) and HIPAA. To meet this challenge, we created an open source modular software toolkit that automates many manual components of biorepository data workflows while also protecting patient privacy. Conversely the modular solutions allow for novel integration points for translational research spanning clinical and genomic data. We believe this work advances the state of the art within the biomedical domain by moving towards modern technologies and architectures to provide translational research resources. Availability of software Project Name: The Biorepository Portal Toolkit Project home page: http://www.brptoolkit.com Code Repositories:Electronic Honest Broker Service: https://github.com/chop-dbhi/ehb-service Electronic Honest Broker Client: https://github.com/chop-dbhi/ehb-client Biorepository Portal: https://github.com/chop-dbhi/biorepo-portal Data Sources: https://github.com/chop-dbhi/ehb-datasources Operating Systems: Linux Other Requirements: Docker (optional but recommended) License: We believe in open source software and open-source our work. The Biorepository Toolkit and all integration work with non-proprietary systems is licensed under BSD 2-clause License. Restrictions to use by non-academics: None. Ethics approval: No ethics approval was required for this work. A demonstration of this software is available at http://www.brptoolkit.com. This website also contains documentation, webinars, descriptions and pointers to code repositories in context. Software discussed in this paper is available on the CHOP Department of Biomedical and Health Informatics github repository at https://github.com/chop-dbhi. A specific implementation of the toolkit is available through the Children’s Brain Tumor Tissue Consortium at http://www.cbttc.org. Declarations The Children’s Hospital of Philadelphia Clinical and Translational Science Award (CTSA) through the Clinical and Translational Research Center (CTRC), U54 KL2RR024132, funded this project and publication. The project is also made possible by the generous philanthropic effort of the Children’s Brain Tumor Foundation (CBTF). This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4. Authors’ contributions ASF was the main author of this manuscript and envisioned this research method along with AJM who is also a main contibutor to this manuscript, envisioned and developed large components of the software. TJR is the current lead developer for the project and a contributor to the manuscript. ACR is the scientific domain contributor, and JWP is a main contributor to design of the project and author of this manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Ethics, consent to participate and consent to publish Not applicable for this research. ==== Refs References 1. Brisson a R Matsui D Rieder MJ Fraser DD Translational research in pediatrics: tissue sampling and biobanking Pediatrics 2012 129 1 153 62 10.1542/peds.2011-0134 22144705 2. Colman E Golden J Roberts M Egan A Weaver J Pharm D Rosebraugh C The path to personalized medicine N Engl J Med 2010 363 4 2012 4 3. Hirtzlin I Dubreuil C Préaubert N Duchier J Jansen B Simon J Lobato De Faria P Perez-Lezaun A Visser B Williams GD Cambon-Thomsen A An empirical survey on biobanking of human genetic material and data in six EU countries Eur J Hum Genet 2003 11 6 475 88 10.1038/sj.ejhg.5201007 12774042 4. Altekruse SF, Rosenfeld GE, Carrick DM, Pressman EJ, Schully SD, Mechanic LE, Cronin KA, Hernandez BY, Lynch CF, Cozen W, Khoury MJ, Penberthy LT. SEER cancer registry biospecimen research: yesterday and tomorrow. Cancer Epidemiol Biomarkers Prev. 2014;23(12):2681–7. 5. Compton C Getting to personalized cancer medicine: taking out the garbage Cancer 2007 110 8 1641 3 10.1002/cncr.22966 17763370 6. Végvári A Welinder C Lindberg H Fehniger TE Marko-Varga G Biobank resources for future patient care: developments, principles and concepts J Clin Bioinf 2011 1 1 24 10.1186/2043-9113-1-24 7. Vaught J Rogers J Carolin T Compton C Biobankonomics: developing a sustainable business model approach for the formation of a human tissue biobank J Natl Cancer Inst Monogr 2011 2011 42 24 31 10.1093/jncimonographs/lgr009 21672892 8. Kamm L Bogdanov D Laur S Vilo J A new way to protect privacy in large-scale genome-wide association studies Bioinformatics 2013 29 7 886 93 10.1093/bioinformatics/btt066 23413435 9. Wang X Liu L Fackenthal J Cummings S Olopade OI Hope K Silverstein JC Olopade OI Translational integrity and continuity: personalized biomedical data integration J Biomed Inform 2009 42 1 100 12 10.1016/j.jbi.2008.08.002 18760382 10. Wang X Olopade O Foster I Personalized biomedical data integration Pers Biomed Data Integr Biomed Eng Trends Electron Commun Software Mr Anthony Laskovski 2011 1 100 12 11. Dhir R, Patel AA, Winters S, Bisceglia M, Swanson D, Aamodt R, et al. A multidisciplinary approach to honest broker services for tissue banks and clinical data: A pragmatic and practical model. Cancer. 2008;113:1705–15. 12. The Health Insurance Portability and Accountability Act of 1996 (HIPAA) P.L. No. 104-191, 110 Stat. 1938 (1996). 13. The Office for Civil Rights (OCR), B. Malin. Guidance regarding methods for de-identification of protected Health Information in Accordance with the Health Insurance Portability and Accountability Act (HIPAA) Privacy Rule. Heal Inf Priv. 2012;1–32. http://www.hhs.gov/hipaa/for-professionals/privacy/special-topics/de-identification/index.html. 14. Dhir R Patel A a Winters S Bisceglia M Swanson D Aamodt R Becich MJ A multidisciplinary approach to honest broker services for tissue banks and clinical data: a pragmatic and practical model Cancer 2008 113 7 1705 15 10.1002/cncr.23768 18683217 15. National Cancer Institute NCI best practices for biospecimen resources 2010 16. Boyd AD, Saxman PR, Hunscher DA, Smith KA, Morris TD, Kaston M, Bayoff F, Rogers B, Hayes P, Rajeev N, Kline-Rogers E, Eagle K, Clauw D, Greden JF, Green LA, Athey BD. The University of Michigan Honest Broker: a Web-based service for clinical and translational research and practice. J Am Med Inform Assoc. 2009;16(6):784–91. doi:10.1197/jamia.M2985. 17. The Children’s Brain Tumor Tissue Consortium. [Online]. Available: www.cbttc.org. Accessed 25 June 2015. 18. Collins FS Varmus H A new initiative on precision medicine N Engl J Med 2015 372 9 2012 4 10.1056/NEJMp1500523 19. Rainie L Fox S Duggan M The Web at 25 in the U.S Pew Res Cent 2014 1 1 5 20. S. Mukherjee. The emperor of all maladies: a biography of cancer. New York, NY: Scribner. 2011. 21. June CH, Maus MV, Plesa G, Johnson L a, Zhao Y, Levine BL, Grupp A, Porter DL. Engineered T cells for cancer therapy. Cancer Immunol Immunother. 2014;63(9):969–75. 22. Ostrom QT Gittleman H Chen Y Wolinsky Y Barnholtz-Sloan J CBTRUS statistical report: primary brain and central nervous system tumors diagnosed in the United States in 2005–2009 Neuro Oncol 2012 14 suppl 5 1 21 10.1093/neuonc/nos218 23. American Brain Tumor Association Brain tumor information 2014 24. Cancer in Children & Adolescents. 2014. 25. Baldwin RT Preston-Martin S Epidemiology of brain tumors in childhood - a review Toxicol Appl Pharmacol 2004 199 2 118 31 10.1016/j.taap.2003.12.029 15313584 26. Boklan JJ Little patients, losing patience: pediatric cancer drug development Mol Cancer Ther 2006 5 8 1905 8 10.1158/1535-7163.MCT-06-0179 16928809 27. Brastianos PK, Taylor-Weiner A, Manley PE, Jones RT, Dias-Santagata D, Thorner AR, Lawrence MS, Rodriguez FJ, Bernardo LA, Schubert L, Sunkavalli A, Shillingford N, Calicchio ML, Lidov W, Taha H, Martinez-Lage M, Santi M, Storm PB, Lee JYK, Palmer JN, Adappa ND, Scott RM, Dunn IF, Laws ER, Stewart C, Ligon KL, Hoang MP , Van Hummelen P, Hahn WC, Louis DN, Resnick AC, Kieran MW, Getz G, Santagata S. Exome sequencing identifies BRAF mutations in papillary craniopharyngiomas. Nat Genet. 2014;46(2):161–5. 28. Parsons DW, Li M, Zhang X, Jones S, Leary RJ, Lin JC-H, Boca SM, Carter H, Samayoa J, Bettegowda C, Gallia GL, Jallo Gl, Binder ZA, Nikolsky Y, Hartigan J, Smith DR, Gerhard DS, Fults DW, VandenBerg S, Berger MS, Marie SKN, Shinjo SMO, Clara C, Phillips PC, Minturn JE, Biegel JA, Judkins AR, Resnick AC, Storm PB, Curran T, He Y, Rasheed BA, Friedman HS, Keir ST, McLendon R, Northcott PA, Taylor MD, Burger PC, Riggins GJ, Karchin R, Parmigiani G, Bigner DD, Yan H, Papadopoulos N, Vogelstein B, Kinzler KW, Velculescu VE. The genetic landscape of the childhood cancer medulloblastoma. Science. 2011;331:435–9. 29. Chin L, Hahn WC, Getz G, Meyerson M. Making sense of cancer genomic data. Genes & Development. 2011;25:534–555. 30. Harris P a Taylor R Thielke R Payne J Gonzalez N Conde JG Research electronic data capture (REDCap)--a metadata-driven methodology and workflow process for providing translational research informatics support J Biomed Inform 2009 42 2 377 81 10.1016/j.jbi.2008.08.010 18929686 31. Chen PP-S The entity-relationship model---toward a unified view of data ACM Trans Database Syst 1976 1 1 9 36 10.1145/320434.320440 32. Pennington JW Ruth B Italia MJ Miller J Wrazien S Loutrel JG Crenshaw EB White PS Harvest: an open platform for developing web-based biomedical data discovery and reporting applications J Am Med Inform Assoc 2013 21 2 379 83 10.1136/amiajnl-2013-001825 24131510 33. Cerami E Gao J Dogrusoz U Gross BE Sumer SO Aksoy BA Jacobsen A Byrne CJ Heuer ML Larsson E Antipin Y Reva B Goldberg AP Sander C Schultz N The cBio Cancer Genomics Portal: an open platform for exploring multidimensional cancer genomics data Cancer Discov 2012 2 5 401 4 10.1158/2159-8290.CD-12-0095 22588877 34. Tomczak K Czerwińska P Wiznerowicz M The Cancer Genome Atlas (TCGA): an immeasurable source of knowledge Contemp Oncol (Poznań, Poland) 2015 19 1A A68 77 35. Alkhatib H Faraboschi P Frachtenberg E Kasahara H Lange D Laplante P Merchant A Burgess A IEEE CS 2022 Report 2014 36. Goble C Stevens R State of the nation in data integration for bioinformatics J Biomed Inform 2008 41 5 687 93 10.1016/j.jbi.2008.01.008 18358788 37. Goble C Stevens R Hull D Wolstencroft K Lopez R Data curation + process curation=data integration + science Brief Bioinform 2008 9 6 506 10.1093/bib/bbn034 19060304 38. Boyd AD, Hosner C, Hunscher D a, Athey BD, Clauw DJ, Green LA. An ‘Honest Broker’ mechanism to maintain privacy for patient care and academic medical research. Int J Med Inform. 2006;76(5–6):407–11. 39. McConnell P, Dash RC, Chilukuri R, Pietrobon R, Johnson K, Annechiarico R, Cuticchia AJ. The cancer translational research informatics platform. BMC Med Inform Decis Mak. 2008;8:60. 40. Kho AN, Cashy JP, Jackson KL, Pah AR, Goel S, Rn Boehnke J, Humphries JE, Kominers SD, Hota BN, Sims SA, Malin BA, French DD, Walunas TL, Meltzer DO, Kaleba EO, Jones RC, Galanter WL. Design and implementation of a privacy preserving electronic health record linkage tool in Chicago. J Am Med Inform Assoc. 2015;22(5):1072–1080. doi:10.1093/jamia/ocv038. 41. Manion FJ Robbins RJ Weems W a Crowley RS Security and privacy requirements for a multi-institutional cancer research data grid: an interview-based study BMC Med Inform Decis Mak 2009 9 31 10.1186/1472-6947-9-31 19527521 42. Hutton JJ Pediatric biomedical informatics 2012 Dordrecht Springer Netherlands 43. Boyd AD, Hunscher DA, Kramer AJ, Hosner C, Saxman P, Athey BD, Greden JF, Clauw DC. The ‘Honest Broker’ method of integrating interdisciplinary research data. AMIA Annu Symp Proc. 2005;902. 44. Blobel B Nordberg R Davis JM Pharow P Modelling privilege management and access control Int J Med Inform 2006 75 597 623 10.1016/j.ijmedinf.2005.08.010 16199198 45. Ecma International ECMA-404: The JSON data interchange format 2013 1 46. Morris R Thompson K Password security: a case history Commun ACM 1979 22 11 594 7 10.1145/359168.359172 47. Warner JL Maddux SE Hughes KS Krauss JC Yu PP Shulman LN Mayer DK Hogarth M Shafarman M Fiscalini AS Esserman L Alschuler L Koromia GA Gonzaga Z Ambinder EP Development, implementation, and initial evaluation of a foundational open interoperability standard for oncology treatment planning and summarization J Am Med Informatics Assoc 2015 22 3 577 86

PMC005xxxxxx/PMC5001242.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556597291310.1186/s12864-016-2913-xResearchA hierarchical model for clustering m6A methylation peaks in MeRIP-seq data Cui Xiaodong xdchoi@gmail.com 1Meng Jia jia.meng@hotmail.com 2Zhang Shaowu zhangsw@nwpu.edu.cn 3Rao Manjeet K. raom@uthscsa.edu 5Chen Yidong cheny8@uthscsa.edu 45Huang Yufei yufei.huang@utsa.edu 141 Department of Electrical and Computer Engineering, University of Texas, San Antonio, TX 78249 USA 2 Department of Biological Science, Xi’an Jiaotong-liverpool University, Suzhou, 215123 China 3 College of Automation, Northwestern Polytechnical University, Xi’an, 710072 China 4 Depeartment of Epidemiology and Biostatistics, University of Texas Health Science Center, San Antonio, TX 78229 USA 5 Greehey Children’s Cancer Research Institute, University of Texas Health Science Center, San Antonio, TX 78229 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.520© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background The recent advent of the state-of-art high throughput sequencing technology, known as Methylated RNA Immunoprecipitation combined with RNA sequencing (MeRIP-seq) revolutionizes the area of mRNA epigenetics and enables the biologists and biomedical researchers to have a global view of N6-Methyladenosine (m6A) on transcriptome. Yet there is a significant need for new computation tools for processing and analysing MeRIP-Seq data to gain a further insight into the function and m6A mRNA methylation. Results We developed a novel algorithm and an open source R package (http://compgenomics.utsa.edu/metcluster) for uncovering the potential types of m6A methylation by clustering the degree of m6A methylation peaks in MeRIP-Seq data. This algorithm utilizes a hierarchical graphical model to model the reads account variance and the underlying clusters of the methylation peaks. Rigorous statistical inference is performed to estimate the model parameter and detect the number of clusters. MeTCluster is evaluated on both simulated and real MeRIP-seq datasets and the results demonstrate its high accuracy in characterizing the clusters of methylation peaks. Our algorithm was applied to two different sets of real MeRIP-seq datasets and reveals a novel pattern that methylation peaks with less peak enrichment tend to clustered in the 5′ end of both in both mRNAs and lncRNAs, whereas those with higher peak enrichment are more likely to be distributed in CDS and towards the 3′end of mRNAs and lncRNAs. This result might suggest that m6A’s functions could be location specific. Conclusions In this paper, a novel hierarchical graphical model based algorithm was developed for clustering the enrichment of methylation peaks in MeRIP-seq data. MeTCluster is written in R and is publicly available. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2913-x) contains supplementary material, which is available to authorized users. The International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background N6-methyl-adenosine (m6A) is the most abundant modification among 100 types of identified RNA modifications in eukaryotic mRNA/lncRNA [1, 2]. Even though m6A was found existing in mammalian mRNAs in as early as 1970s [3], its biological relevance remains unclear due to the difficulties in identifying global m6A sites in mRNA [4]. In 2013, the m6A demethylase Fat mass and obesity associated protein (FTO) was first discovered [5], to be able to reverse the m6A modification in mRNA and it revived our interests of studying m6A in mRNA. To date, ALKBH5 is identified as another demethylase [6] and the methyltransferase like 3/14 (METTL3/METTL14) and Wilms’ tumor 1-assoicating protein (WTAP) are discovered to be subunits of the m6A methyltransferase complex [7, 8]. All these findings provide strong evidences to show that m6A is a dynamic modification and suggest that it may play a critical role in exerting post-transcriptional functions in mRNA metabolism [9–11]. These new wave of breakthroughs cannot be achieved without the recent development of MeRIP-seq [12, 13], which was successfully developed to reveal the transcriptome-wide distribution of m6A in human and mouse cells. In this essay, mRNA is first chemically fragmented into approximately 100-nucleotide (nt) long before immunoprecipitation with anti-m6A antibody. Then, the immunoprecipitated (IPed) methylated mRNA fragments and the un-immunoprecipited input control mRNA fragments are subjected to high-throughput sequencing [14]. The sequenced IP and input reads are aligned to the transcriptome and reads enrichment of IP out of the combined reads in IP and input samples are examined to predict to loci of methylation sites and infer the degree of methylation. We have previously developed exomePeak [15, 16] and HEPeak [17], two algorithms for detecting m6A peaks in MeRIP-seq. Although MeRIP-seq and subsequent computational peak-calling analysis provide an accurate landscape of m6A methylation in transcriptome, the complete mechanisms of this methylation still remains unclear. Just like gene expression where co-expression might suggest co-regulation or similar gene functions, sites with similar methylation degree could be related to similar methylation mechanisms. Therefore, there is a need to develop algorithms to uncover co-methylation pattern in MeRIP-seq data. In this paper, we model the methylation degrees of m6A peaks as a mixture of the Beta-binomial distributions and propose an expectation-maximization based clustering algorithm to uncover the co-methylation patterns. Methods In this section, we first describe the proposed generative model to define m6A peak clusters and then derive the Expectation-Maximization algorithm for the inference. In the end, we discuss a Bayesian Information Criterion (BIC) [18] for selecting the optimal number of m6A peak clusters. The proposed graphical model for clustering RNA methylation peaks The proposed graphical model for clustering of m6A peaks in MeRIP-seq data is shown in Fig. 1. Suppose we have identified a set of N m6A peaks, by using peak-calling software such as exomePeak or HEPeak. The goal is to cluster these peaks according to their methylation degree, which is defined as IP reads count divided by the total count of IP and control reads. For the nth m6A peak, let Zn ∈ {1, 2,.., K} denote the index of the particular methylation cluster that n-th peak belongs to, with K representing the total number of clusters, then Zn follows a discrete distributionFig. 1 The proposed graphical model for peak clusters 1 PZn|π=∏k=1KπkIZn=k where πk is the unknown probability that an m6A peak belongs to cluster k, where ∑Kπk = 1 and I(⋅) is the indicator function. Also, let the observed reads count in the nth peak of the mth IP replicate sample be Xm,n and that of the mth input replicate denote as Ymn. Under the assumption that reads count follows a Poisson distribution, the reads count Xmn given the total reads account Tmn = Xmn + Ymn can be shown to follow a Binomial distribution 2 PXmn|pn,Zn=∏k=1KTmnXmnpnXmn1−pnYmnIZn=k where is pn represents unknown methylation degree at the nth Peak of the mth replicate. In order to model the variance of the replicates for the nth peak, given cluster assignment Zn, pn is assumed to follow the Beta distribution 3 Pp|Zn=∏k=1KBetaαkβkIZn=k Therefore, after integrating the variable pn, Xmn follows a mixture of Beta-binomial distribution 4 PXmn|Zn;α,β=∑pnPXmn|pn,ZnPpn|α,β=∏k=1KC•ΓXmn+αkΓYmn+βkΓTmn+αk+βkΓαk+βkΓαkΓβkIZn=k where α = [α1, α2,.., αK]T, β = [β1, β2,.., βK]T are the unknown parameters of Beta distribution and C is the normalization constant. Thus, by considering the N m6A peaks in M replicates, the joint distribution is 5 PX,Z|α,β,π=∏n=1N∏m=1M∏k=1KπkBBXmn|ZnIZn=k where BB(Xmn|Zn) represents formula (3). Then, the log-likelihood of the observed data can be expressed as 6 l=lgPX|α,β,π=lg∑ZPX,Z|α,β,π=∑n=1N∑m=1Mlg∑k=1KπkBBXmn|Zn;α,β where Z = [Z1, Z2, …, ZN]T, X = [X1,TX2,…,TXNT]T and Xn=X1nX2n…XMnT. The goal of inference is to predict the cluster index Zn for all the peaks and estimate the unknown model parameters θ=αβπ. Next, we first discuss the maximum likelihood solution for parameter inference, based which an EM algorithm is introduced afterwards to perform model parameters inference and cluster assignment jointly. Parameter inference by the Newton’s method Given that the cluster indices are known, the model parameters can be inferred by the maximum likelihood criterion as 7 θ^ML=argmaxlθ. Given (5–6), the log-likelihood l can be rewritten as 8 l=∑n=1N∑m=1Mlg∑k=1KqZnkπkBBXmn|ZnqZnk≥∑n=1N∑m=1M∑k=1KqZnklgπk+lgBBXmn|Zn−lgqZn=∑n=1N∑k=1KM·qZnklgπk−∑n=1N∑k=1KM·qZnklgqZnk+∑n=1N∑m=1M∑k=1KqZnkΦαk+βk−ΦTmn+αk+βk+ΦXmn+αk+ΦYmn+βk−Φαk−Φβk where Φ=lgΓ· and qZn=PZn=kX. Here, given that qZn is a complex simplex, according to the Jensen’s inequality, the lower bound of l is achieved when qZn=PZnX. With a little abuse of notation, l denotes the lower bound of (7). Given the equality constrain ∑Kπk=1, the parameters of π can be computed by maximizing l and its dual problem with Lagrange multiplier λ is 9 maxgπλ=∑n∑m∑kqZnlgπk+λ∑kπk−1 then λ and π can be calculated as 10 λ=−N·Mπk=1N∑n=1NPZn=k|Xn Due to lack of analytical solution for the derivatives of l with respect to α and β, a Newton’s method is applied and the the gradient can be computed as 11 Jk=∑n=1NqZnkΦαk+βk·M−∑m=1MΦTmn+αk+βk−Φαk·M+∑m=1MΦXmn+αk∑n=1NqZnkΦαk+βk·M−∑m=1MΦTmn+αk+βk−Φβk·M+∑m=1MΦYmn+βk and the Hessian is 12 Hk1,1=∑n=1NqZnkΦ'αk+βk·M−∑m=1MΦ'Tmn+αk+βk−Φ'αk·M+∑m=1MΦ'Xmn+αkHk2,2=∑n=1NqZnkΦ'αk+βk·M−∑m=1MΦ'Tmn+αk+βk−Φ'βk·M+∑m=1MΦ'Ymn+βkHk1,2=Hk2,1=∑n=1NqZnkϕ'αk+βk·M−∑m=1Mϕ'Tmn+αk+βk. Then, the parameters for the kth cluster can be updated iteratively as 13 αknewβknew=αkoldβkold−Hk−1Jk m6A peak cluster assignment Assigning m6A peak to a cluster amounts to inferring cluster index Zn, whose posterior probability given θ can be written as 14 PZn=k|Xn,θ=PZn=k,Xn|θ∑k=1KPZn=k,Xn|θ=πk·∏m=1MBBXmn|Zn=k,θ∑k=1Kπk∏m=1MBBXmn|Zn=k,θ. However, PZn=kXn,θ cannot be computed directly, because parameter θ is also unknown. To circumvent the difficulty, we developed an EM [19] algorithm to infer Zn and estimate the model parameters θ in an iterative fashion. The steps of the proposed EM algorithm are described in the following Repeat until convergence achieved: E-step: use the previous computed parameters θold to update the posterior probability of the hidden states PZn=kXn,θ according to (13). M-step: maximize the lower bound l in (7) and estimate parameters θnew according to (12). Selection of the number of states by Bayesian information criterion (BIC) Note that the total number of states K is also unknown. In order to determine K, the BIC is applied search in the range of 2 to 15. The best number of states is selected by the lowest BIC, which is denoted as 15 BIC=−2l^+2KlgN where l^ is the estimated log-likelihood when the EM algorithm converges. Results Performance evaluation by simulation The performance was evaluated by simulation where the true states of methylation peaks are known. Each peak was simulated independently, where the reads count was generated according to the proposed graphical model in Fig. 1. Notably, from (3), we can determine that the distribution of the methylation degree follows the following mixture Beta distribution 16 Pp=∑k=1KπkBetaαkβk where the kth Beta distribution model the methylation degree in cluster k. In our case, we assume there are K=4 clusters and π=0.30.40.20.1. Note that the degree p may vary vastly when the variance of the Beta distribution is large. In addition, the total reads count Tn of the nth peak can introduce another layer of variance and the larger the Tn is, the smaller the variance is. For simplicity, we only investigate the impact of the variances from the Beta distributions on performance. Here, two cases were considered; in the first case, moderate variances of the methylation degree were simulated where αβ=16216420102510and in the second case, the variances were assumed very high and set as αβ=81411.21910. To best mimic the real MeRIP-Seq data, N=10000 methylation peaks and M=2 replicates were simulated. Also, we let the total count Tn=100 for any methylation peak. The performance of the proposed algorithm in uncovering the clusters of m6A peak methylation degree can be evaluated by examining the goodness-of-fit of the mixture Beta distribution (15). Figure 2a demonstrates that the fitting performance for both moderate and high variance cases both cases and we can see the estimated mixture density is extremely close to the true ones, indicating a good fitting performance by the algorithm. In order to quantify the influence of the number of replicates on the fitting performance, simulated datasets with replicates varying from 1 to 10 were generated. The goodness-of-fit measured by Kullback–Leibler (KL) divergence between the estimated and the true mixture distributions was examined for different number of replicates separately. We can see from Fig. 2b that even with no replicate the fitting performance is very high with a KL divergence less 0.7 %. When there are two or more replicates, further improvement can be obtained, where the KL divergence can be reduce to as low as 0.2 %. Taken together, the results provide strong evidence to support a good fitting performance of the proposed algorithm for different reads variations.Fig. 2 Performance evaluation on simulated m6A peaks. a. The algorithm performs well on both moderate and high variances cases. b. As the number of replicates increases, the performance is enhanced Evaluation on real m6A MeRIP-seq data To further validate the accuracy of the proposed algorithm, we applied it to two real public available m6A MeRIP-seq datasets [5, 8]. One is from the mouse midbrain cells including 3 replicates, download from Gene Expression Omnibus (GEO) (accession number GSE47217) and the other dataset including 4 replicates measures transcriptome-wide m6A in human HeLa cells (accession number GSE46705). The datasets were pre-processed according to the HEPeak pipeline and for midbrain dataset, a total of 18162 m6A peaks were identified, whereas 7243 m6A peaks were reported in the HeLa cells both for FDR < 0.05. Next, we applied our algorithm to uncover the peak clusters in two datasets. 2 m6A peak clusters were determined to exist for the mouse midbrain cells (Fig. 3a), where cluster 1 contains 60 % (10875) of the peaks and cluster 2 includes the remaining 40 % (7287). In contrast, 4 different m6A peak clusters were discovered for HeLa cells (Fig. 3b), with the proportion of peaks as 21 % (1521) for cluster 1, 44 % (3155) for cluster 2, 12 % (886) for cluster 3, and 23 % (1681) for cluster 4, where the cluster is ranked according to a descending order of methylation degree.Fig. 3 Pie chart for the proportion of peaks in each m6A clusters discovered in mouse midbrain and human HeLa cells. a. An illustration of the proportion of m6A peaks in each clusters in mouse midbrain cells. b. An illustration of the proportion of m6A peaks in each clusters in human HeLa cells To evaluate the accuracy of the proposed algorithm in characterizing the true mixture distribution of the methylation degree, the estimated density was next tested against the empirical distribution of peak methylation degrees obtained from MeRIP-Seq data. As illustrated in Figs. 4a and 5a, the estimated mixture distributions capture the real distributions of methylation degrees very well for both mouse and human MeRIP-seq datasets. We further investigated each components of the mixture. Figure 4b shows the empirical peak distributions of the two uncovered clusters in the mouse midbrain, which have distinct patterns. The fitted distributions of each cluster well captured the corresponding empirical distribution (chi-square test, pvalue: 9.2e-14 and 4.4e-4 for cluster 1 and 2). For human HeLa cells Fig. 5b, four distinct empirical distributions of peaks can be clearly seen and high fitting performance was also achieved for all four clusters (chi-square test, pvalue: 5.8e-21, 7.48e-38, 1.1e-15 and 1.2e-8 for cluster 1 to 4).Fig. 4 The estimated mixture density closely characterizes the real distribution of m6A peak in mouse midbrain cells. a. The estimated mixture Beta distribution versus the overall distribution of real m6A peaks in mouse midbrain cells. b. Comparison between the two estimated mixture components and the real distributions of m6A peaks in the corresponding cluster Fig. 5 The estimated mixture density closely characterizes the real distribution of m6A peak in HeLa cells. a. The estimated density versus the overall distribution of methylation degree of m6A peaks in human HeLa cells. b. Comparison between the four estimated mixture components and the real distributions of m6A peaks in the corresponding cluster A novel pattern of m6A distribution is revealed In order to gain insights into different clusters of methylation peaks, peaks in each cluster were mapped to the corresponding mRNA or lncRNA and their distribution was subsequently examined. In mouse midbrain cells, noticeable differences in the distributions of two clusters can be observed on mRNA (Fig. 6a). Peaks in cluster 1 that have higher methylation degree are highly enriched near the stop codon, a distribution similar to the general m6A distribution previously reported in the literature [1, 12, 13, 20], whereas those in cluster 2 that have less degree of methylation are clearly more enriched near the start codon towards the 5′ UTR. Interestingly, m6A peak clusters in lncRNA (Fig. 6b) also show the same pattern where the higher methylated peaks are more likely to be enriched toward its 3′UTR. This phenomenon was further supported by the results in human HeLa cells (Fig. 7a, b). We see once again that the highly methylated peaks tend to locate around the stop codon and the peaks move towards the 5′ end as their methylation degree decreases. This pattern was also verified on additional MeRIP-seq datasets (Additional files 1: Figure S1 and Additional file 2: Figure S2).Fig. 6 Distribution of m6A for different clusters in mRNA and lncRNA in mouse midbrain cells. a. The distribution of m6A peaks for different clusters in mRNA. b. The distribution of m6A peaks for different clusters in lncRNA Fig. 7 Distribution of m6A for different clusters in mRNA and lncRNA in human HeLa cells. a. The distribution of m6A peaks for different clusters in mRNA. b. The distribution of m6A peaks for different clusters in lncRNA To gain additional insights into these m6Aclusters, sequence motifs searching was performed on the sequences of the predicted m6A peaks for each particular cluster. The sequences of peaks were obtained by bedtools2.1 and motif search was done by using DREME [21, 22], with the shuffled sequences as the background. The most enriched consensus motifs are illustrated in the Fig. 8 and Additional file 3: Figure S3 in Additional files. Interestingly, the motifs for the highest methylated cluster in both mouse midbrain cells and human HeLa cells are found to be very similar and this similarity also exists for the lowly methylated cluster. For the highest methylated cluster, the common motif is GGAC, which has been shown by PAR-CLIP experiments as the binding motif of methyltransferase METTL14 [8]. For the lowest methylated peaks, the motif is determined as GGAGGA. This distinct motif has not been reported to be associated with any protein binding and thus requires further investigation.Fig. 8 Motifs detected by DREME in human and mouse cells Discussion and Conclusions In this paper, a novel graphical model based methylation peak clustering algorithm, was developed for discovering the patterns in methylation degrees of m6A peaks in the MeRIP-seq data. The peak cluster is modelled as the mixture Beta-binomial distribution, where the Beta distribution can model the variance of the methylation degree across sample replicates. The evaluation on both simulation and real MeRIP-seq datasets demonstrates the accuracy and robustness of our model. In addition, our algorithm successfully uncovered a unique and novel pattern for m6A peak cluster, providing a new lead for understanding the mechanisms and functions of m6A methylation. Abbreviations BIC, Bayesian Information Criterion; CDS, Coding DNA sequence; EM, Expectation of maximum likelihood method; FDR, False discovery rate; MeRIP-seq, Methylated RNA Immunoprecipatation combined with RNA sequencing; UTR, Untranslated region Additional files Additional MeRIP-seq datasets were further examined. One experiment was conducted by knocking out an m6A demethylase obesity associated protein (KO-FTO) in mouse midbrain cells. The other MeRIP-seq dataset was generated by knocking out m6A methyltransferase METTL14 (KO-METTL14) in human HeLa cells.Additional file 1: Figure S1. Distribution of m6A for different clusters in mRNA and lncRNA in KO-FTO mouse midbrain cells. A. The distribution of m6A peaks for different clusters in mRNA. B. The distribution of m6A peaks for different clusters in lncRNA. (PNG 114 kb) Additional file 2: Figure S2. Distribution of m6A for different clusters in mRNA and lncRNA in KO-METTL14 human HeLa cells. (PNG 173 kb) Additional file 3: Figure S3. Motifs for Cluster 2 and 3 detected by DREME in human HeLa cells. (PNG 21 kb) Acknowledgements We acknowledge the funding support from National Institutes of Health (NIH-NCIP30CA54174, 5 U54 CA113001 to YC and R01GM113245 to YH); National Science Foundation (CCF-1246073 to YH); The William and Ella Medical Research Foundation grant, Thrive Well Foundation and The Max and Minnie Tomerlin Voelcker Fund to MKR; Natural Science Foundation of China (61473232) to SZ. We also thank the computational support from the UTSA Computational System Biology Core, funded by the National Institute on Minority Health and Health Disparities (G12MD007591) from the National Institutes of Health. Declarations Publication charges for this article have been funded by R01GM113245. This article has been published as part of BMC Genomics Volume X Supplement X, 2016: XXXXX. The full contents of the supplement are available online at http://XXXXX. Authors’ contributions XC designed the method and drafted the manuscript. JM and SZ help design the validation experiments. MKR and CY provided biological interpretation of results on real data. YH supervised the work, made critical revisions of the paper, and approved the submission of the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. ==== Refs References 1. Pan T N6-methyl-adenosine modification in messenger and long non-coding RNA Trends Biochem Sci 2013 38 4 204 9 10.1016/j.tibs.2012.12.006 23337769 2. Liu J, Jia G. Methylation Modifications in Eukaryotic Messenger RNA. J Genet Genomics 2014;41(1);21–33. 3. Desrosiers R Friderici K Rottman F Identification of methylated nucleosides in messenger RNA from Novikoff hepatoma cells Proc Natl Acad Sci U S A 1974 71 10 3971 5 10.1073/pnas.71.10.3971 4372599 4. He C Grand challenge commentary: RNA epigenetics? Nat Chem Biol 2010 6 12 863 5 10.1038/nchembio.482 21079590 5. Hess ME Hess S Meyer KD Verhagen LA Koch L Bronneke HS Dietrich MO Jordan SD Saletore Y Elemento O The fat mass and obesity associated gene (Fto) regulates activity of the dopaminergic midbrain circuitry Nat Neurosci 2013 16 8 1042 8 10.1038/nn.3449 23817550 6. Zheng G Dahl JA Niu Y Fedorcsak P Huang CM Li CJ Vågbø CB Shi Y Wang WL Song SH ALKBH5 Is a Mammalian RNA Demethylase that Impacts RNA Metabolism and Mouse Fertility Mol Cell 2013 49 1 18 29 10.1016/j.molcel.2012.10.015 23177736 7. Wang Y Li Y Toth JI Petroski MD Zhang Z Zhao JC N6-methyladenosine modification destabilizes developmental regulators in embryonic stem cells Nat Cell Biol 2014 16 2 191 8 10.1038/ncb2902 24394384 8. Liu J Yue Y Han D Wang X Fu Y Zhang L Jia G Yu M Lu Z Deng X A METTL3-METTL14 complex mediates mammalian nuclear RNA N6-adenosine methylation Nat Chem Biol 2014 10 2 93 5 10.1038/nchembio.1432 24316715 9. Meyer KD Jaffrey SR The dynamic epitranscriptome: N6-methyladenosine and gene expression control Nat Rev Mol Cell Biol 2014 15 5 313 26 10.1038/nrm3785 24713629 10. Jia G Fu Y He C Reversible RNA adenosine methylation in biological regulation Trends Genet 2013 29 2 108 15 10.1016/j.tig.2012.11.003 23218460 11. Hussain S Aleksic J Blanco S Dietmann S Frye M Characterizing 5-methylcytosine in the mammalian epitranscriptome Genome Biol 2013 14 11 215 10.1186/gb4143 24286375 12. Meyer KD Saletore Y Zumbo P Elemento O Mason CE Jaffrey SR Comprehensive analysis of mRNA methylation reveals enrichment in 3′ UTRs and near stop codons Cell 2012 149 7 1635 46 10.1016/j.cell.2012.05.003 22608085 13. Dominissini D Moshitch-Moshkovitz S Schwartz S Salmon-Divon M Ungar L Osenberg S Cesarkas K Jacob-Hirsch J Amariglio N Kupiec M Topology of the human and mouse m6A RNA methylomes revealed by m6A-seq Nature 2012 485 7397 201 6 10.1038/nature11112 22575960 14. Wang Z Gerstein M Snyder M RNA-Seq: a revolutionary tool for transcriptomics Nat Rev Genet 2009 10 1 57 63 10.1038/nrg2484 19015660 15. Meng J, Lu Z, Liu H, Zhang L, Zhang S, Chen Y, Rao MK, Huang Y. A protocol for RNA methylation differential analysis with MeRIP-Seq data and exomePeak R/Bioconductor package. Methods 2014 16. Meng J Cui X Rao MK Chen Y Huang Y Exome-based analysis for RNA epigenome sequencing data Bioinformatics 2013 29 12 1565 7 10.1093/bioinformatics/btt171 23589649 17. Xiaodong Cui JM, Manjeet K. Rao, Yidong Chen, Yufei Huang. HEP: An HMM-based Exome Peak-finding Package for RNA Epigenome Sequencing Data. BMC Genomics 2014 18. Posada D Buckley TR Model selection and model averaging in phylogenetics: advantages of Akaike information criterion and Bayesian approaches over likelihood ratio tests Syst Biol 2004 53 5 793 808 10.1080/10635150490522304 15545256 19. Lindstrom MJ Bates DM Newton—Raphson and EM algorithms for linear mixed-effects models for repeated-measures data J Am Stat Assoc 1988 83 404 1014 22 20. Schwartz S, Agarwala SD, Mumbach MR, Jovanovic M, Mertins P, Shishkin A, Tabach Y, Mikkelsen TS, Satija R, Ruvkun G. High-Resolution Mapping Reveals a Conserved, Widespread, Dynamic mRNA Methylation Program in Yeast Meiosis. Cell 2013 21. Bailey TL DREME: motif discovery in transcription factor ChIP-seq data Bioinformatics 2011 27 12 1653 9 10.1093/bioinformatics/btr261 21543442 22. Machanick P Bailey TL MEME-ChIP: motif analysis of large DNA datasets Bioinformatics 2011 27 12 1696 7 10.1093/bioinformatics/btr189 21486936

PMC005xxxxxx/PMC5001243.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27535232279610.1186/s12864-016-2796-xResearchTopic model-based mass spectrometric data analysis in cancer biomarker discovery studies Wang Minkun 12Tsai Tsung-Heng 1Di Poto Cristina 1Ferrarini Alessia 1Yu Guoqiang 2Ressom Habtom W. hwr@georgetown.edu 11 Department of Oncology, Georgetown University, 4000 Reservoir Rd NW, Washington D.C., USA 2 Department of Electrical and Computer Engineering, Virginia Tech, 900 N Glebe Rd, Arlington, VA USA 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.545© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background A fundamental challenge in quantitation of biomolecules for cancer biomarker discovery is owing to the heterogeneous nature of human biospecimens. Although this issue has been a subject of discussion in cancer genomic studies, it has not yet been rigorously investigated in mass spectrometry based proteomic and metabolomic studies. Purification of mass spectometric data is highly desired prior to subsequent analysis, e.g., quantitative comparison of the abundance of biomolecules in biological samples. Methods We investigated topic models to computationally analyze mass spectrometric data considering both integrated peak intensities and scan-level features, i.e., extracted ion chromatograms (EICs). Probabilistic generative models enable flexible representation in data structure and infer sample-specific pure resources. Scan-level modeling helps alleviate information loss during data preprocessing. We evaluated the capability of the proposed models in capturing mixture proportions of contaminants and cancer profiles on LC-MS based serum proteomic and GC-MS based tissue metabolomic datasets acquired from patients with hepatocellular carcinoma (HCC) and liver cirrhosis as well as synthetic data we generated based on the serum proteomic data. Results The results we obtained by analysis of the synthetic data demonstrated that both intensity-level and scan-level purification models can accurately infer the mixture proportions and the underlying true cancerous sources with small average error ratios (<7 %) between estimation and ground truth. By applying the topic model-based purification to mass spectrometric data, we found more proteins and metabolites with significant changes between HCC cases and cirrhotic controls. Candidate biomarkers selected after purification yielded biologically meaningful pathway analysis results and improved disease discrimination power in terms of the area under ROC curve compared to the results found prior to purification. Conclusions We investigated topic model-based inference methods to computationally address the heterogeneity issue in samples analyzed by LC/GC-MS. We observed that incorporation of scan-level features have the potential to lead to more accurate purification results by alleviating the loss in information as a result of integrating peaks. We believe cancer biomarker discovery studies that use mass spectrometric analysis of human biospecimens can greatly benefit from topic model-based purification of the data prior to statistical and pathway analyses. Keywords Bayesian inferenceTopic modelPurificationLC-MSGC-MSExtracted ion chromatogramMetabolomicsProteomicsBiomarker discoveryIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Identification of disease-related alterations in molecular and cellular mechanisms may reveal useful disease biomarkers. Discovery of clinically relevant biomarkers has potentially far reaching implications for disease management and patient treatment [1–4]. High-throughput omic technologies have facilitated the search for changes in the levels of various biomolecules (proteins, glycoproteins, metabolites, lipids, etc.) in biological samples [5, 6]. In particular, liquid (or gas) chromatography coupled with mass spectrometry (LC/GC-MS) has become an essential tool for profiling biomolecules in a variety of large-scale omic studies. Briefly, biomolecules are separated, fragmented, ionized and detected in LC/GC-MS instruments. Abundances of ions with various retention time and mass values are recorded for downstream data processing. While the capability of high-throughput technology to yield comprehensive profiling and quantification offers a unique advantage in biomedical research, the heterogeneous nature of the biological samples poses a fundamental challenge in data analysis and interpretation. Specimens, such as tumor tissues and human blood, are typically mixtures of cells with distinct states and types, and usually only part of the constituent cell populations is relevant to the biological question of interest [7, 8]. In some cancer studies, heterogeneity is also observed within the malignant cell population, where multiple cancerous subtypes co-exist [9]. Ideally in a biomarker discovery study, one would perform between-group (cancer versus related disease, cancer versus healthy samples) differential expression analysis for type-specific constituents in samples [10]. However, biospecimens collected from patients usually exhibit some degree of heterogeneity. Moreover, the proportion of cancerous, other disease-related, and healthy components varies across individual samples pre-selected using pathological estimates. Therefore, the biomolecular measurements in expression profiles are attributed to multiple sites of origins with various mixture proportions. The cancerous profiles of interest are typically contaminated by other components, leading to unreliable results in differential analyses. Purification of samples is hence highly desired to remove the effects of heterogeneity. Experimental methods for cleaning samples and isolating type-specific constituents are costly and time-consuming. Computational purification methods offer an attractive alternative that is inexpensive and efficient to implement, and can be applied to data already generated without any modifications on experimental procedures. Multiple approaches have been developed to deconvolute gene expression profiles in the past years, varying from linear regression based models [11, 12] to generative probabilistic models [13, 14]. Among these approaches, topic model based methods, e.g., ISOLATE [15] and ISOpure [8], showed promising performance in estimating the proportion of mixtures and inferring sample-specific purified profiles in heterogeneous genomic data. However, to the best of our knowledge, in omic studies involving quantitative analysis of proteins or metabolites, no such purification approaches have been applied to deal with the sample heterogeneity issue. With the increasing volume of these data generated by LC/GC-MS, it is necessary to implement the purification of data before downstream differential analyses. In this research, we first apply ISOpure, a topic model based purification approach to both synthetic and experimental data acquired from human sera and liver tissues by LC-MS and GC-MS, respectively. The purpose of this investigation is to test if sample heterogeneity issue in various biomolecular expression profiles can be addressed by adjusting ion intensities through topic models as in genomic studies. Also, we investigate the use of scan-level features, i.e. extracted ion chromatograms (EICs) instead of integrated peak intensities, to alleviate the information loss during the LC/GC-MS data preprocessing. Methods In this section, we introduce topic model-based intensity-level and scan-level purification methods. Assumptions and strategies in the topic models are elaborated. Mass spectrometric datasets from cancer biomarker discovery studies are described. Intensity-level purification model The LC/GC-MS instruments provide ion intensity values by counting the ions detected at specific m/z and retention time points. Due to the existence of heterogeneity, multiple constituents in the sample contribute to the observed expression profile. Therefore, we can model the expression profile of a heterogeneous sample t as a weighted mixture of expression profiles of multiple sources, including a cancerous origin γ and non-cancerous contaminants β. The expression distribution for every biomolecule in each of the sources plays a role as a “topic” contributing to the mixed expression profile. Basically, each ion in the observed profile is associated with a specific topic, i.e. a multinomial distribution of ion counts over biomolecules, determined by the corresponding source profile. In this model, expression profiles refer to integrated peak intensities. The purification procedure can be realized through a set of topic models, which are generative probabilistic models typically applied to text corpora mining. Specifically, each expression profile is characterized by a probability distribution across topics. Topics are probability distributions across biomolecules. These distributions can be inferred based on the analysis of a collection of expression profiles through topic models. These hierarchical Bayesian models are variants of latent Dirichlet allocation (LDA) [16], a topic model that can 1) infer the posterior probability of topics given observed profiles, and 2) estimate the parameters that generate the latent mixture proportion and topic panel. These topic models have been adapted and applied to gene expression profiles in genomic studies [8, 15]. We use a modified topic model to purify the molecular expression profiles in cancer. Basically, three assumptions are made in developing the model. First, the source contaminants in each expression profile {td}d=1,⋯,D are coming from the control group {βm}m=1,⋯,M (i.e., healthy, non-cancerous profiles, etc.). It is commonly observed that the cancerous tissues are surrounded by adjacent non-cancerous tissues, which are typically used as controls in differential expression analysis. Second, the corresponding cancerous origins {γd}d=1,⋯,D share an average cancer profile γ′. Individual cancerous profile can be treated as a noisy version of the average cancer profile. Third, the average cancer profile γ′ has similar patterns as non-cancerous profiles {β}, except for some sites (biomolecules) which are differentially expressed between case and control groups. Thus, the cancerous profile can be treated as a similar non-cancerous profile with several sites altered. The complete likelihood function in (1) describes how the profiles {td}(d=1,…,D) are generated. Specifically, we have two observable variables indicating D expression profiles in case group: {td}d=1,⋯,D,td∈ℝN, and M non-cancerous profiles in control group: {βm}m=1,⋯,M,βm∈ℝL. In our analysis, we normalize all profiles to have identical total ion counts of N and consider L biomolecules that are consistently detected in all the samples. For convenience, we represent the normalized profiles in two ways. Each heterogeneous cancer profile td is represented via N ions, with td,n={1,2,⋯,L} denoting the biomolecule corresponding to the nth ion. Each non-cancerous profile βm is represented via L biomolecules, with βm,l denoting the ion counts of the lth biomolecule, and ∑l=1Lβm,l=N. The second expression can be further normalized by N to give a representation of multinomial distribution as a topic. 1 ℒ(t,z,θ,γ,γ′|α,β,η,κ,κ′)=pγ′|β,η,κ′·∏d=1Dpθd|α·pγd|γ′,κd×∏n=1Npzd,n|θd·ptd,n|zd,n,θd,β,γd The model also incorporates the following latent variables: the average cancer profile γ′∈ℝL, sample-specific pure cancer profiles {γd}d=1,⋯,D,γd∈ℝL, sample-specific mixture proportions of topics {θd}d=1,⋯,D,θd∈ℝM+1, and sample-specific topic indicators {zd}d=1,⋯,D,zd∈ℝN,zd,n={1,⋯,M,M+1}. Their relationships with observations and parameters are given as below. 2 pθd|α=Dirichletθd|α,1 3 pγ′|β,η,κ′=Dirichletγ′|ηTβ,κ′ 4 pγd|γ′,κd=Dirichletγd|γ′,κd 5 pzd,n|θd=Multinomialzd,n|θd 6 ptd,n|zd,n≤M,θd,β,γd=Multinomialtd,n|βzd,n 7 ptd,n|zd,n=M+1,θd,β,γd=Multinomialtd,n|γd The average cancer profile γ′ is sampled from a Dirichlet distribution parameterized by a weighted mixture of non-cancerous profiles. Each pure cancer profile γd together with M contaminants {βm} forms a sample-specific topic panel. The mixture proportion of topics determines zd,n, indicating which source (i.e., γd or {βm}) each ion originates from. We infer the latent variables γ′,{γd}d=1,⋯,D,{θd}d=1,⋯,D, and estimate the parameters using the two-step learning approach developed based on variational EM algorithm (ISOpure package [8], version 1.4). The graphical model representing the above topic model is shown in Fig. 1. This three-level model allows a single profile to be associated with multiple topics (i.e., cancerous and non-cancerous origins). Such property of the LDA-framed models enable more flexible representation in data structure than that by other unigram models or mixture of unigrams [16]. Also in contrast to linear regression models, these methods use a multinomial noise model that is a better fit to noise measurement in biomolecular expression data [13]. Fig. 1 Graphical representation of the generative probabilistic model. Hyperparameters η, κ ′ together with sources of contaminants {β m} determine an average cancer profile γ ′. Each of the D profiles is associated with a mixture proportion θ d (regularized by hyperparameter α) and a topic panel consisting of {β m} and γ ′ (generated from the average cancer profile). Each of the N ions in a profile t n,d is sampled from a topic indicated by z n,d Scan-level purification model Here, we extend the topic model to utilize the scan-level measurements instead of the integrated peak intensities. During LC/GC-MS data preprocessing, ion intensity is obtained by integrating the scan-level measurements of a detected chromatographic peakl within a specified retention time (RT) interval. This integration or truncation, however, inevitably brings in variances which interfere with original sample heterogeneity. Therefore, we propose to investigate LC/GC-MS data purification with scan-level measurements based on extracted ion chromatogram (EIC), which preserves scan-level peak shape information. We hypothesize that purification at the scan level leads to more accurate results and offers the opportunity to extend the model to characterize both ion abundance and peak shape. After ion tracing and missing value interpolation, we can obtain a list of EICs for each sample. EIC is characterized by its retention time (corresponding to multiple scans), mass value, and ion abundance. In this scenario, the observed data {td} (same for {βm}) consists of multiple EIC peaks. Each peak is represented by ion abundances across S scans with a certain elution profile shape F(·) as shown in Fig. 2. Using these scan-level features, we model each EIC peak as shown in Eq. (8): Fig. 2 Extracted ion chromatography and peak shape function. Example of Gaussian (red) and exponentially modified Gaussian (green) peak shapes fitted to an experimental EIC involving 13 scans (blue) 8 td,n(s)=xd,n·δd,n(s)·Fs,ϕd,n+ed,n(s),s=1,⋯,S where, xd,n is the ion abundance for nth compound of dth sample; δd,n(s) is a latent indicator to model the missing scans; the chromatographic peak shape is characterized by the exponentially modified Gaussian (EMG) function [17] parameterized by ϕ, as described in Eq. (9), and ed,n(s) is the random noise. 9 Fs,ϕ=12ζexp12ζ2μ+ζσ2−2s×(1−erfμ+ζσ2−s2σ,ϕ≐{μ,ζ,σ} We hypothesize that the data heterogeneity in td,n corresponds to the shape of the EIC (characterized by ϕ) as well as ion abundance xd,n. We extend the purification model we used for integrated peaks by adding a lower layer to characterize the scan-level information as illustrated in Fig. 3. The three assumptions are maintained in this model in terms of the dependancy of ion abundance variables. That is, Eqs. (2)–(7) still hold for ion abundances xt, xβ, xγ′, and xγ. We assume error terms in intensity measurements in Eq. (8) are independent random variables generated by a normal distribution with conjugate prior following an inverse-Gamma distribution: 10 ed,n(s)|σed2∼N0,σed2,σed2∼IGae,be. Fig. 3 Graphical representation of the scan-level topic model. A lower layer to characterize the scan-level information is added. Ion abundances x t, x β, x γ′, and x γ together with peak shape (parameterized in ϕ) determined the observed feature list t, β The missing scan indicator variable δd,n(s) follows a Bernoulli distribution, parameterized by qd with a prior of Beta distribution: 11 pδd,n(s)|qd=Bernoulliδd,n(s)|qd,pqd|aq,bq=Betaqd|aq,bq. The observed data point therefore follows the distribution: 12 td,n(s)|xtd,n,qd,ϕd,n,σed2∼qdNxtd,nF(s,ϕd,n),σed2+(1−qd)N0,σed2. The peak shape parameters ϕ are considered to have a normal distribution and its detailed priors are described in [17]. The extended model contains variables that are mutually coupled, providing no analytical form for the posterior distribution in calculation. As a variational approximation, we can split the model into two components: 1) mixture model of underlying ion abundances, and 2) scan-level feature generation. We adopt a two-phase approach to iteratively update the latent variables and estimate the parameters between the two parts. Specifically, we use a Markov chain Monte Carlo (MCMC) sampling method [17] to infer the peak shape model parameters of the second part (i.e., ion abundance xt, xβ, and shape function parameters ϕ). We then treat xt, xβ as observed variables to implement the inference on the first part using the same algorithm [8] employed in the intensity-level purification. Once converged, the model outputs the sample-specific mixture proportion θ, pure ion abundance xγ, shape function parameters ϕ and related parameters. After purification is performed, ion intensity may be calculated by applying peak detection algorithms [18, 19] to the pure EIC peaks {γd,n} recovered using Eq. (8). Mass spectrometric datasets The experimental data were acquired by analyses of tissue and blood samples from patients with hepatocellular carcinoma (i.e., HCC, case group) and liver cirrhosis (control group) [1–4]. HCC is a highly heterogeneous disease both at the molecular and clinical levels [20]. Whereas all patients in this study were diagnosed with liver cirrhosis, about half of them were also diagnosed with HCC. Contamination occurs due to the influence from cirrhotic constituents in HCC samples. In this study, we used GC-MS data acquired by analysis of metabolites in 15 tissues and LC-MS data acquired by analysis of proteins in sera from 116 subjects. GC-MS based metabolomic dataset Fifteen liver tissues were collected from 10 participants recruited at MedStar Georgetown University Hospital. As shown in Fig. 4, the tissues were collected from 5 HCC cases (5 tumor and 5 adjacent cirrhotic tissues) and 5 patients with liver cirrhosis. Samples were profiled through Agilent 7890A gas chromatography coupled with LECO’s time-of-flight mass spectrometer to characterize the metabolome alterations associated with HCC development in cirrhotic patients. We identified 559 metabolites after preprocessing the GC-MS raw data by ChromaTOF GC software with True Signal Deconvolution package (Leco Corporation). Two types of purification are investigated on the data. One is to purify HCC profiles by removing contaminants from cirrhotic profiles. The other is to purify adjacent cirrhotic profiles by reducing the impact of the profiles attributed to HCC. Fig. 4 Fifteen tissue samples collected from 10 subjects (5 HCC cases and 5 cirrhotic controls). Five tumor and five adjacent cirrhotic tissues were obtained from the 5 HCC cases. Additional 5 cirrhotic tissues were obtained from the 5 independent subjects with liver cirrhosis LC-MS based proteomic dataset We acquired 116 proteomic data by analysis of sera from 57 HCC cases and 59 patients with liver cirrhosis recruited from the hepatology clinics at MedStar Georgetown University Hospital. Following depletion and digestion, proteins extracted from sera were injected into a 3000 Ultimate nano-LC system interfaced to LTQ Orbitrap Velos and TSQ Vantage mass spectrometers in untargeted and targeted analyses, respectively. Proteins were identified and quantified by MaxQuant [21] and Skyline [22] in preprocessing untargeted and targeted LC-MS data, respectively. Finally, 101 proteins that were consistently identified across 116 samples were selected as intensity-level features in expression profiles (i.e., L=101). All profiles were normalized to the mean total-ion-counts at N=1.68×108. It is still not clear how the development of tumor in liver directly affect the alterations in blood. We hypothesize that there are some impacts from cirrhotic constituents contributing to the HCC profile in serum. The contamination may occur in an indirect way through, for example, secreted biomolecules instead of adjacent tissue cells. We apply the purification to remove the influence from cirrhotic contaminants. Synthetic datasets Before applying the models to experimental data, we generated synthetic datasets by artificially mixing real LC-MS data on both intensity and scan levels, and evaluated the model based on its performance of deconvolving the mixed data. We generated synthetic data based on the 116 LC-MS profiled serum proteomic dataset. We assume here that human sera are homogeneous specimens. Hence we can mix them to simulate heterogeneous cancer profiles. Figure 5 shows the generative process of 30 synthetic cancer profiles with contamination, following the steps below: (i) Average the profiles of HCC group, {λs}s=1,⋯,57, to obtain an average cancer profile γ′, which is close to the real cancerous profile for HCC. (ii) Sample 30 individual pure cancer profiles {γd}d=1,⋯,30 from a Dirichlet distribution, as in (4), parameterized by γ′ and κd=1minl(γl′). (iii) Randomly select a subset of cirrhotic profiles {βm}m=1,⋯,M (M=9 in this simulation) as sources of contamination. Normalize them into form of multinomial distribution. (iv) Combine M cirrhotic profiles with each of the pure cancer profiles to create 30 topic panels, each consisting of M+1=10 profiles. (v) Sample 30 mixture proportions {θd}d=1,⋯,30 from a Dirichlet distribution, as in (2), parameterized by α=[1,⋯,1,5], which is uniform for the first nine constituents (contaminants) and with a larger value assigned to last constituent (cancer origin). This ensures a larger proportion of cancerous component in final cancer profile. (vi) Sample a topic indicator zd,n from θd using (5), and sample a td,n from βz if z≤M or γ′d otherwise, as in (6), (7). Repeat the sampling for N=1.68×108 times to generate a synthetic cancer profile td. Fig. 5 Generative process of heterogeneous cancer profiles. (i) average cancer profiles in case group; (ii) generate sample-specific pure cancer profile; (iii) select sources of contaminants in control group; (iv) form topic panels; (v) generate sample-specific mixture proportions; (vi) generate synthetic cancer profiles Each of these 30 heterogeneous cancer profiles is a mixture of a pure cancer profile and multiple contaminants. The intensity-level purification procedure will help retrieve the pure cancer profile and estimate the sample purity as well as proportions of contaminants. Similar to intensity-level simulation, we generated heterogeneous dataset using scan-level features, i.e. EICs, exported from Skyline [22]. Corresponding to 101 proteins, 187 peptides with 561 scan-level features were extracted in each of the 116 samples. Each feature contains 60 scans representing a chromatographic peak as illustrated in Fig. 6. We followed the same steps (i-vi) except that we average and blend EIC peaks instead of protein intensities. Finally, 30 heterogeneous cancerous samples, each characterized by a list of 561 EICs, are generated. Fig. 6 Extracted ion chromatograms from LC-MS based serum proteomic data. Extracted ion chromatogram is characterized by m/z, retention time, and ion abundance Evaluation methods We evaluated the performances of our proposed models on both synthetic and real experimental LC/GC-MS datasets in consideration of the following three goals: 1) to test on intensity level if the model can reasonably estimate the proportion of mixtures in each of the synthetic profiles and recover the pure cancer profiles underneath; 2) to demonstrate if the scan-level purification model gives more accurate estimation on synthetic data; 3) to investigate the benefits of using these models to purify samples from cancer patients collected in our previous differential analysis studies. Outputs of intensity-level model include the sample-specific mixture proportions {θd∗}, pure cancer profiles {γd∗}, and the estimated average cancer profile γ′∗. Whereas, we expect outputs of sample-specific mixture proportion {θd∗}, pure ion abundance {xγ∗}, peak shape function parameters ϕ∗ from extended model. For synthetic datasets, we compare the estimated proportions of mixtures {θd∗} with the true ones ({θd}) used to generate the synthetic data. Estimation error ratio for a single sample is defined in Eq. (13). 13 ξdθ∗,θ=∥θd∗−θd∥1∥θd∥1×100%,d=1,⋯,30 Different from point-wise intensities, the scan-level estimation error ratio for a single sample is defined in Eq. (14) 14 ξdγ∗,γ=∑s=1Sγd∗(s)−γd(s)1∑s=1Sγd(s)1×100%,d=1,⋯,30 For experimental datasets, we evaluated the performances in multiple aspects including statistical significance of the candidate biomarkers, ROC curves in distinguishing the biological groups, and pathway analysis results. Results and discussions Synthetic datasets We applied current model and the extended model to the synthetic intensity-level and scan-level LC-MS datasets, respectively. By incorporating peak detection algorithms, we can further compare the purification performances between the two topic models. Intensity-level purification We obtained an average error ratio of mixture proportion ξ¯d(θ∗,θ) at 2.33 %, indicating a good characterization of original proportions. The comparison of proportion parameters for the first six profiles is depicted in Fig. 7 using radar charts and scatter plots. As shown in the figure, the estimation in each profile has captured consistent patterns as the ground truth in each of the 10 components. We achieved an average correlation coefficient between θd and θd∗ at 0.975. The model accurately recognized those non-cancerous constituents contributed as small as 5 % in each sample. The proportion of cancerous origin is overestimated in some samples due to the smaller contributions from the contaminants. The differences between θd and θd∗ are also related to the recovered pure cancer profiles {γd∗}. Similarly, we have the average estimation error ratio for sample-specific pure cancer profiles ξd¯(γ∗,γ)=6.51%, which is smaller than ξd¯(t,γ)=16.57%, i.e., the error ratio between unpurified cancer profile and true cancer profile. Figure 8 shows scatter plots of 101 proteins in unpurified cancer profile {td}d=1,⋯,6 versus true cancer profile (blue) and in purified cancer profile versus true cancer profile (orange). The average correlation coefficient increases from 0.986 to 0.999 after purification. The effects of purification are illustrated in Fig. 9 by projecting the high-dimensional (dim=101) profiles onto their top three principal components. We observe that the purified cancer profiles were more distant from non-cancerous profiles, and regularized towards an average cancer profile. Fig. 7 Similarity evaluation on θ. Comparison between estimated θ ∗ and true mixture proportions θ for the first six profiles. Top: radar charts with 10 spokes, each representing a source in topic panel. The proportion of each source is depicted by the length of lines with color (orange for estimation θ ∗ and blue for ground truth θ). Bottom: scatter plots of corresponding proportions in ground truth θ and estimation θ ∗. The correlation coefficients ρ are given on the left-top Fig. 8 Similarity evaluation on γ. The first six out of 30 scatter plots of unpurified cancer profiles versus true cancer profiles (blue) and corresponding scatter plots of purified cancer profiles versus true cancer profiles (orange). The correlation coefficients ρ between each pair of profiles are given on the left-top Fig. 9 PCA analysis on simulated dataset. Thirty cancer profiles {t d} (red square), 30 purified cancer profiles {γ ∗ d} (yellow circle), and 9 sources of cirrhotic contaminants {β m} (blue triangle) Scan-level purification We first evaluated the purification power in the case of scan-level features. The average estimation error ratio of mixture proportions is 3.57 % by Eq. (13). In terms of recovering the underneath pure feature list, we achieved the average estimation error ratio for sample-specific pure cancerous feature list ξd¯(γ∗,γ)=3.12%, which is smaller than ξd¯(t,γ)=9.61%, i.e., the error ratio between unpurified cancerous feature list and ground truth. The purification with scan-level features works to some extent but it is also interesting to prove the extended model works in a more accurate way than intensity-level topic model. To allow intensity-level purification model to handle scan-level synthetic dataset, we employed peak detection algorithms (i.e., through successive convolution with a 4th order Savitzky-Golay smoothing filter and a first-order derivative of a Gaussian kernel with window width of 25 scans, standard deviation of 3) to transfer EIC peaks into intensities using area under curve. The same algorithm is applied for transferring purified peak list resulted from the extended model. We obtained a greater distance of mixture proportion with ξdI¯(θ∗,θ) at 7.23 % if using intensity-level purification model, compared to half (ξdS¯(θ∗,θ)=3.57%) achieved by extended scan-level purification model. LC-MS based proteomic dataset We treated all 59 cirrhotic profiles as origins of contaminants to purify 57 HCC profiles. We plotted these profiles using their first three principal components in Fig. 10. Fig. 10 PCA analysis on proteomic dataset. Fifty seven HCC profiles {t d} (red square), 57 purified HCC profiles {γ ∗ d} (yellow circle), and 59 sources of cirrhotic contaminants {β m} (blue triangle) Similar to the simulation result, we observed a clearer distinction between HCC and cirrhotic profiles after purification. To further understand the improvements, we carried out the following analyses on both purified and unpurified profiles. Firstly, in statistical analysis, the most relevant proteins with differential intensities between HCC cases and cirrhotic controls were selected using t-test, and the associated p-values were adjusted based on multiple testing correction (FDR ≤0.05). We found 43 proteins with significant change in expression between the two groups. The number of reported significant proteins under the same testing method increased from 43 to 75 after purification. The majority of the proteins identified in original profiles (40 out of 43) remained significant after purification. If purified based on scan-level features, the number of significant proteins also increased to 69, among which 38 and 61 are overlapped with unpurification and intensity-level purification results, respectively. Figure 11a, b, and c show ROC curves for each of the 43, 75, and 69 significant proteins, respectively. A bootstrap method (1000 bootstrap replicates) was used to compute the 95 % confidence interval (CI) of the area under each ROC curve. After intensity-level and scan-level purification we respectively achieved an average AUC of 0.793 (with 95 % CI at [0.700, 0.863]) and 0.811(with 95 % CI at [0.719, 0.890]), both higher than 0.706 (with 95 % CI at [0.606, 0.795]) for original biomarkers. More powerful biomarkers were selected after scan-level purification. Fig. 11 ROC curves of significant proteins. a ROC curves for each of 43 significant proteins before purification (AUC¯=0.706,95%CI[0.606,0.795]). b ROC curves for each of 75 significant proteins after intensity-level purification (AUC¯=0.793,95%CI[0.700,0.863]). c ROC curves for each of 69 significant proteins after scan-level purification (AUC¯=0.811,95%CI[0.719,0.890]) Finally, we used DAVID [23] (version 6.7) to identify significant signaling pathways, where the UniProt IDs of the significant proteins were mapped to the KEGG [24] database. As shown in Table 1, three pathways were reported from the original list of significant proteins. Following intensity-level and scan-level purifications, we found peroxisome proliferator-activated receptor (PPAR) signaling pathway with five and six significant proteins involved in addition to the three pathways (complement and coagulation casades, systemic lupus erythematosus, and prion disease) identified without purification. This is interesting in light of previous reports linking cancer and PPARs expressed in human liver [25]. Table 1 Signaling Pathways (number of significant proteins involved in the pathway) Without purification Intensity-level purification Scan-level purification Complement and coagulation cascades (13) Complement and coagulation cascades (18) Complement and coagulation cascades (19) Systemic lupus erythematosus (5) Systemic lupus erythematosus (6) PPAR signaling pathway (6) Prion diseases (4) PPAR signaling pathway (5) Systemic lupus erythematosus (4) - Prion diseases (4) Prion diseases (4) GC-MS based metabolomic dataset Heterogeneity issue is more intuitive in tissue samples, where the contaminations originate from the neighboring non-homogeneous cells. We first purified the HCC profiles {td}d=1,⋯,5 using independent cirrhotic profiles {βm}m=1,⋯,5 as the sources of contamination. Without purification, none of the 559 metabolites passed the statistical test as significant (FDR adjusted p-value ≤0.05). However, seven metabolites were identified as significant after the profiles were purified. For the adjacent cirrhotic profiles {ψd}d=1,⋯,5, we applied the model to remove contaminations from any neighboring cancerous cells. We expected to observe that the purified adjacent cirrhotic profiles became close to independent cirrhotic profiles. The dissimilarity, defined in (8), between independent and adjacent cirrhotic profiles is ξ¯(ψ,β)=28.3%, and goes down to ξ¯ψ∗,β=24.9% after purification. The improvements are less substantial compared to the previous datasets, presumably due to the limited sample size and potential overfitting issue. Conclusions In this paper, we investigate topic model-based inference methods to computationally address heterogeneity issue in samples analyzed by LC/GC-MS. The topic model gives a probabilistic explanation on the corpus of LC/GC-MS based profiles on both integrated peak and scan-level ion intensity levels. The performances of our models in estimating mixture proportion and retrieving underlying true cancer profile are evaluated through well-designed synthetic data. We observed that incorporation of scan-level features gives more accurate purification results by alleviating the loss in information caused as a result of integrating peak intensity values. Through GC-MS metabolomic and LC-MS proteomic datasets we acquired from tissues and blood samples, respectively, we showed the benefit of applying topic-model based purification of the data prior to statistical and pathway analyses. Specifically, we observed improved discrimination between case and control groups and biologically meaningful pathway analysis results. Future studies will focus on cross-validation of the findings either computationally through mass spectrometric data from large-scale cancer biomarker discovery studies or by using ground-truth information from pathology reports and literature survey. Abbreviations EIC, extracted ion chromatogram; GC, gas chromatography; HCC, hepatocellular carcinoma; LC, liquid chromatography; MS, mass spectrum; RT, retention time ᅟ ᅟ Declarations This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4; Funding Publication of this article was funded in part by NIH Grants R01GM086746 and U01CA185188 awarded to HWR. Authors’ contributions MW developed the methods and implemented the algorithms. MW, THT, and HWR designed the study. CDP and AF performed the sample collection, preparation, and data acquisition. GY and HWR supervised the analysis. MW and HWR wrote the manuscript with contributions from all other authors. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. ==== Refs References 1 Xiao JF Varghese RS Zhou B Nezami Ranjbar MR Zhao Y Tsai TH Di Poto C Wang J Goerlitz D Luo Y LC–MS based serum metabolomics for identification of hepatocellular carcinoma biomarkers in Egyptian cohort J Proteome Res 2012 11 12 5914 23 23078175 2 Tsai TH Wang M Di Poto C Hu Y Zhou S Zhao Y Varghese RS Luo Y Tadesse MG Ziada DH LC–MS profiling of N-glycans derived from human serum samples for biomarker discovery in hepatocellular carcinoma J Proteome Res 2014 13 11 4859 68 10.1021/pr500460k 25077556 3 Tsai TH Song E Zhu R Di Poto C Wang M Luo Y Varghese RS Tadesse MG Ziada DH Desai CS LC-MS/MS-based serum proteomics for identification of candidate biomarkers for hepatocellular carcinoma Proteomics 2015 15 13 2369 81 10.1002/pmic.201400364 25778709 4 Ranjbar MRN Luo Y Di Poto C Varghese RS Ferrarini A Zhang C Sarhan NI Soliman H Tadesse MG Ziada DH GC-MS based plasma metabolomics for identification of candidate biomarkers for hepatocellular carcinoma in Egyptian cohort PloS ONE 2015 10 6 0127299 5 Diamandis EP Mass spectrometry as a diagnostic and a cancer biomarker discovery tool opportunities and potential limitations Mol Cell Proteomics 2004 3 4 367 78 10.1074/mcp.R400007-MCP200 14990683 6 Metzker ML Sequencing technologies—the next generation Nat Rev Genet 2010 11 1 31 46 10.1038/nrg2626 19997069 7 Gerlinger M Rowan AJ Horswell S Larkin J Endesfelder D Gronroos E Martinez P Matthews N Stewart A Tarpey P Intratumor heterogeneity and branched evolution revealed by multiregion sequencing N Engl J Med 2012 366 10 883 92 10.1056/NEJMoa1113205 22397650 8 Quon G Haider S Deshwar AG Cui A Boutros PC Morris Q Computational purification of individual tumor gene expression profiles leads to significant improvements in prognostic prediction Genome Med 2013 5 3 29 10.1186/gm433 23537167 9 Garcia-Closas M Hall P Nevanlinna H Pooley K Morrison J Richesson DA Bojesen SE Nordestgaard BG Axelsson CK Arias JI Heterogeneity of breast cancer associations with five susceptibility loci by clinical and pathological characteristics PLoS Genet 2008 4 4 1000054 10.1371/journal.pgen.1000054 10 Shen-Orr SS Tibshirani R Khatri P Bodian DL Staedtler F Perry NM Hastie T Sarwal MM Davis MM Butte AJ Cell type–specific gene expression differences in complex tissues Nat Methods 2010 7 4 287 9 10.1038/nmeth.1439 20208531 11 Wang N Gong T Clarke R Chen L Shih IM Zhang Z Levine DA Xuan J Wang Y Undo: a bioconductor r package for unsupervised deconvolution of mixed gene expressions in tumor samples Bioinformatics 2015 31 1 137 9 10.1093/bioinformatics/btu607 25212756 12 Lawson CL Hanson RJ Solving Least Squares Problems. vol. 161 1974 Englewood Cliffs, NJ Prentice-hall 13 Qiao W Quon G Csaszar E Yu M Morris Q Zandstra PW Pert: a method for expression deconvolution of human blood samples from varied microenvironmental and developmental conditions PLoS Comput Biol 2012 8 12 1002838 10.1371/journal.pcbi.1002838 14 Posekany A Felsenstein K Sykacek P Biological assessment of robust noise models in microarray data analysis Bioinformatics 2011 27 6 807 14 10.1093/bioinformatics/btr018 21252077 15 Quon G Morris Q ISOLATE: a computational strategy for identifying the primary origin of cancers using high-throughput sequencing Bioinformatics 2009 25 21 2882 9 10.1093/bioinformatics/btp378 19542156 16 Blei DM Ng AY Jordan MI Latent dirichlet allocation J Mach Learn Res 2003 3 993 1022 17 Nezami Ranjbar MR Tadesse MG Wang Y Ressom HW Bayesian normalization model for label-free quantitative analysis by lc-ms IEEE/ACM Trans Comput Biol Bioinforma 2015 12 4 914 27 10.1109/TCBB.2014.2377723 18 Wang M Yu G Mechref Y Ressom HW GPA: An algorithm for LC/MS based glycan profile annotation the Proceedings of International Conference on Bioinformatics and Biomedicine (BIBM) 2013 Shanghai, China IEEE 19 Monroe ME Shaw JL Daly DS Adkins JN Smith RD Masic: A software program for fast quantitation and flexible visualization of chromatographic profiles from detected lc–ms (/ms) features Comput Biol Chem 2008 32 3 215 7 10.1016/j.compbiolchem.2008.02.006 18440872 20 Nault JC Villanueva A Intratumor molecular and phenotypic diversity in hepatocellular carcinoma Clin Cancer Res 2015 21 8 1786 88 10.1158/1078-0432.CCR-14-2602 25628398 21 Cox J Mann M Maxquant enables high peptide identification rates, individualized ppb-range mass accuracies and proteome-wide protein quantification Nat Biotechnol 2008 26 12 1367 72 10.1038/nbt.1511 19029910 22 MacLean B Tomazela DM Shulman N Chambers M Finney GL Frewen B Kern R Tabb DL Liebler DC MacCoss MJ Skyline: an open source document editor for creating and analyzing targeted proteomics experiments Bioinformatics 2010 26 7 966 8 10.1093/bioinformatics/btq054 20147306 23 Huang DW Sherman BT Lempicki RA Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources Nat Protoc 2008 4 1 44 57 10.1038/nprot.2008.211 24 Kanehisa M Goto S KEGG: kyoto encyclopedia of genes and genomes Nucleic Acids Res 2000 28 1 27 30 10.1093/nar/28.1.27 10592173 25 Tachibana K, Yamasaki D, Ishimoto K, Doi T. The role of PPARs in cancer: PPAR Res. vol. 2008. Article ID 102737; 2008, p. 15.

PMC005xxxxxx/PMC5001244.txt

==== Front Scoliosis Spinal DisordScoliosis Spinal DisordScoliosis and Spinal Disorders2397-1789BioMed Central London 7710.1186/s13013-016-0077-8Meeting Abstracts12th International Conference on Conservative Management of Spinal Deformities – SOSORT 2015 Annual Meeting Katowice, Poland, 7-9 May 2015Parent Eric 1Richter Alan 1Aulisa Angelo Gabriele 2Guzzanti Vincenzo 34Pizzetti Paolo 5Poscia Andrea 6Aulisa Lorenzo 7Simony Ane 8Christensen Steen Bach 8Andersen Mikkel O. 8Negrini Alessandra 9Donzelli Sabrina 9Maserati Laura 10Zaina Fabio 9Villafane Jorge H 11Negrini Stefano 1011Fortin Carole 12Grunstein Erin 12Labelle Hubert 12Parent Stefan 12Feldman Debbie Ehrmann 13Lou Edmond 14Zheng Rui 14Hill Doug 15Donauer Andreas 15Tilburn Melissa 15Raso Jim 14Schreiber Sanja 16Parent Eric 16Kawchuk Greg 16Hedden Douglas 17Sánchez-Raya Judith 18Adrover Antonia Matamalas 18D’Agata Elisabetta 18Granell Joan Bagó 18Kluszczynski Marek 19Kluszczyńska Anna 19Wąsik Jacek 20Motow-Czyż Marta 20Kluszczyński Adam 19Simony Ane 21Hansen Karen Hojmark 21Thomsen Hanne 22Andersen Mikkel Meyer 23Vuust Morten 22Blicharska Irmina 24Durmała Jacek 24Wnuk Bartosz 24Matyja Małgorzata 25Szopa Andrzej 26Domagalska-Szopa Małgorzata 27Gallert-Kopyto Weronika 28Łosień Tomasz 27Plintla Ryszard 29Landauer Franz 30Vanas Karl 30Gur Gozde 31Altun Necdet Sukru 32Yakut Yavuz 31Gawda Piotr 33Majcher Piotr 33Sulam Lior Neuhaus 34Bradley Michael 35Glynn David 36Hughes Alex 35Maude Erika 35Pilcher Christine 35Lebel Andrea 37Lebel Victoria Ashley 3738Orbán Judit 37Stępień Agnieszka 39Graff Krzysztof 39Speers D. 40Aulisa Angelo Gabriele 41Guzzanti Vincenzo 41Mastantuoni Giuseppe 41Poscia Andrea 42Aulisa Lorenzo 43Aulisa Angelo Gabriele 44Guzzanti Vincenzo 4546Falciglia Francesco 44Poscia Andrea 47Aulisa Lorenzo 48Karavidas Nikos 49Etemadifar Mohammadreza 50Donzelli Sabrina 51Zaina Fabio 51Lusini Monia 51Minnella Salvatore 51Balzarini Luca 52Respizzi Stefano 52Negrini Stefano 5354Güttinger Kathrin 55Durmała Jacek 56Blicharska Irmina 56Drosdzol–Cop Agnieszka 57Skrzypulec–Plinta Violetta 57D’Agata Elisabetta 58Sánchez-Raya Judith 59Sánchez-Raya Judith 60D’Agata Elisabetta 61Paśko Sławomir 62Glinkowski Wojciech 63Michoński Jakub 64Walesiak Katarzyna 65Pakuła Anna 64Sitnik Robert 64Glinkowski Wojciech 65Diers Helmut 66Majcher Piotr 67Gawda Piotr 67Lebel Andrea 68Lebel Victoria Ashley 6869van Loon Piet 70van Erve Ruud 70Grotenhuis Andre 71Zapata Karina 72Parent Eric 73Sucato Dan 72Korbel Krzysztof 74Kozinoga Mateusz 75Stoliński Łukasz 74Kotwicki Tomasz 75Lebel Andrea 76Lebel Victoria Ashley 7677Diers Helmut 78Berdishevsky Hagit 79Berdishevsky Hagit 801 Department of Physical Therapy, University of Alberta, Edmonton, Alberta Canada 2 U.O.C. of Orthopedics and Traumatology, Children’s Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy 3 U.O.C. of Orthopedics and Traumatology, Children’s Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio, Rome 4, Italy 4 University of Cassino, Cassino, Italy 5 Independent practitioner, Rome, Italy 6 Institute of public health, University Hospital “Agostino Gemelli”, Catholic University of the Sacred Heart School of Medicine, Rome, Italy 7 Department of Orthopedics, University Hospital “Agostino Gemelli”, Catholic University of the Sacred Heart School of Medicine, Rome, Italy 8 Sector for Spine Surgery & Research, Middelfart Hospital, Middelfart, Denmark 9 ISICO Italian Scientific Spine Institute, Milan, Italy 10 University of Brescia, Brescia, Italy 11 IRCCS Don Gnocchi, Milan, Italy 12 Université de Montréal, Research centre CHU Sainte-Justine, Montréal, Canada 13 Université de Montréal, Institut de Recherche en santé publique de l’Université de Montréal, Montréal, Canada 14 University of Alberta, Alberta, Canada 15 Alberta Health Services, Alberta, Canada 16 University of Alberta, Edmonton, Canada 17 University of Alberta, Alberta Health Services, Edmonton, Canada 18 Vall d’Hebron Hospital, Barcelona, Spain 19 “Troniny” Children Rehabilitation Center in Częstochowa, Częstochowa, Poland 20 Institute of Physiotherapy, Jan Długosz University, Czestochowa, Poland 21 Sector for Spine Surgery & Research, Middelfart Hospital, Middelfart, Denmark 22 Department of Radiology, Frederikshavn Sygehus, Frederikshavn, Denmark 23 Department of Mathematics and Statistics, Aalborg University, Aalborg, Denmark 24 School of Health Sciences in Katowice, Medical University of Silesia, Chair and Department of Rehabilitation, Katowice, Poland 25 Chair of Kinesitherapy and Special Methods of Physiotherapy, Academy of Physical Education, Katowice, Poland 26 School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Physiotherapy, Katowice, Poland 27 School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Medical Rehabilitation, Katowice, Poland 28 School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Physiotherapy, Katowice, Poland 29 School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Adapted Physical Activity and Sport, Katowice, Poland 30 University Clinic of Orthopedics (PMU), Salzburg, Austria 31 Hacettepe University, Çankaya, Ankara, Turkey 32 Akay Hospital, Ankara, Turkey 33 Uniwersytet Medyczny w Lublinie, Lublin, Poland 34 The Israeli Scoliosis Center, Tel Aviv, Israel 35 Scoliosis SOS Clinic, London, Great Britain 36 University of York, York, Great Britain 37 Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8 Canada 38 Saba University School of Medicine, Saba, Dutch Caribbean Netherlands 39 Józef Piłsudski University of Physical Education, Warsaw, Poland 40 Scheck and Siress, Chicago, IL USA 41 U.O.C. of Orthopedics and Traumatology, Children’s Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy 42 Institute of public health, University Hospital “Agostino Gemelli”, Catholic University of the Sacred Heart School of Medicine, Rome, Italy 43 Department of Orthopedics, University Hospital “Agostino Gemelli”, Catholic University of the Sacred Heart School of Medicine, Rome, Italy 44 U.O.C. of Orthopedics and Traumatology, Children’s Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy 45 U.O.C. of Orthopedics and Traumatology, Children’s Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy 46 University of Cassino, Cassino, Italy 47 Institute of public health, University Hospital “Agostino Gemelli”, Catholic University of the Sacred Heart School of Medicine, Rome, Italy 48 Department of Orthopedics, University Hospital “Agostino Gemelli”, Catholic University of the Sacred Heart School of Medicine, Rome, Italy 49 Scoliosis Spine Laser Centre, Athens, Greece 50 Orthopedic Department, Esfahan University of Medical Sciences, Isfahan, Iran 51 ISICO Italian Scientific Spine Institute, Milan, Italy 52 ICH Istituto Clinico Humanitas, Milan, Italy 53 IRCCS Don Gnocchi, Milan, Italy 54 University of Brescia, Brescia, Italy 55 Zürcher Hochschule für angewandte Wissenschaften (ZHAW), Winterthur, Switzerland 56 School of Health Sciences in Katowice, Medical University of Silesia, Chair and Department of Rehabilitation, Katowice, Poland 57 School of Health Sciences in Katowice, Medical University of Silesia, Chair of Woman’s Health, Katowice, Poland 58 Vall d’Hebron Hospital Institute, Barcelona, Spain 59 Vall d’Hebron Hospital, Barcelona, Spain 60 Vall d’Hebron Hospital, Barcelona, Spain 61 Research Institut Vall d’Hebron Hospital, Barcelona, Spain 62 Warsaw University of Technology, Institute of Micromechanics and Photonics, Warsaw, Poland 63 Chair and Department of Orthopedics and Traumatology of the Locomotor System, Baby Jesus Clinical Hospital, Center of Excellence “TeleOrto” for Telediagnostics and Treatment of Disorders and Injuries of the Locomotor System, Medical University of Warsaw, Warsaw, Poland 64 Warsaw University of Technology, Warsaw, Poland 65 Department of Orthopaedics and Traumology of Locomotor System, Center of Excellence “TeleOrto”, Medical University of Warsaw, Warsaw, Poland 66 Research & Development, Schlangenbad, Germany 67 Uniwersytet Medyczny w Lublinie, Lublin, Poland 68 Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8 Canada 69 Saba University School of Medicine, Saba, Dutch Caribbean Netherlands 70 Care to Move Orthopedics, Deventer, The Netherlands 71 Neurosurgery, UMC Radboud University, Nijmegen, The Netherlands 72 Texas Scottish Rite Hospital, Dallas, Texas USA 73 University of Alberta, Edmonton, Canada 74 Rehasport Clinic, Poznań, Poland 75 Spine Disorders Unit-Departament of Pediatric Orthopaedics and Traumatology, Poznań University of Medical Sciences, Poznań, Poland 76 Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8 Canada 77 Saba University School of Medicine, Saba, Dutch Caribbean Netherlands 78 Research & Development, Schlangenbad, Germany 79 SchrothNYC, New York, NY USA 80 SchrothNYC, New York, NY USA 23 8 2016 23 8 2016 2016 11 Suppl 1 Publication charges for this supplement were funded by the conference.23© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.O1 The functional properties of paraspinal muscles in adolescents with idiopathic scoliosis (AIS): A systematic review of the literature Eric Parent, Alan Richter O2 The importance of the lateral profile in the treatment of idiopathic scoliosis Angelo Gabriele Aulisa, Vincenzo Guzzanti, Paolo Pizzetti, Andrea Poscia, Lorenzo Aulisa O3 Radiological outcome in Adolescent idiopathic scoliosis patients 20 years after treatment Ane Simony, Steen Bach Christensen, Mikkel O Andersen O4 Junctional Kyphosis, how can we detect and monitor it during growth? Alessandra Negrini, Sabrina Donzelli, Laura Maserati, Fabio Zaina, Jorge H Villafane, Stefano Negrini O5 Usefulness of the clinical measure of trunk imbalance in adolescent idiopathic scoliosis Carole Fortin, Erin Grunstein, Hubert Labelle, Stefan Parent, Debbie Ehrmann Feldman O6 Can ultrasound imaging be used to determine curve flexibility when designing spinal orthoses? Edmond Lou, Rui Zheng, Doug Hill, Andreas Donauer, Melissa Tilburn, Jim Raso O7 Reliability of the Schroth curve type classification in adolescents with idiopathic scoliosis (AIS) Sanja Schreiber, Eric Parent, Greg Kawchuk, Douglas Hedden O8 Can Trunk Appearance Perception Scale (TAPS) be used as a descriptive tool of scoliosis severity? Judith Sánchez-Raya, Antonia Matamalas Adrover, Elisabetta D’Agata, Joan Bagó Granell O9 Magnitude of the Cobb angle on an X-ray in relation to the angle of trunk rotation in children who come to the “Troniny” Scoliosis Treatment Centre Marek Kluszczynski, Anna Kluszczyńska, Jacek Wąsik, Marta Motow-Czyż, Adam Kluszczyński O10 Cobb angel measurement without X-ray, a novel method Ane Simony, Karen Hojmark Hansen; Hanne Thomsen; Mikkel Meyer Andersen; Morten Vuust O11 The postural tone magnitude and distribution in patients diagnosed with an adolescent idiopathic scoliosis: a preliminary study Irmina Blicharska, Jacek Durmała, Bartosz Wnuk, Małgorzata Matyja O12 From studies on the function of the respiratory system in children with body posture defects Andrzej Szopa, Małgorzata Domagalska-Szopa, Weronika Gallert-Kopyto, Tomasz Łosień, Ryszard Plintla O13 Scoliosis as the “first” sign of various diseases Franz Landauer, Karl Vanas O14 The effectiveness of core stabilization exercises versus conventional exercises in addition to brace wearing in patients with adolescent idiopathic acoliosis Gozde Gur, Necdet Sukru Altun, Yavuz Yakut O15 The effect of physiotherapy techniques on the body balance in patients with scoliosis treated with corrective appliances Piotr Gawda, Piotr Majcher O16 New combine method treating AIS – preliminary results Lior Neuhaus Sulam O17 Does a 4-week intensive course of ScolioGold therapy reduce angle of trunk rotation in scoliotic patients: a retrospective case series. Michael Bradley, David Glynn, Alex Hughes, Erika Maude, Christine Pilcher O18 Schroth physiotherapy method without bracing is an effective treatment for scoliosis in improving curves and avoiding surgery and should be offered as a treatment option for scoliosis in Canada: case series Andrea Lebel, Victoria Ashley Lebel, Judit Orbán O19 Rotation of the trunk and pelvis and coupled movements in the sagittal plane in double support stance in adolescent girls with idiopathic scoliosis Agnieszka Stępień, Krzysztof Graff O20 Curve progression analysis in Risser 0 patients orthotically managed with compliance monitors D. Speers O21 Conservative treatment in Scheuermann’s kyphosis: comparison between lateral curve and variation of the vertebral geometry Angelo Gabriele Aulisa, Vincenzo Guzzanti, Giuseppe Mastantuoni, Andrea Poscia, Lorenzo Aulisa O22 The plaster cast in the conservative treatment of idiopathic scoliosis can still play a positive role? Angelo Gabriele Aulisa, Vincenzo Guzzanti, Francesco Falciglia, Andrea Poscia, Lorenzo Aulisa O23 Bracing for Adolescent Idiopathic Scoliosis (AIS) and Scheuermann Kyphosis : The issue of overtreatment in Greece Nikos Karavidas O24 Efficacy of Milwaukee brace for correction of scheurmann kyphosis Mohammadreza Etemadifar O25 The three dimensional analysis of the Sforzesco brace correction Sabrina Donzelli, Fabio Zaina, Monia Lusini, Salvatore Minnella, Luca Balzarini, Stefano Respizzi, Stefano Negrini O26 Quality of Life in adolescents with idiopathic scoliosis: A comparison measured by the Kidscreen 27 between scoliotic patients and healthy controls Kathrin Güttinger O27 The degree of illness acceptance in young women with idiopathic scoliosis treated with orthopedic braces: a preliminary study Jacek Durmała, Irmina Blicharska, Agnieszka Drosdzol–Cop, Violetta Skrzypulec–Plinta O28 Which are the personality traits of the patients with Adolescent Idiopathic Scoliosis? Elisabetta D’Agata, Judith Sánchez-Raya O29 How many Scolioses do exist in the same person? A zoom vision on the perception of the patient Judith Sánchez-Raya, Elisabetta D’Agata P1 The algorithm for the automatic detection of the pelvic obliquity based on analysis of the PA viev of the x-ray image Sławomir Paśko, Wojciech Glinkowski P2 Monitoring of spine curvatures and posture during pregnancy using surface topography – case study and method assessment Jakub Michoński, Katarzyna Walesiak, Anna Pakuła, Robert Sitnik, Wojciech Glinkowski P3 Spinal rotation under static and dynamic conditions: a prospective study comparing normative data vs. scoliosis Helmut Diers P4 The principle of non-surgical treatment of idiopathic scoliosis right-sided breast depending on the volatility of the formation of the intervertebral discs and vertebral bodies Piotr Majcher, Piotr Gawda P5 Unexpected late progression of adolescent idiopathic scoliosis treated with short-term, aggressive, full-time bracing and Schroth physiotherapy with excellent preliminary result: case study Andrea Lebel, Victoria Ashley Lebel P6 Visible posture in relation to the neuroanatomical and neurodynamical features in spinal deformations Piet van Loon, Ruud van Erve, Andre Grotenhuis P7 Immediate effects of scoliosis-specific corrective exercises on the Cobb angle after 1 week and after 1 year of practice Karina Zapata, Eric Parent, Dan Sucato P8 Retrospective analysis of idiopathic scoliosis medical records coming from one out-patient clinic for compatibility with Scoliosis Research Society criteria of brace treatment studies Krzysztof Korbel, Mateusz Kozinoga, Łukasz Stoliński, Tomasz Kotwicki P9 Adult female with severe progressive scoliosis possibly secondary to benign tumor removal at age 3 treated with scoliosis specific Schroth physiotherapy after refusing surgery: case study Andrea Lebel, Victoria Ashley Lebel P10 New aspects of scoliosis therapy planning and monitoring Helmut Diers P11 Outcome of intensive outpatient rehabilitation in an adult patient with M. Scheuermann evaluated by radiologic imaging – a case report Hagit Berdishevsky P12 The effectiveness of a Scoliosis Specific Home Exercise Program and bracing to reduce an idiopathic scoliosis curve with more than 90 % success in less than a year of exercises. Case report. Hagit Berdishevsky 12th International Conference on Conservative Management of Spinal Deformities – SOSORT 2015 Annual Meeting Katowice, Poland 7-9 May 2015 issue-copyright-statement© The Author(s) 2016 ==== Body ORAL PRESENTATIONS O1 The functional properties of paraspinal muscles in adolescents with idiopathic scoliosis (AIS): A systematic review of the literature Eric Parent, Alan Richter Department of Physical Therapy, University of Alberta, Edmonton, Alberta, Canada Introduction: Exercise-based approaches exist; however, it is unclear whether these approaches are based on scientific findings in the literature on trunk muscle deficits in scoliosis that could be targeted by exercises. The aims of this study were to systematically review the literature to understand the functional muscular properties of paraspinal muscles in AIS to determine: 1) differences in functional outcomes between patients with AIS and controls, 2) differences in functional outcomes between sides (concave and convex) in patients compared to controls 3) differences between concave and convex sides as well as levels in subjects with AIS, 4) differences in functional outcomes between different curve types. 5) Associations between functional outcomes and curve characteristics, and 6) associations between functional characteristics and progression. Design: Systematic review Methods: A search was conducted in EMBASE, MEDLINE, SPORTdiscus, CINAHL, SCOPUS, and Web of Science, for keywords describing functional properties of paraspinal muscles and measurement tools including: scoliosis, spinal deformity, spinal muscles, erector, rotatores, longissimus, spinalis, illiocostalis, force, strength, endurance, fatigability, and muscle fatigue. Two reviewers independently reviewed abstracts and then full-text articles to determine if they met selection criteria. Two reviewers used an extraction form to extract information and appraise the quality during the full-text review. Levels of evidence were determined for summarized results for each of the 6 objectives. Results: Our search yielded 316 unique records. Inter-reviewer agreement for abstract selection was Kappa = 0.73 and was 0.77 for full-text inclusion. Full-text review was done for 48 papers and 24 were included. A large amount of heterogeneity was observed in sample studied and assessment methodology. Quality appraisal revealed that no study met a minimum of 50 % of the relevant quality criteria. Studies recruited consistently low sample sizes and samples were largely heterogeneous. Limited evidence was noted supporting, a prolonged bilateral EMG activation during gait between AIS and controls; elevated heterolateral:homolateral activity ratios during side-bending; overall weakness in those with scoliosis compared to controls; no asymmetry in normalized muscle activity during submaximal isometric contractions; prolonged latencies on the side of the spine opposite of the curve and bilaterally in response to an unloading reflex; strength & muscle volume differences are most commonly pronounced in double curves; Axial rotation of the UEV is correlated with a high convex:concave activity ratio at the LEV; no correlation between latency and curve severity, but a correlation between latency and progression and higher convex:concave EMG ratios and progression, this is more pronounced in sitting positions. Conclusions: Evidence is limited on most of our six objectives due to low quality evidence and lack of research about muscle impairments in scoliosis. Current exercise-based interventions cannot yet be said to be based on a strong understanding of muscle impairments in scoliosis. Research is needed using large, homogenous samples allowing for a comparison between curve types and examining relation to the risk of progression. While many exercise-based programs focus on addressing endurance deficits using high repetitions and long holds, no studies were found on endurance deficits in AIS. O2 The importance of the lateral profile in the treatment of idiopathic scoliosis Angelo Gabriele Aulisa1, Vincenzo Guzzanti2, Paolo Pizzetti3, Andrea Poscia4, Lorenzo Aulisa5 1U.O.C. of Orthopedics and Traumatology, Children's Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy; 2U.O.C. of Orthopedics and Traumatology, Children's Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio, Rome 4, Italy ; University of Cassino, Cassino, Italy; 3Independent practitioner, Rome, Italy; 4Institute of public health, University Hospital "Agostino Gemelli", Catholic University of the Sacred Heart School of Medicine, Rome, Italy; 5Department of Orthopedics, University Hospital "Agostino Gemelli", Catholic University of the Sacred Heart School of Medicine, Rome, Italy Background Adolescent idiopathic scoliosis (AIS) is a 3-dimensional spinal deformity. Thoracic sagittal malalignment has been thought to play an important role in the development of thoracic idiopathic scoliosis. Thoracic hypokyphosis with increasing axial rotational instability is claimed to be a primary factor for the initiation of Idiopathic Scoliosis (IS) according to some authors. Other authors have shown that Thoracic hypokyphosis is strongly associated with curve progression in thoracic IS. Moreover, the studies evaluating the impact of hypokyphosis on conservative treatment are limited in number and fragmentary. Design In previous studies, we have evaluated the impact of rotation on the conservative treatment of idiopathic scoliosis. The purpose of the present study was to determine the trend of hypokyphosis during conservative treatment and its interaction with the lateral curve and rotation. Material and methods From a prospective database, we selected all patients with adolescent thoracic idiopathic scoliosis, with Risser 0-3 and lateral radiographs performed at the beginning and at the end of treatment.From this group, we excluded all cases in which X-rays did not allow a correct measurement of kyphosis angle. 107 patient with Lyon brace, 97 female and 10 male, mean age 12.4 ± 1.81 years fulfilled the inclusion criteria. The minimum follow-up was 24 months. Postero-Anterior and lateral radiographs were used to estimate the lateral curve magnitude (CM), the torsion of the apical vertebra (TA) and the degree of kyphosis (KM) at 2 time points: beginning of treatment (t1) and 2-year minimum follow-up from the end of treatment (t5). Three outcomes were distinguished in agreement with SRS criteria: correction, stabilization and progression. Statistical analyses was performed. Results The results from our study showed that of the 107 patients CM mean value was 34.5 ± 9.9 SD at t1 and 21.7 ± 12.5 SD at t5. The difference in cobb degrees between t5-t1 was -12.8° (p < 0.01) in CM, -4.5° (p < 0.01) in TA and -2.0°(p < 0.01) in KM. The regression analysis shows that the evolution of the scoliosis is not associated with the initial level of kyphosis, both in terms of Cobb (p = 0.31; R2 = 0.009) and Pedriolle degrees (p = 0.74; R2 = 0.001). Confirms, furthermore, that the evolution of the scoliosis is not associated with the evolution of kyphosis, both in terms of Cobb (p = 0.989; R2 = 0.000)and Pedriolle degrees(p = 0.788; R2 = 0.001). Curve correction was accomplished in 87 patients (81.3 %), whereas a curve stabilization was obtained in 17 patients (15.8 %). 3 patients (3 %) had a curve progression and of these, in only one surgical treatment was recommended. Discussion Our results confirm that a well-designed brace with correctly push can limit the feared worsening of hypo-kyphosis. Moreover, contrary to what is commonly thought, it doesn’t appear that the hypo-kyphosis modifies the results of conservative treatment. It is confirmed, instead, as reported in previous papers both clinical and biomechanical: rotation significantly affects the outcome of conservative treatment. A rotation of more than 20 °, leads to the hysteresis of the intervertebral discs, hinders the transmission of the corrective forces to the vertebrae and leads to any brace corrective action. O3 Radiological outcome in Adolescent idiopathic scoliosis patients 20 years after treatment Ane Simony, Steen Bach Christensen, Mikkel O Andersen Sector for Spine Surgery & Research, Middelfart Hospital, Middelfart, Denmark Purpose The purpose of this study was to evaluate the long term radiological outcome, curve progression and adjacent level degeneration 20 years after scoliosis treatment. Method 219 patients treated with Boston brace or posterior spinal fusion a. m Harrington were invited to participate in a long term evaluation with clinical examination and x-ray evaluation. The old medical charts and x-ray descriptions where available.Standing X-ray was examined, the Cobb angel measured and compared to the patient’s prior x-rays and the adjacent levels where evaluated for any signs of adjacent level disease or local kyphosis. Results 159 patients participated (78 %). 66 patients treated with Boston brace and 92 patients, treated with posterior spinal fusion a.m. Harrington from 1983-1990 at University Hospital Copenhagen.In the Brace group, the Cobb angel prior to treatment was 37.5° (35.1°-40.0°), after treatment 34.7° (31.9°-37.5°). Cobb angel after 20 years was 40.2° (36.7°-43.6°).In the surgical group the Cobb angel prior to treatment was 54.5° (50.4°-58.8°), 1 year postoperative 29.5 ° (25.7°-33.9°).Cobb angel after 20 years 32.35° (27.9°-39.5°).26 patients had distal segment degeneration in x-rays (16.5 %), 4 patients treated with Brace and 22 patients with posterior spinal fusion.8 patients had proximal segment degeneration (5 %), 2 treated with brace and 6 patients with posterior spinal fusion. 4 patients were treated with posterior fusion of the distal adjacent segment (2.6 %), 1 treated with Brace and 3 treated with posterior spinal fusion. Conclusion The average follow up was 24.5 years (range 24-31 years). The Brace group had a small reduction of the spinal deformity during the treatment period, and X-rays shows a small progression of the deformity, Cobb angel increasing 5.5° within 20 years. The surgically treated patients had a large correction during surgery and there is no progression or loss of correction over a 20 year period.Only 4 patients in the Brace group have distal segment degeneration and 1 was treated with a one level spinal fusion.The surgically treated groups had a significant deformity correction during surgery and have maintained the correction after 20 years. 22 patients have distal degeneration and 3 patients were treated with distal adding on surgery. O4 Junctional Kyphosis, how can we detect and monitor it during growth? Alessandra Negrini1, Sabrina Donzelli1, Laura Maserati2, Fabio Zaina1, Jorge H Villafane3, Stefano Negrini2,3 1ISICO Italian Scientific Spine Institute, Milan, Italy; 2University of Brescia, Brescia, Italy; 3IRCCS Don Gnocchi, Milan, Italy Introduction Despite its importance in affecting adult pain, and disability, there is a lack of universal criteria for the diagnosis and evaluation of junctional kyphosis (JK) and a gold standard measurement and diagnostic system does not exist. Aim To verify the sensibility and specificity of clinical, and Formetric data in identifying junctional kyphosis in respect to the radiographical standard references. Material and methods Design: This is a cross sectional study from a prospective database started in March 2003. Participants: 52 patients: 29 with JK, and 23 with thoracic hyperkyphosis (TK). Inclusion criteria: patients affected by JK or TK at first visit with a complete clinical, radiographical and surface topography evaluation. Groups. JK: lower limit of kyphosis below T12. Control group: subjects with a thoracic kyphosis radiographic measure exceeding 50°Cobb. Diagnostic tests used to detect JK:Clinical: plumbline distances: T12 < S1. Formetric criteria included the % of thoraco-lumbar inflexion point in trunk length over 60 %. Statistics: sensitivity, specificity, positive (PPV) and negative predictive values (NPV), by using diagnostic test vs the actual gold standard were calculated using a 2x2 table. Results The sensititvity of the plumbline distances of T12 < S1, in detecting JK in respect to radiographic criteria, resulted 55 %, with an accuracy of 46 %. The specificity of the test was 65 %, PPV 67 % and NPV 33 %. The sensitivity of the surface topography test resulted 73 %, as of the 29 patients with a JK x-rays diagnosis 22 showed a positive test, and only 7 without JK resulted negative. Therefore the specificity of the test was only 32 %. PPV and NPV resulted respectively of 40 % and 59 %. Conclusion The need for a useful criteria able to characterize JK to allow diagnosis and monitoring of the deformity is still lacking, and further studies will deepen this issue. O5 Usefulness of the clinical measure of trunk imbalance in adolescent idiopathic scoliosis Carole Fortin1, Erin Grunstein1, Hubert Labelle1, Stefan Parent1, Debbie Ehrmann Feldman2 1Université de Montréal, Research centre CHU Sainte-Justine, Montréal, Canada; 2Université de Montréal, Institut de Recherche en santé publique de l’Université de Montréal, Montréal, Canada Introduction Trunk imbalance, defined as a shift of the trunk in the frontal plane and measured with a plumbline from C7 to S1, is part of the clinical evaluation in adolescent idiopathic scoliosis (AIS). To date, little is known about evidences of this clinical measure in AIS. Objectives 1) To determine the reliability and validity of the clinical measure of trunk imbalance, 2) to assess its prevalence and 3) to explore the relationship between trunk imbalance and Cobb angle and with back pain in adolescents with AIS. Materials and methods Trunk imbalance measurements of 55 participants aged 10 to 19 years old with AIS (Cobb angle: 15° to 60°) were assessed by a physical therapist on two separate occasions. Markers placed on spinous processes of C7 and S1 were used to measure the horizontal distance between the plumbline placed at C7 and S1 with a rigid ruler. Cobb angle and trunk imbalance were measured on radiographs taken on the same day and the pain level was determined using the Numerical Pain Rating Scale (NPRS) and the Scoliosis Research Society-22 (SRS-22) pain score. Generalizability theory (f) was used to estimate the reliability and standard error of measurement (SEM) for the overall, test-retest design. Prevalence of trunk imbalance was given in percentage using the cut-off of 6.1 mm (minimal detectable change value). Pearson correlation coefficients (r) were used to assess validity of trunk imbalance compared with the radiographic method and to explore the association with Cobb angle and back pain. Logistic regression models also served to describe trunk imbalance (as a dichotomous outcome using several cutoffs: 10 mm, 15 mm and 20 mm) as a function of back pain. Results Trunk imbalance measured with a plumbline demonstrated high test-retest reliability (f : 0 .98 and SEM : 2.2 mm) and good correlation with measurements on radiographs (r :0.83, p : < 0.005).Trunk imbalance prevalence was 85 %. We found fair to moderate significant positive correlation between trunk imbalance and Cobb angle (r : 0.32 to 0.66, p < 0.05) but not with back pain. In the logistic regression model, there was a trend for trunk imbalance > 20 mm to be related with lower back pain. Conclusions The good psychometric properties and the high prevalence of trunk imbalance provide evidence for the usefulness of this clinical measure in AIS. Trunk imbalance can be easily measured in a clinical setting. A longitudinal study with a larger cohort is still needed to document the association of trunk imbalance with curve progression and back pain as well as the implications of the treatment of trunk imbalance on both Cobb angle and back pain. O6 Can ultrasound imaging be used to determine curve flexibility when designing spinal orthoses? Edmond Lou1, Rui Zheng1, Doug Hill2, Andreas Donauer2, Melissa Tilburn2, Jim Raso1 1University of Alberta, Alberta, Canada; 2Alberta Health Services, Alberta, Canada Background Spinal flexibility of a patient with Adolescent Idiopathic Scoliosis (AIS) affects in-orthosis correction. More flexible spines are better corrected in an orthosis which in turn should lead to better longterm outcomes. Curently, it is difficult to know how flexible a curve is during the orthosis design stage because radiographs are not done in order to minimize radiation exposure in growing children. Recently, ultrasound spinal imaging has been shown capable of measuring proxy Cobb angles and vertebral rotations reliably and repeatably in AIS patients. Design and level of evidence This pilot study investigated ultrasound imaging as a tool to provide spine flexibility in real time to assist the orthotists goal for the final in-orthosis correction. This is a Level III of Evidence study. Materials and methods AIS subjects prescribed full-time TLSO were asked to participate to this study. The inclusion criteria used the SOSORT brace management guidelines. Local ethics approval was received and all participants signed consent forms prior to participation. During the casting clinic, participants were scanned with ultrasound in the prone position. During scanning, participants with single right thoracic curves bent maximally to the right side while keeping their hips level and both shoulders in contact with the bed. An ultrasound (US) system with built-in position tracking was used to scan the spine along the spinal processes. An in-house program was used to reconstruct, display and measure proxy Cobb angles in real-time. Spinal bending correction (flexibility) was calculated as [Pre-orthosis X-ray Cobb – US Bending Proxy Cobb] / Pre-orthosis X-ray Cobb x 100 %. The final in-orthosis correction was calculated as [Pre-orthosis Cobb - In-orthosis Cobb] / Pre-orthosis Cobb x 100 %. Results Six participants (age: 13.9 ± 1.2 years) with 9 curves were recuited. The largest treated Cobb angles measured 39° ± 7° on a posterioanterior radiograph prior to casting. Curve flexibility averaged 74 ± 12 % (range: 56 % – 82 %). During casting, the orthotist used individual flexibility measures as targets for acceptable in-orthosis correction. Approximately 8 minutes were added to the clinic visit time: 1 minute to scan, 3 minutes to process and 4 minutes to measure the parameters. Recuritment rate was 100 %. At the following in-orthosis clinic, the orthotist used acheivment of 50 % – 70 % of the curves’ flexibility to determine the need for readjustment. To date, 3 participants have returned to clinic, the major treated flexibility versus the in-orthosis correction was 77 % vs 46 %, 81 % vs 36 % and 82 % vs 57 %. None of these subjects required adjustments to the orthosis. Discussion & conclusion Ultrasound imaging has the potential to provide radiation-free real-time measures of spinal flexibility but more study is requird to validate the process before it can be widely used. O7 Reliability of the Schroth curve type classification in adolescents with idiopathic scoliosis (AIS) Sanja Schreiber1, Eric Parent1, Greg Kawchuk1, Douglas Hedden2 1University of Alberta, Edmonton, Canada; 2University of Alberta, Alberta Health Services, Edmonton, Canada Introduction Schroth exercises are scoliosis-specific exercises aiming to improve postural alignment, control and stability of the spine. A classification system with four curve types is used to guide Schroth therapists in prescribing specific exercises for patients with scoliosis. This classification should be reliable to assure appropriate therapy delivery. We developed a rule-based algorithm to assist in reliably classifying patients. The aim of this study was to determine the intra- and inter-therapist reliability in classifying patients with AIS using our proposed classification algorithm. Design An international intra- and inter-rater reliability study. Material and methods We recruited 44 consecutive volunteers with AIS, aged 10 to 18, with curves between 10°-50° from a scoliosis clinic and 10 consecutive English-speaking volunteers from the international registry of certified Schroth therapists. The patients’ standing posture from each side and the Adam’s forward bend test were videotaped by the primary author. Therapists reviewed a manual with operational definitions, rated and reviewed four practice cases streamed from the study website illustrating each of the four Schroth curve types as training for using the algorithm before the study started. After the training period, the therapists, blinded to participants’ identity, rated video assessments presented randomly on two occasions at least seven days apart. The intra- and inter-rater reliability estimates were calculated for the entire sample of therapists (N = 10), the therapists who reported having full understanding of the algorithm (well-trained, N = 6), and the therapists who had conceptualized and used the algorithm in a randomized controlled trial (experienced, N = 2). Gwet’s AC1 and weighted AC1 coefficients were used to calculate the reliability. A weighted analysis was justified. The 3c and 4c Schroth curve patterns share thoracic curves and balanced pelvis and their exercise prescription does not differ as drastically (assigned weight =0.5) as between 3cp vs. 4cp (opposite pelvis corrections), 4cp vs. 3c (emphasis on pelvis and lumbar vs. thoracic curves) or 3cp vs. 4c curve patterns (emphasis on pelvis and thoracic vs. thoracic and lumbar curves only), for which pairs the assigned weight was 0. Results Patient’s age was 14.2 ± 2.0 years and their mean largest Cobb angle was 25.8° ± 10.0o. Based on the experienced rater’s ratings, there were nine 3c, 12 3cp, six 4c and 17 4cp curve types. The overall intra-rater AC1 was 0.64 (95 % CI 0.53-0.73), 0.70 (95 % CI 0.60-0.78) among well-trained raters, and 0.81 (95 % CI 0.77-0.85) in experienced raters. The weighted intra-rater AC1 averaged 0.75 (95 % CI 0.63-0.84) overall, 0.82 (95 % CI 0.73-0.88) in well-trained raters, and 0.89 (95 % CI 0.80-0.94) in experienced raters. Inter-rater AC1 was 0.43 (95 % CI 0.28-0.58) overall, 0.50 (95 % CI 0.38-0.61) for well-trained raters, and 0.67 (95 % CI 0.50-0.85) for experienced raters. The weighted inter-rater AC1 was 0.48 (95 % CI 0.29-0.67) overall, 0.61 (95 % CI 0.49-0.72) among well-trained, and 0.79 (95 % CI 0.64-0.94) among experienced raters. Conclusions A high level of understanding of the algorithm improved the intra- and inter-rater reliability justifying future refinement of the training. Weighted analysis demonstrated adequate intra- and inter-rater reliability, suuporting usage of the proposed algorithm in raters reporting full understanding. O8 Can Trunk Appearance Perception Scale (TAPS) be used as a descriptive tool of scoliosis severity? Judith Sánchez-Raya, Antonia Matamalas Adrover, Elisabetta D'Agata, Joan Bagó Granell Vall d'Hebron Hospital, Barcelona, Spain Background The Trunk Appearance Perception Scale (TAPS) is a valid instrument for evaluating the patient perception of their trunk deformity. There are no studies that evaluate the validity of TAPS when used by doctors to assess trunk deformity. Design and level of evidence Correlational study to assess inter and intra-observer reliability for usage of TAPS instrument. Material and methods The sample consisted of 32 patients (26 females), with a mean age of 15.56 and a mean Cobb Angle of 40.17° (ranging from18° to 74°). Mean TAPS was 3.33. Patients were also given access to the TAPS instrument too. For each patient, three photographies were made (anterior, posterior and Adam’s test position). Three specialists in scoliosis evaluated the three photographies, according to the TAPS scoring. One week later, the three evaluators assessed again the photographies. Results The mean results of the first TAPS evaluations were: 3.67 (E1), 3.67 (E2), 3.78 (E3). The ICC (average measure) result for inter-observer reliability was .89 (p < 0.001). The Maximum Cobb Angle correlation was -.55 (E1), -50 (E2), -.51 (E3). The means of the second evaluations were: 3.85 (E1), 3.71 (E2), 3.79 (E3). ICC for intra observer reliability was ICC = 0.96. Correlation between Cobb angle and Patient TAPS was r = -.433 (p < 0.05), while no significant correlations were found between Patient TAPS and Observer TAPS evaluations. Significant and high correlations were found between Maximum Cobb Angles and Observer TAPS, ranging from r = -0.46 and r = -0.58 (p < 0.005). Conclusion When scored by clinicians, TAPS presented a satisfactory inter-observer and intra observer reliability and good correlation with Maximum Cobb Angle. This being said, this tool proves to be an effective and accurate tool when describing curve severity. O9 Magnitude of the Cobb angle on an X-ray in relation to the angle of trunk rotation in children who come to the “Troniny” Scoliosis Treatment Centre Marek Kluszczynski 1, Anna Kluszczyńska 1, Jacek Wąsik 2, Marta Motow-Czyż 2, Adam Kluszczyński 1 1“Troniny” Children Rehabilitation Center in Częstochowa, Częstochowa, Poland, 2Institute of Physiotherapy, Jan Długosz University, Częstochowa, Poland Introduction The limiting value of a back asymmetry measured in a child, authorising a therapist to send the child to have an radiograph taken, is 7o ,according to the SOSORT consensus. When a child comes to a scoliosis therapy centre with a radiograph, then it is possible to determine a value of the Cobb angle and compare it to the ATR angle. The objective of the study is evaluation of Cobb angles in a group of children with ATR angles within the 4-6o range examined in the clinic for the first time. The analysis provided below is to draw attention to the rightfulness of taking up prevention and X-ray testing in the group of children with back asymmetry below 7o. Design and level of evidence Cross-sectional Study with verification by reference (gold) standard. Material and methods The material were 117 children (30 %) from among 351 treated in the Centre within the last 5 years, who had a radiograph of their backs taken on their medical appointment. On the basis of the radiograph of the spine, the Cobb’s angle, rotation of the vertebra, acc. to Cobb were determined and the degree of skeletal maturity was evaluated by means of the Risser sign. Moreover, the examination included: evaluation of the angle of lumbar lordosis and thoracic kyphosis with the use of the Saunders inclinometer as well as evaluation of the angle of trunk rotation (ATR) with the use of the Bunnell scoliometer. The screened children were 6-17 years of age, on average 12.6 +/- 1.9 y. o. a. . The group included 64 % girls and 36 % boys, respectively. Results of the measured Cobb angle values and ATR angle values for both groups are compared. Materials have been statistically analysed. Contingency tables have been made and chi-square tests have been calculated. Statistically significant results was assumed to be p < 0.05. Results. The group of children with ATR angle values lower than 7o included 59 % (69) members of the entire group, whereas the group of children with ATR angle values higher than 7o included 41 % (48) members of the examined group. In each group, the number of children having curvatures within the 3 ranges of magnitude of the Cobb angle 10-14o, 15-20o, and over 20o, was determined. Thus, the group with the ATR angle lower than 7o included 24 (34.8 %) children with Cobb angles 10-14o, 23 (33.3 %) children with Cobb angles 15-20o, and 22 (31.9 %) children with Cobb angles over 20o. The group with the ATR angle higher than 7o included 10 (20.8 %) children with Cobb angles 10-14o, 8 (16.7 %) children with Cobb angles 15-20o, and 30 (62.5 %) children with Cobb angles over 20o. There were significant statistical differences in distribution between individual groups (Chi-square = 10.833; df = 2; p ° = 0.004). A higher proportion of group members within the in the 10-14o and 15-20o Cobb angle ranges was observed in the group of children with the ATR lower than 7o. There was a higher proportion of >20o Cobb angles in the group of children with the ATR higher than 7o. Conclusions.The frequency of incidence of scoliosis with the value of the Cobb angle 15o and higher in the group of children referred to the Scoliosis Treatment Centre due to spinal asymmetry lower than 7o ATR was 37 %. Decisions about starting scoliosis prophylaxis and taking a radiograph of spine of children younger than 12 years of age with diagnosed spinal asymmetry of ATR 4-6o value should be considered individually. O10 Cobb angel measurement without X-ray, a novel method Ane Simony1, Karen Hojmark Hansen1, Hanne Thomsen2, Mikkel Meyer Andersen3, Morten Vuust2 1Sector for Spine Surgery & Research, Middelfart Hospital, Middelfart, Denmark; 2Department of Radiology, Frederikshavn Sygehus, Frederikshavn, Denmark; 3Department of Mathematics and Statistics, Aalborg University, Aalborg, Denmark Background X rays have been the Golden Standard for evaluation of development and progression of scoliosis, for many years. Cobb angel measurement is the most important tool, to determine curve progression and effect of treatment. The patients are children or adolescent and standard x-rays of the spine expose the breast area, the thyroid, and the gonads, with ionizing radiation. Increased incidence of cancer is observed among patients, treated for adolescent idiopathic scoliosis [1,2]. Previous studies with measurement of the trunk rotation, rotation of the spine etc. has not been to create a method to determine Cobb. Aim To validate the accuracy of The Manual Method, against convention radiographs. Methods We hypothesised, that by manually marking the spinous process and take a photograph of the full spine, we where able to measure Cobb. This study is a validation of this method, and a comparison of this method, and Cobb measurements in X-ray.130 consecutive patients, referred to standing x-ray of the spine, were invited to participate in this study. 78 patients fulfilled the inclusion criteria. Before x-ray, the Spinous processes where manually palpated from T1 to S1, and marked with a pen. The patient was placed for X-ray, and the photo was taken with the patient standing in exactly the same position, as the AP X-ray. Marking and photographs where taken by a Research nurse, and staff from Radiological Department. X-rays were evaluated by 2 independent doctors, and the photographs were evaluated by the same 2 doctors, 2 weeks later. The measurements where evaluated by an independent statistician. Inter and intra observer variation was evaluated, and the difference between X-ray and Photo was evaluated. Results For the thoracic curves, the mean difference was 6.9 (p value < 0.0001), such that on average, the angle measured with x-ray was 6.9 degrees larger than that measured with photo. The Pearson correlation between x-ray and photo angle was 0.58 (p value < 0.0001).For the thoracolumbal curves, the mean difference was 5.2 (p value < 0.0001). The Pearson correlation between x-ray and photo angle was 0.66 (p value < 0.0001).In the lumbar group, only 7 patients participated. This is not enough to evaluate the methods feasibility, and these results are not presented. Conclusion By this study, is seems possible to evaluated Cobb, by a manually method. The method has been proven successful in thoracic and the thoraco-lumbar region. Further examination is needed, to evaluate if this method is useable in the lumbar region as well.Furthermore studies with continuous measurements are needed, to ensure this method can be used to determine progression as well. References 1. Bone CM, Hsieh GH. The risk of carcinogenesis from radiograhs to pediatric orthopaedic patients. J Pediatr Orthop. 2000 Mar-Apr;20(2):251-4. 2. Levy AR, Goldberg MS, Hanley JA, Mayo, Poitras B. Projecting the lifetime risk of cancer from O11 The postural tone magnitude and distribution in patients diagnosed with an adolescent idiopathic scoliosis: a preliminary study Irmina Blicharska1, Jacek Durmała1, Bartosz Wnuk1, Małgorzata Matyja2 1School of Health Sciences in Katowice, Medical University of Silesia, Chair and Department of Rehabilitation, Katowice, Poland; 2Chair of Kinesitherapy and Special Methods of Physiotherapy, Academy of Physical Education, Katowice, Poland Background The scientific researches confirm the relationship between abnormal postural patterns in infants and defects occurring in the later stages of life. During the child’s development, the central nervous system is developed and integrated. Manifestation of its functioning is reflected in postural and motor patterns as well as parameters characterizing postural tone. An accurate postural tone is strictly connected with the central nervous system functioning, the body stabilization and the relation between mobility and stability. It also determines an accurate arrangement of the body segments. Due to its disorder, there are compensation changes. By adjusting the child’s movement possibilities to his/her own abilities, conditioned by abnormally functioning CNS, it leads to the secondary, neuro-orthopedic changes in the body posture. Material and methods 37 subjects of both sexes, aged x = 12.8 ± 1.53 were qualified to a prospective study. The study group (A) was constituted by patients diagnosed with an adolescent idiopathic scoliosis, who have not been treated yet. The average value of the primary curvature was x = 21.4°. The criterion for exclusion from this group was the scoliosis etiology other than unknown or intensive, regular rehabilitation. The control group (B) was formed by healthy children, who underwent a screening test that did not show any signs of scoliosis and other significant deviations within the body posture. The quantitative assessment of parameters associated with the postural tone was measured with the calculator of the magnitude and distribution of the postural tone. Twenty-one parameters, i.e. a functional length of muscle groups or the body posture components, were entered into the calculator. On the basis of such parameters, the program automatically determines the values of the Postural Tone Index (that define magnitude of postural tone), also the values of the Spastoidal Tone Index and Athetoidal Tone Index (that define distribution of the postural tone). The angle of trunk rotation (ATR) was determined with Bunell’s scoliometer. This tool was also used in a screening test (in the control group) in order to exclude the scoliosis occurrence. The body posture quality was assessed by means of Kasperczyk’s Scale. The results were analyzed in the Statistica v.10. A compliance distribution of the variables with a normal distribution was performed with Shapiro-Wilk’s test. The correlation between parameters was determined on the basis of R-Pearson’s test. The comparison of the measurable parameters between groups was performed by means of U-Mann-Whitney’s test. The value p <0.05 was determined as a level of statistical significance. Results In the group of patients with scoliosis, the average value of the Postural Tone Index was x = 0.349, while in the control group x = 0.230. Values closer to 0 indicate the better quality of postural tone. The difference between groups is statistically significant (p <0.05). There was no correlation between the value of the Cobb’s angle in the primary curvature and the value of the Postural Tone Index. Conclusions The subjects, with the adolescent idiopathic scoliosis, demonstrated a significantly reduced postural tone. The study requires a follow-up and further supplementation. The study group is too small to assign occurring disturbances to the population of people with scoliosis. O12 From studies on the function of the respiratory system in children with body posture defects Andrzej Szopa1, Małgorzata Domagalska-Szopa2, Weronika Gallert-Kopyto3, Tomasz Łosień2, Ryszard Plintla4 1School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Physiotherapy, Katowice, Poland; 2School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Medical Rehabilitation, Katowice, Poland; 3School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Physiotherapy, Katowice, Poland; 4School of Health Sciences in Katowice, Medical University of Silesia in Katowice, Department of Adapted Physical Activity and Sport, Katowice, Poland Background Scoliosis, despite of being a deviation of spine’s anatomical axis from its mechanical axis especially in the lateral direction, also constitutes a faulty posture in the frontal plane. Although the lateral curvature dominates, multi-layered and multi-segmented nature of scoliosis is emphasized. Signs of scoliosis also pertain changes in regard to anteroposterior curvatures, rotation and torsion of vertebrae in the transversal plane, and they also directly impact the positioning and shape of other sections of the locomotor system. Numerous studies which concern with functioning of the respiratory system in scoliosis, were focused mainly on the value of the spine's lateral curvature angle. Design and level of evidence Replacing measurements of the angle value of the spine's lateral curvature with analysis of body posture based on the moiré topography (MT) allows the assessment of nearly all scoliosis symptoms combined with spirometric measurements. The purpose of presented study was to provide answers to the following questions: 1. Do restrictive ventilatory defects occur in children with scoliosis? 2. Which features of body posture have a decisive influence on the functioning of the respiratory system? Material and methods The study involved 68 children with idiopathic scoliosis, aged 9–15. All subjects met the following criteria: (1) older than 7 years of age, (2) able to follow verbal directions, (3) mild scoliosis (angle of vertebral lateral curvature < 25°), (4) no previous surgical procedures. Basic elements of the conducted research included: body posture examination, based on moiré topography (MT) in standing and spirometric measurement. The MT examination was performed by using a CQ Electronic System (Poland). The spirometric examination was performed on a Micro Lab MK8 Viasys spirometer, while chest expansion was assessed by means of chest circumference measurement using measuring tape. Results For the majority of patients with mild scoliosis, the obtained results, including the values of the basic ventilation indicator (i.e. the percentage of due lung vital capacity- VC%), were within norm and did not confirm the existence of features characteristic for restrictive ventilatory defects. Moreover, no dependency was stated between basic functioning parameters of the respiratory system, and features characteristic for scoliosis, i.e. the primary curvature angle value and degree of rotation. Meanwhile, the obtained results indicated that in these children the VC values depend on the shape of thoracic kyphosis, determined by: depth (TKD), length (TKL) as well as the length/depth indicator (TKI%). The calculated correlation coefficients (r- Pearson) showed that the ventilation efficiency was not deteriorated due to deepened thoracic kyphosis but, to the contrary, VC% decreased together with shallowing of the curvature, which usually accompanies scoliosis. Conclusions Research results, as well as their analyses presented in this study, are not meant to negate the existence of function defects of the respiratory system in idiopathic scoliosis, but merely provide some evidence which undermine the unambiguity of presented problem. O13 Scoliosis as the “first” sign of various diseases Franz Landauer, Karl Vanas University Clinic of Orthopedics (PMU), Salzburg, Austria Background The success or failure of any brace treatment is closely linked with thecause of scoliosis and concomitant diseases. Design and level of evidence This observational study is designed like a case control study to differentiate causes of scoliosis. Material and methods 285 patients referred as adolescent “idiopathic” scoliosis and indicated for bracing (SOSORT criteria) were followed over 5 years. They were searched at every appointment for changes in the spine (bones, connective tissue, nerves and muscles) and for diagnoses with impact on scoliosis (hormonal disorders, operations in the early childhood, etc.). Results Pathologies with a direct or indirect effect on the development or progression on scoliosis were found in nearly 19,3 % (n-55) of the patients. In 17 cases leg-length discrepancy >1 cm was the cause (2 after trauma of the growth plate, 1 fibrous dysplasia, 1 hypoplasia of the fibula, 1 hypoplasia of the femur, 2 chronic slipped capital femoris, 1 unknown Perthes disease could be found).In 10 cases a syndrome could be found (3 Marfan Sy., 1 Ehlers Danlos Sy., 2 Neurofibromatosis, 4 Prader Willi Sy., 1 Moebius Sy.). An operation in the early childhood was done in 4 cases (3 heart operations, 1 atresia of the esophagus). Malformations could be dedected in 7 patients (5 Hemisacralisation, 1 Patella Nail Syndrome, 1 M. pectoralis major hypoplasia). 4 patients showed a spondylolysis or developed a spondylolisthesis. Different tumor or tumor like lesions could be found in 3 cases (1 Medulloplastoma, 1 aplastic anemia, 1 eosinophilic granuloma). Hormonal disorders could be dedected in 7 cases (3 Hashimoto thyreoiditis, 1 Hypothyreoiditis, 3 growth hormone treatment). Also 1 case with a massive disc herniation, 1 CRMO and 1 case after meningitis in the early childhood could be dedected. Conclusion The compensation of leg length difference is the basic requirement before brace treatment.The problem of lumbosacral transition fault is still too little attention. The current MRI-technique is not sensitive enough to differentiate all causes. But the MRI-scan is necessary to detect different diagnosis in the spine.At every appointment history of the patient and her family should be part of the examination. Sometimes it takes years to find the right diagnosis. Especially to diagnose different syndromes can also be expensive.The improvement of the examination lowers the number of “idiopathic” scoliosis. O14 The effectiveness of core stabilization exercises versus conventional exercises in addition to brace wearing in patients with adolescent idiopathic acoliosis Gozde Gur1, Necdet Sukru Altun2, Yavuz Yakut1 1Hacettepe University, Çankaya, Ankara, Turkey; 2Akay Hospital, Ankara, Turkey Introduction Conservative treatment of Adolescent idiopathic scoliosis (AIS) involves a variety of physical exercises and bracing. For some of these treatments there is insufficient evidence. The aim of this study was to investigate the effects of core stabilization exercises versus conventional exercises in addition to brace wearing on trunk asymmetry, perception of deformity and health related quality of life in patients with AIS. Material and methods Nineteen female subjects with AIS were randomly placed into two treatment groups: Group 1 (n = 9, mean age 14,4 ± 2,1 years) core stabilization exercises (CSE) and brace; Group 2 (n = 10, mean age 14,1 ± 1,5 years) conventional exercises (CE) and brace. The average Cobb angle of the major curve was 32,2 ± 11,8° for thoracic (range: 14°–54°), 27,5 ± 6,9° for lumbar regions (range: 16°–34°) for the first group, 32,6 ± 7,2° for thoracic (range: 19°–40°), and 36,8 ± 7,9° for lumbar regions (range: 23°–50°) for the second group. The trunk asymmetry was assessed by Posterior Trunk Symmetry Index (POTSI), perception of deformity of physiotherapist and patient by Walter Reed Visual Assessment Scale (WRVAS), spinal rotation degree by scoliometer with Adam’s forward bend test, and health related quality of life by SRS-22. Measurements were carried out at baseline examination, following the completion of treatment. Exercises were performed one hour a day and seven days per week for ten weeks. Two sessions of exercise treatment were performed in the clinic by a physical therapist with three day intervals and between these intervals patients performed the same exercises at home. Patients were instructed to wear brace 23 hours each day. Results were analyzed using The Wilcoxon rank-sum test to compare repeated measurements and Mann-Whitney U test to compare two groups. Results Both thoracic and lumbar rotation was improved in CSE group (p < 0,05). In CE group, only lumbar rotation was improved after a ten-week treatment period. Also, improvement in lumbar rotation was better in CSE group (p < 0,05). Physiotherapists’ WRVAS total scores were better after treatment in both groups (p < 0,05). But there was no difference between groups. Patient’s WRVAS total scores did not change after treatment (p > 0,05). POTSI scores were improved in both group but there was no significant difference between groups (p > 0,05). There was no difference in both groups after treatment in terms of SRS-22 subgroups, self-image, mental health and total score except for function and pain in CSE group. Function and pain improved with CSE treatment. Conclusion We concluded that core stabilization exercises therapy improves trunk symmetry, spinal rotation and rib hump, function and pain in addition to brace treatment in patients with AIS. Further studies with more cases are needed to demonstrate effectiveness of this method in AIS conservative treatment. O15 The effect of physiotherapy techniques on the body balance in patients with scoliosis treated with corrective appliances Piotr Gawda, Piotr Majcher Uniwersytet Medyczny w Lublinie, Lublin, Poland Background The body vertical orientation in the upright standing position is maintained by keeping the body center of gravity (COG) upright above the base of support by a dynamic interplay of visual, vestibular, and somatosensory control systems. The relationship between balance control and independent mobility is particularly important in the population of individuals with severe scoliosis who have difficulties in balance maintenance and a decreased confidence in performance of daily activities. The control of upright stance can change during conditions of increased postural tension which limits the synergy of muscles system. Central nervous system receives a very large number of stimuli information coming from myofascial structures. The application of external forces that arise using corrective appliances leads to stimulation of mechanoreceptors of these structures, which affects the current balance control of the body. The force platform technique is one of the most frequently used quantitative techniques for postural control assessment that enables the measurement of the COG sway velocity . Objectives of this study are to point out the impact of the braces treatment on the postural control strategy in patients with scoliosis and to estimate the influence of the stretching therapy on the postural control strategy in patients with scoliosis treated with corrective appliances. Materials and methods The studied group consists of 24 girls between the ages of 10 to 15 years old. All diagnosed with idiopathic scoliosis and qualified to treat with corrective appliances. Twelve of them (group A) had been prepared for appliances corrections by rehabilitation with the use of stretching techniques, the rest (group B) without. Cobb angle of the respondents ranged from 25 degrees to 46 degrees. Postural characteristics of the subjects were measured with the use of a computerized force platform. The software program filters the center of pressure data and then calculates COG. The mCTSIB assesses a person’s ability to use sensory inputs for balance and distinguishes between normal and abnormal balance performance. This test measures COG sway velocity while standing in the different conditions. All evaluations were compared in two groups. All variables had normal distribution and were analyzed with parametric statistical tests. The differences of continuous variables among patients’ groups were determined with T test. The Fisher exact test was used to test the association for categorical data. Results with p-values <0.05 were regarded as statistically significant. Results The individuals with corrective appliances showed a significantly higher mean COG sway velocity as compared to the same patients before appliances. The increase in speed of COG deviation was positively correlated with the size of the angular correction of the curvature. Selected rehabilitation treatment with using stretching techniques reduces COG sway velocity increase after the appliance of the corrective braces. Conclusion Postural stability is decreased in patients with scoliosis just after applying the corrective braces. In order to maintain postural stability after the appliances a series of physiotherapy techniques, exercises of the antigravity muscles to prepare for the correction of soft tissue are necessary. O16 New combine method treating AIS – preliminary results Lior Neuhaus Sulam The Israeli Scoliosis Center, Tel Aviv, Israel Background The ApiFix is a novel non-fusion system of treatment for AIS. The system involves an expendable rod attached to the spine by only 2 screws, inserted around the apex of the major curve. The system is designed to gradually increase its length when the patient is doing exercises which increases the distance between the two screws considering the other curves. The goal of the system is to act as an “Internal Brace”, reduce the curve to below 35 degrees and maintain it at that level. The surgical procedure is significantly less invasive compared to the long fusion. The incision is around 10 cm, operation time around 1 hour and after 2-3 days the patient goes home. 40 patients were enrolled so far, 20 of them in controlled clinical trials in Europe. Aims To present a new method for treating AIS patients classified as Lenke 1 with a flexible curve, up to 60 degrees Cobb angle. The method comprises a small surgery combined with specific physiotherapy. Preliminary results of 9 cases. Design Retrospective case report of 9 girls treated along the past 2 year. Material and methods 9 females, age 12.5-24, most passed growth spurt, with main Rt thoracic 410-640 cobb curve had Apifix operation with specific PT program 1 month before op and 6 month after the op including home exercise program. Assessment made by x-rays in standing and side flex, clinical assessment before op, and standing xrays with clinical assessment after the op, 2 weeks, 4 weeks, 3 m and 6 m. Results The improvement of the cobb was between 10 %-56 %, average of 31 % of correction, The improvement of the TRACE was 28 %-45 %, average 38 %, The improvement of the ATR was 9 %-42 %, average 27 %. Conclusion The main purpose of the this treatment is to stabilize and support the spine preventing progression in adult life without limiting the normal movement of the spine, decrees torsion forces, less muscle power against gravity, better cosmetics and less surgery complications. Reducing the Cobb under 350, the TRACE and the ATR, will help the patient to achieve this aims. It is important to follow the inclusions criteria to this procedure to have the best results. In addition it is crucial to start Physiotherapy treatment at least one month before op, the 2 girls in the bottom of the table which were the first operated in Israel start the PT treatment only after the op including the clinical assessment. There is a need to have longer follow up to validate stable results in adulthood. O17 Does a 4-week intensive course of ScolioGold therapy reduce angle of trunk rotation in scoliotic patients: a retrospective case series Michael Bradley1, David Glynn2, Alex Hughes1, Erika Maude1, Christine Pilcher1 1Scoliosis SOS Clinic, London, Great Britain; 2University of York, York, Great Britain Background Angle of Trunk Rotation (ATR) affects rib and lumbar prominences, which can significantly influence a patient's back shape, and may negatively impact body image and self-esteem. Various studies have also linked ATR to Cobb angle for prediction of curve progression without the need for repeated radiation exposure from X-rays, with varying degrees of success. Therefore, ATR remains an important clinical outcome measure in practice. The Bunnell Scoliometer is widely used as a basic method for measurement of ATR. ATR is the angle between the horizontal and the plane across the back at the greatest elevation of a rib prominence or lumbar prominence. Design and level of evidence The study was a retrospective case series in which 305 patients (47 males and 258 females) treated between December 2011 and June 2014 were measured for vertebral rotation by a ScolioGold Therapist using the Scoliometer in a standardised forward-bend position at the beginning and end of a 4-week intensive course of ScolioGold treatment. Patients were aged between 8 and 76 years old (mean 26.6), and were only included if at least one curve had rotation > =5o as measured by Scoliometer at start of treatment. There was no randomisation of patients. Material and methods Scoliometer readings were taken with each patient at the start and end of their 4-week ScolioGold treatment course. Each measurement was taken in a standardised seated forward-bend position, with knee height being consistent at both time points, and the same Scoliometer used. These results were then documented and anonymised before the values were analysed by an independent statistician. Paired t-tests were used to evaluate the difference between sum of ATR at start and end of treatment. ATR is a reliable measurement with good reproducibility. Previous studies have suggested that the Scoliometer has excellent intra- and inter-observer agreement, with a change of 2 degrees reported previously to be clinically significant. Results In the cohort measured, average sum of total ATR reduced from 18.04 degrees (SD 6.70) to 14.30 degrees (SD = 6.15). Single thoracic curvatures (n = 48): Mean reduction in ATR of 2.33 degrees (SD 0.42, P < 0.05). Thoracolumbar curvatures (n = 54): Mean reduction in ATR of 3.05 degrees (SD 0.56, P < 0.05). Double curvatures (n = 190): Mean reduction of ATR of 1.71 degrees (SD 0.25) and 1.95 degrees (SD 0.31) in thoracic and lumbar respectively (P < 0.05). Post-spinal fusion (n = 13): Mean reduction of ATR of 4.19 degrees (SD 1.54, P < 0.05). There is a statistically significant difference in the sum of ATR before and after treatment with the ScolioGold method. Conclusion ScolioGold therapy proved effective at reducing ATR magnitude for this case series to both a clinically and statistically significant degree in single thoracic or thoracolumbar curvatures. There was a statistically significant reduction of sum of ATR in double curvatures, but the literature is divided as to whether this change was clinically significant or not. These results were not statistically affected by whether treatment was completed as one 4-week block, or if they were split into two 2-week treatment blocks. O18 Schroth physiotherapy method without bracing is an effective treatment for scoliosis in improving curves and avoiding surgery and should be offered as a treatment option for scoliosis in Canada: case series Andrea Lebel1†, Victoria Ashley Lebel1,2†, Judit Orbán1 1Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8, Canada; 2Saba University School of Medicine, Saba, Dutch Caribbean, Netherlands Correspondence: Andrea Lebel – Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8, Canada †These authors contributed equally to this work Background Idiopathic scoliosis (IS) is a complex multifactorial three-dimensional (3D) spinal deformity of unknown cause. The treatment options offered for IS by orthopedic surgeons in Canada are observation for curve progression, bracing, and spinal fusion surgery. Physiotherapeutic scoliosis-specific exercises (PSSE) are currently not recommended by orthopaedic surgeons in Canada, even though it has been proven effective in preventing scoliosis curve progression and in a number of cases, reducing scoliosis curve angles (measured in Cobb degrees). Schroth physiotherapy can be effective in reducing the number of braces prescribed in Canada and in reducing the number of spinal fusion surgeries. The purpose of this study series is to evaluate the effect of an outpatient Schroth physiotherapy program in patients with IS and high risk of progression and who have not received bracing as treatment, by following primary curve degree Cobb angles and the angle of trunk rotation (ATR) measurements based on initial and follow-up radiographs and scoliometre measurments. Methods This retrospcetive case series includes 6 female patients ages 5-23 years diagnosed with IS. All study patients required an initial diagnostic radiograph, taken no earlier than 6 months before beginning Schroth physiotherapy, and a follow-up radiograph, taken within 6 months of completing Schroth physiotherapy. None of the case series patients received any form of bracing treatment prior to or during this study. Results In the period between initial diagnostic radiographs and follow-up radiographs, primary scoliosis curves showed an improvement of 10-19 degrees Cobb angle and ATR measurements improved by 2-8 degrees. Two years after the follow-up radiographs were taken, the ATR measurements of all 6 patients have remained stable. None of the patients required surgery. Conclusions Schroth physiotherapy without bracing is an effective treatment option for IS even in patients with a high risk of curve progression. A Schroth physiotherapy exercise program can improve scoliosis curve Cobb angles and ATR measurements, eliminating the need for surgery and/or bracing in a number of cases, as well as decrease the risk of scoliosis curve progression into adulthood. O19 Rotation of the trunk and pelvis and coupled movements in the sagittal plane in double support stance in adolescent girls with idiopathic scoliosis Agnieszka Stępień, Krzysztof Graff Józef Piłsudski University of Physical Education, Warsaw, Poland Introduction Scientists are still looking for causes of scoliosis and its progression. Trunk and pelvic movements in the transverse plane were evaluated only in a few studies [1,2]. A few authors indicated gait pattern as a reason of scoliosis progression [3]. Aim The aim of this study was to determine the trunk and pelvis rotation range of motion (TR - trunk rotation, PR - pelvic rotation) in adolescent girls with idiopathic scoliosis (AIS) in a position imitating the double-support phase (DSP) of gait. The additional aim was to describe angular motions in the sagittal plane (MSP) occurring during rotation. Methods 59 AIS girls (age 10-18, average 14,4) with the right thoracic curve or/ and the left lumbar curve were subsequently qualified to the study. Four groups including girls with different types of scoliosis were formed. Measurements were taken in the standing position imitating DSP. A special designed prototype axial rotation tester with the computer system was used to assess TR and PR and coupled MSP. The shoulder girdle with upper part of the trunk was stabilized during pelvic movements. The pelvis was fixed during trunk rotation motions. Special sensors were used to control feet motions. The number and order of motions were precisely determined. Right TR in the position with the right lower limb in front was compared to left TR in the position with the left lower limb in front. Right PR in the position with the left lower limb in front was compared with left PR with the right lower limb in front. 30 healthy girls without scoliosis were tested as the control group. ANOVA test and T-test were used for statistical analyzes. Results Significant difference between the right and left TR was found in girls with double curve scoliosis with the dominant thoracic curve. Left TR was significant lower than TR to the right. Differences between the right and left PR were not observed in groups. PR to the right and left in girls with the lumbar curve was significant larger than in other groups. TR was coupled with characteristic MSP. Left TR was coupled with forward trunk movement and right TR was associated with trunk backward movement in the majority of participants. The posterior pelvic tilt was observed during the left PR in girls with the lumbar curve. The increased anterior pelvic tilt appeared in girls with the single thoracic curve during the right PR. Conclusions TR / PR values and coupled MSP depend on a scoliosis type and direction of rotation. It is important to pay attention during physiotherapy to coupled spine movements which occur during rotation. Observed differences can be one of causes of gait pattern asymmetry in AIS. This hypothesis needs confirmation. References 1. McIntire Kevin L, Asher Marc A, Burton Douglas C, Liu Wen. Trunk rotational strength asymmetry in adolescents with idiopathic scoliosis: an observational study. Scoliosis. 2007; 2: 9. 2. Stępień A. A range of rotation of the trunk and pelvis in girls with idiopathic scoliosis. Advences in Rehabilitation 2011, (3), 5-12. (article in Polish) 3. Burwell RG, Cole AA, Cook TA, Grivas TB, Kiel AW, Moulton A, Thirwall AS, Upadhay SS, Webb JK, Wemyss-Holden SA, Whitwell DJ, Wojcik AS, Wythers DJ: Pathogenesis of idiopathic scoliosis. The Nottingham concept. Acta Orthop Belg 1992, 58:33-58. O20 Curve progression analysis in Risser 0 patients orthotically managed with compliance monitors D. Speers Scheck and Siress, Chicago, IL, USA Background Bracing for adolescent idiopathic scoliosis is the major modality for conservative care [1]. Recognizing, and treating, scoliosis at a young age has shown to be beneficial for successfully, conservatively managing scoliosis and not letting it progress to surgical levels. Increased dosing has shown to proportionally correlate to success levels with bracing. Aim To see if curves remained stable or decreased in the Risser 0 population while undergoing brace treatment in out of brace x-rays (72 hours out of brace). Methods Eight children with idiopathic scoliosis and Risser 0 maturity managed with a tlso had out of brace x-rays taken (72 hrs out of brace) to document curve status. The curves were categorized as increased (greater than 5 degrees progression), within error (within 5 degrees above or below initial curve measurement), or decreased (greater than 5 degrees regression). Average time in brace 14.01 hours. Results For the eight Risser 0 patients, two had curves that decreased, four were within error, one increased and one is to have out of brace x-rays within the next month. Conclusion Bracing the Risser 0 population appears to slow progression and even reverse scoliotic curves. References: 1. Effects of bracing in adolescents with idiopathic scoliosis. Weinstein SL, Dolan LA, Wright JG, Dobbs MB. 2. Brace wear control of curve progression in adolescent idiopathic scoliosis. Katz DE, Herring JA, Browne RH, Kelly DM, Birch JG. 3. Validation of a miniature thermochron for monitoring thoracolumbosacral orthosis wear time. Benish BM, Smith KJ, Schwartz MH. O21 Conservative treatment in Scheuermann's kyphosis: comparison between lateral curve and variation of the vertebral geometry Angelo Gabriele Aulisa1, Vincenzo Guzzanti1, Giuseppe Mastantuoni1, Andrea Poscia2, Lorenzo Aulisa3 1U.O.C. of Orthopedics and Traumatology, Children's Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy; 2Institute of public health, University Hospital "Agostino Gemelli", Catholic University of the Sacred Heart School of Medicine, Rome, Italy; 3Department of Orthopedics, University Hospital "Agostino Gemelli", Catholic University of the Sacred Heart School of Medicine, Rome, Italy. Background Conservative treatment of vertebral deformity promotes with the application, of external forces to obtain, via appropriate geometry orthosis, during skeletal growth, remodelling of the deformed vertebras. In a previous paper on Scheuermann's kyphosis, we have studied the geometry variations of all vertebrae included in the curve, before and after the treatment. Design The purpose of this prospective study was to confirm the effectiveness of conservative treatment in Scheuermann's kyphosis and was to evaluate and compare the variation of the vertebral geometry with the curve trend in Cobb degrees, before and after conservative treatment. Material and Methods This prospective study was conducted on 90 patients with thoracic Scheuermann's kyphosis, treated using anti-gravity brace: 59 male, 31 female. The mean age at the beginning of the treatment was 14 years. Radiographical measurements were performed on radiographs from a lateral projection, at the beginning (t1) and at the end of the treatment (t5). To avoid the great variance in the range of curve angles in thoracic kyphosis that rely on the radiological position, x-rays were performed all at our Radiology Department observing the following position: standing with head straight, arms bent at 45° and hands placed on a support. Vertebral geometry modifications at t1 and t5 were analysed according to the following parameters and evaluated by three independent observers: Anterior wedging angle (ALFA) of the apex vertebra and Posterior wall inclination (APOS) of the limiting lower vertebra. These parameters were chosen because they had shown to be the most significant in a previous study. The curve was measured in Cobb degrees. Statistical analyses was performed. Results The results from our study showed that of the 90 patients with a thoracic curve mean value of Cobb degrees was 57.8 ± 6.0 SD at t1 and 41.3 ± 5.6 SD at t5. The differences between t1(angle at baseline) and t5 (end of treatment) were calculated for Cobb, alpha and Apos angle and were respectively -16.4 ± 4.5, -6.4 ± 1.4 and -2.7 ± 1.2; tested with paired t-test were significative (p < 0.01). The results of the regression analysis to test the relationship between the three measures for the kyphosis (cobb degree, alpha and Apos) showed that the best association was between Cobb t5 and Alpha t5 (p < 0.01) and between Cobb t1 and Apos t1 (p < 0.01). No significative association was found between the difference between alpha and Apos. Conclusion Our results confirm that conservative treatment in Scheuermann's kyphosis can remodelling the deformed vertebras. We sustain that using new parameters to study vertebral remodelling allows us to reach a better comprehension of Scheuermann spine response to anti-gravity brace treatment. Furthermore, the evaluation of the alpha angle of the apex vertebra confirms to be more reliable than Cobb’s angle because it cannot be affected by the radiological position O22 The plaster cast in the conservative treatment of idiopathic scoliosis can still play a positive role? Angelo Gabriele Aulisa1, Vincenzo Guzzanti2, Francesco Falciglia1, Andrea Poscia3, Lorenzo Aulisa4 1U.O.C. of Orthopedics and Traumatology, Children's Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy; 2U.O.C. of Orthopedics and Traumatology, Children's Hospital Bambino Gesù, Institute of Scientific Research, P.zza S. Onofrio 4, Rome, Italy; University of Cassino, Cassino, Italy; 3Institute of public health, University Hospital "Agostino Gemelli", Catholic University of the Sacred Heart School of Medicine, Rome, Italy; 4Department of Orthopedics, University Hospital "Agostino Gemelli", Catholic University of the Sacred Heart School of Medicine, Rome, Italy Background The current treatment of idiopathic scoliosis is based on the development of protocols and guidelines that show the way and the time needed until the conservative treatment results effective. The first of these protocols included the use of three corrective plaster casts before applying the brace. However, in the last years, many schools have abandoned the plaster cast both to increase compliance and because convinced of the effectiveness of the brace. Design The purpose of the present study was to evaluate whether the corrective plaster cast positively affects on the outcome and if its use can still play a positive role in the conservative treatment of scoliosis. Material and methods From a consecutive series of patients, included in a prospective database, to whom had been proposed the indication for a corrective plaster cast, to improve the effectiveness of conservative treatment, 128 scoliosis were selected: 78 thoracic (12.6 ± 1.8 years) treated with Lyon brace and 50 lumbar or thoracolumbar (12.9 ± 1.8 years) treated with PASB. The 50 % of patients had accepted the treatment with plaster cast before the application of the brace. In these cases 2 corrective plaster casts for about 20 days each were applied. X-rays were used to estimate the curve magnitude (CM) and the torsion of the apical vertebra (TA) at 3 time points: beginning of treatment (t1), four months after the beginning (t2), 2-year minimum follow-up (t3). Three outcomes were distinguished in agreement with SRS criteria: curve correction, curve stabilization and curve progression. Results The results from our study showed that of the 78 patients with a thoracic curve CM mean value was 40.9 ± 7.1 SD at t1 and 27.1 ± 9.4 SD at t3. The 50 patients with a lumbar curve CM mean value was 39.4 ± 6.8 SD at t1 and 23.3 ± 9.3 SD at t3. The difference in cobb degrees between t3-t1 was 16.5° in thoracic curves treated with plaster cast and 11.2° in those treated without plaster cast while in lumbar curves was respectively 19.7° and 12.5°.Therefore the patients who used the plaster cast before starting treatment with brace achieved an higher improvement in cobb degrees than those who have not used it (5.71 cobb degrees, P < 0.01). The patients with higher Cobb degree at baseline showed a better evolution of the scoliosis at T2 if they used the plaster cast (-22.4° vs -11.6°; p < 0.01), even if at t3 they showed a little worsening (-17.5° vs -12°; p < 0.01). Similarly, even patients with lower Cobb degree at baseline showed a better evolution of the scoliosis at t2 if they used the plaster cast (-19.9° vs -9.0°; p < 0.01), with a little worsening at t3 (-17.5° vs -11.8°; p < 0.01). Curve correction was accomplished in 114 patients (89.1 %), stabilization in 10 patients (7.8 %). 4 patients (3.1 %) had a curve progression. Conclusion In all cases, the treatment with corrective plaster cast showed a better outcome compared to the non-plaster, even in the lower curves. Therefore, considering the results in greater curves, to ensure a better QOL, the treatment with plaster cast should be the first choice. While in the lower curves, to improve compliance, no indication to the plaster cast can be given even if outcomes are better. O23 Bracing for Adolescent Idiopathic Scoliosis (AIS) and Scheuermann Kyphosis : The issue of overtreatment in Greece Nikos Karavidas Scoliosis Spine Laser Centre, Athens, Greece Introduction Most recent evidence has proved the efficacy of brace in the treatment of spinal deformities for young adolescents. Scoliosis Research Society (SRS) and Society on Scoliosis Orthopedic Treatment (SOSORT) have produced guidelines to indicate when brace treatment must be applied. The purpose of this study was to evaluate the rate of overtreatment for AIS and Kyphosis in Greece, according to SOSORT and SRS guidelines. To date, this is the first study to investigate overtreatment percentage in a group of patients with spinal deformities. Material and methods Cross-sectional study design and data analysis were performed in a group of patients that received treatment in a private clinic, in 2014. Of 289 treated patients, 167 young adolescents (128 females - 41 males, mean age 15, 7 years) were eligible for inclusion criteria (age 9-18 years, brace wearing). Overtreatment was defined as the unnecessary use of brace according to SRS and SOSORT indications for brace treatment, and referred to individuals that should have never started wearing a brace or to those that brace weaning was very prolonged. Overtreatment was assessed by a Schroth certified Physiotherapist, by estimating Cobb angle, Risser sign, age of menarche in girls, and vertebral wedging in Scheuermann’s kyphosis, alongside with a subsequent analysis of risk prognostic factors (family history, angle trunk rotation, thoracic hypokyphosis, and curve type). The braces were prescribed by 34 medical doctors (MD) from different geographical areas of Greece. Results The data analysis revealed that 71 out of 167 subjects (42,5 %) had received some kind of overtreatment. The percentage of overtreatment was similar for AIS (51/118 patients, 43,2 %) and kyphosis (20/ 49 patients, 40,8 %). A further analysis showed that in the AIS subgroup, 20 subjects (16,9 %) had Cobb angle < 20o , 7 subjects (5,9 %) had Cobb angle 20o – 25o but good prognosis, 12 subjects (10,2 %) started bracing after Risser 4 or 5, and 12 subjects (10,2 %) had not reached brace weaning even a long time after skeletal maturity. It is noticeable that 8 subjects (6,8 %) were at Risser 5 with Cobb angle < 20o and were prescribed a brace. In the Kyphosis subgroup, 11 subjects (22,5 %) showed no signs of Scheuermann’s disease and no clinical rigidity, 3 subjects (6,1 %) started bracing after Risser 4 or 5, and 6 subjects (12,2 %) should have reached brace weaning much earlier. Conclusions An extremely high rate of overtreatment (42,5 %) was identified in a random group of adolescents treated with a brace for AIS and Kyphosis. This is probably attributed to lack of knowledge in the field of conservative treatment of spinal deformities. It also seems that the majority of MDs ignore the role of the Physiotherapeutic Scoliosis Specific Exercises (PSSE) in the treatment of scoliosis. Overtreating a child with a brace can cause social, financial and psychological problems to them. The present study pinpoints the need for an evidence-based approach to conservative treatment of idiopathic scoliosis and kyphosis, according to SOSORT and SRS guidelines, in order to avoid overtreatment and to enhance clinical outcomes. O24 Efficacy of Milwaukee brace for correction of scheurmann kyphosis Mohammadreza Etemadifar Orthopedic Department, Esfahan University of Medical Sciences, Isfahan, Iran Introduction Scheurmann's kyphosis is a relatively common sagittal plane mal alignment inadolescents, which could be corrected relatively easily provided that is diagnosed earlyand treated with brace. There are several kinds of braces useful in treatment of thisdeformity in which Milwaukee is the most popular one. Few studies recently havedemonstrated brace efficacy in treating scheurmann kyphosis. The aim of our study wasto investigate effects of a low-profile Milwaukee brace for treatment of this deformity. Material and methods All adolescent patients with diagnosis of scheurmann kyphosis with at least one year ofgrowth remaining were included in our study. Standing AP/LAT (T1- S1) X-rays wasperformed in all cases. Thoracic kyphosis Cobb angle and vertebral wedging, end platesclerosis and irregularity were identified and recorded. A modified low profile Milwaukee brace (without neck ring) was administered for all patients. Follow up X-rayswere performed at 2, 6 and12 months of brace wear and one year after brace weaning. T-test was used to compare average pre and post- brace kyphosis. Results A total of 158 patients were enrolled in our study. 32 cases were lost at final follow up (including 5 cases who were shifted to surgery). 126 cases were remained for evaluation.Average age of the patients was 13.5 years (10 to 16 years). There were 78 females and 48 males. Average primary thoracic kyphosis was 67 degrees and after 2, 6 and 12 months brace wear it was 50, 41 and 33 degrees respectively. Average full time bracewear was 18 months and at the last follow up mean kyphosis angle was 45 degrees,which was statistically improved compared to primary kyphosis (p < 0.05). All patientsand parents were satisfied with results. Conclusion Low profile Milwaukee brace can be effective for conservative treatment of scheurmann kyphosis provided that it is well fitted and used regularly by the patient. Although there issome correction loss, final thoracic kyphosis is acceptable and most patients are satisfied with results. O25 The three dimensional analysis of the Sforzesco brace correction Sabrina Donzelli1, Fabio Zaina1, Monia Lusini1, Salvatore Minnella1, Luca Balzarini2, Stefano Respizzi2, Stefano Negrini3,4 1ISICO Italian Scientific Spine Institute, Milan, Italy; 2 ICH Istituto Clinico Humanitas, Milan, Italy; 3IRCCS Don Gnocchi, Milan, Italy; 4University of Brescia, Brescia, Italy Background Scoliosis is a three dimensional deformity, and brace correction should be 3D too. There is a lack of knowledge of the effect of braces, particularly in the sagittal and transverse plane. The aim of this study is to analyse the Sforzesco Brace correction, through all the parameters provided by Eos 3D imaging system. Method Design: This is a cross sectional study from a prospective database started in March 2003. Participants: 16 AIS girls (mean age 14.01 ) in Sforzesco brace treatment, with EOS x-rays , at start, in brace after 1 month and out of brace at 4 months. Outcome measures: All the parameters and the Torsio-Index obtained from 3D Eos System, in and out of brace, in the three planes. Statistical analysis: the variability of the parameters and the mean differences were analysed and compared using paired T test. ANOVA was used for multiple comparisons. P value was set below 0.05. Results In the comparison in brace vs start of treatment the mean Cobb angle change significantly from 36.44 + -4 to 28.99 + -3.9 (p = 0.01). Significant changes in all the sagittal parameters were found (p = 0.02). In the axial plane, the TorsioIndex, changed significantly in brace, only for thoracolumbar and lumbar curves (P < 0.05). The analysis of the single vertebral tilt, demonstrated that the effect of brace are mostly concentrated to some segments:T4-T5; T10-T12, L1 and L5 in the AXIAL plane and T3-T6; T10-L1 in the frontal plane. Conclusion Sforzesco brace mostly modify the middle of the spine, and preserve the sagittal balance. The single vertebral orientation in each plane, should be considered together with the typically used values to assess brace effect. O26 Quality of Life in adolescents with idiopathic scoliosis: A comparison measured by the Kidscreen 27 between scoliotic patients and healthy controls Kathrin Güttinger Zürcher Hochschule für angewandte Wissenschaften (ZHAW), Winterthur, Switzerland Background Different questionnaires evaluate the quality of life in scoliotic patients. The questionnairs which were developed especially for scoliotic patients are: Brace Questionnaire (BrQ) , Bad Sobernheimer Stress Questionaire Brace and Deformity (BSSQ-Brace/ BSSQ-Deformity), SRS-22 and Scoliosis Quality of Life Index (SQLI). In addition there are the questionnaires which measure the general quality of life in children and adolescents: SF 36, Kidscreen 52 & 27. Aim The aim is an evaluation of the quality of life and the self-efficacy of adolescents with idiopathic scoliosis (IS). The goal of the study is to compare the results of the adolescents with scoliosis with the results of the adolescents without scoliosis. Design It is a cross-sectional design study. 30 scoliotic patients and 30 healthy adolescents participated in the study. The inclusion criteria for the patients group was: IS, age between 12 & 18 years, girls and Cobbangle > 25°. The inclusion criteria for the control group was: girls, age between 12 & 18 years and no known scoliosis. Material & methods 30 patients were included (mean age 14. 4 years, mean Cobbangle 36.1 °) and 30 healthy controls (mean age 14.8 years). Both groups filled out the kidscreen 27 and special additional questions. The additional questions were: 1. the subgroup About yourself of the kidscreen 52, 2. special questions formulated by the author, 3. questions about the self-efficacy, also formulated by the author and 4. questions about the sport activities and the school type (both groups) and the brace wearing time and the therapy intensity (in the patientsgroup). The Kidscreen 27 consists of 5 categories: Physical Well-Being (5 items), Psychological Well-being (7 items), Autonomy & Parents (7 items), Peers & Social Support (4items) and School Environment (4 items). The higher the score the better the quality of life. To make the score generally comparabel, the score was transfered with an Raschanalysis. Results In the dimension Psychological Well-Being the scoliotic patients had higher T-Scores than the control (48,9 +/- 7,42 versus 46,13 +/-8,40). Also in the categorie About yourself the patients had higher scores (47,65 +/-7.64 versus 43.4 +/-6.23). In the categorie Autonomy and Parents the patient group had lower T-Scores than the controls (52.09 +/- 6.96 versus 55.03 +/-8.95). This is the same for the category Peers and Social Support (51.1 +/- 7.7 versus 53.06 +/- 7.92). But in the comparison these two categories between the scoliosis patients and the standard value of switzerland the scoliosis patients are not below the standard value (52, 51). Conclusion Adolescents with scoliosis showed better scores in the categorie Psychological Well-Being and About yourself than the controls. In the cateories Autonomy & Parents and Peers and Social they had lower scores. O27 The degree of illness acceptance in young women with idiopathic scoliosis treated with orthopedic braces: a preliminary study Jacek Durmała 1, Irmina Blicharska 1, Agnieszka Drosdzol–Cop 2, Violetta Skrzypulec–Plinta 2 1School of Health Sciences in Katowice, Medical University of Silesia, Chair and Department of Rehabilitation, Katowice, Poland, 2School of Health Sciences in Katowice, Medical University of Silesia, Chair of Woman’s Health, Katowice, Poland Background The degree of illness acceptance is one of the factors that have an impact on the patient’s quality of life. At present, this topic is frequently discussed due to ideological transformations in medicine, where the comprehensive approach towards the patient’s heath is taken into consideration. In the holistic model of medical care, the subjective dimension of ailment is very significant. The acceptance, however, has an influence on the self-esteem and it determines particular attitude towards the therapy.People with a higher degree of acceptance show better adaptation and lower intensification of negative emotions. Material and methods 36 women (aged 20.7 ± 1.89 and diagnosed with adolescent idiopathic scoliosis) were qualified for the study. An average value of the Cobb’s angle in the primary curvature was 31.2° ± 12.39°. Patients were treated by means of kinesotherapy (within the range of DoboMed technique) in connection with the Chaneau brace. For the acceptance evaluation,the Acceptance of Illness Scale (AIS) was used. It is a tool developed by B. J. Felton and his colleagues (in the Polish adaptation by Zygfryd Juczyński). The scale contains eight statements (measured in points from 1-5), which describe negative consequences of poor health condition. The degree of acceptance is calculated by the amount of points within the range of 8 to 40. A low result indicates non-acceptance and ailment accommodation as well as strong feeling of psychological discomfort. A high score attests to the self-acceptance. The apical vertebral rotation (AVR) was determined on the basis of a current radiograph. The degree of trunk deformations was based on the topography of a surface of the body and the Posterior Trunk Symmetry Index (POTSI).The aim of the prospective research with randomization, conducted by means of double blind testing, was an evaluation of acceptance in women treated with scoliosis braces. The obtained results were analyzed by means of Statistica v.10. The assessment of variables concurrence with the typical layout was conducted with the test invented by Shapiro-Wilk. The dependence between parameters was determined on the basis of the Pearson’s R Correlation Test. The value p <0.05 was determined as a level of statistical significance. Results An average amount of points obtained from the Acceptance of Illness Scale was 33.1 ± 6.63. This indicates a high acceptance levelin a group of tested women with scoliosis. The considerable dependence between: (1) the amount of points in AIS, (2) the Cobb angle primary scoliosis (R = -0.3, p = 0.06), as well as (3) the AVR amount (R = -0.2, p = 0.24), hasn’t been observed. Statistically considerable correlation was determined between the degree of acceptance and the POTSI index (R = -0.4, p = 0.02). Women with a lower degree of trunk deformations present a higher level of scoliosis acceptance. Conclusions Women, diagnosed with scoliosis, indicate a high degree of its acceptance. Factors, which may determine the perception of a particular disease, are mainly trunk deformations and distortions. An angular curvature value does not considerably affect the level of scoliosis acceptance. It can be correlated with the lack of an accurate correspondence between the Cobb angle and trunk deformations in the coronal plane. The investigation requires further examination and supplementation. O28 Which are the personality traits of the patients with Adolescent Idiopathic Scoliosis? Elisabetta D'Agata1, Judith Sánchez-Raya2 1Vall d'Hebron Hospital Institute, Barcelona, Spain, 2Vall d'Hebron Hospital, Barcelona, Spain Background According to the bio-socio-psychological model, for a more complex understanding of the patient, biological aspects have to be integrated with psychological dimensions. This field in scoliosis is now emerging. This study is about the main traits of personality in patients with Adolescent Idiopathic Scoliosis (AIS). Material and methods 27 patients with AIS (aged 14.6 years; Mean Cobb Angle: 31°.5; 40.7 % braced) answered a Socio-Clinical Questionnaire, a Quality of Life tool (SRS-22), the Trunks Apperception Scale (TAPS) and a Personality Questionnaire (16PF- Adolescent Personality Questionnaire). 16 PF-APQ Questionnaire presents 16 scales and identifies 5 global dimensions. Results Results for SRS-22 subscales were: Function = 5; Pain = 4; Self Image = 3.6; Mental Health = 3.5. In 16 PF-APQ Personality Questionnaire, the personality scales with a percentile value > 50 were: Dominance (72 PCTL), Rule consciousness (63° PCTL), Vigilance (65 PCTL), Privateness (61 PCTL) Openness to change (60 PCTL) and Self- reliance (70 PCTL). With reference to the global dimensions, Independency corresponded to percentile 65 while Extraversion to 28.4. Neither the scales nor the dimensions presented any extreme percentiles. School note mean was 7.5. Rule-consciousness related significantly (p < 0.001) to SRS-pain (Spearman r = 0.5) and to SRS-Self Image (Spearman r = 0.7). Conclusion Patients with Idiopathic Scoliosis did not present any psycho-pathological features. They appeared as introverted people with a tendency to be independent and assertive and with a good scholar success. O29 How many Scolioses do exist in the same person? A zoom vision on the perception of the patient Judith Sánchez-Raya1, Elisabetta D'Agata2 1Vall d'Hebron Hospital, Barcelona, Spain; 2Research Institut Vall d'Hebron Hospital, Barcelona, Spain Introduction The Trunk Appearance Perception Scale (TAPS) is a valid instrument for evaluating the perception patients have of their trunk deformity. There are no studies about the correlation among TAPS scored in each case by a physician, a patient and his /her parents. The object of the study is to compare the different perceptions of scoliosis and how the patient perception affects his/her quality of life. Material and methods The sample consisted of 64 patients (51females), mean age 15.25, mean Cobb Angle 30.6 (ranging from10°- 55°). 29 were not treated, 26 braced, 6 treated with physiotherapy. For each case, TAPS was scored individually by patient, his/her parents and the same doctor. Patients also scored Quality of Life Questionnaire (SRS-22). The sample was split into two groups according to the age (1st group: 9-14; 2nd group: 15-34). Spearman correlation Index was calculated for the three TAPS, Cobb Angle and SRS-22 4 subscales (Function, Pain, Self Image, Mental Health). Results Correlations between parents’ TAPS and doctor’s TAPS was r = 0.5 (p < 0.001). Parents’ TAPS and doctor’s TAPS were statistically different, regarding the younger patients (Wilcoxon Signed Ranked Test, p = 0.02); besides, with relation to the older group, the correlation between parents and doctor TAPS was low (r = 0.4, p < 0.05). Correlations between patients and their parents as well as between patients and the doctor were moderate (r < 0.5). The correlation between patient TAPS and SRS-22 Self Image had a moderate value in the younger group (r = 0.5) and a low one in the older group (r = 0.4, p < 0.05). The correlation between Body Image and Mental health was significative only for the younger subjects (r = 0.4, p < 0.005). Conclusion Doctor’s and parents’ perceptions are a bit discordant: in the younger group they are different, while in the older one their relation is moderate. Besides, in a patient the relation between his/her trunk perception with Body Image was moderate and associated to his/her age. ORAL POSTER PRESENTATIONS P1 The algorithm for the automatic detection of the pelvic obliquity based on analysis of the PA viev of the x-ray image Sławomir Paśko1, Wojciech Glinkowski2 1Warsaw University of Technology, Institute of Micromechanics and Photonics, Warsaw, Poland; 2Chair and Department of Orthopedics and Traumatology of the Locomotor System, Baby Jesus Clinical Hospital, Center of Excellence “TeleOrto” for Telediagnostics and Treatment of Disorders and Injuries of the Locomotor System, Medical University of Warsaw, Warsaw, Poland Background Information The oblique position of the pelvis frequently coincides with scoliosis. Asymmetrical location iliac crests may result in unequal pressure distribution while sitting and pain. It may reduce the patient’s tolerance of sitting position. The disease affects on knees, hip joints, feet and spine and can contribute to the emergence of diseases in these organs. Smaller degrees of the pelvic obliquity are almost to be accepted. It requires observation over the time; The problem may require relief before it becomes a greater clinical problem. The physicians can diagnose the problem of the pelvic obliquity based upon the analysis of PA view of the X-ray image. It is detected as the angular difference between the inter iliac crest line and the horizontal line. The value of this angle reflects the severity of the pelvis obliquity. Purpose The algorithm is intended to aid the automatic process of analyzing an X-ray image taken in PA view. It allows detecting horizontal positions of the upper edge of iliac crests. If the calculated positions are different then the physician is noted by the system of the possibility of occurrence of the pelvic obliquity. Greater degrees of the pelvic obliquity are easy to notice from the analysis of the X-ray image. Smaller degrees are difficult recognize without a referential horizontal line drawn on the image in the proximity of the pelvic bones. This process can be improved by the proposed algorithm that makes possible the detection of the small difference between the pelvic bones. Methods The multilevel algorithm processes an X-ray image based on the image processing. On the first level, the area of the pelvic girdle is roughly delimited. Due to this the upper part of the trunk is omitted from further analysis. Next they are designated few candidate areas in which pixels distribution may indicate that in one of them is located the searched edge of the pelvic bones. During the following level of processing, this set is narrowed to the maximum four candidates down. Finally, the last part of the algorithm is started where the best candidate chosen. Result The study was done on a collection of X-ray from scoliotic patients with or without the incidence of the pelvic obliquity. The set of images consisted of 10 digital PA view X-rays. The algorithm run for each image and results were saved. The same set was analyzed by specialist for the expert opinion. Positive matches were noted in analyzed cases. Conclusions and Discussion Algorithm based supported diagnosis can simplify the process of an X-ray image analysis and improve the reliability of this procedure. The physician can remain focused on the spinal curvature without the risk to skip minor radiologic signs on the radiographic image including the oblique position of the pelvis. The bigger radiographic collection should be tested for more extensive research of the algorithm to calculate its reliability. Acknowledgement: This research was supported by project NR13-0109-10/2010, founded by National Center for Research and Development. P2 Monitoring of spine curvatures and posture during pregnancy using surface topography – case study and method assessment Jakub Michoński1, Katarzyna Walesiak2, Anna Pakuła1, Robert Sitnik1, Wojciech Glinkowski2 1Warsaw University of Technology, Warsaw, Poland; 2Department of Orthopaedics and Traumology of Locomotor System, Center of Excellence ”TeleOrto”, Medical University of Warsaw, Warsaw, Poland Background Lower back pain during pregnancy is a well-known problem. Ostgaard et al. report that almost 50 % pregnant women suffer from back pain [1]. Gutke et al. find that the frequency of lower back pain in pregnant women can be up to 4 times higher than in non-pregnant women [2]. In the prospective study of Kristiansson et al. 30 % women with the highest pain score report great difficulties with normal activities [3]. Moore at al. performed a study on the postural changes in pregnant women in 1989 [4], however did not use surface topography to achieve this goal. Design and level of Evidence A pilot study was conducted to test the potential of monitoring the change of spine curvatures and posture during pregnancy using surface topography. A single case was studied to test the methodology and preliminarily assess the usefulness of the procedure before performing a randomized trial. The apparatus used in this study was metrologically tested and utilized in scoliosis screening. Material and Methods The subject was measured using a custom-made structured light illumination scanner with accuracy of 0.2 mm. Measurement was taken every 2 weeks, between 17th and 37th week of pregnancy, 11 measurements in total. The subject was 34 years old at the beginning of the study, no systemic disorder, no drug use, no previous trauma or surgery of spine or lower limbs. From the measurement the thoracic kyphosis and lumbar lordosis angles, and vertical balance were extracted automatically. All measurements were repeated three times and the median calculated values were chosen. Oswestry Low Back Pain Disability Questionnaire (ODI) was done with every measurement. Results The values were correctly extracted from the measurement without any user interaction. From the registered values mean and standard deviation were calculated. The registered change was 4 degrees in kyphosis angle, 5 degrees in lordosis angle and 4 degrees in vertical balance angle. The calculated ODI index was between moderate disability and crippling back pain (20 % to 73 %). Conclusions We have found that surface topography is suitable for monitoring of spinal curvature and posture change in pregnant women. Automatic calculation of curvature angles and extraction of vertical balance angle allows to obtain high reliability of such a study without the observer errors introduced otherwise. No concrete conclusions can be drawn from the study because of insufficient amount of data, however flattening of lordosis in the evaluated case matches the conclusions of Moore et al. References 1. Prevalence of Back Pain in Pregnancy. OSTGAARD, H C MD; ANDERSSON, G B J MD, PhD; KARLSSON, K MD 2. Predicting persistent pregnancy-related low back pain. A. Gutke, H.C. Ostgaard, B. Oberg. Spine, 33 (12) (2008), pp. E386–E393 3. Back Pain During Pregnancy: A Prospective Study. Kristiansson, Per MD*†; Svärdsudd, Kurt MD, PhD*; von Schoultz, Bo MD, PhD‡ 4. Postural changes associated with pregnancy and their relationship with low-back pain. K. Moore, Msca, G.A. Dumas, PhD, J.G. Reid, PhD P3 Spinal rotation under static and dynamic conditions: a prospective study comparing normative data vs. scoliosis Helmut Diers Research & Development, Schlangenbad, Germany Background Using SST (Spine & Surface Topography) can significantly reduce the amount of harmful radiation in scoliosis treatment and follow-up during therapy. The SST under dynamic condition shoes the body movement, body mechanics and the activity of the muscles. People with scoliosis show under dynamic conditions reduced mobility and asymmetric lateral rotation. Aim The improvement of the spinal rotation during walking on the treadmill should be verified and this additional information should be integrated in the therapy. Design Patients according to different age groups and gender were measured with static 4D (habitual standing) and dynamic 4D motion (walking on the treadmill). The age of the patients was between 25-50 years. 20 Patients were measured two times. On measurement was under static conditions and the second under dynamic conditions. The focus was on the parameter “vertebral rotation”. Methods SST uses surface topography imaging for 3D back scanning and techniques creating 3D models of the spine without exposing patients to any ionizing radiation. The spine reconstruction model has been used successfully for more than 20 years in evaluation and treatment of patients with spinal deformities such as scoliosis, lordosis and kyphosis. Dynamic spine analysis is using similar algorithms as used in static measurement. Results Based on the degree of scoliosis there is a significant improvement of the vertebral rotation during walking on the treadmill. The rotation becomes more equivalent. In this study the parameter of the “vertebral rotation” was picked out and the differences between static and dynamic were worked out to get the output for the therapy. Conclusion The information of the spinal rotation is important for the therapy of patients with scoliosis. Regular monitoring of patients with scoliosis is under static conditions. The additional information of the dynamic and functional results provides important information towards personalized care and the organization of the therapy. The therapy should be adapted on these parameters to get better results. P4 The principle of non-surgical treatment of idiopathic scoliosis right-sided breast depending on the volatility of the formation of the intervertebral discs and vertebral bodies Piotr Majcher, Piotr Gawda Uniwersytet Medyczny w Lublinie, Lublin, Poland Introduction Kinesitherapy procedure in the treatment of children and adolescents with idiopathic scoliosis will vary depending on the size of the angular curvature of the spine and intervertebral krążkó sklinowania size and vertebral bodies, and of whether the child will be treated conservatively, and will prepare for surgery. Rights Delpech-Wolff and Hueter-Volkmann's talk about changes in the growth and remodeling of bone asymmetrically loaded, and the so-called hypothesis. "vicious circle" described by Stroeks talks about the role of the vertebrae in the emerging scoliosis. Is ignored in the literature Polish and international significance of the intervertebral discs in forming the bend. Materials and methods Research on radiographs of 200 children and adolescents divided into groups at different stages of the creation of the right-hand treatment of thoracic idiopathic scoliosis. Chocking evaluated intervertebral discs and vertebral bodies on the left side compared to the right side of the right-hand idiopathic scoliosis prove that wedged intervertebral discs is essential and may be a prognostic factor in this progresji.Stosunek the amount left to right intervertebral discs and vertebral allows the calculation rate their height. Results The authors present the principle of non-operative treatment depending on the set values and the degree of curvature angular wedged intervertebral disc and vertebral body in idiopathic scoliosis. Treatment program for children and adolescents with idiopathic scoliosis can be divided according to the value measured by Cobb angle and / or the value of wedged intervertebral disc and vertebral body on: 1. The bending of the spine from 0 to 10 (defect attitude - the attitude of idiopathic scoliosis) no wedged intervertebral disc and vertebral body, recommended only significant physical activity. 2. Curvature of the spine of the angle from 10 to 20-25 and / or wedged vertebral and non-vertebral wedged - recommended intensive individual therapy. 3. Curvature of the spine with a value in excess of 20-25 and / or wedged vertebrae with a value of 0.7 and wedged vertebral body - recommended corrective orthopedic corset, individual therapy. 4. Curvature of the spine above 40-45 and / or wedged vertebrae with a value of 0.54 and wedged vertebral body - recommended surgery, follow-up corrective orthopedic corset, individual therapy fails to improve. ConclusionsThe treatment of scoliosis is based on their pathomechanics. The study should continue in many centers. P5 Unexpected late progression of adolescent idiopathic scoliosis treated with short-term, aggressive, full-time bracing and Schroth physiotherapy with excellent preliminary result: case study Andrea Lebel1†, Victoria Ashley Lebel1,2† 1Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8, Canada; 2Saba University School of Medicine, Saba, Dutch Caribbean, Netherlands. Correspondence: Andrea Lebel – Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8, Canada †These authors contributed equally to this work Background Adolescent idiopathic scoliosis (AIS) is a complex three-dimensional (3D) spinal deformity of unknown cause diagnosed between the ages of 10 to 18 years. The Scoliosis Research Society (SRS) goal of treatment is to halt curve progression before it reaches surgical magnitude. The risk of scoliosis curve progression decreases with age as the vertical growth potential slows in female AIS patients post-menarche. However, relatively late, unexpected progression of AIS, 1-2 years post-menarche, still has the potential to lead to irreversible structural changes of the spine and torso and to spinal surgery if these patients are only managed with the “wait and see” observation method for scoliosis. The “try and see” method of conservative scoliosis management involving physiotherapeutic scoliosis-specific exercises (PSSE) and 3D bracing should be offered to AIS patients with late, unexpected curve progression, even 1-2 years post-menarche, until the very end of growth. The purpose of this study was to evaluate the effectiveness of conservative management of scoliosis (Schroth physiotherapy and bracing) in a female adolescent patient with late progression of AIS. Methods This prospective case study follows a 13-year and 10-month old female, 1 year post-menarche, with late progressive AIS, who was treated with short-term, aggressive, full-time bracing and PSSE. The patient was followed by the Schroth physiotherapist from July 2013 to February 2015. The patient’s curve progression and improvement were monitored by measuring vital capacity, chest expansion, and height, in addition to scoliosis curve angle measurements obtained from follow-up radiographs. Results The case study patient showed significant improvement in vital capacity, chest expansion, and scoliosis curve Cobb angle over the course of this study with short-term, aggressive, full-time bracing and Schroth physiotherapy. Conclusions In this case study, short-term, aggressive, full-time bracing combined with daily Schroth physiotherapy proved effective in halting curve progression, reducing scoliosis curve Cobb angles, and improving vital capacity, chest expansion, and future quality of life. Consent Written informed consent was obtained from the patient for publication of this abstract and any accompanying images. A copy of the written consent is available for review by the Editor of this journal. P6 Visible posture in relation to the neuroanatomical and neurodynamical features in spinal deformations Piet van Loon1, Ruud van Erve1, Andre Grotenhuis2 1Care to Move Orthopedics, Deventer, The Netherlands; 2Neurosurgery, UMC Radboud University, Nijmegen, The Netherlands Introduction All deformities are natural deviations of the optimal posture a human body can achieve. The visible feature of the individual morphology in a standing person is called posture. There is no discussion a normal posture will be formed in gradual ways by forces of neuromuscular origin. So also deformities are caused by these natural forces of growth. The neuro-osseous growth relation and the reciprocal feedback , based on signaling stretch by the nervous cells and moulding capacity of muscular forces on the cartilage ( and young bone) is in depth described by prof. Milan Roth in the period 1965-1985. It reflects the wellknown paradigm: form follows function, which kept its force mainly in paramedical practice and in those Health systems, where prevention of “bad postures”, from the day of birth, kept its place. The relationship between postural deviations and the characteristics of the CNS on MRI is scarcely described. Method and material In a series of 4 cases with a proven scoliosis or hyperkyphosis a comparison will be shown of all the diagnostic features available to visualize the relationship between external and internal features in spinal deformations. Besides clinical photographs in standing ( and sitting) position and of clinical tests to show the functional signs in postural problems based on neuromuscular thightness, surface topography derived models ( Diers GmbH, Schlangenbad, Germany), standing radiographs and MRI are put alongside to explain the relationships between the main systems involved in deformations during growth: the CNS and the musculoskeletal system. Results In all cases there is direct relationship between the external form of the spine and the internal features of the central cord/roots complex. Especially the position and the caliber of the central cord and roots in relation to the osseous boundaries of the canal will be shown and explained. The ever present and in fact pathological contact zones between the cord and roots complex at the apices of pathological curves and at the cervicothoracic and the lumbosacral areas present at early ages will be highlighted and related to physiological and pathophysiological signs and symptoms that can occur during adulthood. Discussion No epidemiologic or cohort study is necessary for understanding pathogenesis , but this visualization of an understandable relationship between external and internal features (most never described in literature) in the involved tissues can invite other institutions with a great number of cases and resources for research on a bigger scale to provide also statistical evidence of the relationships. Concomitant features of a “short cord” like Arnold Chiari malformation and syringomyelie can be explained out of the neuro-osseous growth discongruency. Conclusion The skeleton and the CNS are strongly related in anatomical and physiological features. An attempt is made to visualize with available diagnostic tools that the anatomic neuro-osseous growth relations that will end-up in spinal deformities if discongruency between the two types of growth occur. The outside always reflects what is going on inside the body. P7 Immediate effects of scoliosis-specific corrective exercises on the Cobb angle after 1 week and after 1 year of practice Karina Zapata1, Eric Parent2, Dan Sucato1 1Texas Scottish Rite Hospital, Dallas, Texas, USA; 2University of Alberta, Edmonton, Canada Background We are unaware of any studies describing the immediate effects of scoliosis-specific exercises on the Cobb angle measured by radiograph. This study aimed to describe the differences between radiographs obtained with and without corrective exercises after initial training and after one year. Design Case Study, Level 5 Methods A female with adolescent idiopathic scoliosis was first seen at 13 + 0 years of age with a Risser 0, 2 months post-menarche. She had a 43o left lumbar, 15o right thoracic curve. She had worn a Providence night-time brace for three months. She was seen again after 6, 12 and 24 months and performed exercises from 12 to 24 months. She taught Barcelona Scoliosis Physical Therapy School (BSPTS) exercises for a four-curve type (lumbar dominant with pelvis deviation to the lumbar concave side). She attended 8 visits for 2 hours each over one week of intensive instruction by a BSPTS-certified therapist with 7 months of experience. She was asked to alternate performing 5 of 8 home exercises (semi-hanging, prone-on-knees, prone-on-stool, sidelying, side-sitting, rotational-sitting, sitting, standing) for 30 minutes per day, 5 times a week. No direct care was provided during this interval due to not having access to a trained PT. At 12 and 24 months, x-rays were obtained with and without performing corrective exercises without the PT present. Results At 6 months, her lumbar and thoracic curves measured 41o and 28o, respectively. Her Risser was 2-3 and she had grown 3 cm. At 12 months, her lumbar and thoracic curves measured 47 o and 30o, respectively. Her Risser was 4. The brace was discontinued. Also at 12 months, immediately after her x-ray in the relaxed standing position, she performed her corrective exercises in standing with arms lowered for a second x-ray. The corrections included: pelvis corrections for her curve type, auto-elongation, opening her concavities, depressing her convexities, and shoulder counter-traction. Her lumbar and thoracic curves remained similar and measured 43o and 32o, respectively. At 24 months, one year after exercise instruction, her lumbar and thoracic curves measured 26o and 41o, respectively. When asked about her improved x-ray, she reported using corrective exercises during the x-ray (without being asked). Another x-ray was obtained in relaxed position during the same visit. Her lumbar and thoracic curves measured 39o and 35o, respectively. She was Risser 4 and had grown 3.5 cm the past year. The patient and parents reported home exercise compliance at 1-3 times a week the past year. The immediate effect of corrective exercises after a year of training was a 33 % improvement at the lumbar spine compared to only a 9 % improvement the previous year. Conclusion After initial training, corrective exercises during a standing x-ray did not significantly improve the Cobb angle for the major lumbar curve compared to the relaxed standing x-ray. However, a year after performing exercises, unsolicited corrective exercises resulted in a significantly improved Cobb angle compared to relaxed standing for the curve primarily targeted by the exercise program. Improved exercise ability and spinal flexibility may have contributed to the improved Cobb angle. Consent Written informed consent was obtained from the patient for publication of this abstract and any accompanying images. A copy of the written consent is available for review by the Editor of this journal. P8 Retrospective analysis of idiopathic scoliosis medical records coming from one out-patient clinic for compatibility with Scoliosis Research Society criteria of brace treatment studies Krzysztof Korbel1, Mateusz Kozinoga2, Łukasz Stoliński1, Tomasz Kotwicki2 1Rehasport Clinic, Poznań, Poland, 2Spine Disorders Unit-Departament of Pediatric Orthopaedics and Traumatology, Poznań University of Medical Sciences, Poznań, Poland Background According to Scoliosis Research Society (SRS) idiopathic scoliosis (IS) is a spine deformity of more than 10° Cobb angle value with rotation seen on the standing radiograph. In 2005, the SRS experts proposed inclusion criteria for studies on the efficacy of IS corrective brace treatment: a) age ≥10 years, b) Risser 0-2, c) Cobb angle of 25-40°, d) no previous treatment, e) patients before menarche or less than one year after menarche. Design and Level of Evidence Retrospective study. Material and methods A retrospective review of the consecutive medical histories of girls with IS treated in the outpatient clinic of the University Hospital in Poznań from 1989 to 2002 was carried out. The outpatient clinic was dedicated to scoliosis problems in general, including congenital and neuromuscular patients. Patients’ age, Cobb angle, Risser sign and menarche status at the beginning of treatment were noted from the charts. The number of patients eligible according to SRS criteria was searched. Results Total number of 2705 medical charts of consecutive patients (girls only) was checked. 183 out of 2705 girls (6.8 %) undergoing brace treatment were aged ≥10 years and presented Cobb angle in the range 25°-40°. However, considering the maturation status, the number of girls fully following the SRS criteria diminished to 102 girls (3.8 %): 42 presented single thoracic curve and 60 presented double thoracic and lumbar curve. The remaining 81 girls had to be excluded for the following reasons: a) 9 single thoracic; Risser 0-2, 1st menstruation > 1 year, b) 9 single thoracic: Risser sign > 2, 1st menstruation ≤ 1 year, c) 11 single thoracic: Risser sign > 2, 1st menstruation > 1 year, d) 9 double thoracic and lumbar; Risser sign 0-2, 1st menstruation > 1 year, e) 15 double thoracic and lumbar; Risser sign > 2, 1st menstruation ≤ 1 year, f) 28 double thoracic and lumbar; Risser sign > 2, 1st menstruation > 1 year. The mean age at first visit was 12.6 ± 1.7 years, the mean age at start of brace treatment was 12.9 ± 1.7 years. Discussion and conclusion In the years 1989-2002, the majority of IS brace treated girls were older, more mature and with bigger Cobb angle comparing to current SRS criteria for brace efficacy studies. This study certifies about the changing criteria for conservative treatment of IS. In the light of our findings it seems difficult to directly compare the nowadays results with historical series of cases. P9 Adult female with severe progressive scoliosis possibly secondary to benign tumor removal at age 3 treated with scoliosis specific Schroth physiotherapy after refusing surgery: case study Andrea Lebel1†, Victoria Ashley Lebel1,2† 1Ottawa & District Physiotherapy Clinic, Scoliosis Physiotherapy and Posture Centre, McLeod Street, Ottawa, K2P 0Z8, Canada; 2Saba University School of Medicine, Saba, Dutch Caribbean, Netherlands †These authors contributed equally to this work Background Scoliosis is a complex three-dimensional (3D) spinal deformity. Acquired scoliosis in early childhood may progress into adulthood and pose an increased risk of health problems and reduction in quality of life. In Canada, adult patients with scoliosis are not referred for physiotherapeutic scoliosis-specific exercises (PSSE) despite the fact that Schroth physiotherapy, a scoliosis-specific 3D posture training and exercise program, can be effective in reducing pain and improving scoliosis curves, vital capacity, and overall quality of life in scoliosis patients. This case study shows that indeed adult curve progression can be stopped and even reversed with scoliosis specific schroth physiotherapy (SSSPT) in an adult patient with scoliosis. Methods This is a retrospective case study involving a 23-year-old female scoliosis patient who began an outpatient Schroth physiotherapy exercise program and was initially monitored monthly and then annually for improvement in measurements of angle of trunk rotation (ATR) and chest expansion and improvement in vital capacity measured with incentive spirometry. Photos were taken to document body image periodically throughout Schroth physiotherapy treatment. Additionally, the patient completed SRS-22 quality of life questionnaires every 2 years to evaluate daily function, pain, self-imagine, mental health, and scoliosis management satisfaction. Results Within one month of beginning SSSPT, the patient reported no more back pain and within 2 months, reported improved breathing. The patient also benefitted from improved chest expansion, scoliosis curve angles (measured in Cobb degrees), vital capacity, ATR, and SRS-22 scores. She became more active and resumed all athletic activity within 8 months of beginning Schroth physiotherapy. Conclusions Adult scoliosis patients are not routinely referred for PSSE in Canada, even though Schroth physiotherapy, a kind of PSSE, is shown to be effective in this case study. This patient was successfully treated with Schroth physiotherapy. Long-term comprehensive Schroth physiotherapy, to help correct and maintain proper posture in all aspects of daily living, should be part of scoliosis management for adult scoliosis patients in Canada to stop and reverse curve progression and improve overall quality of life. Consent Written informed consent was obtained from the patient for publication of this abstract and any accompanying images. A copy of the written consent is available for review by the Editor of this journal. P10 New aspects of scoliosis therapy planning and monitoring Helmut Diers Research & Development, Schlangenbad, Germany Background Using SST (Spine & Surface Topography) can significantly reduce the amount of harmful radiation in scoliosis treatment and follow-up during therapy. The examinations are possible during standing and walking conditions. Aim Purpose of this study is the therapy monitoring based on stato-dynamic measured values to show new aspects of therapy planning, especially scoliosis treatment, and monitoring. Design Patients according to different age groups and gender were measured with static 3D (habitual standing). The age of the patients was between 18-99. The different age groups have a minimum of 100 participants: 18-45, 46-65, 66-99. The focus was on the following parameters: coronal imbalance, pelvic obliquity, pelvic torsion, vertebral rotation and apical deviation. Methods SST uses surface topography imaging for 3D back scanning and techniques creating 3D models of the spine without exposing patients to any ionizing radiation. The spine reconstruction model has been used successfully for more than 20 years in evaluation and treatment of patients with spinal deformities such as scoliosis, lordosis and kyphosis. Dynamic spine analysis is using similar algorithms as used in static measurement. Results Creating new perspectives trough continuous innovation, steady improvement of products and patient specific solutions are important aspects in the organization of therapy and for patient specific solutions. These three conditions are intermeshed like cogs in a machine, and all play a part in maintaining a continuous flow of reliable monitoring mechanisms for personalized care. Conclusion Normative data offer guidance. The different parameters are closely interacting and when one value changes, there will be influences to others. Monitoring of these sometimes small changes in comparison to the normative data are essential parts when talking about patient individualized supply. Close supervisions and monitoring guarantees quality in therapy organization. P11 Outcome of intensive outpatient rehabilitation in an adult patient with M. Scheuermann evaluated by radiologic imaging – a case report Hagit Berdishevsky SchrothNYC, New York, NY, USA Background No studies examine the efficacy of intensive specific physical therapy (PT) exercises along with brace for the adult with Scheuermann’s kyphosis (SK). The aim of this study was to examine the effects of intensive PT based on the Barcelona Scoliosis Physical Therapy School (BSPTS) and SpinoMed brace on a 76-year-old female with SK. Case Description J, 76-year-old female, diagnosed with SK as an adolescent, presented in October 2014 with thoracic hyperkyphosis T1 to T12 Cobb angle of 85° and lumbar hyper lordosis L1 to L5 Cobb angle 700. Lumbar scoliosis T12-L5 with 210 Cobb and vertebral rotation 2. Trunk translation in the sagittal plan was 4.5 cm. Intermittent low back pain 6/10 at worst. Quality-of-life score was 3.8 (SRS 22 questionnaire). Method The PT regimen included one-hour Schroth exercise sessions three times per week for six months. In addition, a home exercise program (HEP) was recommended. Patient also wore a SpinoMed brace for two hours per day. All tests and measurements were recorded before and after treatment. Results After a six-month treatment period the kyphosis Cobb angle was reduced to 70° and lordosis improved to 57°. A recent x-ray (October 2015) showed another improvement in the sagittal plane with thoracic kyphosis measuring 640 and lumbar lordosis 550. Lumbar curvature decreased to 120 and vertebral rotation to 1. The quality-of-life score showed improvement with a score of 4.5 on the SRS 22. Pain score diminished to 2. Trunk deviation improved by 2.2 cm. Conclusion These findings suggest that intensive and specific PT and bracing was successful for the treatment of this adult patient with SK. Consent Written informed consent was obtained from the patient for treatment, photos, and publication of this case study. A copy of the written consent is available for review by the Editor of this journal. References. 1. Iemolo B: Seven Criteria to Treat Scheuermann’s Disease: Scoliosis. 2007 2(suppl1):S39 2. Wikipedia: Kyphosis. https://en.wikipedia.org/wiki/Kyphosis. 3. Weiss HR, Dieckmann J, Gerner HJ: Outcome of In-patent Rehabilitation in Patients with M. Scheuermann Evaluated by Surface Topography: Studies in Health Technology and Informatics. 2002, 88:246-249 4. Weiss HR, Dieckmann J, Gerner HJ: Effect of Intensive Rehabilitation on Pain in Patients with Scheuermann's disease. Topography: Studies in Health Technology and Informatics. 2002, 88:254-257 5. Sinaki M, Itoi E, Rogers JW, Bergstralh EJ, Wahner HW: Correlation of Back Extensor Strength with Thoracic Kyphosis and Lumbar Lordosis in Estrogen-Deficient Women: Amer J Phys Med & Rehab. 1996, 75(5):370-375 6. Pfeifer M, Begerow B, Minne HW: Effects of a new spinal orthosis on posture, trunk strength, and quality of life in women with postmenopausal osteoporosis: A randomized trial: Am J Phys Med Rehabil. 2004, 83:177–186 P12 The effectiveness of a Scoliosis Specific Home Exercise Program and bracing to reduce an idiopathic scoliosis curve with more than 90 % success in less than a year of exercises. Case report Hagit Berdishevsky SchrothNYC, New York, NY, USA Background There have been no specific studies regarding the effect of an intensive home exercise program (HEP) using the Barcelona Scoliosis Physical Therapy School (BSPTS) method based on the principles of Katharine Schroth. The aim of this study was to examine the efficacy of an eight-month scoliosis specific HEP combined with the Wood Cheneau Rigo (WCR) brace on an eleven-year-old adolescent girl, A’, with adolescent idiopathic scoliosis (AIS). Case Description An eleven-year-old girl diagnosed with AIS in 2013. Curvatures were measured: 24.80 Upper thoracic, 410 thoracic, and 21.70 lumbar, Risser 0, angle of trunk rotation (ATR) 100. Trunk imbalance to the thoracic convex side (right), Pelvic translation to the thoracic concave side (left), No pain reported on initial evaluation. Physical therapy using the BSPTS/Schroth method and WCR brace initiated within three and a half months of diagnosis. Method Patient received ten hours of intensive outpatient physical therapy to learn BSPTS/Schroth-based exercises prior to bracing. This program was followed by an additional 10 hours of instructional sessions. HEP 30-60 min/day, 5 days/week. WCR brace worn 22 hours/day. Results By the conclusion of the eight-month treatment period, the patient had experienced significant and measurable improvement: Upper thoracic 9.60, Thoracic 00 and lumbar 80. The patient ATR reduced from 100to 20in less than two years. Conclusion This case study demonstrates that bracing and self-discipline HEP A was able to significantly improve her curve magnitude and clinical appearance in eight months. Consent Written informed consent was obtained from the patient’s parent for treatment, photos, and publication of this case study. A copy of the written consent is available for review by the Editor of this journal. References 1. Rigo M, Quera-Salvá G, Villagrasa M, Ferrer M, Casas A, Corbella C, Urrutia A, Martínez S, Puigdevall N: Scoliosis intensive out-patient rehabilitation based on Schroth method. Stud Health Technol Inform. 2008, 135:208-27 2. Schreiber S, Parent EC, Hedden DM, Hill D, Moreau MJ, Lou E, Watkins EM,Southon SC. The effect of Schroth exercises added to the standard of care on the quality of life and muscle endurance in adolescents with idiopathic scoliosis—an assessor and statistician blinded randomized controlled trial: “SOSORT 2015 Award Winner”. Scoliosis. 2015;10:24. 3. Kuru T, Yeldan İ, Dereli EE, Özdinçler AR, Dikici F, Çolak İ. The efficacy of three-dimensional Schroth exercises in adolescent idiopathic scoliosis: A randomised controlled clinical trial. Clinil Rehabil. 2015. 4. Rigo M ,Villagrasa M, Galo DA: Specific Scoliosis Classification Correlating with Brace Treatment: Description and Reliability. Scoliosis. 2010, 5:1 5. Monticone M, Ambrosini E, Cazzaniga D, Rocca B, Ferrante S: Active Self-Correction and Task-Oriented Exercises Reduce Spinal Deformity and Improved Quality of Life in Subjects with Mild Adolscent Idiopathic Scoliosis. Results of Randomised Controlled Trial. Eur Spine J. 2014, 23(6):1204-14 ᅟ

PMC005xxxxxx/PMC5001245.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556804290510.1186/s12864-016-2905-xResearchStatistical modeling for sensitive detection of low-frequency single nucleotide variants Hao Yangyang haoyan@iupui.edu 12Zhang Pengyue zhangpe@imail.iu.edu 23Xuei Xiaoling xxuei@iupui.edu 45Nakshatri Harikrishna hnakshat@iupui.edu 67Edenberg Howard J. edenberg@iu.edu 145Li Lang lali@iu.edu 12Liu Yunlong yunliu@iupui.edu 12571 Department of Medical and Molecular Genetics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 2 Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 3 Department of Biostatistics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 4 Department of Biochemistry and Molecular Biology, Indiana University School of Medicine, Indianapolis, IN 46202 USA 5 Center for Medical Genomics, Indiana University School of Medicine, Indianapolis, IN 46202 USA 6 Department of Surgery, Indiana University School of Medicine, Indianapolis, IN 46202 USA 7 IU Simon Cancer Center, Indiana University School of Medicine, Indianapolis, IN 46202 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.514© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Sensitive detection of low-frequency single nucleotide variants carries great significance in many applications. In cancer genetics research, tumor biopsies are a mixture of normal and tumor cells from various subpopulations due to tumor heterogeneity. Thus the frequencies of somatic variants from a subpopulation tend to be low. Liquid biopsies, which monitor circulating tumor DNA in blood to detect metastatic potential, also face the challenge of detecting low-frequency variants due to the small percentage of the circulating tumor DNA in blood. Moreover, in population genetics research, although pooled sequencing of a large number of individuals is cost-effective, pooling dilutes the signals of variants from any individual. Detection of low frequency variants is difficult and can be cofounded by sequencing artifacts. Existing methods are limited in sensitivity and mainly focus on frequencies around 2 % to 5 %; most fail to consider differential sequencing artifacts. Results We aimed to push down the frequency detection limit close to the position specific sequencing error rates by modeling the observed erroneous read counts with respect to genomic sequence contexts. 4 distributions suitable for count data modeling (using generalized linear models) were extensively characterized in terms of their goodness-of-fit as well as the performances on real sequencing data benchmarks, which were specifically designed for testing detection of low-frequency variants; two sequencing technologies with significantly different chemistry mechanisms were used to explore systematic errors. We found the zero-inflated negative binomial distribution generalized linear mode is superior to the other models tested, and the advantage is most evident at 0.5 % to 1 % range. This method is also generalizable to different sequencing technologies. Under standard sequencing protocols and depth given in the testing benchmarks, 95.3 % recall and 79.9 % precision for Ion Proton data, 95.6 % recall and 97.0 % precision for Illumina MiSeq data were achieved for SNVs with frequency > = 1 %, while the detection limit is around 0.5 %. Conclusions Our method enables sensitive detection of low-frequency single nucleotide variants across different sequencing platforms and will facilitate research and clinical applications such as pooled sequencing, cancer early detection, prognostic assessment, metastatic monitoring, and relapses or acquired resistance identification. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2905-x) contains supplementary material, which is available to authorized users. The International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background In 2005, the first next-generation sequencing (NGS) technology was released by 454 Life Sciences (now Roche) [1]. Within the past ten years, different sequencing technologies and platforms, including Illumina, SOLiD, Ion Torrent, Complete Genomics, were released to the public. The much faster sequencing speed, high-throughput capacity and now up to several hundred bases read length, together with a greatly reduced cost, revolutionized the scope and efficiency of biomedical related field researches [2]. Paired with the increasingly diverse range of biological application of NGS technologies, numerous computational and informatics tools, frameworks and pipelines emerged to enable researchers to harness the power of NGS technologies. Statistical models suitable for count data modeling gained much attention in NGS data analysis due to the discrete count nature of the data generated by NGS sequencers. Such models were broadly applied in DNA sequencing (DNA-Seq) based variants identification such as samtools [3], VarScan2 [4], and SNVMix [5]. For DNA sequencing based single nucleotide variant (SNV) identification, emerging new applications bring challenges to refine the statistical modeling methods and pushing the limit of NGS technologies. In cancer genetics research, low frequency tumor somatic SNV identification is crucial due to the inevitable normal tissue contamination [6, 7] and the highly heterogeneous, constantly evolving nature of tumors [8]. Accurate and sensitive identification of low frequency SNVs also carries clinical significance, since it enables the early diagnosis, cancer progression monitor and relapse identification. The recent discovery of circulating tumor DNA (ctDNA) also gained much attention. Contrast to traditional tumor biopsies, which is invasive and can only offer a snapshot of the tumor genetics landscape at certain checkpoints, ctDNA based ‘liquid biopsy’ [9] is non-invasive and can be done repeatedly for close monitoring of early sign of relapse or metastasis. However, ctDNA only takes a small percentage of all blood sample DNA, a previous research [10] reported for some advanced cancers, ctDNA is about 1 ~ 10 % of blood DNA. The difficulty for low-frequency SNV identification using NGS technologies is due to the relatively high sequencing artifacts or error rates, which is around 0.1 ~ 1 % for most platforms. Further, such error rates differ significantly under various genome contexts. For example, Illumina sequencing data are prone to have mismatches while Ion Torren and Ion Proton data contain more homopolymer related indels and consequently, mismatches near homopolymer loci [11–13]. For somatic SNV identification paired tumor-normal design, some existing methods derive the sequencing error probability from base qualities followed by error likelihood ratio test of tumor and normal sample at the same location, for example in Mutect [7], Strelka [14]. While VarScan2 applies a Fisher’s exact test on the paired samples, treating non-reference read counts from the normal sample as background error rate. The former failed to consider differential error rates for substitution types while the latter only utilized information in one location thus the background error rate estimation is off. For one sample low-frequency SNV calling, UDT-Seq [15] tabulated the error rate based on substitution types, strand and location on the read to derive an empirical background error rate, then use binomial model to distinguish signal from error, and the candidate SNVs are further refined by 7 filters. This method is context-aware but also ad-hoc, thus the ability to adapt to different sequencing technologies is limited. A brief summary of the tools mentioned above is included in Additional file 1. By analyzing previous efforts, our group proposed a framework (Yangyang Hao XX, Li L, Nakshatri H, Edenberg HJ, Liu Y. RareVar: A Framework for Detecting Low Frequency Single Nucleotide Variants, submitted) to first generated position specific error model (PSEM) using genome sequence contexts for candidate SNV identification and then apply a machine-learning model to refine the candidates. Testing on an Ion Proton benchmark dataset, our framework outperforms existing methods, especially at 0.5 % to 3 % frequencies. However, the potential to improve PSEM performances on SNVs with close to sequencing error rates by implementing more sophisticated statistical modeling and the generalizability and adaptiveness of PSEM remain untested. In this research, we explored what distributions fit the DNA-Seq data error rates modeling as well as the possibility of improved position specific error rate prediction for higher precision and recall on SNVs down to 0.5 % frequency. Further, we evaluated how different sequencing technologies affect the behavior of PSEM and the generalizability and adaptiveness of the PSEM framework. Results In the Result section, we first briefly summarized the benchmark datasets used for PSEM. Then we compared the testing benchmark dataset from Ion Proton with Illumina MiSeq sequencing data in terms of allele frequency composition and depth distribution. Utilizing count data visualization plots and tabulation, we selected the candidate distributions that may fit the data. Since the Ion Proton dataset contains 3 times of the number of benchmark SNVs from Illumina MiSeq and also is enriched with SNVs of ≤ 1 % allele frequency, we mainly focused on Ion Proton data set for model development and evaluation. To test the generalizability of the PSEM, we further trained and evaluated it on Illumina MiSeq dataset. Benchmarks overview and comparison Two sets of designed benchmarks targeting low-frequency SNVs from both Ion Proton and Illumina MiSeq [15] sequencing technologies were included. The details of these 2 datasets are described in Methods section and Additional files 2 and 3. Briefly, the Ion Proton training benchmark is the sequencing data from a single individual with known genotypes, while the test benchmark was designed to mimic the paired normal-tumor design for somatic SNVs identification applications. The Illumina MiSeq benchmark data were generated by mixing 4 individuals at 4 different percentages and then permuted the mixing percentage assignment 4 times to generate 4 calibration datasets – CAL_A, CAL_B, CAL_C and CAL_D. Since the 4 calibration data sets were generated with the same procedures, without loss of generality, we used CAL_A as training benchmark and treated the others as testing benchmark. Comparing the two testing benchmarks, Ion Proton contains a total of 1557 somatic SNVs while Illumina MiSeq contains 514 SNV – mixed allele frequency pairs, with 175 unique SNVs. More importantly, Ion Proton benchmark was designed to comprehensively characterize the SNV caller performance on close to sequencing error allele frequencies, thus it is enriched with SNVs of < = 3 % allele frequencies, with 0.5 % as the lowest targeted frequency. Plotting the cumulative percentages of SNV numbers at different allele frequencies (Fig. 1) from the two test benchmarks, it is clear the major components of Ion Proton benchmark SNV allele frequencies are at 0.5 %, 1 %, 2 % to 5 %, followed by continuous frequencies until 46 %, the maximum somatic SNV frequency designed in the dataset. Whereas MiSeq data set includes roughly equal percentages of SNVs at 4 discrete allele frequency levels.Fig. 1 Allele frequency composition of Ion Proton and Illumina MiSeq testing benchmark SNVs Except for allele frequency composition, sequencing depth is also a crucial factor affecting the performances of the SNV callers, especially at the low-frequency ranges. The average depth for Ion Proton sequencing testing benchmark is about 4000x and about 1500x for MiSeq. In addition, despite the amplicon-based capture assay was applied on benchmark datasets from both technologies, the evenness of the depth across the targeted regions is different. When comparing the depth on known testing benchmark SNV loci of two technologies (Fig. 2), the depth distribution for Ion Proton is skewed while the distribution profile for Illumina MiSeq data displays a bell shape. Further, the average depth at SNV loci from both benchmarks are around 3000x, despite the much higher overall depth for Ion Proton. Thus, we speculate lowered recall for some Ion Proton benchmark SNVs, particularly for the ≤ 1 % ones, the identifiable power of which are more sensitive to the depth and read count number sampling variances.Fig. 2 SNV loci depth distribution by allele frequency for Ion Proton and Illumina MiSeq. The dashed lines show the 3000x depth Candidate distributions selection To model error rate based on count data, 3 most common distribution choices are binomial, Poisson and negative binomial (NB) distributions. We applied a graphical exploratory plot – distplot [16–18] on the model response – number of reads containing non-reference bases – to get visual intuition about the overall fit of response data on different distributions. Intuitively, if an assumed distribution fits the data well, the data points should follow a straight line determined by the distribution metameters. As shown in Fig. 3, the obvious curve for binomial distribution plot suggests binomial distribution is not appropriate. The Poisson and NB plots show better agreement with the straight line although both curves deviate more from the straight line when the x-axis approaches 0. Tabulating the percentages of zero in the model responses show for Ion Proton training dataset, 85 % is 0 while 80 % for MiSeq training data. Thus zero-inflated models should be considered. In the modeling step, we included Poisson, NB and their zero-inflated counterparts (zero-inflated Poisson [19] or ZIP and zero-inflated negative binomial [20] or ZINB) as the candidate distributions under generalized linear model (GLM) framework.Fig. 3 Distplot on binomial, Poisson and negative binomial distributions. The y-axis is the distribution metameter calculated by the method distplot used. The open points show the observed count metameters; the filled points show the confidence interval centers and the dashed lines show the confidence intervals for each point. 95 % confidence interval is used Comparing the goodness-of-fit of different distributions 9 genomic sequence context covariates, totaling 24 degrees of freedom, were included in the GLM models (Methods and Additional file 4). Since ZIP and ZINB GLM require covariates for both the ‘zero’ and ‘count’ parts, the same covariates were provided for both, resulting in doubled degrees of freedom of those included in Poisson and NB GLM. To compare the goodness-of-fit of models based on different distributions, we used Vuong’s non-nested hypothesis test [21]. BIC-corrected Vuong z-statistic was used to impose stronger penalty on additional parameters. The pairwise comparison results are summarized in Table 1. Poisson distribution GLM is treated as the reference distribution to compare to, given its simple configuration. As expected, NB GLM is superior to Poisson GLM, since NB models dispersion of the data, and this is also supported by dispersion test [22] (z = 68.5881, p value < 2.2e-16). The necessity of modeling zero-inflation is supported by the Vuong’s test comparing ZIP with Poisson GLM. When comparing ZIP with NB, NB fits the data better. However, it is worth noting the evidence of superiority – the absolute value of BIC-corrected Vuong z-statistic – is much smaller than the other tests. The merit of considering both dispersion and zero-inflation is further emphasized by the comparisons of ZIP with ZINB and NB with ZINB. In conclusion, based on Vuong’s test, for Ion Proton sequencing dataset, the most appropriate distribution is ZINB, followed by NB, ZIP and Poisson.Table 1 Vuong’s non-nested tests on 4 distributions applied to Ion Proton training data Model 1 Model 2 Vuong z-statistic BIC-corrected Hypothesis P value Poisson NB −122.67 model2 > model1 <2.22e-16 Poisson ZIP −143.73 model2 > model1 <2.22e-16 NB ZIP 36.81 model1 > model2 <2.22e-16 ZIP ZINB −92.16 model2 > model1 <2.22e-16 NB ZINB −119.51 model2 > model1 <2.22e-16 Performance evaluation on Ion Proton testing benchmark We first evaluated the overall precision and recall values of all models on the test benchmark. From Table 2, it is observed the Poisson GLM achieves the highest recall while ZINB GLM has the highest precision. F1 score, the harmonic mean of precision and recall, is used to evaluate the overall performance. The conclusion from F1 score is consistent with that of Vuong’s test, with ZINB performs the best, followed by NB, ZIP and Poisson GLM. However, the precision values listed in Table 2 are lower than the ones reported previously [7, 14, 15]. There are 2 major reasons: 1. the Ion Proton test benchmark dataset is designed to enrich with low-frequency SNVs, with 68.9 % of all SNVs of allele frequency < = 3 %, in which 17.3 % at 0.5 % frequency and 19.8 % at 1 % frequency, whereas the majority of previous studies focused on SNVs of > = 5 % allele frequency; 2. one popular paradigm of SNV calling is a two-step procedure, first generating SNV candidates and then applying multiple sequencing quality filters to refine the SNV call. The PSEM aims to efficiently recover high quality SNV candidates to facilitate the filtering step, thus it is only fair to compare the performance of PSEM with other candidate generating methods. The result from VarScan2 before applying sequencing quality filters was included in Table 2. It is evident that except for Poisson GLM, the other methods outperformed VarScan2 in both recall and precision. Therefore, choosing appropriate statistical modeling method enables us to recover more true SNVs without any loss of precision in candidate generating step.Table 2 Overall performance comparison for Ion Proton testing benchmark Poisson NB ZIP ZINB VarScan2 Recall 0.98 0.89 0.95 0.90 0.83 Precision 0.25 0.62 0.54 0.71 0.53 F1 Score 0.40 0.73 0.69 0.79 0.65 Next, for all distributions, we explored the performance profiles on different allele frequencies. As shown in Fig. 4, the well-separated F1 score levels clearly show that SNVs of lower allele frequencies are more difficult to identify, no matter what distributions were used. In addition, the significant separation of 0.5 % from the other allele frequencies indicate the detection limit is around 0.5 % under current sequencing platform and depth. Meanwhile, the power of appropriate modeling is evident when comparing the performances of all distributions on SNVs of 0.5 % allele frequency. Relative to Poisson GLM, considering either zero-inflation or dispersion boosted the F1 score by about 0.2 at 0.5 %, while considering both by ZINB further increased F1 score by about 0.1. Interestingly, compared with the second best model – NB GLM, both precision and recall increased in ZINB GLM, which pinpoints the necessity of modeling zero-inflation to derive more accurate error rates estimation. Furthermore, for SNVs with allele frequency greater than 1 %, the average recall is 97.5 % with 82.3 % average precision for ZINB GLM. To summarize, the performance evaluation results on low-frequency SNV identification also support the conclusion from Vuong’s non-nested test, with ZINB being the most appropriate model. Further, the necessity of modeling both dispersion and zero-inflation is exemplified by the much-elevated performance at close to sequencing error rate allele frequency, which is important for pushing down the detection limit of low-frequency SNV callers.Fig. 4 Performance by allele frequency summary on Ion Proton testing benchmark Application of ZINB PSEM on Illumina MiSeq data To evaluate the generalizability and adaptiveness of the GLM based PSEM, the same modeling strategies were applied to the Illumina MiSeq sequencing data sets. The same genomic sequence context features from Ion Proton modeling were applied to the Illumina MiSeq CAL_A dataset. Similar to the analysis on Ion Proton data set, paired Vuong’s non-nested hypothesis tests were conducted on the 4 candidate distributions, with details summarized in Additional file 5. The test conclusions remained the same except for the NB (model 1) and ZIP (model 2) comparison, where the BIC-corrected Vuong z-statistic is −0.47 resulting in p value = 0.318. Therefore the goodness-of-fit for these two distributions on MiSeq dataset are not significantly different. Despite similar statistical modeling schema can be readily generalized to Illumina MiSeq data set, Illumina MiSeq and Ion Proton sequencers differ significantly in terms of sequencing chemistry. The former is based on sequencing-by-synthesis (SBS) that relies on high-resolution optic systems, whereas the latter is based on Ion semiconductor sequencing where no modified nucleotides or optics are required. The differences in sequencing mechanisms make Ion Proton sequencers run faster but are prone to homopolymer related errors. Comparing the NB GLM regression coefficients on both datasets (Additional file 6), homopolymer related features significant in Ion Proton data set regression are either insignificant (hmer_len, hmer_dist) or show opposite effect (hmer_op, hmer_den) on the error rate. The same trend was also observed in ZIP and ZINB models comparing Ion Proton with Illumina MiSeq (Additional files 7, 8, 9 and 10). To evaluate whether the differences in GLM coefficients affect the performance profiles on various allele frequencies, we applied the 4 GLM models trained on CAL_A to the other 3 calibration datasets and conducted the recall, precision and F1 score analyses by allele frequency on the combined dataset. As shown in Fig. 5, similar to the Ion Proton data set, SNVs of lower allele frequencies are more difficult to identify. However, when comparing the performances of ZIP with NB GLM on 0.5 % ~ 1 % allele frequency, different from Ion Proton dataset, NB demonstrated a much higher F1 score compared with ZIP. A closer look at the performance profiles shows the noticeable drop in recall comparing NB with ZIP in Ion Proton is absent in MiSeq data. Examination on the benchmark SNVs missed by NB but recovered by ZIP showed lower depth for the missed ones. While the absent of recall drop in MiSeq is due to its relatively even depth contrast to the Ion Proton dataset (Fig. 2). For SNVs with > 1 % allele frequency, the F1 scores are all greater than 0.9 and clustered together for all distributions.Fig. 5 Performance by allele frequency summary on Illumina MiSeq testing benchmark Comparing with the results from UDT-Seq [15], which reported approximately 90 % recall and >95 % precision (no specific number was given, the precision was inferred by the precision for the other data UDT-Seq tested - Illumina GAII benchmark data at 1500x depth), ZINB GLM demonstrates higher overall recall (95.1 %) and high precision (93.4 %). Discussion The PSEM model aims to predict the position specific error rates associated with various genomic sequence contexts, under which the specific sequencing technology is prone to error. Based on publications evaluating features associated with sequencing errors and experiences from our previous effort, 9 types of significant features are considered. With the features fixed, using GLM, we evaluated the appropriateness of distributions with different mean – variance relationships and the ability to consider zero-inflation. Consistent with the computational tool EdgeR [23] for RNA-Seq data, we found the ability to model over-dispersion by NB distribution necessary for DNA-Seq data as well. Additionally, for DNA-Seq erroneous read counts modeling, zero-inflation is also a key factor for accurate prediction and inference. The much-elevated F1 score for 0.5 % allele frequency SNVs as well as the highest overall performance by ZINB GLM highlighted the importance of choosing suitable statistical models. Moreover, comparing with VarScan2, which conducts the Fisher’s exact test for each targeted location on paired normal-tumor sequencing data, the significance of applying the correct reference error model is exemplified by higher recalls as well as precisions for 0.5 % and 1 % frequency SNVs. In theory, for low frequency SNV loci, VarScan2 treated the sequencing reads with non-reference bases from normal as the background error, which is essentially point estimation based on one location. Whereas PSEM collectively considers all loci with similar context features and thus is able to generate more accurate error estimation. The evaluation of PSEM modeling on Illumina MiSeq dataset and the performance comparison with Ion Proton dataset show the generality of the PSEM framework as well as its adaptiveness to different technologies. Moreover, except for the established importance of choosing appropriate statistical model, the sequencing depth evenness is also an important factor affecting low-frequency SNVs calling performances. The current GLM-based PSEM framework only considers 9 types of genome sequence context features. To further improve the performances, more informative features associated with sequencing errors should be included and tested. In addition, from the modeling aspect, exploration of the potential to further increase the performances by applying more sophisticated computational models are desired. To better understand its generalizability and adaptiveness, tests on other sequencing technologies, such us SOLiD and Complete Genomics, are necessary. Besides, since the capture assay for the two benchmarks is amplicon-based, hybridization-based approach should be tested to compare the performance profiles. Differentiating low frequency SNVs from sequencing artifacts is the key for identifying SNVs at frequencies close to sequencing error rates. Our PSEM approach tried to push the limit toward the sequencing error rates. Based on the analyses on benchmarks from standard sequencing protocols and the given sequencing depth, we speculate the detection limit is around 0.5 % on the regions covering all exons of hundred of genes, with a total size up to millions of bases. However, with high accuracy sequencing protocols, such as duplex sequencing [24] and ultra-deep target enrichment assay [25], the researchers reported identification of SNVs around 0.1 % on a single gene scale. Despite the promising results, more efforts to make such protocols applicable on larger regions are required for broader applications. Conclusion Our method enables sensitive detection of low-frequency single nucleotide variants across different sequencing platforms down to 0.5 % frequency. Thus will facilitate research and clinical applications such as pooled sequencing, cancer early detection, prognostic assessment, metastatic monitoring, and relapses or acquired resistance identification. Methods Overall workflow For position specific error model training, we used the invariant loci from training benchmark. Genomic sequence context features were extracted for each locus and then fed to the generalized linear models using 4 different distributions. Then testing benchmark paired tumor and normal sequencing data went through the PSEM and the candidate SNVs were derived. Additional file 11 provides a diagram illustrating this procedure. In the following method section, we first introduced the benchmark datasets from both Ion Proton and Illumina MiSeq. Then we described the application of generalized linear models for PSEM. Last, we described the performance evaluation metrics. Benchmark dataset Both Ion Proton and Illumina MiSeq datasets were generated from amplicon-based targeted sequencing. The targeted region for Ion Proton datasets included all exons of 409 known cancer-related genes, totaling about 1.7 million bases covered by about 16,000 amplicon primer pairs from Ion AmpliSeq™ Comprehensive Cancer Panel. The training benchmark is the DNA sequencing data of NA11993. The testing benchmark mimic the paired normal-tumor design, where the normal sample is the DNA sequencing data of NA12878 while tumor sample is a mixture of 17 individuals from 1000 Genomics Project plus NA12878. The mixing percentage assignment is listed in Additional file 2. The sequencing data were aligned with TMAP from Torrent Suite software. Reads with mapping quality less than 40 were filtered out. The length of targeted regions for Illumina MiSeq datasets is 23.2 kb, covered by 158 amplicons. The design details can be found in the paper [15] and Additional file 3. The raw reads were downloaded from NCBI Short Read Archive (SRP009487.1) and processed as the paper described. Reads with mapping quality less than 30 were filtered out. Generalized linear models The details of the 9 genomic sequence contexts considered in GLM were summarized in Additional file 4. Briefly, general contexts including substitution types, immediate upstream and downstream bases, GC content, and homopolymer related features: whether the locus is within a homopolymer, the closest homopolymer length, the distance to the closest homopolymer, the local homopolymer base percentages and whether the alternative base is the same as the immediate upstream or downstream base are considered. These 9 features are the covariates included in the GLMs. The Poisson GLM for erroneous sequencing read counts with log link function is expressed in eq. (1), where Ns,b,l is the observed number of erroneous reads for strand s (forward or reverse) with alternative base b (three possible values other than the reference) at location l, λs,b,l represents the expected mean for Ns,b,l, cs,b,l is the vector of genomic sequence context covariates, and β is the vector of fitted coefficients. The sequencing depth for strand s at location l is treated as the offset. 1 logλs,b,l=logENs,b,l|cs,b,l=logds,l+β'cs,b,l The negative binomial distribution GLM with log link function can be expressed in eq. (2), where μs,b,l represents the expected mean for Ns,b,l and θ is the dispersion parameter (the shape parameter of the gamma mixing distribution). The mean E(Ns,b,l) = μs,b,l and variance VAR(Ns,b,l) = μs,b,l + θμs,b,l2 can be estimated from GLM shown below. 2 logμs,b,l=logENs,b,l|cs,b,l=logds,l+β'cs,b,l The zero-inflated Poisson distribution can be written as: 3 PNs,b,l=ns,b,l|πs,b,l,λs,b,l,θ={πs,b,l+1−πs,b,lPoisλs,b,l0ifns,b,l=01−πs,b,lPoisλs,b,lns,b,lifns,b,l>0 Parameters of the zero-inflated Poisson distribution (3) can be estimated by generalized linear model as shown in (4), where zs,b,l is the vector of genomic sequence context covariates for the zero part, and γ is the vector of fitted coefficients. 4 logitπs,b,l1−πs,b,l=γ'zs,b,l logλs,b,l=β'cs,b,l The zero-inflated negative binomial distribution can be written as: 5 PNs,b,l=ns,b,l|cs,b,l,zs,b,l={πs,b,l+1−πs,b,lNBμs,b,l,θ;0ifns,b,l=01−πs,b,lNBμs,b,lθns,b,lifns,b,l>0 Parameters of the zero-inflated negative binomial distribution (5) can be estimated by generalized linear model as shown in (6). 6 logitπs,b,l1−πs,b,l=γ'zs,b,l logμs,b,l=β'cs,b,l A location with a certain alternative base is called as a candidate SNV if the numbers of reads from both strands are significantly greater than the predicted error rates. The p values were corrected using Benjamini–Hochberg procedure. The corrected p value cut-off is 0.01. Performance evaluation measurements Precision, recall and F1 score are defined below. precision=numberofrecoveredbenchmarkSNVsnumberofpredictedSNVs recall=numberofrecoveredbenchmarkSNVsexpectednumberofbenchmarkSNVs F1=2*precision*recallprecision+recall For Ion Proton dataset, loci with at least 5 reads supporting alternative base are included in the evaluation. For Illumina MiSeq dataset, filter 2 used by UDT-Seq was applied which requires > = 0.2 % frequency for alternative bases. However, the other filters were not used. We relied on the PSEM framework to properly address sequencing problems, for example, uneven depth and local sequence context induced errors. Additional files Additional file 1: Summary of major somatic callers, their methods of detecting SNV candidates and the limitations. (PDF 60 kb) Additional file 2: Ion Proton testing benchmark design. (PDF 58 kb) Additional file 3: Illumina MiSeq benchmark design. (PDF 58 kb) Additional file 4: Genomic sequence context features used in GLM. (PDF 71 kb) Additional file 5: Vuong’s non-nested test on 4 distributions applied to Illumina MiSeq training data. (PDF 65 kb) Additional file 6: Negative binomial GLM coefficients for Ion Proton and Illumina MiSeq training datasets. (PDF 71 kb) Additional file 7: Zero-inflated Poisson GLM coefficients for Ion Proton training datasets. (PDF 69 kb) Additional file 8: Zero-inflated Negative Binomial GLM coefficients for Ion Proton training datasets. (PDF 69 kb) Additional file 9: Zero-inflated Poisson GLM coefficients for Illumina MiSeq training datasets. (PDF 71 kb) Additional file 10: Zero-inflated Negative Binomial GLM coefficients for Illumina MiSeq training datasets. (PDF 70 kb) Additional file 11: Overall workflow. This diagram illustrates the training and testing steps. The training data and the position specific error model derived from it are highlighted with dashed lines. After training, testing benchmark paired normal and tumor samples go through the PSEM model and the candidate SNVs are derived. (PDF 115 kb) This study is supported by the funds from the US National Institutes of Health U10AA008401 (Collaborative Study on the Genetics of Alcoholism, to H.E.), Pilot funds from the breast cancer program of IUSCC (Indiana University Simon Cancer Center) and Zeta Tau Sorority and Susan G Komen for the Cure grant SAC110025 (to Y.L. and H.N.). The sequencing was performed in the Center for Medical Genomics (CMG) sequencing core at Indiana University School of Medicine. Declarations The publication costs for this article were funded by the corresponding author. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Availability of data and materials The datasets supporting the conclusions of this article are included within the article and its Additional files 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 and 11. Authors' contributions YL and LL conceived the project. YH and PZ conducted the analysis. LL provided statistical support. XX performed the sequencing experiment. YH, YL, and HE wrote the manuscript. All authors reviewed the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare no competing financial interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Margulies M Egholm M Altman WE Attiya S Bader JS Bemben LA Berka J Braverman MS Chen YJ Chen Z Genome sequencing in microfabricated high-density picolitre reactors Nature 2005 437 7057 376 380 16056220 2. van Dijk EL Auger H Jaszczyszyn Y Thermes C Ten years of next-generation sequencing technology Trends in Genetics: TIG 2014 30 9 418 426 10.1016/j.tig.2014.07.001 25108476 3. Li H A statistical framework for SNP calling, mutation discovery, association mapping and population genetical parameter estimation from sequencing data Bioinformatics 2011 27 21 2987 2993 10.1093/bioinformatics/btr509 21903627 4. Koboldt DC Zhang Q Larson DE Shen D McLellan MD Lin L Miller CA Mardis ER Ding L Wilson RK VarScan 2: somatic mutation and copy number alteration discovery in cancer by exome sequencing Genome Res 2012 22 3 568 576 10.1101/gr.129684.111 22300766 5. Goya R Sun MG Morin RD Leung G Ha G Wiegand KC Senz J Crisan A Marra MA Hirst M SNVMix: predicting single nucleotide variants from next-generation sequencing of tumors Bioinformatics 2010 26 6 730 736 10.1093/bioinformatics/btq040 20130035 6. Carter SL Cibulskis K Helman E McKenna A Shen H Zack T Laird PW Onofrio RC Winckler W Weir BA Absolute quantification of somatic DNA alterations in human cancer Nat Biotechnol 2012 30 5 413 421 10.1038/nbt.2203 22544022 7. Cibulskis K Lawrence MS Carter SL Sivachenko A Jaffe D Sougnez C Gabriel S Meyerson M Lander ES Getz G Sensitive detection of somatic point mutations in impure and heterogeneous cancer samples Nat Biotechnol 2013 31 3 213 219 10.1038/nbt.2514 23396013 8. Meacham CE Morrison SJ Tumour heterogeneity and cancer cell plasticity Nature 2013 501 7467 328 337 10.1038/nature12624 24048065 9. Crowley E Di Nicolantonio F Loupakis F Bardelli A Liquid biopsy: monitoring cancer-genetics in the blood Nat Rev Clin Oncol 2013 10 8 472 484 10.1038/nrclinonc.2013.110 23836314 10. Diehl F Li M Dressman D He Y Shen D Szabo S Diaz LA Jr Goodman SN David KA Juhl H Detection and quantification of mutations in the plasma of patients with colorectal tumors Proc Natl Acad Sci U S A 2005 102 45 16368 16373 10.1073/pnas.0507904102 16258065 11. McElroy KE Luciani F Thomas T GemSIM: general, error-model based simulator of next-generation sequencing data BMC Genomics 2012 13 74 10.1186/1471-2164-13-74 22336055 12. Bragg LM Stone G Butler MK Hugenholtz P Tyson GW Shining a light on dark sequencing: characterising errors in Ion Torrent PGM data PLoS Comput Biol 2013 9 4 e1003031 10.1371/journal.pcbi.1003031 23592973 13. Ross MG Russ C Costello M Hollinger A Lennon NJ Hegarty R Nusbaum C Jaffe DB Characterizing and measuring bias in sequence data Genome Biol 2013 14 5 R51 10.1186/gb-2013-14-5-r51 23718773 14. Saunders CT Wong WS Swamy S Becq J Murray LJ Cheetham RK Strelka: accurate somatic small-variant calling from sequenced tumor-normal sample pairs Bioinformatics 2012 28 14 1811 1817 10.1093/bioinformatics/bts271 22581179 15. Harismendy O Schwab RB Bao L Olson J Rozenzhak S Kotsopoulos SK Pond S Crain B Chee MS Messer K Detection of low prevalence somatic mutations in solid tumors with ultra-deep targeted sequencing Genome Biol 2011 12 12 R124 10.1186/gb-2011-12-12-r124 22185227 16. Hoaglin DC A poissonness plot Am Stat 1980 34 No.3 146 149 17. Hoaglin DC Mosteller F Tukey JW Hoaglin DC Mosteller F Tukey JW Checking the Shape of Discrete Distributions Checking the Shape of Discrete Distributions, in Exploring Data Tables, Trends, and Shapes 2011 Hoboken John Wiley & Sons, Inc 18. Friendly M Visualizing Categorical Data 2000 Cary SAS Institute 19. Lambert D Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing Technometrics 1992 34 1 1 14 10.2307/1269547 20. Greene WH. Accounting for Excess Zeros and Sample Selection in Poisson and Negative Binomial Regression Models. In: NYU Working Paper No. EC-94-10; 1994. 21. Vuong QH Likelihood Ratio Tests for Model Selection and Non-Nested Hypotheses Econometrica 1989 57 2 307 333 10.2307/1912557 22. Cameron AC Trivedi PK Regression-Based Tests for Overdispersion in the Poisson Model J Econometrics 1990 46 3 347 364 10.1016/0304-4076(90)90014-K 23. Robinson MD McCarthy DJ Smyth GK edgeR: a Bioconductor package for differential expression analysis of digital gene expression data Bioinformatics 2010 26 1 139 140 10.1093/bioinformatics/btp616 19910308 24. Kennedy SR Schmitt MW Fox EJ Kohrn BF Salk JJ Ahn EH Prindle MJ Kuong KJ Shen JC Risques RA Detecting ultralow-frequency mutations by Duplex Sequencing Nat Protoc 2014 9 11 2586 2606 10.1038/nprot.2014.170 25299156 25. Schmitt MW Fox EJ Prindle MJ Reid-Bayliss KS True LD Radich JP Loeb LA Sequencing small genomic targets with high efficiency and extreme accuracy Nat Methods 2015 12 5 423 425 10.1038/nmeth.3351 25849638

PMC005xxxxxx/PMC5001246.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556634290410.1186/s12864-016-2904-yResearchConcordance of copy number loss and down-regulation of tumor suppressor genes: a pan-cancer study Zhao Min mzhao@usc.edu.au 1Zhao Zhongming zhongming.zhao@uth.tmc.edu 23451 School of Engineering, Faculty of Science, Health, Education and Engineering, University of the Sunshine Coast, Maroochydore DC, QLD 4558 Australia 2 Department of Biomedical Informatics, Vanderbilt University School of Medicine, Nashville, TN 37203 USA 3 Department of Cancer Biology, Vanderbilt University School of Medicine, Nashville, TN 37232 USA 4 Department of Psychiatry, Vanderbilt University School of Medicine, Nashville, TN 37212 USA 5 Center for Precision Health, School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, TX USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.532© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Tumor suppressor genes (TSGs) encode the guardian molecules to control cell growth. The genomic alteration of TSGs may cause tumorigenesis and promote cancer progression. So far, investigators have mainly studied the functional effects of somatic single nucleotide variants in TSGs. Copy number variation (CNV) is another important form of genetic variation, and is often involved in cancer biology and drug treatment, but studies of CNV in TSGs are less represented in literature. In addition, there is a lack of a combinatory analysis of gene expression and CNV in this important gene set. Such a study may provide more insights into the relationship between gene dosage and tumorigenesis. To meet this demand, we performed a systematic analysis of CNVs and gene expression in TSGs to provide a systematic view of CNV and gene expression change in TSGs in pan-cancer. Results We identified 1170 TSGs with copy number gain or loss in 5846 tumor samples. Among them, 207 TSGs tended to have copy number loss (CNL), from which fifteen CNL hotspot regions were identified. The functional enrichment analysis revealed that the 207 TSGs were enriched in cancer-related pathways such as P53 signaling pathway and the P53 interactome. We further performed integrative analyses of CNV with gene expression using the data from the matched tumor samples. We found 81 TSGs with concordant CNL events and decreased gene expression in the tumor samples we examined. Remarkably, seven TSGs displayed concordant CNL and gene down-regulation in at least 50 tumor samples: MTAP (212 samples), PTEN (139), MCPH1 (85), FBXO25 (67), SMAD4 (64), TRIM35 (57), and RB1 (54). Specifically to MTAP, this concordance was found in 14 cancer types, an observation that is not much reported in literature yet. Further network-based analysis revealed that these TSGs with concordant CNL and gene down-regulation were highly connected. Conclusions This study provides a draft landscape of CNV in pan-cancer. Our findings of systematic concordance between CNL and down-regulation of gene expression may help better understand the TSG biology in tumorigenesis and cancer progression. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2904-y) contains supplementary material, which is available to authorized users. Keywords Tumor suppressor genePan-cancerCopy number variationCopy number lossGene expressionThe International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Cancer is characterized by unconstrained cell proliferation. In the normal cell, there is precise control of cell division such as cell cycle check points [1]. In cellular system, tumor suppressor genes (TSGs) are important guardian genes that protect a normal cell from one step on the path to uncontrolled growth [2, 3]. In cancer cells, TSGs may lose their normal functions because of mutations occur at its critical sites. For single nucleotide or small insertions/deletions (indels), these mutations often lead to truncation of transcripts or proteins, including nonsense mutations, splicing site mutations, or frameshift mutations. Similar effects can be caused by larger scale mutations, such as copy number variations (CNVs), gene fusions, or structural variants (SVs) [4, 5]. The mutated TSGs often coordinate with oncogenes for cancer progression [4, 6, 7]. Therefore, the identification and understanding of TSGs have profound influence to develop the diagnosis biomarkers and effective drugs for cancer therapies. CNVs are the variable number of DNA fragments in the human genome. Their lengths typically range from a kilo base pairs to a mega base pairs [8]. CNVs are divided into two major groups: copy number loss (CNL) and copy number gain (CNG). CNL denotes the decreased gene (or sequence fragment) copies in the genome while CNG denotes the gain of additional gene copies in the human genome. With the development of high throughput technologies such as Comparative genomic hybridization (CGH) array and next-generation sequencing, a very large number of CNVs, as well as other types of mutations and genomics data (e.g., gene expression) have been unveiled, especially in cancer genomes [9, 10]. This allows us to systematically study cancer mutation signatures, heterogeneity, and other molecular features [11]. For CNV, such deleted or duplicated DNA fragments often have profound effects on gene expression, which subsequently affects gene’s function [12]. Despite a number of studies have explored CNVs and gene expression in various cancers [13], there has been no systematic study of the features in TSGs yet. Moreover, the results from single cancer type may not be representative in other types of cancer, or they may vary among the subtypes of the cancer. To overcome these limitations, we conducted a pan-cancer CNV analysis on TSGs to explore the landscape of CNV features and cross-validate some observations. This study may help us better elucidate the relationship between CNV and gene expression change in this important gene category in cancer. Methods The curated TSGs from thousands of literatures To conduct a systematic CNV survey of TSGs, we downloaded all the 1207 curated human TSGs from TSGene database in a plain text format with all the Entrez Gene IDs and official symbols (version 2.0) [2]. In this version of TSGene database, there were 1088 protein-coding and 198 non-coding TSGs. All these TSGs were manually curated from over 9000 PubMed abstracts by us. To annotate TSGs with CNVs, it requires the genomic location for mapping. Therefore, we downloaded the corresponding RefSeq mapping information for TSGs from RefSeq database. We implemented an in-house script to extract all the genomic location information from the completed human genome RefSeq sequences (accession number starting with NC). In total, 1207 TSGs were annotated with accurate genomic locations in GRCH 38. The pan-cancer CNV data from The Cancer Genome Atlas (TCGA) To explore the CNVs in pan-cancer systematically, we downloaded all the prepared TCGA CNV data with the GRCH 38 genomic coordinates from the Catalogue of Somatic Mutations in Cancer (COSMIC) database (V73). When integrating TCGA data, COSMIC introduced a few thresholds to define the copy number loss and gain. CNG was obtained by the following criteria: (the average genome ploidy < =2.7 AND total DNA segment copy number > =5) OR (average genome ploidy >2.7 AND total DNA segment copy number > =9). Similarly, the criteria for CNL were: (the average genome ploidy < =2.7 AND total DNA segment copy number =0) OR (average genome ploidy >2.7 AND total DNA segment copy number < (average genome ploidy – 2.7)). In this study, we followed COSMIC criteria and overlapped all the CNV regions with TSGs using the GRCH 38 coordinates. By intersecting all the CNV gain and loss information to all the 1207 TSGs with GRCH 38 coordinates, we annotated 1170 TSGs with precise gain or loss information. For each cancer type, we calculate the number of samples with CNL and CNG, respectively. Since TSGs are often in loss-of-function in cancer progression, we pulled out those TSGs with higher frequency of CNLs than that with CNGs. Specifically, we set a cut-off of 2 to filter out those TSGs without having at least twice of tumor samples with CNLs as tumor samples with CNGs. This process resulted in 207 TSGs with the evidence of an overall loss of CNVs. These genes were used for the following gene expression analysis. Gene expression analysis for TSGs with CNL To check the CNV-correlated gene expression changes on TSGs, we downloaded the TCGA pan-cancer gene expression data from the COSMIC database (V73). In this study, we focused on only those gene expression changes in the matched TCGA samples with TSG CNLs. For the gene expression quantification, COSMIC started from FPKM calculated using trimmed short reads generate by the RNA-Seq platform and the RSEM quantification results from the RNAseq V2 platform. Here, FPKM denotes Fragments Per Kilobase of transcript per Million mapped reads, which is used to indicate the relative expression of a transcript. And RSEM is one of the popular measures for accurate transcript quantification of RNA-Seq data. The average and the standard deviation of expression were computed using those tumor samples that are diploid for each corresponding gene. The standard Z score was used to characterize whether a TSG is over or under expressed. The Z-score with absolute value 2 was used as the threshold value. The Z-score over 2 was defined as over expression while the Z-score less than −2 represented the decreased gene expression. For those 81 TSGs with CNL-associated gene expression change, we further systematically examined their somatic CNV patterns in pan-cancer of TCGA samples using cBio portal [14]. Sub-network extraction for the TSGs with high frequency CNLs To explore the relevant biological mechanisms related to TSGs with frequently observed CNLs and consistent gene down-regulation, we extracted a PPI network to connect 81 TSGs with the remaining human genes. To this goal, we started from a non-redundant human interactome extracted from the Pathway Commons database [15, 16], containing 3629 proteins and 36,034 PPIs. It is worth noting that this integrated human interactome is based on well-curated pathway databases (HumanCyc, Reactome, and KEGG pathway database [17]). Therefore, those links in the interactome have biological meaning rather than physical interactions. Based on the pathway-based interactions, we used the similar approach implemented in our previous study to extract a sub-network related to our 81 TSGs [16, 18, 19]. In this sub-network extraction strategy, all the 81 seed TSGs were overlapped to the human pathway-based interactome. Then, a sub-network with the maximum number of the seed TSGs was formed by connecting each TSG through the shortest path. To characterize the function of the network, we relied on the network topological properties (degree and shortest path) calculated from the network. In practice, we utilized NetworkAnalyzer plugin in Cytoscape 2.8 to compute topological properties in the TSG network [20]. The degree is defined as the number of direct connections of each node with other nodes in the TSG network [21, 22]. The network layout was conducted based on Cytoscape 2.8 [20]. Results Genomic regions with frequent copy number loss in TSGs in multiple cancer types To systematically survey the somatic CNVs in TSGs, our pipeline started with a list of 1207 human TSGs from the TSGene 2.0 database [2, 23] (Fig. 1). These genes have multiple lines of evidence in literature and other data, and have been manually curated. To provide an unbiased global view of CNVs in major types of cancer, we overlapped all these 1207 TSGs with the somatic CNVs identified from TCGA, which is the largest cancer genomics data source. This resulted in a list of non-redundant 1170 TSGs, which are annotated with CNVs (Additional file 1: Table S1). However, the majority CNVs are not informative due to the lack of matched control tissue. In this study, we only focused on those TSGs with precise gain or loss data using the normal tissue as control (see Methods). By counting the number of samples with gain or loss of gene copies, we set a threshold to prioritize most informative CNV events for TSGs. Since TSGs typically play their roles by loss-of-function, we used only those TSGs that tend to have copy number loss. To this end, we required CNVs were observed in at least as twice the samples with copy number loss as those with copy number gain. The process resulted in a total of 207 TSGs. We named them as TSGs with CNL in cancer, and used them for the follow up functional enrichment and integrative analyses. The list is provided in Additional file 2: Table S2.Fig. 1 Pipeline for the identification of concordant copy number loss and down-regulation of tumor suppressor genes in human cancer. This figure shows the pipeline for identifying the tumor suppressor genes (TSGs) with concordant copy number variations (CNVs) and gene expression. It involves four main steps. 1) Downloading TSGs from the TSGene 2.0 database and overlapping to the TCGA pan-cancer CNV data. 2) The resulted 1170 TSGs with CNV overlapping information were further extracted and calculated the precise copy number gain (CNG) and loss (CNL). 3) Based on the number of samples with CNGs and CNLs in the pan-cancer CNV data, we collected 207 TSGs with frequent CNLs. 4) Using the gene expression data from the matched TCGA cancer samples, we identified 81 TSGs with consistent CNLs and decreased gene expression in the same samples We performed functional enrichment analysis of these 207 genes using Gene Ontology (GO) terms as functional units. Figure 2 displays the main features of GO-related functional features, and their clusters. Overall, they are enriched with cell proliferation, apoptosis, cell cycle, and growth control; all are important features of cancer cells. The TSGs with CNL also have fundamental roles in development such as embryonic morphogenesis and reproduction. In addition, they involve in negative regulation of cell metabolism and protein phosphorylation. Some TSGs may influence cell communication, cell junction assembly, and response to the extracellular stimulus.Fig. 2 Gene Ontology (GO) analysis of 207 human tumor suppressor genes (TSGs) with frequent copy number losses (CNLs). The scatterplot shows the GO clusters for the 207 TSGs with CNLs in a two-dimensional space derived by applying multidimensional scaling to a matrix of the GO terms' semantic similarities. Bubble color represents the frequency of the GO term in the GOA database (more general terms are toward red). Bubble size indicates the log of corrected p-value (the smaller corrected p-value, the larger bubble) Interestingly, the 207 TSGs could highly cluster into fifteen chromosome regions. All these regions had the corrected enrichment p-values less than 0.01 (Table 1). Among the 15 regions, eight could be further clustered into three genomics locations: 3p21, 8p21-22, and 17p13.1-3. In 3p21, we found four enriched cytobands with a total of 27 TSGs. For example, the 3p21.3 cytoband is enriched with 14 TSGs (CTDSPL, CYB561D2, LIMD1, MST1R, NPRL2, PTPN23, RASSF1, RBM5, RBM6, RHOA, SEMA3B, SEMA3F, TUSC2, and ZMYND10). Another six TSGs (CDCP1, LTF, MIR1226, SETD2, SMARCC1, and TDGF1) were clustered in 3p21.31. The remaining 7 TSGs were located close to 3p21. These 27 TSGs in 3p21 had CNLs in 129 TCGA samples covering 20 different cancer types. Specific to the tissue site, there were 27, 20, 15, 15, and 11 samples from lung, central nervous systems, kidney, breast, and large intestine, respectively. A similar observation of the high frequency of loss involving the short arm of chromosome 3 was reported as a tumor suppressor locus in a variety of histologically different neoplasms more than twenty years ago [24, 25]. However, our survey provides precise locations in various cancer samples.Table 1 The 15 genomics regions associated with 207 tumor suppressor genes (TSGs) with frequent copy number losses (CNLs) Cytoband p-value q-value # TSGs TSG list 3p21.3 5.98E-19 3.48E-16 14 CTDSPL, CYB561D2, LIMD1, MST1R, NPRL2, PTPN23, RASSF1, RBM5, RBM6, RHOA, SEMA3B, SEMA3F, TUSC2, ZMYND10 8p22 4.48E-09 9.78E-07 7 CCAR2, DLC1, LZTS1, MIR383, MTUS1, SOX7, ZDHHC2 11p15.5 5.05E-09 9.78E-07 10 CDKN1C, H19, MIR210, MIR483, NUP98, RNH1, SIRT3, TRIM3, TSPAN32, TSSC4 17p13.1 5.30E-08 7.71E-06 9 ALOX15B, BCL6B, GABARAP, MIR195, MIR497, TNK1, TP53, XAF1, ZBTB4 19p13.3 2.52E-06 2.93E-04 10 AMH, DAPK3, DIRAS1, FZR1, GADD45B, PLK5, SIRT6, STK11, TCF3, TNFSF9 8p21.3 3.45E-06 3.34E-04 5 DOK2, MIR320A, PIWIL2, PPP3CC, RHOBTB2 3p21.31 3.63E-05 3.01E-03 6 CDCP1, LTF, MIR1226, SETD2, SMARCC1, TDGF1 3p21.1 1.49E-04 1.02E-02 4 ACY1, CACNA2D3, MIR135A1, MIRLET7G 6q26 1.58E-04 1.02E-02 3 MAP3K4, IGF2R, PACRG 10q24-q25 2.09E-04 1.16E-02 2 CHUK, MXI1 8p21 2.19E-04 1.16E-02 3 BNIP3L, EXTL3, TNFRSF10A 17p13.3 2.40E-04 1.16E-02 5 ALOX15, MNT, MYBBP1A, PAFAH1B1, VPS53 22q13.31 2.62E-04 1.17E-02 4 FBLN1, MIRLET7B, MIRLET7A3, PPARA 9p21 7.46E-04 3.10E-02 3 CDKN2A, CDKN2B, MTAP 3p21 1.08E-03 4.19E-02 3 GNAT1, MST1, PBRM1 q-values were calculated by Benjamini-Hochberg multiple testing correction of the raw p-values, which were calculated by the hypergeometric test On chromosome 8, the 8p22 locus contained 7 neighbouring TSGs (CCAR2, DLC1, LZTS1, MIR383, MTUS1, SOX7, and ZDHHC2), while another 8 TSGs (BNIP3L, DOK2, EXTL3, MIR320A, PIWIL2, PPP3CC, RHOBTB2, and TNFRSF10A) clustered at 8p21. These 15 TSGs at 8p21-22 had CNL detected in 219 TCGA patients. The cancer tissues that had most frequent CNLs in TSGs at this locus are breast (61 samples), lung (42 samples), large intestine (30 samples), ovary (23 samples), and prostate (11 samples). Another CNL hot region is at 17p13.1-3, which covers 14 TSGs, including the most studied TSG TP53. This region on chromosome 17 had detectable CNLs in a total of 50 TCGA tumor samples. Interestingly, the above three genomic regions with frequent CNLs in TSGs harbour not only well-known protein-coding TSGs such as TP53, but also six microRNAs (MIRLET7G, MIR135A1, MIR195, MIR320A, MIR383, and MIR497). By overlapping to TSGene database, we found all the six microRNAs are tumor suppressor microRNAs. Collectively, our systematic examination on CNL in TSG cluster regions provides precise information of such CNL in multiple cancers. The results may be useful for further studying the similar or different roles of CNL in differential cancer types as well as cancer heterogeneity. Correlation of CNL with gene expression decrease using the matched tumor samples Through incorporating the gene expression change of the TCGA samples with the CNL on TSGs, we examined the correlation between CNL and TSG gene down-regulation. We utilized the Z-score to assess whether a TSG is up-regulated or down-regulated in specific TCGA samples. Here, Z-score is a transformation of the p value calculated by the formula as below: Z=x‐μσ Where μ represents the mean expression of a gene across multiple TCGA samples; σ represents the standard deviation of the expression scores of the gene in TCGA samples. Specifically, we used the Z-score threshold value −2 to identify down-regulated TSGs in specific TCGA samples. After examining the same TCGA tumor samples for both expression and CNV loss, we found 81 TSGs that had concordant decreased gene expression and loss of gene copy numbers in tumor samples (Additional file 3: Table S3). The functional enrichment analyses revealed that the 81 TSGs are mainly associated with cancer-related pathways such as cell cycle (adjusted P-value = 1.15E-6) (Additional file 4: Table S4). Interestingly, they are also related to a number of cancer-related phenotypes such as hamartomatous polyposis (adjusted P-value = 1.05E-6) and intussusception (adjusted P-value = 3.98E-6). The CNV mutational patterns across multiple cancers are plotted in Fig. 2. In terms of their CNVs, these 81 TSGs are highly mutated. For example, in TCGA esophageal carcinoma cohort, there were 142 cases (77.2 %) that had at least one gene with copy number change (Additional file 5: Table S5). More than 50 % of the esophageal carcinoma patients had at least one deletion event in one of the 81 TSGs. The similar prevalence of copy number alteration (>60 % cases, including both CNLs and CNGs) was found in other 11 cancer datasets from 9 cancer types: metastatic prostate cancer, malignant peripheral nerve sheath tumor, sarcoma, ovarian serous cystadenocarcinoma, lymphoid neoplasm diffuse large B-cell lymphoma, bladder urothelial carcinoma, glioblastoma multiforme, uterine carcinosarcoma, and stomach adenocarcinoma. However, specific to CNLs, only the glioblastoma multiforme (GBM) cohort had over 60 % patients with CNLs (Fig. 3, the blue bar represent the CNLs in different cancer types). Although the other cancer cohorts also possess CNLs in the majority of the affected patients, a small portion of patients had CNGs rather than CNLs. This may imply the importance of the 81 TSGs in cancer progression of GBM via the massive copy number losses.Fig. 3 A pan-cancer global view of copy number variation (CNV) features based on 81 tumor suppressor genes (TSGs) with decreased gene expression potentially induced by copy number losses (CNLs) To further explore the potential CNL-induced gene expression change, we specifically checked four TSGs with the most frequently observed CNLs; all these genes were observed in more than 50 tumor samples. As shown in Fig. 4, TSG MTAP had CNL in more than 40 % cases in TCGA GBM cohort. The TSG MCPH1 was deleted in more than 14 % patients in a prostate cancer dataset. PTEN showed similar frequent CNV loss in prostate cancer samples. The gene loss of SMAD4 was prominent in the pancreatic cancer. Furthermore, we found consistent, low gene expression of these four genes in the tumor samples with CNL (Fig. 5). The results suggested that CNL might induce gene expression decrease as a common mechanism in cancer.Fig. 4 A pan-cancer view of copy number of variation (CNV) distribution in four tumor suppressor genes: MTAP, MCPH1, PTEN, and SMAD4. The CNV mutational landscape for (a) MTAP, (b) MCPH1, (c) PTEN, and (d) SMAD4 Fig. 5 The correlation of copy number variation (CNV) and gene expression in four tumor suppressor genes: MCPH1, MTAP, PTEN, and SMAD4. a MTAP using TCGA glioblastoma data, (b) MCPH1 using TCGA breast cancer data, (c) PTEN using TCGA lung squamous carcinoma (LUSC), and (d) SMAD4 using TCGA colorectal cancer data A connected biological map of TSGs with concordant CNL and decreased gene expression To further investigate the common functional regulation and enhance our understanding of the cellular events of these 81 TSGs with decreased expression and CNL, we conducted a pathway-based protein-protein interaction (PPI) analysis using the pathway annotation data from Pathway Commons database [15]. These reliable interactions are based on evidence from known biological pathways such as the KEGG and Reactome pathway databases. Therefore, this pathway-based interactome is useful for the pathway reconstruction because such pathways may avoid high-level noises, sparseness, and highly skewed degree distribution, which are often observed in physical interaction-based PPI networks. By applying the Klein-Ravi algorithm for module searching [18], we first mapped the 81 TSGs to the human pathway interactome. Then, a subnetwork was extracted, allowing to connect as many as the 81 TSGs as possible. The reconstructed TSG network contains 54 nodes and 56 links (Fig. 6a). Among the 54 nodes, 35 are from the 81 TSGs. The remaining 19 are the linker genes to bridge those TSGs so that a fully connected map could be built. The degrees of the nodes in this reconstructed map potentially follow a power law distribution P(k) ~ k-b, where P(k) is the probability that a node has connections with other k nodes and b is an exponent with an estimated value of 1.4 (Fig. 6b). Moreover, most of the genes in the network can be connected by three to five steps on average, as measure by the shortest path (Fig. 6c). These two topological features (degree and shortest path) indicated that most TSGs in this map were closely connected with high modularity. Considering the tight connection of the map, the nodes with multiple connections are likely to play critical and concordant roles to mediate biological regulation such as signalling transduction in cellular system. In the network, there are 7 nodes with four or more connections: TP53 (12 connections), SMAD4 (6), TGFBR2 (6), MAP3K1 (5), HSP90AA1 (5), ATM (4), and SP1 (4). It is interestingly that TP53 is the node with most connection in the network. SMAD4 is also in the centre of the map with six connections. In summary, our reconstructed map for the 81 TSGs with potential CNL-driven gene down-regulation contains some interesting features such as the TSGs with potential CNL-induced dysregulation.Fig. 6 Reconstructed interaction map for the 81 tumor suppressor genes (TSGs) with decreased gene expression potentially induced by copy number losses (CNLs). a The network includes 35 genes (in yellow) from the 81 TSGs with decreased expression potentially induced by CNLs and 19 linker genes (in blue) that connect these 35 TSGs. The node size reflects the number of connection. A bigger size means more connections associated with the gene. b The degree distribution of the nodes (genes) in the network (a). c The distribution of the shortest path length Discussion This study revealed some important somatic mutational features of TSGs in multiple cancer types, particularly with respect to the CNVs and their effects on gene expression. Since the loss-of-function is the typical mechanism that TSGs involve in cancer initiation and progression, a large-scale change of gene copy number may induce gene expression alteration. In this scenario, a critical regulation change is that CNL in a TSG leads to the over-expression of its guardian genes. Although previous studies have explored the balance of germline CNVs and gene expression, there still lack of direct links of somatic CNVs on gene expression dosage compensation. In this study, we only focused on the concordant patterns between CNL and gene down-regulation because TSGs often play functions in a manner of loss-of-function. Our results only provided the insight of correlation between gene dosage and somatic CNV; more systematic examination of the expression quantitative trait locus may provide more depth on the relationship between CNV and gene expression. This study was mainly based on the TCGA genomic data. The cohort size of some cancer types is relatively small (e.g., ~100 samples). A small sample size may filter out many low-frequency CNVs. In addition, TCGA mainly relies on the CGH array between normal and tumor tissues to characterize CNVs, which may lose signals outside of pre-designed probes. These undetected CNVs may also contribute to TSGs functionality on cancer progression. Another limitation in this study is that we only incorporate the protein-coding gene expression, not including non-coding gene expression. The further integration of large-scale CNV data and gene expression of noncoding RNA (microRNA and long non-coding RNA) may provide new insight into the roles of the non-coding TSGs. In this study, we made an effort to construct a biological map for the genes with consistent CNL and gene down-regulation in cancer. Although the majority of genes in the reconstructed map are linked with each other, the size of the network is relatively small. Therefore, it has limited power to explore the overall network functions based on the topological features. For example, we found the degree of the network might follow the power-law distribution. This feature is different from the whole human PPI network, in which the majority nodes (genes) are sparsely connected with exponent b as 2.9 [26]. It is not sufficient to impose the scale-free properties on this constructed small network due to the small size. For the same reason, it is not good for us to define the hub nodes based on the high connectivity. Nevertheless, the nodes with multiple connections in our network should provide some clues for the common CNL events related to gene down-regulation. The further experimental validation may provide more insight into the potential molecular mechanisms for those CNL events that were detected in multiple cancers. Conclusions In conclusion, our systematic exploration of copy number variations on human TSGs revealed that the copy number loss of TSGs cluster in a few genomics regions. These TSGs with frequent copy number loss often have profound roles in cancer-related pathways. The loss of copy number in a number of TSGs may contribute to the gene expression change involving tumorigenesis. Abbreviations CNG, copy number gain; CNL, copy number loss; CNV, copy number variation; TCGA, The Cancer Genome Atlas; TSG, tumor suppressor gene Additional files Additional file 1: Table S1. The 207 tumor suppress genes (TSGs) with frequent copy number losses (CNLs). (XLSX 18 kb) Additional file 2: Table S2. The enriched pathways and interactors for the 207 tumor suppress genes (TSGs) with frequent copy number loss (CNL). (XLSX 11 kb) Additional file 3: Table S3. The 81 tumor suppress genes (TSGs) with decreased gene expression potentially induced by copy number losses (CNLs). (XLSX 15 kb) Additional file 4: Table S4. Functional enrichment results of 81 tumor suppress genes (TSGs) with decreased gene expression potentially induced by copy number losses (CNLs). (XLSX 24 kb) Additional file 5: Table S5. The CNV frequency of 81 tumor suppress genes (TSGs) with decreased gene expression potentially induced by copy number losses (CNLs) in multiple cancers. (XLSX 12 kb) Acknowledgements We thank the investigators of The Cancer Genome Atlas (TCGA) whose effort of data generation and analyses made this work possible. Declaration Publication of this article was charged from the faculty retention funds to Z.Z. from Vanderbilt University. This article has been published as part of BMC Genomics Volume 17 Supplement 7, 2016: Selected articles from the International Conference on Intelligent Biology and Medicine (ICIBM) 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-7. Funding This work was partially supported by National Institutes of Health (NIH) grant (R01LM011177) and Ingram Professorship Funds (to Z.Z.). The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. Availability of data and materials Datasets supporting the results of this article are also included in the Additional files 1, 2, 3, 4 and 5: Tables S1 to S5. Authors’ contributions MZ and ZZ conceived the project. MZ collected the data and carried out all the analyses. MZ wrote the manuscript draft and MZ and ZZ finalized the manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Ethics approval and consent to participate Not applicable. ==== Refs References 1. Malumbres M Barbacid M Cell cycle, CDKs and cancer: a changing paradigm Nat Rev Cancer 2009 9 3 153 166 10.1038/nrc2602 19238148 2. Zhao M, Kim P, Mitra R, Zhao J, Zhao Z. TSGene 2.0: an updated literature-based knowledgebase for tumor suppressor genes. Nucleic Acids Res. 2016;44(D1):D1023-1031. 3. Sherr CJ Principles of tumor suppression Cell 2004 116 2 235 246 10.1016/S0092-8674(03)01075-4 14744434 4. Haber DA Settleman J Cancer: drivers and passengers Nature 2007 446 7132 145 146 10.1038/446145a 17344839 5. Pellman D Cell biology: aneuploidy and cancer Nature 2007 446 7131 38 39 10.1038/446038a 17330036 6. Zhao M Sun J Zhao Z Distinct and competitive regulatory patterns of tumor suppressor genes and oncogenes in ovarian cancer PLoS One 2012 7 8 e44175 10.1371/journal.pone.0044175 22952919 7. Zhao M Sun J Zhao Z Synergetic regulatory networks mediated by oncogene-driven microRNAs and transcription factors in serous ovarian cancer Mol Biosyst 2013 9 12 3187 3198 10.1039/c3mb70172g 24129674 8. Alkan C Coe BP Eichler EE Genome structural variation discovery and genotyping Nat Rev Genet 2011 12 5 363 376 10.1038/nrg2958 21358748 9. Zhao M Wang Q Wang Q Jia P Zhao Z Computational tools for copy number variation (CNV) detection using next-generation sequencing data: features and perspectives BMC Bioinformatics 2013 14 Suppl 11 S1 10.1186/1471-2105-14-S11-S1 10. Zhao M Zhao Z CNVannotator: a comprehensive annotation server for copy number variation in the human genome PLoS One 2013 8 11 e80170 10.1371/journal.pone.0080170 24244640 11. Jia P Pao W Zhao Z Patterns and processes of somatic mutations in nine major cancers BMC Med Genomics 2014 7 11 10.1186/1755-8794-7-11 24552141 12. Henrichsen CN Chaignat E Reymond A Copy number variants, diseases and gene expression Hum Mol Genet 2009 18 R1 R1 8 10.1093/hmg/ddp011 19297395 13. Lu TP Lai LC Tsai MH Chen PC Hsu CP Lee JM Hsiao CK Chuang EY Integrated analyses of copy number variations and gene expression in lung adenocarcinoma PLoS One 2011 6 9 e24829 10.1371/journal.pone.0024829 21935476 14. Gao J Aksoy BA Dogrusoz U Dresdner G Gross B Sumer SO Sun Y Jacobsen A Sinha R Larsson E Integrative analysis of complex cancer genomics and clinical profiles using the cBioPortal Sci Signal 2013 6 269 l1 10.1126/scisignal.2004088 15. Cerami EG Gross BE Demir E Rodchenkov I Babur O Anwar N Schultz N Bader GD Sander C Pathway Commons, a web resource for biological pathway data Nucleic Acids Res 2011 39 Database issue D685 690 10.1093/nar/gkq1039 21071392 16. Zhao M Austin ED Hemnes AR Loyd JE Zhao Z An evidence-based knowledgebase of pulmonary arterial hypertension to identify genes and pathways relevant to pathogenesis Mol Biosyst 2014 10 4 732 740 10.1039/C3MB70496C 24448676 17. Kanehisa M Araki M Goto S Hattori M Hirakawa M Itoh M Katayama T Kawashima S Okuda S Tokimatsu T KEGG for linking genomes to life and the environment Nucleic Acids Res 2008 36 Database issue D480 484 18077471 18. Zhao M Li X Qu H EDdb: a web resource for eating disorder and its application to identify an extended adipocytokine signaling pathway related to eating disorder Sci China Life Sci 2013 56 12 1086 1096 10.1007/s11427-013-4573-2 24302289 19. Zhao M Kong L Qu H A systems biology approach to identify intelligence quotient score-related genomic regions, and pathways relevant to potential therapeutic treatments Sci Rep 2014 4 4176 24566931 20. Smoot ME Ono K Ruscheinski J Wang PL Ideker T Cytoscape 2.8: new features for data integration and network visualization Bioinformatics 2011 27 3 431 432 10.1093/bioinformatics/btq675 21149340 21. Barabasi AL Oltvai ZN Network biology: understanding the cell's functional organization Nat Rev Genet 2004 5 2 101 113 10.1038/nrg1272 14735121 22. Zhao M Qu H High similarity of phylogenetic profiles of rate-limiting enzymes with inhibitory relation in Human, Mouse, Rat, budding Yeast and E. coli BMC Genomics 2011 12 Suppl 3 S10 10.1186/1471-2164-12-S3-S10 22369203 23. Zhao M Sun J Zhao Z TSGene: a web resource for tumor suppressor genes Nucleic Acids Res 2013 41 Database issue D970 976 10.1093/nar/gks937 23066107 24. Rowley H Jones A Spandidos D Field J Definition of a tumor suppressor gene locus on the short arm of chromosome 3 in squamous cell carcinoma of the head and neck by means of microsatellite markers Arch Otolaryngol Head Neck Surg 1996 122 5 497 501 10.1001/archotol.1996.01890170031007 8615966 25. Killary AM Wolf ME Giambernardi TA Naylor SL Definition of a tumor suppressor locus within human chromosome 3p21-p22 Proc Natl Acad Sci U S A 1992 89 22 10877 10881 10.1073/pnas.89.22.10877 1438292 26. Jin Y Turaev D Weinmaier T Rattei T Makse HA The evolutionary dynamics of protein-protein interaction networks inferred from the reconstruction of ancient networks PLoS One 2013 8 3 e58134 10.1371/journal.pone.0058134 23526967

PMC005xxxxxx/PMC5001247.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27557076290710.1186/s12864-016-2907-8ResearchA method for identifying discriminative isoform-specific peptides for clinical proteomics application Zhang Fan fan.zhang@unthsc.edu 1Chen Jake Y. JakeChen@iupui.edu 23451 Department of Molecular and Medical Genetics, University of North Texas Health Science Center, Fort Worth, TX 76107 USA 2 Wenzhou Medical University 1st Affiliate Hospital, Zhejiang Province, China 3 Institute of Biopharmaceutical Informatics and Technology, Wenzhou Medical University, Zhejiang Province, China 4 School of Informatics and Computing, Indiana University, Indianapolis, IN 46202 USA 5 Indiana Center for Systems Biology and Personalized Medicine, Indianapolis, IN 46202 USA 22 8 2016 22 8 2016 2016 17 Suppl 7 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editors declare that they have no competing interests.522© The Author(s). 2016 Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background Clinical proteomics application aims at solving a specific clinical problem within the context of a clinical study. It has been growing rapidly in the field of biomarker discovery, especially in the area of cancer diagnostics. Until recently, protein isoform has not been viewed as a new class of early diagnostic biomarkers for clinical proteomics. A protein isoform is one of different forms of the same protein. Different forms of a protein may be produced from single-nucleotide polymorphisms (SNPs), alternative splicing, or post-translational modifications (PTMs). Previous studies have shown that protein isoforms play critical roles in tumorigenesis, disease diagnosis, and prognosis. Identifying and characterizing protein isoforms are essential to the study of molecular mechanisms and early detection of complex diseases such as breast cancer. However, there are limitations with traditional methods such as EST sequencing, Microarray profiling (exon array, Exon-exon junction array), mRNA next-generation sequencing used for protein isoform determination: 1) not in the protein level, 2) no connectivity about connection of nonadjacent exons, 3) no SNPs and PTMs, and 4) low reproducibility. Moreover, there exist the computational challenges of clinical proteomics studies: 1) low sensitivity of instruments, 2) high data noise, and 3) high variability and low repeatability, although recent advances in clinical proteomics technology, LC-MS/MS proteomics, have been used to identify candidate molecular biomarkers in diverse range of samples, including cells, tissues, serum/plasma, and other types of body fluids. Results Therefore, in the paper, we presented a peptidomics method for identifying cancer-related and isoform-specific peptide for clinical proteomics application from LC-MS/MS. First, we built a Peptidomic Database of Human Protein Isoforms, then created a peptidomics approach to perform large-scale screen of breast cancer-associated alternative splicing isoform markers in clinical proteomics, and lastly performed four kinds of validations: biological validation (explainable index), exon array, statistical validation of independent samples, and extensive pathway analysis. Conclusions Our results showed that alternative splicing isoform makers can act as independent markers of breast cancer and that the method for identifying cancer-specific protein isoform biomarkers from clinical proteomics application is an effective one for increasing the number of identified alternative splicing isoform markers in clinical proteomics. Electronic supplementary material The online version of this article (doi:10.1186/s12864-016-2907-8) contains supplementary material, which is available to authorized users. The International Conference on Intelligent Biology and Medicine (ICIBM) 2015 Indianapolis, IN, USA 13-15 November 2015 http://watson.compbio.iupui.edu/yunliu/icibm/issue-copyright-statement© The Author(s) 2016 ==== Body Background Clinical proteomics is the application of proteomic techniques to the field of medicine with the aim of solving a specific clinical problem within the context of a clinical study. In the past year significant commitments from research institute and development of clinical proteomics has been witnessed. The application of clinical proteomic research is growing rapidly in the field of biomarker discovery, especially in the area of cancer diagnostics. Clinical proteomics holds the potential of taking a snapshot of the total protein complement of a cell, or body fluid, and identifying proteins as potential biomarkers for the differentiation of disease and health [1]. The study of clinical proteomic may provide us with opportunities in more effective strategies for early disease detection and monitoring, more effective therapies, and developing a better understanding of disease pathogenesis [2]. Such studies may aim at earlier or more accurate diagnosis, improvement of therapeutic strategies, and better evaluation of prognosis and/or prevention of the disease. Although clinical proteomics currently mainly focuses on diagnostics and biomarker discovery, it includes the identification of new therapeutic targets, drugs and vaccines for better therapeutic outcomes and successful disease prevention. In addition, success for a clinical proteomics requires the communication among clinicians, statisticians/bioinformaticians and biologists [3]. Until recently, researches have viewed protein isoform as any of several different forms of the same protein, not as a new class of early diagnostic biomarkers for clinical proteomics. Protein isoforms are an essential mechanism employed by human cells to enhance molecular functional diversity encoded by the genome. For protein isoforms, we refer to proteins derived from allellic polymorphisms, mRNA alternative splicing, or post-translational modifications (PTM). Allellic polymorphisms in protein-coding genes commonly take the form of single nucleotide polymorphisms (SNPs) of genes. Alternative splicing occurs in 40–60 % human genes and works by selecting specific exons and sometimes even intronic regions of the gene into mature mRNAs. Posttranslational modifications of proteins include all chemical modifications after protein translation, e.g., phosphorylation, glycosylation, and ubiquination. Approximately 8 % of these isoforms, including both SNPs and alternative splicing, are generated during the process of transcribing the coding genes into mRNA. More than 90 % of protein isoforms are created through PTMs after the mRNA is translated into a protein. Traditional methods have been used for protein isoform determination such as EST sequencing [4], Microarray profiling [5] (exon array [6], Exon-exon junction array [7]), mRNA next-generation sequencing [8, 9]. However, there are several limitations with these traditional methods. First, they all identify isoforms in the transcript level, not in the protein level. Therefore, they cannot determinate isoform quantitatively in protein level, especially for measurement of low concentrations in biological specimens. Second, they give no connectivity information about the connection of nonadjacent exons. Third, they give no SNPs information about each exon and intron. Fourth, they give no information about posttranslational modifications of peptides. Last, the biggest challenge for the analysis of protein isoform with traditional methods is their low reproducibility by other methods such as RT-PCR. Only very few events are identified with high confidence. Indeed, the typical output is usually in the order of 10 validated alternative splicing events, which cannot meet the requirement of high through identification of protein isoform. Recent advances in clinical proteomics technology, for example, LC-MS/MS, have enabled it possible to detect complex mixtures of proteins, peptides, carbohydrates, DNA, drugs, and many other biologically relevant molecules unique to disease processes [10] in parallel in biological samples. A modern mass spectrometry (MS) instrument consists of three essential modules: an ion source module that can transform molecules to be detected in a sample into ionized fragments, a mass analyzer module that can sort ions by their masses, charges, or shapes by applying electric and magnetic fields, and a detector module that can measure the intensity or abundance of each ion fragment separated earlier. Tandem mass spectrometry (MS/MS) has the additional analytical modules for bombarding peptide ions into fragment peptide ions by pipeline two MS modules together, therefore providing peptide sequencing potentials for selected peptide ions in real time. Recent developments of new generations of mass spectrometers and improvements in the field of chromatography have revolutionized protein analytics. Particularly the combination of liquid chromatography as a separation tool for proteins and peptides with tandem mass spectrometry as an identification tool referred to as LC-MS/MS has generated a powerful and broadly used technique in the field of proteomics [11]. LC-MS/MS proteomics have been used to identify candidate molecular biomarkers in diverse range of samples, including cells, tissues, serum/plasma, and other types of body fluids. For example, Flaubert et al. discovered highly secreted protein biomarkers which changed significantly in abundance corresponding with aggressiveness by using LC-MS/MS to analyze the secreted proteomes from a series of isogenic breast cancer cell lines varying in aggressiveness: non-tumorigenic MCF10A, premalignant/tumorigenic MCF10AT, tumorigenic/locally invasive MCF10 DCIS.com and tumorigenic/ metastatic MCF 10CA cl. D. They obtained proteomes from conditioned serum-free media, analyzed the tryptic peptide digests of the secreted proteins using a Waters capillary liquid chromatograph coupled to the nanoflow electrospray source of a Waters Q-TOF Ultima API-US mass spectrometer, and separated peptide on a C18 reverse phase column [12]. Although clinical proteomics can provide better evaluation of prognosis and prevention of the disease, there exist the computational challenges of clinical proteomics studies: 1) low sensitivity of instruments leads to many false negatives detection of molecules, especially when the molecules exists in low abundance and is unable to monitor specific molecules at will that can be associated with key phenotypes (typical in Genomics or functional genomics assays); 2) high data noise (false positives) introduced by limitation of accuracy of instruments causes false identification of peptides or assignment to proteins based on single peptide evidence brings uncertainty to the value of individual peptides; and 3) high variability and low repeatability of proteomics experiments exists even in high-abundance proteins (variability within individuals under different physiological conditions, worse across individuals), and the degree of variability differs for different proteins. Therefore, in the paper, we presented a peptidomics method for identifying cancer-related and isoform-specific peptide for clinical proteomics application from LC-MS/MS which can provide hopes for improving both the sensitivity (many abundant proteins could generate alternative splicing isoforms in a cancer) and the specificity (particular types of protein isoforms may be uniquely regulated in a given condition) of candidate cancer biomarkers for clinical proteomics. First, we built a Peptidomic Database of Human Protein Isoforms, then created a peptidomics approach to perform large-scale screen of breast cancer-associated alternative splicing isoform markers in clinical proteomics, and last performed four kinds of validations: biological validation (explainable index), exon array, statistical validation of independent samples, and extensive pathway analysis. Our results showed that alternative splicing isoform makers can act as independent markers of breast cancer and that the method we presented is an effective one for increasing the number of identified alternative splicing isoform markers in clinical proteomics. Methods Reagents Ammonium carbonate, ammonium bicarbonate, urea, formic acid, lysozyme, 2-Iodoethanol, and triethylphosphine were all purchased from Sigma-Aldrich (St. Louis, MO, USA). Acetonitrile and MS grade water were purchased from Honey Burdick & Jackson (Morristown, NJ, USA). Trypsin was purchased from Worthington Biochemical Corporation (Lakewood, NJ, USA). Seppro tip IgY-12 and reagent kit were purchased from GenWay Biotech (San Diego, CA, USA). Human plasma samples Plasma protein profiles were collected by the Hoosier Oncology Group (HOG) (Indianapolis, IN, USA) in two batches, which we refer to as Study II and III (each contained 40 plasma samples from women with breast cancer and 40 plasma samples from healthy age-matched volunteer women as control). Most of patients involved in the two studies were diagnosed with a stage II or III or earlier breast cancer. Most patients had previously been treated with chemotherapy. All samples were collected with the same standard operating procedure and stored in a central repository in Indianapolis, IN, USA. The demography and clinical distribution of breast cancer stages/subtypes for Study II and III are comparable (Additional file 1: Table S1). In Study II, there are 9 metastasis and 30 non-metastasis, 30 INV and 10 DCIS, mean tumor size 1.56, 8GI, 11 GII and 15 GIII. In Study III, there are 1 metastasis and 20 non-metastasis, 23 INV and 8 DCIS, mean tumor size 1.93, 3GI, 9GII and 18GIII. Proteomics methods Biomarker identification and characterization holds great promise for more precise diagnoses and for tailored therapies. The heterogeneity of human cancers and unmet medical needs in these diseases provides a compelling argument to focus biomarker development in cancer. Mass Spectrometry (MS)- based proteomics approaches have provided insight into biomarkers of cancer and other diseases with femtomole sensitivity and high analytical precision. We presented a four steps pipeline for the identification and validation of isoform-specific peptide biomarkers from breast cancer proteomics: Peptide Search Database Construction, Peptide Identification and Quantification, Statistical Identification of Isoform Markers, and Validation. Peptide search database construction A comprehensive database of human peptides characteristic of all known and theoretic protein isoforms was developed in three steps: 1) obtaining gene structures of all protein-coding genes in the human genome, 2) compiling in silico isoform junction peptides, and 3) validating those peptides in current protein knowledgebase. First, we downloaded all information about human genes in the Ensemble [13]. We retrieved gene information such as name, position, exon phase, exon/intron coordinates, and annotation. Exons which overlap with each other were classified into a group, and a serial number was assigned to each group according to its order in the sequence. For instance, the first group in the gene would be marked as group one, and the second as group two, etc. Introns can be obtained by the sequence between two exons. Then, we generated in silico Isoform Junction Peptides (IJP), which contains two types of peptides: the peptides translated from all exons and the ones that are virtually translated from all possible exon/intron junction regions. Four types of exon/intron sequence joining types are considered when generating IJPs: intron-exon (I_E_TH, left intron retention junction), exon-intron (E_I_TH, right intron retention junction), knowledgebase validated exon-exon (E_E_KB, exon-exon junctions which can be found in Ensembl transcripts) and theoretical exon-exon (E_E_TH, exon-exon junctions which cannot be found in Ensembl transcripts). For those exons with the phase information in Ensemble transcript, we directly used the phase to translate the sequence. For those exons without the phase information in Ensemble transcript, we designed an artificial translation method as follows. This sequence is used to generate three peptides, each of which has a different opening reading frame (ORF) and a maximal length of 140 amino acid residues (longer than the longest possible peptide fragments directly obtained from a MS/MS spectrum). The three ORFs are estimated in a validation procedure, where the ORF will be discarded if a stop codon is found in exon, knowledgebase validated exon-exon, or theoretical intron-exon, or if a stop codon is found in the first exon in theoretical exon-exon or theoretical exon-intron. In the third and final step, we validated each IJP in the ensemble transcript database. Those Ensemble predicted transcripts have been mapped by Ensemble to full-length or near-full-length protein sequence already available in the public sequence databases [13]. We labeled the IJP as knowledge based (_KB) if it can be matched as a substring of any ensemble transcript of the same ensemble gene; otherwise, as theoretic (_TH). Peptide identification and quantification Proteins were prepared and subjected to LC/MS/MS analysis. Samples were run on a Surveyor HPLC (ThermoFinnigan) with a C18 microbore column (Zorbax 300SBC18, 1 mm × 5 cm). All tryptic peptides (100 μL or 20 μg) were injected onto the column in random order. Peptides were eluted with a linear gradient from 5 to 45 % acetonitrile developed over 120 min at a flow rate of 50 μL/min, and the eluant was introduced into a ThermoFinnigan LTQ linear ion-trap mass spectrometer. The data were collected in the “triple-play” mode (MS scan, Zoom scan, and MS/MS scan). We searched the OMSSA against the protein isoform database we created to identify peptide. Peptide quantification was carried out using the LC/MS-based label-free protein quantification software licensed from Eli Lilly and Company. Label-free peptide identification and peptide quantitative analysis services were performed by professionals at the Protein Analysis and Research Center/Proteomics Core of Indiana University School of Medicine, co-located at Monarch Life Sciences, Inc, Indianapolis. For a thorough review of the principle and method developed and used, refer to the review by Wang et al [14]. Briefly, once the raw files were acquired from the LTQ, all extracted ion chromatograms (XIC) were aligned by retention time. Each aligned peak should match parent ion, charge state, daughter ions (MS/MS data) and retention time (within a 1-min window). If any of these parameters were not matched, the peak was disqualified from the quantification analysis. After alignment, the area-under-the-curve (AUC) from individually aligned peak was measured, normalized, and compared for their relative abundance using methods described in [15]. All peak intensities were transformed to a log2 scale before quantile normalization. Peptides with intensity lower than preset quality threshold are marked as present; otherwise, as absent. Statistical identification of isoform markers Statistical Significance was measured by a three-step method. First, we conducted a Chi-Square Goodness-of-Fit Test to calculate the p value (also called false discovery rate). Then we calculated the FDR adjusted p value. Last, we calculated the FDR q value using the Storey-Tibshirani method [16]. We chose a significance screening filters (q < 0.05) to select peptides of which we estimated significant differences in the health and breast cancer samples. The False Positive Rate (FPR) or expected proportion of false positive among the proteins with declared changes is FPR = qvalue × number of the proteins with declared changes. Validation Four validation methods including biological, statistical, Exon Array and pathway validation methods were used to validate our results. Biological validation was carried out with Explainable Index. For gene, we define “Explainable Index” as α=#conC+1#incC+1⋅#conH+1#incH+1, where #con is the number of consistent peptide markers and #inc is the number of inconsistent peptide markers, C be cancer marker set and H be health marker set. If α > 1, we define the gene to be “more explainable”; and if α ≤ 1, we define the gene to be “less-explainable”. The “consistent” is defined as one of following three conditions:∃i, j, k:Ej ∊ H&Ei_Ek ∊ C(i < j < k) ∃j:Ej ∊ H and ∄ Ei_Ek(i < j < k) ∃i, k:Ei_Ek ∊ C(i < k) and ∄ Ej(i < j < k); And the “inconsistent” is defined as one of following three conditions:∃i, j, k:Ej ∊ C&Ei_Ek ∊ H(i < j < k) ∃j:Ej ∊ C and ∄ Ei_Ek(i < j < k) ∃i, k:Ei_Ek ∊ H(i < k) and ∄ Ej(i < j < k). For statistical validation, we used forward feedback neural network to train Study II and then test Study III. We chose each combination of N (N = 5 for five-marker panel or N = 10 for ten-marker panel or N = 26 for twenty-six-marker panel) out of all the 26 differentially expressed isoforms common in both Study II (90) and B (79) as inputs to the FFNN. The training sets are 40 healthy and 40 cancer samples from Study II. The testing sets are independent 40 healthy and 40 cancer samples from Study III. For a neural network, the output data has to be transformed into binary or numerical data. A two-variable outcome encoding scheme, i.e., healthy = (0,1), cancer = (1,0) was used. In this scheme, it is theoretically possible to have (1,1) or (0,0) as outcomes although extremely rare. For the two variable outcome encoding scheme, we constructed the input layer as N nodes (corresponding to a N-marker panel outcome), the hidden layer as 7 nodes, and the output layer as 2 nodes. In order to find the optimal classifier, we presented an optimization method that measures the area under the curve (AUC) for Receiver Operating Characteristics (ROC). In this scheme, we first trained neural network for each combination using Study II results. Then, we measured the AUC for each combination using Study III results for testing. Lastly, the optimal combination C* was determined by C*=argmincAUCNETC,P, where AUC is the area under the ROC curve of neural network’s prediction result, NET is the trained neural network, C is combination of picking N out of the 26 isoforms, and P is the testing set of Study III. The Exon Array for validation was downloaded from GSE19154 in Gene Expression Omnibus. R and BioConductor libraries were used to perform Exon Array analysis. For pathway validation, the 90 alternative splicing biomarkers in Study II and the 79 alternative splicing biomarkers in Study III were used to perform pathway analysis using the Kyoto Encyclopedia of Genes and Genomes (http://www.genome.ad.jp/kegg/) [17]. Significance level for pathway comparisons was set by hit number >2 due to results of small counts. This allows avoiding any assumptions about the shape of sampling distribution of population. Results The statistics of IJP are shown in Table 1. Among the 5060822 peptides we derived, there are 208269 exon sequences, 222731 validated exon-exon junctions, 4109197 hypothetical exon-exon junctions, 413761 exon-intron junctions and 106864 intron-exon junctions. There are 367956 normal exon-exon junctions, in which the combined exons are continuous on the gene sequence, and 3963972 skipping exon-exon junctions. The longest exon peptide length is 6057 amino acids (aa), and the average exon length is 48 aa; the longest junction peptide length is 140 aa, and the average junction peptide length is 64 aa.Table 1 The statistics of peptide database Peptide Type Number of Peptides E_E Type Number of Peptides EXON_KB 208269 Normal E_E 367956 E_E_KB 222731 Skipping E_E 3963972 E_E_TH 4109197 Peptide Length (aa) E_I_TH 413761 Longest Exon 6057 I_E_TH 106864 Average Exon 48 Longest Junction 140 Total 5060822 Average Junction 64 Among the total 5060822 peptides, intron-exon junctions account for the largest proportion, and theoretical exon-exon junctions the smallest proportion. Majority of exon-exon junctions are normal, while the minority are exon skipping. The average lengths are 64 and 48, for junction and exon, respectively. The maximum of length are 140 and 6057, for junction and exon, respectively. The peptide types are exon region (EXON_KB), annotated exon-exon junctions (E_E_KB), hypothetical exon-exon junctions (E_E_TH), hypothetical exon-intron junctions (E_I_TH), and hypothetical intron-exon junctions (I_E_TH) Alternative splicing peptides searching In order to identify the tumor-specific alternative splicing isoform patterns, we ran OMSSA search engine with the peptide database against 40 normal plasma and 40 breast cancer plasma in Study II. Maximum 1 missed cleavage, Maximum 10 peptide hitlist length per spectrum, and Evalue cutoff 1.0 were chosen for filtering peptides. Statistics analysis With the statistics analysis in the method section, 90 alternative splicing isoforms in 38 genes were found, which showed statistically significant (q < 0.05) differences between normal breast and breast cancer samples in Study II (Additional file 2: Table S2; Fig. 1). Four out of five kinds of alternative splicing isoforms: exon splicing, single Exon, intron retention (left intron), and intron retention (right intron) were identified (Fig. 2) except for the normal exon for which we fail to reject the null hypothesis that there is no difference between normal and cancer samples since the p-value is not less than the significance level. Among the 90 alternative splicing isoforms, 57 are exon splicing, 23 single Exon and 10 intron retention. Those exon splicing and intron retention markers are more likely to be present in cancer samples than in normal samples and those single exon markers are more likely to be present in normal samples (χ2 = 53, df = 1, pvalue = 3.2e-13, Table 2). Another interesting finding is that Alternative Splicing isoform markers could be more likely to be found for genes with two or more than two transcript variants encoding different isoforms than genes with only one transcript (Chisquare Pvalue =1.35e-11 between genome and Study II’s markers, Chisquare Pvalue =0 between genome and Study III’s markers, Fig. 3). The human genome contains totally 30370 genes with only one transcript and 19136 genes with two or more than two transcript variants. Isoform Markers in Study II contains totally 3 genes with only one transcript and 35 genes with two or more than two transcript variants. Isoform Markers in Study III contains totally 2 genes with only one transcript and 53 genes with two or more than two transcript variants.Fig. 1 Heatmap of 90 alternative splicing isoform markers differentiating the normal and cancer samples of Study II. X axis is 90 alternative splicing isoform markers. Y-axis shows the cancer and normal samples ordered by unsupervised clustering. The top are cancer samples and bottom normal samples (H, health, green; C, cancer, blue). Red squares stand for presence, and white ones for absence Fig. 2 Five splicing types. Red, blue and green boxes are exon. Pink boxes are retained intron. Black lines are intron Table 2 number of alternative splicing and normal markers between the normal and cancer samples Health cancer Total Alternative Splicing 7 60 67 Normal 22 1 23 total 29 61 Fig. 3 Densities for genes with single transcript and multiple transcripts across whole genome, Study II’s markers and Study III’s markers. It shows that alternative splicing isoform markers could be more likely to be found for genes with two or more than two transcript variants encoding different isoforms than genes with only one transcript (Chisquare Pvalue =1.35e-11 between genome and Study II’s markers, Chisquare Pvalue =0 between genome and Study III’s markers) Four validation methods We presented four validation methods to validate our results. First we used the explainable index defined in method section to perform biological validation for the 38 gene markers. 36 out of 38 genes are “more explainable” except for two genes:JAK1 and KTN1 with explainable index of 1. The mean explainable index is 3.526316, the median explainable index 2, and maximum 12. We then performed the validation using the Human Exon 1.0 ST Array we downloaded from GSE19154 in Gene Expression Omnibus. The experiments include six mRNA samples which were extracted from human breast cancer cell line MCF7, and MCF10A, a nontumorigenic human breast epithelial cell line. Array analysis was performed using R and BioConductor libraries. Probeset in the exon array to the peptide sequence in our database was performed using the exon’s starting and ending positions in each transcript. Because of the limitation of the exon array, we can only validate the 23 single exon markers and test if those markers are more likely to be expressed in the same group as in our proteomics result. The validation results show that 21 of 23 single exon markers were confirmed by the exon array (Additional file 2: Table S2). The two unconfirmed markers were identified with not very significant pvalue (LLPNQNLPLDITLQSPTGAGPFPPIR 0.166; AAMKPGWEDLVRR 0.0895). The mutations that alter a splice site or a nearby regulatory sequence may have subtle effects by shifting the ratio of the resulting proteins without entirely eliminating any form, as a result of alternate splicing. Next, we performed the statistical validation using the independent 40 healthy and 40 cancer samples from Study III as testing set (Fig. 4). 66 (82.5 %) out of 80 samples are correctly predicted. For the 40 cancer samples, the prediction accuracy is 37/40 = 92.5 %.Fig. 4 Heatmap of 26 alternative splicing isoform markers in Study II differentiating the normal and cancer samples of Study III. X axis is 26 alternative splicing isoform markers from Study II. Y-axis shows the cancer and normal samples in Study III ordered by unsupervised clustering. The top are health samples and bottom cancer samples. The prediction results are green for health and blue for cancer. Red squares stand for presence, and white ones for absence Pathway analysis Last, we performed extensive pathway analysis to discover highly significant pathways from a set of cancer vs healthy samples. The knowledge of activation of these processes may lead to novel assays identifying their proteomic signatures in plasma of patient at high risk for cancer disease. In Study II, of the 24 significant pathways we observed, at least 23 of these pathways were involved cancers, signal transduction, diseases, and cellular processes (Additional file 3: Table S3). The top pathways include Pathways in cancer (8), MAPK signaling pathway (3), Cell cycle (3), Apoptosis (3), Focal adhesion (3), Adherens junction (3), Jak-STAT signaling pathway (3), Prostate cancer (3). All are also significant pathways in Study III except for Adherens junction (Additional file 4: Table S4). And ‘pathways in cancer’ are listed top 1 in both Study II and Study III. Discussion In this study we developed a peptidomics approach to identifying novel protein isoforms for clinical proteomics application. First, we built a Peptidomic Database of Human Protein Isoforms, then created a peptidomics approach to perform large-scale screen of breast cancer-associated alternative splicing isoform markers in clinical proteomics, and last performed four kinds of validations: biological validation (explainable index), exon array, statistical validation of independent samples, and extensive pathway analysis. Our results showed that alternative splicing isoform makers can act as independent markers of breast cancer and that the method we presented is an effective one for increasing the number of identified alternative splicing isoform markers in clinical proteomics. The combination of protein isoform database, statistical analysis, and statistical and biological validations has the potential for extremely high-resolution signatures to better resolve tumor subtypes and determining optimal therapies. Protein isoform database With the advances in mass spectrometry (MS) and large-scale generation of MS/MS (tandem MS)-based proteomics data, it has become clear that MS-based peptide sequence data can be mined to identify and validate isoforms in the protein level rather than in the transcript level where traditional methods such as EST sequencing [4], exon array [6], Exon-exon junction array [7]), and mRNA next-generation sequencing [8, 9] do. Moreover, it can eliminate limitations with these traditional methods for protein isoform determination such as no connectivity about connection of nonadjacent exons, no SNPs and PTMs, and low reproducibility. However, there are some limitations in identifying protein isoforms using current MS proteomics search database. For example, traditional mass spectrometry search database using isoforms of well-known proteins is biased. Using ESTs and a sequence database compression strategy to identify peptide isoforms existing in the EST database from MS data [18] is also defective because of the inherent characteristics of ESTs, such as transcript redundancy, low sequence quality and high error rates. 282 novel open reading frames were identified by searching six-frame translation of the human genome against MS spectrum [19]. But it only takes into account a small portion of alternative splicing isoform. Although there are several general-purpose alternative splicing mRNA transcript databases including ASTD [20], EID [21, 22], ASPicDB [23], and ECgene [24], they cannot be used for searching uncharacterized protein isoforms. And also the coverage of splicing junctions in all the databases are small. The new PEPPI database [25] contains the five types of combinations of exon and intron: EXON_KB, E_E_KB, E_E_TH, E_I_TH and I_E_TH, and makes it easy for different types of biomedical users to search for and identify alternative splicing isoform from proteomics experiment. We believe that it will be useful in the ongoing analysis of proteomics data, particularly those with clinical application potentials. The current PEPPI database contains only alternative splicing isoform. We will add SNP protein isoform and PTM protein isoform in the future so that the database of virtual peptides will be expanded to accommodate the amino acid alterations introduced by each SNP and PTM. Biological significance of isoform-specific peptides In this study, we have shown that isoform-specific peptides can distinguish normal breast from breast cancer. The number and type of splicing peptides identified exceeds the average number of events that is normally identified by splicing microarray profiling [26]. The accuracy and applicability of the newly identified alternative splicing signature was shown by its capacity to identify breast cancer sample (Fig. 4). The signature identified 92.5 % cancer samples and 72.5 % of normal samples in an independent set of 40 normal samples and 40 breast cancer samples. All cancer samples that were identified as normal could be either of the complexity of the proteome in plasma samples where the low abundance expected for specific markers of cancer are hindered, or of false positive associations that occur with analysis of high dimensional database. We observed that there appeared to be a higher proportion of alternative splicing markers in cancer samples (58 out of 65 alternative splicing are predominant in the cancer samples) and a higher proportion of normal markers in normal samples (22 out of 23 normal splicing are predominant in normal samples). Those exon splicing and intron retention markers are more likely to be present in cancer samples than in normal samples and those single exon markers are more likely to be present in normal samples. The strong correlation of alternative splicing isoform with cancer suggests the potential value of alternative splicing as prospective markers for the early detection and treatment of cancer. Interestingly, we also found that alternative splicing isoform markers could be more likely to be found for genes with two or more than two transcript variants encoding different isoforms than genes with only one transcript (Fig. 3). Previously, many alternative splicing variants had been observed in cancer, for examples, EGFR, CD44, NER and BRCA1. In our 38 gene markers, alternative splicing events of at least 5 genes were previously reported to occur in cancer. Two single exon markers and nine alternative splicing markers for ATM were identified in our results. This gene and the closely related kinase ATR are master controllers of cell cycle checkpoint signaling pathways that are required for cell response to DNA damage and for genome stability. Three alterations, del exon 4, deletion exon 29–34 and insertion of 137 bp in exon 46/47 were commonly observed in 8 HL cell lines and 7 clinical cases [27]. Katzenberger etc. presented the evidence that the ATM/CHK2 and ATR/CHK1 signaling pathways control gene expression by regulating alternative splicing [28]. Ho et. al. used ATM sequence alterations located within exons or in short intron regions flanking each exon that encompass putative splice site regions as predictor for late normal tissue responses in breast cancer patients treated with radiotherapy [29]. ATM allelic variants were reported to be associated to hereditary breast cancer in 94 Chilean women [30]. ATM SNPs have been associated with increased risk of breast, prostate, leukaemia, colon and lung cancer. Nguyen etc used two exons of ATM, both containing an SNP interfering with standard mutation scanning to screen 1356 subjects from an international breast cancer genetics Study IInd improved identification of rare known and unknown variants, while dramatically reducing the sequencing effort [31]. Three splicing markers (E3_E10 exon splicing, i6 intron retention and E5 single exon) for MET were identified in our results. The first two were predominant in cancer samples and the last one was predominant in health samples. Lee etc. had detected a novel type of structural variant of the tyrosine kinase receptor for MET, also known as the hepatocyte growth factor receptor, in mouse tissues, and demonstrated that a tyrosine kinase receptor could achieve additional diversity by alternative splicing at a key regulatory site in its cytoplasmic domain [32]. The cDNA of the variant transcript of MET lacks 141 base pairs and causes an in-frame deletion of 47 amino acids in the juxtamembrane region of the cytoplasmic domain. Extensive evidences indicate that MET signaling is involved in the progression and spread of several cancers such as breast, liver, lung, ovary, kidney, and thyroid [33]. And understanding of its role in disease has led to the development of Met as a major target in cancer drug and the development of a variety of MET pathway antagonists with potential clinical applications. Various mutations in the MET gene were reported to be associated with cancers. Zohar Tiran etc. identified a novel splice variant of the Met receptor, which encodes a truncated soluble form of the receptor [34]. This variant was produced as a recombinant Fc-fused protein named Cgen-241A and significantly inhibited HGF/SF-induced MET phosphorylation as well as cell proliferation, survival, and a profound inhibitory effect on cell scattering, invasion, and urokinase up-regulation. CREBBP is ubiquitously expressed and is involved in the transcriptional coactivation of many different transcription factors. First isolated as a nuclear protein that binds to cAMP-response element binding protein (CREB), this gene is now known to play critical roles in embryonic development, growth control, and homeostasis by coupling chromatin remodeling to transcription factor recognition. Its alternative splicing results in multiple transcript variants encoding different isoforms. It was reported that Co-regulator expression of CREBBP/p300 had been associated with lower tumor grade [35]. We identified intron 14 of JAK1 was retained through translation, which might be related to the mutation of 14 exons of JAK1 [36]. Xie found 12 cases (14 %) found to have single nucleotide polymorphism in exon 14. Somatic mutations in the SMO gene have also been identified in breast cancer. Recently, two groups have shown that hedgehog signaling may be active in a subset of human breast cancer cell lines, and that SMO antagonists can inhibit breast cancer growth [37, 38]. Conclusions We developed a peptidomics method to discover novel alternative splicing biomarkers from breast cancer proteome. First, we built a Peptidomic Database of Human Protein Isoforms, then created a peptidomics approach to perform large-scale screen of breast cancer-associated alternative splicing isoform markers in clinical proteomics, and last performed four kinds of validations: biological validation (explainable index), exon array, statistical validation of independent samples, and extensive pathway analysis. Our results showed that alternative splicing isoform makers can act as independent markers of breast cancer and that the method we presented is an effective one for increasing the number of identified alternative splicing isoform markers in clinical proteomics. Additional files Additional file 1: Table S1. Breast cancer plasma sources. (PDF 172 kb) Additional file 2: Table S2. 90 alternative splicing isoforms with statistically significant (q < 0.05) differences between normal breast and breast cancer samples in Study II (PDF 184 kb) Additional file 3: Table S3. Pathway analysis for Study II (XLS 34 kb) Additional file 4: Table S4. Pathway analysis for Study III (XLS 34 kb) Acknowledgements This work was supported in part by a grant from the National Cancer Institute (U24CA126480-01), part of NCI’s Clinical Proteomic Technologies Initiative (http://proteomics.cancer.gov), awarded to Dr. Fred Regnier (PI) and Dr. Jake Chen (co-PI). We thank Hoosier Oncology Group for collecting breast cancer plasma samples and Dr. Mu Wang for providing LC/MS/MS proteomics experimental data for this analysis. We also thank Indiana Center for Systems Biology and Personalized Medicine for its support. Declarations The publication charges for this article have been funded by the grant from the National Cancer Institute (U24CA126480-01. This article has been published as part of BMC Genomics Volume X Supplement X, 2016: XXXXX. The full contents of the supplement are available online at http://XXXXX. Authors’ contributions JYC conceived the initial work and designed the method. FZ developed the distance method and performed the statistical analyses. All authors are involved in the drafting and revisions of the manuscript. Both authors read and approved the final manuscript. Competing interests The authors declare that they have no competing financial interests. ==== Refs References 1. Chan D Clinical proteomics Clin Proteomic 2006 2 1 1 4 10.1385/CP:2:1:1 2. Hanash S Moving forward with clinical proteomics Clin Proteomic 2004 1 1 3 5 10.1385/CP:1:1:003 3. Mischak H Apweiler R Banks RE Conaway M Coon J Dominiczak A Ehrich JHH Fliser D Girolami M Hermjakob H Clinical proteomics: A need to define the field and to begin to set adequate standards Proteomic Clin Appl 2007 1 2 148 156 10.1002/prca.200600771 4. Adams MD Kelley JM Gocayne JD Dubnick M Polymeropoulos MH Xiao H Merril CR Wu A Olde B Moreno RF Complementary DNA sequencing: expressed sequence tags and human genome project Science 1991 252 5013 1651 1656 10.1126/science.2047873 2047873 5. Lee C Roy M Analysis of alternative splicing with microarrays: successes and challenges Genome Biol 2004 5 7 231 10.1186/gb-2004-5-7-231 15239822 6. Kapur K Xing Y Ouyang Z Wong WH Exon arrays provide accurate assessments of gene expression Genome Biol 2007 8 5 R82 10.1186/gb-2007-8-5-r82 17504534 7. Johnson JM Castle J Garrett-Engele P Kan Z Loerch PM Armour CD Santos R Schadt EE Stoughton R Shoemaker DD Genome-wide survey of human alternative pre-mRNA splicing with exon junction microarrays Science 2003 302 5653 2141 2144 10.1126/science.1090100 14684825 8. Pan Q Shai O Lee LJ Frey BJ Blencowe BJ Deep surveying of alternative splicing complexity in the human transcriptome by high-throughput sequencing Nat Genet 2008 40 12 1413 1415 10.1038/ng.259 18978789 9. Pando MP Kotraiah V McGowan K Bracco L Einstein R Alternative isoform discrimination by the next generation of expression profiling microarrays Expert Opin Ther Targets 2006 10 4 613 625 10.1517/14728222.10.4.613 16848696 10. Klampfl CW Review coupling of capillary electrochromatography to mass spectrometry J Chromatogr A 2004 1044 1-2 131 144 10.1016/j.chroma.2004.04.072 15354433 11. Frohlich T Arnold GJ Proteome research based on modern liquid chromatography--tandem mass spectrometry: separation, identification and quantification J Neural Transm 2006 113 8 973 994 10.1007/s00702-006-0509-3 16835695 12. Mbeunkui F Metge BJ Shevde LA Pannell LK Identification of differentially secreted biomarkers using LC-MS/MS in isogenic cell lines representing a progression of breast cancer J Proteome Res 2007 6 8 2993 3002 10.1021/pr060629m 17608509 13. Hubbard TJ Aken BL Ayling S Ballester B Beal K Bragin E Brent S Chen Y Clapham P Clarke L Ensembl 2009 Nucleic Acids Res 2009 37 Database issue D690 697 10.1093/nar/gkn828 19033362 14. Wang M You J Bemis KG Tegeler TJ Brown DP Label-free mass spectrometry-based protein quantification technologies in proteomic analysis Brief Funct Genomic Proteomic 2008 7 5 329 339 10.1093/bfgp/eln031 18579615 15. Higgs RE Knierman MD Gelfanova V Butler JP Hale JE Comprehensive label-free method for the relative quantification of proteins from biological samples J Proteome Res 2005 4 4 1442 1450 10.1021/pr050109b 16083298 16. Storey JD Tibshirani R Statistical significance for genomewide studies Proc Natl Acad Sci U S A 2003 100 16 9440 9445 10.1073/pnas.1530509100 12883005 17. Kanehisa M Araki M Goto S Hattori M Hirakawa M Itoh M Katayama T Kawashima S Okuda S Tokimatsu T KEGG for linking genomes to life and the environment Nucleic Acids Res 2008 36 Database issue D480 484 18077471 18. Edwards NJ Novel peptide identification from tandem mass spectra using ESTs and sequence database compression Mol Syst Biol 2007 3 102 17437027 19. Fermin D Allen BB Blackwell TW Menon R Adamski M Xu Y Ulintz P Omenn GS States DJ Novel gene and gene model detection using a whole genome open reading frame analysis in proteomics Genome Biol 2006 7 4 R35 10.1186/gb-2006-7-4-r35 16646984 20. Koscielny G, Le Texier V, Gopalakrishnan C, Kumanduri V, Riethoven JJ, Nardone F, Stanley E, Fallsehr C, Hofmann O, Kull M, et al. ASTD: The Alternative Splicing and Transcript Diversity database. Genomics. 2009;93(3):213–20. 21. Shepelev V Fedorov A Advances in the Exon-Intron Database (EID) Brief Bioinform 2006 7 2 178 185 10.1093/bib/bbl003 16772261 22. Saxonov S Daizadeh I Fedorov A Gilbert W EID: the Exon-Intron Database-an exhaustive database of protein-coding intron-containing genes Nucleic Acids Res 2000 28 1 185 190 10.1093/nar/28.1.185 10592221 23. Castrignano T D’Antonio M Anselmo A Carrabino D D’Onorio De Meo A D’Erchia AM Licciulli F Mangiulli M Mignone F Pavesi G ASPicDB: a database resource for alternative splicing analysis Bioinformatics 2008 24 10 1300 1304 10.1093/bioinformatics/btn113 18388144 24. Lee Y Kim B Shin Y Nam S Kim P Kim N Chung WH Kim J Lee S ECgene: an alternative splicing database update Nucleic Acids Res 2007 35 Database issue D99 103 10.1093/nar/gkl992 17132829 25. Zhou A Zhang F Chen JY PEPPI: a peptidomic database of human protein isoforms for proteomics experiments BMC Bioinformatics 2010 11 Suppl 6 S7 10.1186/1471-2105-11-S6-S7 20946618 26. Gardina PJ Clark TA Shimada B Staples MK Yang Q Veitch J Schweitzer A Awad T Sugnet C Dee S Alternative splicing and differential gene expression in colon cancer detected by a whole genome exon array BMC Genomics 2006 7 325 10.1186/1471-2164-7-325 17192196 27. Liu A Takakuwa T Fujita S Luo WJ Tresnasari K Van den Berg A Poppema S Aozasa K ATR alterations in Hodgkin’s lymphoma Oncol Rep 2008 19 4 999 1005 18357388 28. Katzenberger RJ Marengo MS Wassarman DA ATM and ATR pathways signal alternative splicing of Drosophila TAF1 pre-mRNA in response to DNA damage Mol Cell Biol 2006 26 24 9256 9267 10.1128/MCB.01125-06 17030624 29. Ho AY Fan G Atencio DP Green S Formenti SC Haffty BG Iyengar P Bernstein JL Stock RG Cesaretti JA Possession of ATM sequence variants as predictor for late normal tissue responses in breast cancer patients treated with radiotherapy Int J Radiat Oncol Biol Phys 2007 69 3 677 684 10.1016/j.ijrobp.2007.04.012 17517479 30. Tapia T Sanchez A Vallejos M Alvarez C Moraga M Smalley S Camus M Alvarez M Carvallo P ATM allelic variants associated to hereditary breast cancer in 94 Chilean women: susceptibility or ethnic influences? Breast Cancer Res Treat 2008 107 2 281 288 10.1007/s10549-007-9544-5 17351744 31. Nguyen-Dumont T Calvez-Kelm FL Forey N McKay-Chopin S Garritano S Gioia-Patricola L De Silva D Weigel R Sangrajrang S Lesueur F Description and validation of high-throughput simultaneous genotyping and mutation scanning by high-resolution melting curve analysis Hum Mutat 2009 30 6 884 890 10.1002/humu.20949 19347964 32. Lee CC Yamada KM Identification of a novel type of alternative splicing of a tyrosine kinase receptor. Juxtamembrane deletion of the c-met protein kinase C serine phosphorylation regulatory site J Biol Chem 1994 269 30 19457 19461 7518457 33. Eder JP Vande Woude GF Boerner SA LoRusso PM Novel therapeutic inhibitors of the c-Met signaling pathway in cancer Clin Cancer Res 2009 15 7 2207 2214 10.1158/1078-0432.CCR-08-1306 19318488 34. Tiran Z Oren A Hermesh C Rotman G Levine Z Amitai H Handelsman T Beiman M Chen A Landesman-Milo D A novel recombinant soluble splice variant of Met is a potent antagonist of the hepatocyte growth factor/scatter factor-Met pathway Clin Cancer Res 2008 14 14 4612 4621 10.1158/1078-0432.CCR-08-0108 18628476 35. Green AR Burney C Granger CJ Paish EC El-Sheikh S Rakha EA Powe DG Macmillan RD Ellis IO Stylianou E The prognostic significance of steroid receptor co-regulators in breast cancer: co-repressor NCOR2/SMRT is an independent indicator of poor outcome Breast Cancer Res Treat 2008 110 3 427 437 10.1007/s10549-007-9737-y 17902051 36. Xie HJ Bae HJ Noh JH Eun JW Kim JK Jung KH Ryu JC Ahn YM Kim SY Lee SH Mutational analysis of JAK1 gene in human hepatocellular carcinoma Neoplasma 2009 56 2 136 140 10.4149/neo_2009_02_136 19239328 37. Kubo M Nakamura M Tasaki A Yamanaka N Nakashima H Nomura M Kuroki S Katano M Hedgehog signaling pathway is a new therapeutic target for patients with breast cancer Cancer Res 2004 64 17 6071 6074 10.1158/0008-5472.CAN-04-0416 15342389 38. Mukherjee S Frolova N Sadlonova A Novak Z Steg A Page GP Welch DR Lobo-Ruppert SM Ruppert JM Johnson MR Hedgehog signaling and response to cyclopamine differ in epithelial and stromal cells in benign breast and breast cancer Cancer Biol Ther 2006 5 6 674 683 10.4161/cbt.5.6.2906 16855373

PMC005xxxxxx/PMC5001248.txt

==== Front BMC GenomicsBMC GenomicsBMC Genomics1471-2164BioMed Central London 27556803279310.1186/s12864-016-2793-0ResearchA new algorithm for “the LCS problem” with application in compressing genome resequencing data Beal Richard r.beal@computer.org Afrin Tazin Farheen Aliya Adjeroh Donald Lane Department of Computer Science and Electrical Engineering, West Virginia University, Morgantown, WV USA 18 8 2016 18 8 2016 2016 17 Suppl 4 Publication of this supplement has not been supported by sponsorship. Information about the source of funding for publication charges can be found in the individual articles. The articles have undergone the journal's standard peer review process for supplements. The Supplement Editor declares that they have no competing interests.544© The Author(s) 2016 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Background The longest common subsequence (LCS) problem is a classical problem in computer science, and forms the basis of the current best-performing reference-based compression schemes for genome resequencing data. Methods First, we present a new algorithm for the LCS problem. Using the generalized suffix tree, we identify the common substrings shared between the two input sequences. Using the maximal common substrings, we construct a directed acyclic graph (DAG), based on which we determine the LCS as the longest path in the DAG. Then, we introduce an LCS-motivated reference-based compression scheme using the components of the LCS, rather than the LCS itself. Results Our basic scheme compressed the Homo sapiens genome (with an original size of 3,080,436,051 bytes) to 15,460,478 bytes. An improvement on the basic method further reduced this to 8,556,708 bytes, or an overall compression ratio of 360. This can be compared to the previous state-of-the-art compression ratios of 157 (Wang and Zhang, 2011) and 171 (Pinho, Pratas, and Garcia, 2011). Conclusion We propose a new algorithm to address the longest common subsequence problem. Motivated by our LCS algorithm, we introduce a new reference-based compression scheme for genome resequencing data. Comparative results against state-of-the-art reference-based compression algorithms demonstrate the performance of the proposed method. Keywords Longest common subsequenceLCSLongest previous factorLPFCompressionBiologyGenome resequencingIEEE International Conference on Bioinformatics and Biomedicine 2015 Washington, DC, USA 9-12 November 2015 http://cci.drexel.edu/ieeebibm/bibm2015/issue-copyright-statement© The Author(s) 2016 ==== Body Background Measuring similarity between sequences, be it DNA, RNA, or protein sequences, is at the core of various problems in molecular biology. An important approach to this problem is computing the longest common subsequence (LCS) between two strings S1 and S2, i.e. the longest ordered list of symbols common between S1 and S2. For example, when S1=abba and S2=abab, we have the following LCSs: abb and aba. The LCS has been used to study various areas (see [2, 3]), such as text analysis, pattern recognition, file comparison, efficient tree matching [4], etc. Biological applications of the LCS and similarity measurement are varied, from sequence alignment [5] in comparative genomics [6], to phylogenetic construction and analysis, to rapid search in huge biological sequences [7], to compression and efficient storage of the rapidly expanding genomic data sets [8, 9], to re-sequencing a set of strings given a target string [10], an important step in efficient genome assembly. The basic approach to compute the LCS, between the n-length S1 and m-length S2, is via dynamic programming. Using LCS to denote the dynamic programming (DP) table, the basic formulation is as follows, given 0≤i≤n and 0≤j≤m: LCS(i,j)=0,ifi=0∨j=01+LCS(i−1,j−1),ifS1[i]=S2[j]maxLCS(i,j−1),LCS(i−1,j),ifS1[i]≠S2[j] The above computes the length of the LCS in the last position of the table (LCS(n,m)). As with the edit distance computation, the actual string forming the LCS can be obtained by using a trace back on the DP table. This requires O(nm) time and O(nm) space. The LCS matrix has some interesting properties: the entries in any row or in any column are monotonically increasing, and between any two consecutive entries in any row or column, the difference is either 0 or 1. An example LCS matrix and trace are shown in Fig. 1. Fig. 1 LCS dynamic programming table for S 1=A A C C T T A A and S 2=A G G T C G T A. A sample LCS trace (ACTA) is highlighted Alternatively, we can formulate the problem as a two-dimensional grid, where the goal is to find the minimal cost (or maximal cost, depending on the formulation) path, from the start position on the grid (typically, (0,0)), to the end position (n,m). Myers et al. [11] and Ukkonen [12] used this idea to propose a minimum cost path determination problem on the grid, where the path takes a diagonal line from (i−1,j−1) to (i,j) if S1[ i]=S2[ j] with cost 0, and takes a horizontal or vertical line with a cost of 1, corresponding respectively to insert or delete operations. Hunt and Szymanski [13] earlier used an essentially similar approach to solve the LCS problem in (r+n) logn time, with n≪m, where r is the number of pairwise symbol matches (S1[ i]=S2[ j]). When two non-similar files are compared, we will have r≪nm, or r in O(n), leading to a practical O(n logn) time algorithm. However, for very similar files, we have r≈nm, or an O(nm logn) algorithm. This worst-case occurs, for instance, when S1=an and S2=am. Hirschberg [14] proposed space-efficient approaches to compute the LCS using DP in O(nm) time and O(n+m) space, rather than O(nm). More recently, Yang et al. [15] used the observation on monotonically increasing values in the LCS table to identify the “corner points”, where the values on the diagonals change from one row to the next. The corners define a more sparse 2D grid, based on which they determine the LCS. A generalization of the LCS problem is to find the LCS for a set of two or more sequences. This is the multiple longest common subsequence problem, which is known to be NP-hard for an arbitrary number of sequences [16]. Another interesting view of the LCS problem is in terms of the longest increasing subsequence (LIS) problem, suggested earlier in [17–19], and described in detail in [2]. The LIS approach also solves the LCS problem in O(r logn) time (where m≤n). In most practical scenarios, r<nm. The LCS has been used in some recent algorithms to compress genome resequencing data [20, 21]. Compression of biological sequences is an important and difficult problem, which has been studied for decades by various authors [22–24]. See [9, 25, 26] for recent surveys. Most of the earlier studies focused on lossless compression because it was believed that biological sequences should not admit any data loss, since that would impact later use of the compressed data. The earlier methods also generally exploited self-contained redundancies, without using a reference sequence. The advent of high-throughput next generation sequencing, with massive datasets that are easily generated for one experiment, have challenged both compression paradigms. Lossy compression of high-throughput sequences admitting limited errors have been proposed in [27, 28] for significant compression. With the compilation of several reference genomes for different species, more recent methods have considered lossless compression of re-sequencing data by exploiting the significant redundancies between the genomes of related species. This observation is the basis of various recently proposed methods for reference-based lossless compression [20, 21], whereby some available standard reference genome is used as the dictionary. Compression ratios in the order of 80 to 18,000 without loss have been reported [20, 21]. The LCS is the hallmark of these reference-based approaches. In this work, we first introduce a new algorithm for the LCS problem, using suffix trees and shortest-path graph algorithms. Motivated by our LCS algorithm, we introduce an improved reference-based compression scheme for resequencing data using the longest previous factor (LPF) data structure [29–31]. Methods Preliminaries A string T is a sequence of symbols from some alphabet Σ. We append a terminal symbol $∉Σ to strings for completeness. A string or data structure D has length- |D|, and its ith element is indexed by D[i], where 1≤i≤|D|. A prefix of a string T is T[1…i] and a suffix is T[i…|T|], where 1≤i≤|T|. The suffix tree (ST) on the n-length T is a compact trie (with O(n) nodes constructed in O(n) time [3]) that represents all of the suffixes of T. Suffixes with common prefixes share nodes in the tree until the suffixes differentiate and ultimately, each suffix T[ i…n] will have its own leaf node to denote i. A generalized suffix tree (GST) is an ST for a set of strings. A substring of T is T[ i…j], where 1≤i≤j≤n. The longest common subsequence is defined below in terms of length-1 common substrings. Definition 1. Longest common subsequence (LCS): For the n-length S1 and m-length S2, the LCS between S1 and S2 is the length of the longest sequence of pairs ℳ={m1,…,mM}, where mi=(u,v) such that S1[mh.u]=S2[ mh.v] for 1≤h≤M and mi.u<mi+1.u ∧ mi.v<mi+1.v for 1≤i<M. LCS algorithm Below, we compute the LCS between S1 and S2 in the following way. (i) We use the GST to compute the common substrings (CSSs) shared between S1 and S2. (ii) We use the CSSs to construct a directed acyclic graph (DAG) of maximal CSSs. (iii) We compute LCS by finding the longest path in the DAG. Steps (i) and (iii) are standard tasks. For step (ii), we develop new algorithms and data structures. Computing the CSSs We now briefly describe finding the common substrings (CSSs) between S1 and S2. In our LCS algorithm, for simplicity of discussion, we will only use CSSs of length-1. Let A=∅. Compute the GST on S1$1∘S2$2, for terminals {$1,$2}. Consider a preorder traversal of the GST. When at depth-1 for a node N, let S=∅. During the preorder traversal from N, we collect in S all of the suffix index leaves descending from N, which represent the suffixes that share the same first symbol. Let S1=S2=∅. For s∈S, if s≤|S1|, then store s in S1. Otherwise, store s in S2. We represent all of our length-1 matches in the following structure: MATCH {id, p1, p2}. The id is a unique number for the MATCH, and p1 and p2 are respectively the positions in S1 and S2 where the CSS exists. Let id=2. Now, for each s1∈S1, we create a new MATCH m=(id++,s1,s2) for each s2∈S2. Store each m in A. The running time is clearly the maximum of the GST construction and the number of length-1 CSSs. Lemma 2. Say n= |S1| and m= |S2|, then computing the η CSSs of length-1 between S1 and S2 requires O(max{n+m,η}) time. DAG construction Given all of the MATCHes found in A, our task now is to construct the DAG for A. For all paths of the DAG to start and end at a common node, we make MATCHes S and E to respectively precede and succeed the MATCHes in A. (Let S have id=1 and E have id=|A+2| and then store S and E in A.) The goal of the DAG is to represent all maximal CSSs between S1 and S2 as paths from S to E. We will later find the LCS, the longest such path. In the DAG, the nodes will be the MATCH ids and the edges between MATCHes, say m1 and m2, represent that S1[ m1.p1]=S2[ m1.p2] is chosen in the maximal common subsequence followed by S1[ m2.p1]=S2[ m2.p2]. The DAG is acyclic because, by Definition 1, the LCS is a list of ordered MATCHes. Since we cannot choose mi∈ℳ and then mh∈ℳ with h<i, then no cycle can exist. Our DAG construction, displayed in Algorithm 1, operates in the following way. We initialize the DAG dag by first declaring dag.gr of size |A|, since gr will represent all of the nodes. All outgoing edges for say the node N∈A are represented by dag.gr[N.id][1…dag.sz[N.id]]. By setting dag.sz={0,…,0}, we clear the edges in our dag. Now, setting these edges is the main task of our algorithm. We can easily construct the edges by assuming that there exists a data structure PREV pv that can tell us the set of parents for each node a∈A. That is, we can call getPrnts(pv,L) to get the set of nodes P that directly precede MATCH L∈A in the final dag. By “directly precede”, we mean that in the final dag, there is connection from each p∈P to a, i.e. each p is in series with a, meaning that both p AND a are chosen in a maximal CSS. Further, no p,p2∈P can be in series with one another, and rather, they are in parallel with one another, meaning that either p OR p2 is chosen in a maximal common subsequence. With P, we can build an edge from a2∈P to a by first allocating a new space in dag.gr[a2.id] by incrementing dag.sz[a2.id] and then making a directed edge from parent to child, i.e. dag.gr[a2.id][dag.sz[a2.id]]=a.id. After computing the incoming edges for each node a∈A, the dag construction is complete. PREV data structure The simplicity of the DAG construction is due to the PREV pv, detailed here. The pv is composed of four attributes. HashMap <int,int >p1. Suppose that all a.p1 values (for a∈A) are placed on an integer number line. It is very unlikely that all a.p1 values will be consecutive and so, there will be unused numbers (gaps) between adjacent values. Since we later declare matrices on the MATCH p1 (and p2) values, these gaps will be wasteful. With a scan of the a.p1 values (say using a Set), we can rename them consecutively without gaps; these renamed values are found by accessing HashMap <int,int >p1 with the original a.p1 value. HashMap <int,int >p2. This is the same as the aforementioned p1, but with respect to the a.p2 values. MATCHtbl1[][]. A fundamental data structure to support the getPrnts function is the tbl1, defined below. Definition 3. Max Table w.r.t.p1(tbl1): Given the set of all MATCH values A and PREV pv on A (with pv.p1 and pv.p2), the tbl1[|pv.p1|][|pv.p2|] is defined such that each tbl1[i][j] is the a∈A with the maximumpv.p1.get(a.p1)≤i, where pv.p2.get(a.p2)≤j. In the case that multiple such a exist, tbl1[ i][j] is the a with the rightmostpv.p2.get(a.p2)≤j. If no such a exists, tbl1[ i][j]=null. In other words, the tbl1[ i][j] stores the “closest” MATCH a with respect to the p1 values (i.e. we maximize a.p1 before a.p2). To construct tbl1, we first declare the table, tbl1[ |pv.p1|][ |pv.p2|] and initialize all elements tbl1[ i][j]=null, signifying that no MATCHes are found. Next, we insert each a∈A into the list by setting tbl1[ pv.p1.get(a.p1)][pv.p2.get(a.p2)]=a. Now, each tbl1[ i][j]=null needs to be set as the rightmost MATCH m with the maximum m.p1 in the subtable tbl1[ 1…i][ 1…j]. This is easily computed by first moving vertically in tbl1 and setting tbl1[ i][j]=tbl1[ i−1][j] if tbl1[ i][j]=null to propagate the maximum values vertically. Finally, we need to move horizontally in tbl1 and store in tbl1[ i][j] the rightmost tbl1[ i][ v](1≤v≤j) with the maximum tbl1[ i][ v].p1. This is done by a left-to-right scan of each row, comparing the adjacent elements, and setting tbl1[ i][ v]=tbl1[ i][ v−1] if tbl1[ i][ v−1].p1>tbl1[ i][ v].p1. MATCHtbl2[][]. The tbl2 is the same as tbl1 except that we define “closest” to mean that the a.p2 value is maximized before the a.p1. Definition 4. Max Table w.r.t.p2(tbl2): Given the set of all MATCH values A and PREV pv on A (with pv.p1 and pv.p2), the tbl2[ |pv.p1|][ |pv.p2|] is defined such that each tbl2[ i][j] is the a∈A with the maximumpv.p2.get(a.p2)≤j, where pv.p1.get(a.p1)≤i. In the case that multiple such a exist, tbl2[ i][j] is the a with the rightmostpv.p1.get(a.p1)≤i. If no such a exists, tbl2[ i][j]=null. The construction of tbl2 is the same as tbl1, except that in the final horizontal scan, we compare tbl2[ i][v].p2 and tbl2[ i][ v−1].p2. In terms of construction time, if we assume that adding and accessing HashMap entries are constant time operations, and the Set is implemented with a HashMap, then the PREV pv on A from the n-length S1 and m-length S2 is constructed in O(|pv.p1|×|pv.p2|) time. While pv.p1 and pv.p2 eliminate the gaps between the respective p1 and p2 values of A, we have |pv.p1|∈O(n) and |pv.p2|∈O(m) in the very worst-case. Theorem 5. Given the n-length S1 and m-length S2, and the set of all MATCHes A, PREV pv on A is constructed in O(nm) time. getPrnts function Given the PREV pv data structure on all MATCHes A, we call getPrnts(pv,L) in line 11 of constructDAG to retrieve the set of parent MATCHes P of the MATCH L∈A. Recall that these parents P of the MATCH L are all MATCHes that directly precede L in the DAG, i.e. each p∈P is in series with L and no p,p2∈P are in series with one another. Using pv, we can compute, for any MATCH c∈A, two direct parents that are closest to c with respect to the p1 and p2 values. Definition 6. Direct Parents: Given the PREV pv on the MATCHes in A between the n-length S1 and the m-length S2, and a MATCH c∈A, let i=pv.p1.get(c.p1) and j=pv.p2.get(c.p2). The direct parent of c w.r.t. p1 is: d1=null,ifi≤1∨j≤1∨i>|pv.p1|∨j>|pv.p2|pv.tbl1[i−1][j−1],otherwise The direct parent of c w.r.t. p2 is: d2=null,ifi≤1∨j≤1∨i>|pv.p1|∨j>|pv.p2|pv.tbl2[i−1][j−1],otherwise The first getDPrnt in Algorithm 2 implements Definition 6 to return the direct parents for any MATCH say L∈A. In cases where we want to find the direct parent for a MATCH at a certain location in the pv.tbl1 or pv.tbl2, say pv.tbl1[ i][j] or pv.tbl2[ i][j], we overload getDPrnt. The direct parents computation (getDPrnt) is the cornerstone of the getPrnts function. The following lemma, implemented in Algorithm 3, proves that the direct parents of c can be used to determine all parents of c. Lemma 7. Given A, the MATCHes between S1 and S2, and a MATCH c∈A, the two direct parents of c can be used to compute the set P with all parents of c. Proof. Let d1 and d2 be the direct parents of c (Definition 6). By Definition 3, d1 is a direct parent because it directly precedes c with the maximum p1 and the rightmost p2 value. Similarly by Definition 4, d2 is a direct parent of c because it directly precedes c with the maximum p2 and the rightmost p1 value. To find the remaining parents of c, we now find other MATCHes that precede c, which are also parallel with d1 and d2. There are three cases. Case (a). When d1=null, then also d2=null since there cannot be another MATCH preceding c. Thus, P=∅. Case (b). When d1=d2, the nearest parents to c are the same MATCH. There are only two types of MATCHes that are parallel with d1. First, we need to consider all MATCHes, say m1, with the same endpoint m1.p1=d1.p1 and m1.p2∈{1,2,…,d1.p2−1}. Second, we need to consider the MATCHes, say m2, with the same endpoint m2.p2=d1.p2 and m2.p2∈{1,2,…,d1.p1−1}. In the LCS computation, suppose that we chose, w.l.o.g., m1 (with m1.p2=d1.p2−2) instead of d1. Then, we cannot choose a MATCH m3 with m3.p1<d1.p1 and m3.p2=d1.p2−1. So, having any m1 or m2 parallel to d1 will only lead to suboptimal CSSs. Thus, only P={d1} is a parent of c. Case (c). Otherwise, d1≠d2 and we have two different direct parents of c. Set P={d1,d2}. Let us collect the endpoints of d1 and d2: i1=d2.p1,i2=d1.p1,j1=d1.p2, and j2=d2.p2. What MATCH, say m3, is parallel to d1 and d2? By Definition 6, there cannot be any MATCH m3 directly preceding c with endpoints after i2 or j2. By (b), we do not need to consider other MATCHes with endpoints on either d1 or d2. So, all the possible MATCHes parallel to d1 and d2 are those with (m3.p1∈w∧m3.p2∈x), where w={i1+1,i1+2,…,i2−1} and x={j1+1,j1+2,…,j2−1}. To find such m3, we only need to find direct parents (by (b)), say dd1 and dd2, for a theoretical MATCH m with (m.p1∈w∧m.p2=j)∨(m.p1=i∧m.p2∈x). Then, when we have i1<dd1.p1<i2 and j1<dd1.p2<j2, this is a possible MATCH parallel with d1 and d2, which is also a possible parent of c, so we add dd1 to P. We do the same process for dd2. Since we computed all the possible parents in P, additional processing on P is needed to ensure that no pair of MATCHes in P are in series; if any are in series, delete the MATCH furthest from c. With the pv and getDPrnt, this task is simple. We simply check the direct parents (say dd1 and dd2) for each y∈P, and remove dd1 if dd1∈P and remove dd2 if dd2∈P. □ Computing the LCS Since our dag has a single source S (and all paths end at E), the longest path between S and E, i.e. the LCS, is computed by giving all edges a weight of −1 and finding the shortest path from S to E via a topological sort [32]. Complexity analysis Our LCS algorithm: (i) finds the length-1 CSSs, (ii) computes the DAG on the CSSs, and (iii) reports the longest DAG path. Here, we analyze the overall time complexity. Step (i) First, we find (and store in A) the η length-1 CSSs in O(max{n+m,η}) time by Lemma 2. Step (ii) We then construct the DAG dag on these a∈A with constructDAG. In constructDAG, we initially compute the newly proposed PREV pv data structure in O(nm) time by Theorem 5. After constructing pv, the computeDAG iterates through each a∈A and creates an incoming edge between the parents of a and a. So, computeDAG executes in time O(max{nm,η×tgetPrnts}), where tgetPrnts is the time of getPrnts. The getPrnts running time is in O((i2−i1)+(j2−j1)), with respect to the local variables i1,i2,j1, and j2. However, it may be the case that i1=j1=1,i2=n, and j2=m, and so O(n+m) time is required by getPrnts. Below we formalize the worst-case result and the case for average strings from a uniform distribution. Lemma 8. For the n-length S1 and the m-length S2, the getPrnts function requires O(n+m) time. Lemma 9. For average case strings S1 and S2 with symbols uniformly drawn from alphabet Σ, the getPrnts function requires O(|Σ|) time. Proof. Since d1 and d2 are the direct parents of c (see Definitions 3, 4 and 6), and since the uniformness of S1 and S2 means that for any symbol say S1[s] we can find every σ∈Σ in S2[s−Δ…s+Δ] with Δ∈O(|Σ|), then (i2−i1)∈O(|Σ|) and (j2−j1)∈O(|Σ|). □ So, the overall constructDAG time follows. Theorem 10. Given A, the length-1 MATCHes in the n-length S1 and the m-length S2, the constructDAG requires O(max{nm,η× max{n,m}}) time in the worst-case and O(max{nm,η×|Σ|}) on average. Step (iii) We find the LCS with a topological sort in time linear to the dag size [32], which cannot require more time than that needed to build the dag (see Theorem 10). Summary Overall, (i) and (iii) do not add to the complexity of (ii). Given the above, the overall running time is as follows. Theorem 11. The LCS between the n-length S1 and the m-length S2 can be computed in O(max{nm,η× max{n,m}}) time in the worst-case and O(max{nm,η×|Σ|}) on average. Compressing resequencing data When data is released, modified, and re-released over a period of time, a large amount of commonality exists between these releases. Rather than maintaining all uncompressed versions of the data, it is possible to keep one uncompressed version, say D, and compress all future versions Di with respect to D. We refer to Di as the target and D as the reference. This idea is used to compress resequencing data in [20, 21], primarily using the LCS. The LCS, however, has two core problems with respect to compression. For very similar sequences, the LCS computation time is almost quadratic, or worse, potentially leading to long compression time. Secondly, the LCS may not always lead to the best compression, especially when some CSS components are very short. Rather than focusing on the LCS, we consider the maximal CSSs that make up the common subsequences. To intelligently choose which of the CSSs are likely to lead to improved compression, we use the longest previous factor (LPF), an important data structure in text compression [33]. Consider compressing the target T with respect to the reference R; let Z=R∘T. Suppose we choose exactly |T| maximal-length CSSs, specifically, for β=Z[i…|Z|] we have α=Z[ h…|Z|] such that (1) CSSs α[1…k]=β[ 1…k] and (2) this is the maximal k for h<i, where |R|+1≤i≤|Z|. These ks are computed in the LPF data structure on Z at LPF[ i]=k and the position of this CSS is at POS[ i]=h [29]. (Note that LPF and POS are constructed in linear time [29–31].) The requirement that h<i suits dictionary compression and compressing resequencing data because the CSS beginning at i is compressed by referencing the same CSS at h, occurring earlier in target T or anywhere in the reference R. Our idea is to use the LPF and POS to represent or encode CSSs that make up the target T with tuples. We will then compress these tuples with standard compression schemes. Our compression scheme We now propose a reference-based compression scheme which scans the LPF and POS on Z in a left-to-right fashion to compress T with respect to R. This scheme is similar to the LZ factorization [29], but differs in how we will encode the CSSs. Our contribution here is (1) using two files to compress T, (2) only encoding CSSs with length at least k, and (3) further compressing the resulting files with standard compression schemes. Initially, the two output files, triples and symbols, are empty. Let i=|R|+1. (!) If LPF[ i]<k, we simply encode the symbol; append the (say 1-byte) char T[ i−|R|] to symbols and increment i. Otherwise LPF[ i]≥k, so we will encode this CSS with the triple (pT,pZ,l), where pT=i−|R| is the starting position of the CSS in T, pZ=POS[i] is the starting position of the CSS in Z[ 1…i−1], and l=LPF[ i] is the length of the CSS. We write three long (say 4-byte integer) words pT, pZ, and l to triples. Since the triple encodes an l-length CSS, we set i=i+l to consider compressing the suffix following the currently encoded CSS. Lastly, if i≤|Z|, continue to (!). The resulting files triples and symbols are binary sequences that can be further compressed with standard compression schemes (so, decompression will start by first reversing this process). The purpose of the k and the two files (one with byte symbols and one with long triples) is to introduce flexibility into the system and encode CSSs with triples (12 bytes) only when beneficial and otherwise, encode a symbol with a byte. For convenience, our implementation encodes each symbol with a byte, but we acknowledge that it is possible to work at the bit-level for small alphabets. The decompression is also a left-to-right scan. Let i=1 and point to the beginning of triples and symbols. (†) Consider the current long word w1 in triples. According to the triple encoding, this will be the position of the CSS in T. If i=w1, then we pick up the next two long words w2 and w3 in triples. We now know T[i…i+w3−1]=Z[w2…w2+w3−1]. Since we only have access to R and T[1…i−1], then we pick up each symbol of Z[w2…w2+w3−1] by picking up R[j] if j≤|R| and picking up T[j−|R|] otherwise, for w2≤j≤w2+w3−1. We next will consider i=i+w3. Else i≠w1, so we pick up the next char c in symbols since T[i]=c; we next consider i++. If i≤|T|, go to (†). The compression and decompression algorithms are detailed in Algorithms 4 and 5, respectively. Results and discussion We implemented the previously described compression scheme, selected and fixed parameter k, and ran our program to compress various DNA corpora. In this section, we describe the selection of k and our final results. Choosing parameter k Recall that the parameter k is a type of threshold used by our compression scheme to determine whether it is more beneficial to encode a symbol verbatim (that is, 1 byte) or encode a CSS as a triple (that is, 12 bytes). Specifically, our compression algorithm works on the LPF (which represents the CSSs of the n-length T) in a left-to-right fashion, selecting the leftmost CSS, say T[i…i+l−1] of length-(LPF[i]=l), and determining whether to encode that CSS as a triple [and then consider the next CSS (T[i+l…i+l+LPF[i+l]−1] of length- LPF[i+l])], or encode the first symbol (T[i]) [and then consider the next CSS (T[i+1…i+LPF[i+1]] of length- LPF[i+1])]. Obviously, it is better to encode a length-(l=1) CSS with a 1-byte symbol, rather than a 12-byte triple. It is clearly the case that for any CSS length 1≤l<12, it is better to encode the first symbol with 1-byte and take a chance that the next CSS to the right will be significantly larger. Why can we afford to take this chance? One LPF property, which also allows for an efficient construction of the data structure (see [29]), is that LPF[i+1]≥LPF[i]−1. That is, if we pass up on encoding the CSS at i of length-(LPF[ i]=l) as a triple, we can encode T[ i] as a symbol and (1) are guaranteed that there is at least a length-(l−1) CSS with a prefix of T[ i+1…n] and (2) the longest CSS common to a prefix of T[ i+1…n] is of length- LPF[ i+1], maybe even larger than LPF[ i]. Clearly, we want to encode most CSSs as triples to take advantage of the concise triple representation. Now, the question becomes: how large should we set k, such that we can afford to take a risk passing up length-(l<k) CSSs in hopes of finding even larger CSSs better suited as triples? For this paper, we decided to select k by studying the impact of the parameter on our compressed results for the Arabidopsis thaliana genome, using target TAIR9 and reference TAIR8. The compression results for various k are shown in Fig. 2; since chromosome 4 does not compress as well as the others, we show it separately in Fig. 3 for improved visualization. For very small k<12, we have a result that basically encodes with triples only; when k=1, we are exclusively encoding CSSs as triples. We see that when k is roughly between 12 and 35, we are encouraging the algorithm to pass up encoding smaller CSSs as triples, which leads to the best compression result. The results stay competitive until say k≥100, where the algorithm becomes too optimistic and passes up the opportunity to encode smaller CSSs as triples in hopes that larger CSSs will exist. Further, we see from Fig. 4 that as k becomes large, it indeed becomes very expensive to pass up encoding these CSSs as triples. Also, we see from Fig. 5 that beyond say k=20, there is minimal compression savings. Thus, we want to balance the expensive symbols files with the space-savings from the triples files. Fig. 2 Total bytes needed by our algorithm to compress the Arabidopsis thaliana genome, i.e. file size sum of symbols and triples Fig. 3 Compressing the Arabidopsis thaliana genome Chromosome 4 Fig. 4 Size of the symbols file when compressing the Arabidopsis thaliana genome Fig. 5 Size of the triples file when compressing the Arabidopsis thaliana genome In Table 1, we show the best compression results and k for the Arabidopsis thaliana genome. Unless otherwise specified, our experiments below fix parameter k as 31, since it is the optimal k common to 4-of-5 of the Arabidopsis thaliana chromosomes and gives a competitive result for the remaining chromosome. This result follows intuition because k should be at least 11 and not too large, so that we can consider CSSs that are worthy of encoding. Table 1 Arabidopsis thaliana genome: Optimal k for compressing chromosome U into the smallest C (in bytes) U k |C| 1 31–35 1086 2 16–1578 504 3 24–39 746 4 18 4418 5 19–91 433 Compression results Like [20, 21], we compress the Arabidopsis thaliana genome chromosomes in TAIR9 (target) with respect to TAIR8 (reference). In Table 2, we display the compression results. We see that all of our results are competitive with the GRS and GReEn systems, except for chromosome 4, which has the smallest average CSS length of about 326K. Nonetheless, we are able to further compress our results using compression schemes from 7-zip, with 𝕃 and ℙ respectively representing lzma2 and ppmd, to achieve even better compression. Table 2 Arabidopsis thaliana genome: Results (in bytes) for compressing chromosome U into C U |U| Our Scheme GRS GReEn |C| |𝕃(C)| |ℙ(C)| [20] [21] 1 30 427 671 1 086 963 1 037 715 1 551 2 19 698 289 504 584 605 385 937 3 23 459 830 746 759 803 2 989 1 097 4 18 585 056 4 555 2 507 3 156 1 951 2 356 5 26 975 502 433 502 520 604 618 Sum 119 146 348 7 324 5 315 6 121 6 644 6 559 Bold signifies the best result In Table 3, we show results for compressing the genome Oryza sativa using the target TIGR6.0 and reference TIGR5.0. After compressing our algorithm’s output with lzma2 or ppmd, our results are better than both GRS [20] and GReEn [21]. Note that for each of the chromosomes 6, 9, and 12, our algorithm’s output is 12 bytes, better than that of GRS [20] (14 bytes) and GReEn [21] (482 bytes, 366 bytes, and 429 bytes respectively). When we compress our result with lzma2 or ppmd, the result is bloated since more bytes are needed. So, we can further improve the overall result by not compressing chromosomes 6, 9, and 12, and further, selecting the best such compression scheme for each individual chromosome. We acknowledge that additional bits would need to be encoded to determine which compression scheme was selected. Table 3 Oryza sativa genome: Results (in bytes) for compressing chromosome U into C U |U| Our Scheme GRS GReEn |C| |𝕃(C)| |ℙ(C)| [20] [21] 1 43 268 879 15 207 4 735 4 551 1 502 040 4 972 2 35 930 381 4 645 1 649 1 517 1 409 1 906 3 36 406 689 54 234 15 693 15 556 47 764 17 890 4 35 278 225 21 474 6 636 6 432 36 145 6 750 5 29 894 789 17 030 5 431 5 359 6 177 5 539 6 31 246 789 12 146 141 14 482 7 29 696 629 5 899 2 064 1 972 4 067 2 448 8 28 439 308 23 126 8 794 10 115 118 246 9 507 9 23 011 239 12 146 141 14 366 10 23 134 759 175 228 49 713 50 277 788 542 60 449 11 28 512 666 41 407 13 006 13 351 2 397 470 14 797 12 27 497 214 12 146 141 14 429 Sum 372 317 567 358 286 108 159 109 553 4 901 902 125 535 Bold signifies the best result In Table 4, we show the compression results for the Homo sapiens genome, using KOREF_20090224 as the target and KOREF_20090131 as the reference. After compressing our computed symbols and triples files with lzma2, we see that most all of our results are better than GRS and GReEn. Recall in previous experiments that sometimes secondary compression with 7-zip does not improve the initial compression achieved by our proposed algorithm. For this genome, we exercise the flexibility of our compression framework. In Table 4, (*) indicates that the M chromosome was not further compressed with lzma2 due to the aforementioned reason. To indicate that M was not compressed, we will simply encode a length-25 bitstring (say 4-byte) header to specify whether or not the lzma2 was applied. There is no need to encode k in the header since it is a fixed value. Thus, the overall compressed files require 15,460,478 bytes, which is only slightly better than GRS and GReEn. Table 4 Homo sapiens genome: Results (in bytes) for compressing chromosome U into C U |U| Our Scheme GRS GReEn |C| |𝕃(C)| [20] [21] 1 247 249 719 2 836 652 1 082 859 1 336 626 1 225 767 2 242 951 149 2 871 186 1 050 170 1 354 059 1 272 105 3 199 501 827 2 115 410 790 444 1 011 124 971 527 4 191 273 063 2 398 432 910 898 1 139 225 1 074 357 5 180 857 866 2 064 874 764 458 988 070 947 378 6 170 899 992 1 902 067 710 355 906 116 865 448 7 158 821 424 2 326 721 844 194 1 096 646 998 482 8 146 274 826 1 617 884 617 996 764 313 729 362 9 140 273 252 1 877 509 704 205 864 222 773 716 10 135 374 737 1 623 010 617 633 768 364 717 305 11 134 452 384 1 586 558 604 901 755 708 716 301 12 132 349 534 1 476 523 566 997 702 040 668 455 13 114 142 980 1 100 576 399 527 520 598 490 888 14 106 368 585 1 026 227 377 695 484 791 451 018 15 100 338 915 1 055 663 398 720 496 215 453 301 16 88 827 254 1 225 378 443 009 567 989 510 254 17 78 774 742 1 081 739 396 371 505 979 464 324 18 76 117 153 865 138 320 361 408 529 378 420 19 63 811 651 862 129 320 789 399 807 369 388 20 62 435 964 605 179 229 418 282 628 266 562 21 46 944 323 488 340 180 096 226 549 203 036 22 49 691 432 568 734 205 244 262 443 230 049 X 154 913 754 7 525 925 2 494 884 3 231 776 2 712 153 Y 57 772 954 1 343 260 429 099 592 791 481 307 M 16 571 151 151(*) 183 127 Sum 3 080 436 051 42 445 265 15 460 474 19 666 791 17 971 030 Bold signifies the best result To improve this result, we exploit the difference between the Homo sapiens genome and those discussed earlier. That is, the Homo sapiens genome uses the extended alphabet {A, C, G, K, M, R, S, T, W, Y, a, c, g, k, m, n, r, s, t, w, y}. The observation is that, the alphabet size decreases roughly in half by converting to one character-case. Such a significant reduction in alphabet size will yield more significant redundancies identified by our compression algorithm. Our new decomposition method will decompose each chromosome into two parts: (1) the payload (ρ), representing the chromosome in one character-case, and (2) the character-case bitstring (α), in which each bit records whether the corresponding position in the target was an upper-case character. Next, we use our previously proposed algorithm to compress ρ into Cρ and α into Cα. Table 5 shows compression via decomposition for the Homo sapiens genome. Note that the |Cρ|, i.e. compressed payload, column corresponds to the results reported in our initial work [1]. We observe that in various scenarios, the character-case of the alphabet symbol is not significant. For example, the IUB/IUPAC amino acid and nucleic acid codes use only upper-case letters (see http://www.bioinformatics.org/sms/iupac.html). Also, some environments and formats (such as FASTA) do not distinguish between lower-case and upper-case. According to the NCBI website for BLAST input formats (see http://blast.ncbi.nlm.nih.gov/blastcgihelp.shtml): “Sequences [in FASTA format] are expected to be represented in the standard IUB/IUPAC amino acid and nucleic acid codes, with these exceptions: lower-case letters are accepted and are mapped into upper-case; …” Further, some programs/environments use character cases for improved visualization, as is the case with the USC Genome Browser, which uses lower-case to show repeats from RepeatMasker and Tandem Repeats Finder (ftp://hgdownload.cse.ucsc.edu/goldenPath/hg38/chromosomes/README.txt). Table 5 Homo sapiens genome: Results (in bytes) for compressing chromosome U via decomposition, i.e. compressing the payload (ρ) into C ρ and compressing the character-case bitstring α into C α U |U| Our Scheme GRS GReEn |C ρ| |𝕃(Cρ)| |C α| |𝕃(Cα)| |𝕃(Cρ)|+|𝕃(Cα)| [20] [21] 1 247 249 719 381 577 161 319 755 092 447 919 609 238 1 336 626 1 225 767 2 242 951 149 356 526 153 805 756 823 452 338 606 143 1 354 059 1 272 105 3 199 501 827 284 096 119 348 553 835 343 213 462 561 1 011 124 971 527 4 191 273 063 330 381 137 301 619 981 383 882 521 183 1 139 225 1 074 357 5 180 857 866 259 922 109 768 550 876 331 063 440 831 988 070 947 378 6 170 899 992 265 222 110 544 508 662 310 029 420 573 906 116 865 448 7 158 821 424 292 797 121 289 611 475 355 616 476 905 1 096 646 998 482 8 146 274 826 222 972 93 378 434 420 261 455 354 833 764 313 729 362 9 140 273 252 309 512 132 957 493 024 276 468 409 425 864 222 773 716 10 135 374 737 245 264 103 115 436 272 257 895 361 010 768 364 717 305 11 134 452 384 222 735 92 471 423 687 254 637 347 108 755 708 716 301 12 132 349 534 214 123 88 447 393 764 239 811 328 258 702 040 668 455 13 114 142 980 148 938 62 730 301 116 183 038 245 768 520 598 490 888 14 106 368 585 141 128 57 354 286 839 170 916 228 270 484 791 451 018 15 100 338 915 138 219 58 777 302 957 173 600 232 377 496 215 453 301 16 88 827 254 151 606 62 779 346 282 191 190 253 969 567 989 510 254 17 78 774 742 136 168 57 030 301 837 171 680 228 710 505 979 464 324 18 76 117 153 113 469 47 122 241 437 140 909 188 031 408 529 378 420 19 63 811 651 130 468 53 531 230 673 134 701 188 232 399 807 369 388 20 62 435 964 94 273 38 689 169 584 99 796 138 485 282 628 266 562 21 46 944 323 71 121 28 744 141 387 79 835 108 579 226 549 203 036 22 49 691 432 81 329 33 663 164 026 89 961 123 624 262 443 230 049 X 154 913 754 523 282 196 868 1 533 249 875 026 1 071 894 3 231 776 2 712 153 Y 57 772 954 152 464 57 002 300 287 153 582 210 584 592 791 481 307 M 16 571 64 64(*) 49 49(*) 113 183 127 Sum 3 080 436 051 5 267 656 2 178 095 10 857 634 6 378 609 8 556 704 19 666 791 17 971 030 Bold signifies the best result Also, we see that further compressing the payload with lzma2 more than doubles the compression ratio. Interestingly, the payload (ρ) compresses much better than the character-case bitstring (α). Nonetheless, the compression via decomposition (in Table 5) yields a compression ratio of 360, a significant improvement over the 199 compression ratio when compressing the genome’s characters in their native character-case (in Table 4). As described earlier, we do not further compress chromosome M after initial coding for the symbols and triplets, and thus encode only a 4-byte header to remember this decision, given that the payload and character-case bitstring k values are fixed. Thus, 8,556,708 bytes are needed, which is an improvement over GRS and GReEn. Theoretically, our compression scheme requires time linear in the length of the uncompressed text, since we perform one scan of the LPF, which is constructed in linear time via the suffix array SA [29]. For the Arabidopsis thaliana and Oryza sativa genomes, we ran our programs on a laptop; for the Homo sapiens genome, we ran our programs in an AWS EC2 m4.4xlarge environment. Consider, for example, the larger chromosomes of the Homo sapiens genome. For a payload (ρ), the SA construction required 2,376 sec and the LPF construction required 399 sec. Note that depending on the application, the SA and LPF may already be available. Given the LPF, our compression algorithm completed in less than one second. Decompression is also fast, and lightweight, since no data structures are required as parameters. Our future plan includes using more efficient SA and LPF constructions. Conclusions We proposed a new algorithm to compute the LCS. Motivated by our algorithm, we introduced a new reference-based compression scheme for genome resequencing data using the LPF. For the Arabidopsis thaliana genome (originally 119,146,348 bytes), our scheme compressed the genome to 5315 bytes, an improvement over the best performing state-of-the-art methods (6644 bytes [20] and 6559 bytes [21]). For the Oryza sativa genome (originally 372,317,567 bytes), our scheme compressed the genome to 108,159 bytes, an improvement over the 4,901,902 bytes in [20] and the 125,535 bytes in [21]. We also experimented with the Homo sapiens genome (originally 3,080,436,051 bytes), which was compressed to 19,666,791 bytes and 17,971,030 bytes in [20] and [21], respectively. By applying our scheme via a decomposition approach, we compress the genome to 8,556,708 bytes, and if alphabet character-case is not significant, we compress the genome to 2,178,095 bytes. Further improvement can be obtained by choosing the k parameter for each specific chromosome, or each specific species. Declarations This article has been published as part of BMC Genomics Vol 17 Suppl 4 2016: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine 2015: genomics. The full contents of the supplement are available online at http://bmcgenomics.biomedcentral.com/articles/supplements/volume-17-supplement-4. Funding This work was supported in part by grants from the US National Science Foundation, #IIS-1552860, #IIS-1236983. Authors’ contributions All authors contributed to the core elements of this work. DA initiated the project. RB, TA, and DA developed the LCS algorithm. DA and RB developed the compression algorithm. RB implemented the LPF and compression methods. AF and RB performed the experiments. DA coordinated the overall project. RB and DA prepared the final manuscript. All authors read and approved the final manuscript. Competing interests The authors declare that they have no competing interests. ==== Refs References 1 Beal R, Afrin T, Farheen A, Adjeroh D. A new algorithm for ‘the LCS problem’ with application in compressing genome resequencing data. In: Bioinformatics and Biomedicine (BIBM), 2015 International, IEEE, Conference on: 2015. p. 69–74. 2 Gusfield D Algorithms on Strings, Trees and Sequences: Computer Science and Computational Biology 1997 New York, NY Cambridge University Press 3 Adjeroh D Bell T Mukherjee A The Burrows-Wheeler Transform: Data Compression, Suffix Arrays, and Pattern Matching 2008 New York, NY Springer 4 Lin Z Wang H McClean S A multidimensional sequence approach to measuring tree similarity IEEE Trans Knowl Data Eng 2012 24 2 197 208 10.1109/TKDE.2010.239 5 Smith TF Waterman MS Identification of common molecular subsequences J Mol Biol 1981 147 195 7 10.1016/0022-2836(81)90087-5 7265238 6 Aach J Bulyk M Church G Comander J Derti A Shendure J Computational comparison of two draft sequences of the human genome Nature 2001 26 1 5 14 7 Wandelt S Leser U FRESCO: Referential compression of highly similar sequences IEEE/ACM Trans Comput Biol Bioinform 2013 10 5 1275 88 10.1109/TCBB.2013.122 24524158 8 Wandelt S Starlinger J Bux M Leser U RCSI: Scalable similarity search in thousand(s) of genomes Proc VLDB Endow 2013 6 13 1534 45 10.14778/2536258.2536265 9 Giancarlo R Scaturro D Utro F Textual data compression in computational biology: Algorithmic techniques Comput Sci Rev 2012 6 1 1 25 10.1016/j.cosrev.2011.11.001 10 Kuo C-E Wang Y-L Liu J-J Ko M-T Resequencing a set of strings based on a target string Algorithmica 2015 72 2 430 49 10.1007/s00453-013-9859-z 11 Myers EW An O(ND) difference algorithm and its variations Algorithmica 1986 1 2 251 66 10.1007/BF01840446 12 Ukkonen E Algorithms for approximate string matching Inform Control 1985 64 100 18 10.1016/S0019-9958(85)80046-2 13 Hunt JW Szymanski TG A fast algorithm for computing longest subsequences Commun ACM 1977 20 5 350 3 10.1145/359581.359603 14 Hirschberg DS A linear space algorithm for computing maximal common subsequences Commun ACM 1975 18 6 341 3 10.1145/360825.360861 15 Yang J Xu Y Shang Y Chen G A space-bounded anytime algorithm for the multiple longest common subsequence problem IEEE Trans Knowl Data Eng 2014 26 11 2599 609 10.1109/TKDE.2014.2304464 25400485 16 Maier D The complexity of some problems on subsequences and supersequences J ACM 1978 25 2 322 36 10.1145/322063.322075 17 Apostolico A Giancarlo R The Boyer-Moore-Galil string searching strategies revisited SIAM J Comput 1986 15 1 98 105 10.1137/0215007 18 Jacobson G Vo K-P Heaviest increasing common subsequence problems Proceedings of the Third Annual Symposium on Combinatorial Pattern Matching, ser. CPM ’92 1992 London Springer-Verlag 19 Pevzner PA Waterman MS A fast filtration algorithm for the substring matching problem LNCS 684 Comb Pattern Matching 1993 684 197 214 20 Wang C Zhang D A novel compression tool for efficient storage of genome resequencing data Nucleic Acids Res 2011 39 7 e45 10.1093/nar/gkr009 21266471 21 Pinho AJ Pratas D Garcia SP GReEn: A tool for efficient compression of genome resequencing data Nucleic Acids Res 2012 40 4 e27 10.1093/nar/gkr1124 22139935 22 Nevill-Manning CG Witten IH Protein is incompressible Proceedings of the Conference on Data Compression, ser. DCC ’99 1999 Washington IEEE Computer 23 Adjeroh D, Nan F. On compressibility of protein sequences. In: DCC.IEEE Computer Society. IEEE: 2006. p. 422–34. 24 Coxm AJ Bauer MJ Jakobi T Rosone G Large-scale compression of genomic sequence databases with the Burrows-Wheeler Transform Bioinformatics 2012 28 11 1415 9 10.1093/bioinformatics/bts173 22556365 25 Giancarlo R Scaturro D Utro F Textual data compression in computational biology: A synopsis Bioinformatics 2009 25 13 1575 86 10.1093/bioinformatics/btp117 19251772 26 Wandelt S Bux M Leser U Trends in genome compression Curr Bioinform 2014 9 3 315 26 10.2174/1574893609666140516010143 27 Fritz M Leinonen R Cochrane G Birney E Efficient storage of high throughput DNA sequencing data using reference-based compression Genome Res 2011 21 734 40 10.1101/gr.114819.110 21245279 28 Hach F Numanagic I Alkan C Sahinalp SC SCALCE: boosting sequence compression algorithms using locally consistent encoding Bioinformatics 2012 28 23 3051 7 10.1093/bioinformatics/bts593 23047557 29 Crochemore M Ilie L Computing longest previous factor in linear time and applications Inf Process Lett 2008 106 2 75 80 10.1016/j.ipl.2007.10.006 30 Beal R Adjeroh D Parameterized longest previous factor Theor Comput Sci 2012 437 21 34 10.1016/j.tcs.2012.02.004 31 Beal R Adjeroh D Variations of the parameterized longest previous factor J Discret Algorithm 2012 16 129 50 10.1016/j.jda.2012.05.004 32 Cormen TH Stein C Rivest RL Leiserson CE Introduction to Algorithms 2001 Cambridge, Massachusetts The MIT Press 33 Crochemore M Ilie L Smyth WF A simple algorithm for computing the Lempel Ziv factorization Proceedings of the Data Compression Conference, ser. DCC ’08 2008 Washington IEEE Computer Society

PMC005xxxxxx/PMC5001249.txt

==== Front Sci RepSci RepScientific Reports2045-2322Nature Publishing Group srep3217410.1038/srep3217427552933ArticleCombined expression of miR-34a and Smac mediated by oncolytic vaccinia virus synergistically promote anti-tumor effects in Multiple Myeloma Lei Wen 1Wang Shibing 2Yang Chunmei 1Huang Xianbo 1Chen Zhenzhen 1He Wei 1Shen Jianping 3Liu Xinyuan 4Qian Wenbin a11 Institute of Hematology, the First Affiliated Hospital, College of Medicine, Zhejiang University, Hangzhou, 310003, P.R. China2 Clinical Research Institute, Zhejiang Provincial People’s Hospital, Hangzhou, 310014, P.R. China3 Department of Hematology, the First Affiliated Hospital of Zhejiang Chinese Medicine University, Hangzhou, 310006, P.R. China4 Institute of Biochemistry and Cell Biology, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences, Shanghai, 200031, P.R. Chinaa qianwenb@hotmail.com24 08 2016 2016 6 3217406 06 2016 03 08 2016 Copyright © 2016, The Author(s)2016The Author(s)This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/Despite great progress made in the treatment of multiple myeloma (MM), it is still incurable. Promising phase II clinical results have been reported recently for oncolytic vaccinia virus (OVV) clinic therapeutics. One reason for this has focused on the critical therapeutic importance of the immune response raised by these viruses. However, few studies have performed their applications as an optimal delivery system for therapeutic gene, especially miRNA in MM. In this study, we constructed two novel OVVs (TK deletion) that express anti-tumor genes, miR-34a and Smac, respectively, in MM cell lines and xenograft model. The results demonstrated that the novel OVV can effectively infect MM cell lines, and forcefully enhance the exogenous gene (miR-34a or Smac) expression. Furthermore, utilization of VV-miR-34a combined with VV-Smac synergistically inhibited tumor growth and induced apoptosis in vitro and in vivo. The underlying mechanism is proposed that blocking of Bcl-2 by VV-miR-34a increases the release of cytochrome c from mitochondria and then synergistically amplifies the antitumor effects of Smac-induced cell apoptosis. Our study is the first to utilize OVV as the vector for miR-34a or Smac expression to treat MM, and lays the groundwork for future clinical therapy for MM. ==== Body Multiple myeloma (MM) is a genetically complex hematologic malignancy characterized by the accumulation of clonal malignant plasma in bone marrow. Despite remarkable progress made in the pathobiology and management of the disease, it is still incurable12. Deregulated expression of microRNAs (miRNAs), small noncoding RNAs that regulate gene expression, plays a key role in the pathogenesis and progression of several human cancers including MM345. miRNAs may act as oncogenes or tumor-suppressors. Loss of a tumor-suppressive miRNA activates oncogenic pathways that promote tumorigenic potential and drug resistance67. Therefore, replacing tumor suppressor miRNAs may be a potential strategy to eradicate cancer cells. In this light, recent studies have demonstrated that a variety of miRNAs such as miR-214, miR-29b, miR-125b, and miR-34a replacement strategies are active in reducing growth of MM cells891011. miR-34a is the first identified tumor-suppressor miRNA121314. Indeed, miR-34a inhibits many oncogenic processes by regulating genes that function in various cellular pathways including cell cycle, p53, and Wnt signaling1415. miR-34a has also been associated with regulation of cancer stem cell function14151617. Overexpression of miR-34a upon treatment with demethylating agent 5-Aza-2′- deoxycytidine inhibits prostate cancer stem cell growth and metastasis by directly targeting CD441718. miR-34a is of special interest for myeloma miRNA therapeutics because the miR-34a promoter is hypermethylated, an epigenetic modification that typically silences gene expression, in MM patient cells and MM cell lines19. Further research from Wang showed that epigenetic silencing of miR-34a is caused by an lncRNA (Lnc34a), which recruits proteins that modify the gene and switches off the production of miR-34a20. Recent studies have demonstrated that replacement of miR-34a exerts anti-tumor activity in vitro and in vivo against human MM xenografts in animal models, which suggests an important potential clinical application112122. However, optimization of the delivery of miR-34a is needed before it can be translated into the clinic setting. Another factor implicated in the pathogenesis and progression of MM is Smac. Smac is a mitochondrial protein that is released into the cytosol to induce apoptosis2324. Prior studies have established that Smac release is critical for anti-myeloma agent-induced apoptosis, and that dysfunctional Smac release contributes to cancer cell drug-resistance242526. Recent report reveals that Smac is able to potently induce appotosis in MM by decreasing the expression of inhibitor of apoptosis (IAP) proteins27. Thus, regulation of Smac expression could be a promising approach to treat MM. Oncolytic vaccinia viruses (OVVs) recently emerge as a promising novel approach to treat cancer because of their potential to infect, replicate in, and lyse tumor cells. OVVs also disseminate through the bloodstream to induce an anti-tumor immune response2829. In fact, JX-594, an OVV with a human granulocyte-macrophage colony-stimulating factor (hGM-CSF) insertion, has moved to clinical trials, where it has shown potential in liver carcinomas and colorectal cancer2830. In addition, recent reports have shown that an apoptin-producing recombinant VV selectively kills human cancer cells in vitro and in vivo31, and that VV expressing p53 effectively infects glioma cells and induces apoptosis32, suggesting that the recombinant “armed” OVV can be designed to express additional therapeutic genes that enhance the efficiency of destruction of the cancer cells. In this study, we constructed OVV expressing miR-34a (VV-miR-34a) or Smac (VV-Smac), and evaluated their anti-myeloma potential as monotherapies and in combination. We found that these OVVs were able to replicate in and kill human MM cells in vitro and in vivo. Furthermore, combination therapy with VV-miR-34a and VV-Smac synergistically improved anti-tumoral efficacy. Results Construction and characterization of the novel OVVs: VV-miR-34a and VV-Smac The attenuated (thymidine kinase [TK]-negative) replication competent OVV vector (Western Reverve strain) was used to generate OVV expressing miR-34a or Smac (Fig. 1A). The miR-34a or Smac gene was inserted into the TK region, disrupting the function of TK. Deletion of the TK gene inhibits viral replication in normal, non-dividing cells33. However, cancer cells have a high concentration of functional nucleotides that enables OVV replication to occur in the absence of viral thymidine kinase. Therefore, disruption of TK results in selective replication of the OVV in tumor cells. The T7 promoter was inserted before the exogenous genes to initiate their expression, and the gpt gene works as a screen gene engineered behind the exogenous genes. The whole expression cassette was constructed into the pCB vector, which is a shuttle plasmid for VV packaging kindly provided by academician Xinyuan Liu. To evaluate the infectious efficiency of the novel OVV on MM cell lines, U266 and RPMI-8226 cell lines were infected with VV-GFP, an OVV carrying GFP instead of miR-34a or Smac, at multiple doses and time points, respectively, and then photographed under inverted fluorescence microscopy (Nikon E300). GFP was readily detectable at 24 hours and peaked at 48 hours (Fig. 1B; upper panel). FACS analysis showed that the proportion of GFP + cells increased in a dose-dependent manner. The infectious efficiency increased to 59.1% in U266 and 49.26% in RPMI-8226, respectively, at 48 hours after infection with VV-GFP with a multiplicity of infection (MOI) of 8 (Fig. 1B; lower panel). To verify expression of the miR-34a and Smac transegenes following OVVs infection, we examined their ectopic expression in hematologic malignancy cells using real time PCR (RT-PCR) and immunoblot. Compared to normal human liver cells QSG-7701, basal levels of miR-34a were relatively low in hematologic cell lines including MM and leukemia cells (Fig. 1C). However, infection with VV-miR-34a and VV-Smac increased the expression of miR-34a and Smac in MM cell lines U266 and RPMI-8226 remarkably on an mRNA and protein level (Fig. 1D). VV-miR-34a and VV-Smac synergistically induce apoptosis through activation of the caspase pathway in MM cells The effect of miR-34a and Smac expression on MM cell viability and proliferation was examined. MM cells were treated with VV-miR-34a alone, VV-Smac alone or both at the indicated doses. Cell death and proliferation was examined at 72 hours later using MTT assay. Compared with VV-miR-34a or VV-Smac treatment alone, proliferation was significantly decreased following co-infection. Taking the RPMI-8226 for example, cells survival rate kept in 26.3% at the highest dose (8 MOI) in the treatment of VV-Smac, but dramatically dropped to 17.5% when exposed to VV-miR34a at the highest dose of 8 MOI. Other two MM cell lines U266 and P3X63Ag8 presented the similar results (Fig. 2A). All of the combination indexes (CI) were less than 1 (data not show). We also infected human peripheral blood mononuclear cells (PBMCs) from healthy donors and a normal liver cell line QSG-7701 with the same doses of the indicated OVVs to verify their safety on the normal cells. Results showed that minimal cytotoxic effects were detected in normal PBMCs and normal liver cell line QSG-7701. These findings suggest that the combined forced expression of Smac and miR-34a has a specific, synergistic suppressive effect on MM cell proliferation and survival. To determine whether the forced expression of miR-34a and Smac increase cell death by activating Smac-induced apoptosis, we examined activation of proteins in the Smac-induced caspase pathway at 48 hours after co-infection with VV-miR-34a and VV-Smac at a MOI of 4. Immunoblot analysis revealed that forced expression of miR-34a and Smac activated the cleavage of caspase-3, caspase-9 and PARP-1 (Fig. 2B). Apoptosis rates following co-infection, as determined by Annexin V/PI double staining, increased to 58.1%, nearly two times more than that of VV-Smac (32.4%) or VV-miR-34a (25.9%) treatment alone (Fig. 2C). Blocking activation of Caspase-9 by Caspase-9 inhibitor Z-LEHD-FMK (BioVison, 40 uM) inhibited apoptosis induction by co-infection with VV-miR-34a and VV-Smac. The proportion of Annexin V positive cell decreased from 52.91% to 37.2% (Fig. 2D). Taken together, the results indicate that co-expression of miR-34a and Smac induce apoptosis by activating the Smac-induced caspase pathway. VV-miR-34a and VV-Smac regulate expression of pro-survival proteins and protein inhibitors of apoptosis To determine the effects of miR-34a and Smac on the expression of pro-survival proteins, we also examined the expressions of miR-34a anti-apoptotic gene targets after infection. The results indicated that treatment with VV-miR-34a alone or in combination with VV-Smac both down regulated Bcl-2 and SIRT1 on an mRNA and protein level in U266 and RPMI-8226 cell lines. Moreover, miR-34a overexpression down-regulated Bcl-2 in a dose-dependent manner (Fig. 3A). A mimics of miR-34a which is a synthetic miR-34a delivered by stable nucleic acid lipid particles was used as control to further confirm that up-regulation of miR-34a resulted in a down-regulation of Bcl-2 by VV-miR-34a. Results in Fig. 3B showed that more than two times of miR-34a and Bcl-2 expression were altered in the treatment of VV-miR-34a compared with synthetic mimics, which suggests OVV is an ideal expression vector for miR-34a delivery in MM cell lines. To further investigate the underlying mechanism of the viruses, we examined the changes in the expression of IAP family proteins (inhibitors of apoptosis proteins) after infection with the indicated viruses. Treatment with VV-Smac alone or in combination with VV-miR-34a resulted in an apparent downregulation of c-IAP1, c-IAP2 and XIAP (Fig. 3C), which demonstrated that VV-Smac inhibited the IAPs anti-apoptosis activity. VV-miR-34a and VV-Smac induce mitochondria-initiated apoptosis through release of cytochrome c Previous studies have reported that suppressing Bcl-2 expression increases release of cytochrome c into the cytoplasm and stimulates apoptosis correspondingly3435. To determine whether miR-34a promotes Smac-induced apoptosis by increasing release of cytochrome c into cytoplasm, we examined changes in cytochrome c whole cell protein levels after infection with VV-miR-34a, VV-Smac or in combination, respectively. Immunoblot analysis showed that VV-miR-34a in combination with VV-Smac and VV-Smac alone stimulated upregulation of cytochrome c. However, infection with VV-miR-34a alone had no effect on cytochrome c whole cell expression (Fig. 4A). It suggests that miR-34a has not the ability to promote the whole protein upexpression of cytochrome c. To determine whether VV-miR-34a stimulates cytochrome c release from the mitochondrial endomembrane, we examined cytochrome c diffusion using laser scanning confocal microscopy (LSCM). We found that VV-miR-34a and VV-Smac infection alone can increase diffusion when compared to untreated or control VV-treated cells, while the combination synergistically increased the diffusion (Fig. 4B). These results suggest that miR-34a promotes Smac regulated apoptosis by increasing the diffusion of cytochrome c from the mitochondrial endomembrane to cytoplasm. In conclusion, our findings demonstrated that during OVVs infecting the tumor cells and eliciting the tumor specific oncolysis, enforced expression of miR-34a and Smac by VV-miR-34a and VV-Samc induced significant apoptosis in MM cells through blocking function of Bcl-2, and thus activating the release of cytochrome c from the mitochondrial endomembrane and finally synergistically enhancing the Smac-induced cytochrome c/Apaf-1/Caspase-9 signaling cascade (Fig. 5). Combined treatment with VV-miR-34a and VV-Smac inhibits tumor growth of MM xenograft To examine the therapeutic effects of treatment with VV-miR-34a and VV-Smac in vivo, animal experiments were performed using a MM tumor xenograft model established by RPMI-8226 cells. Tumors were measured up to postimplantation day 49 to monitor the effects of the viruses. As shown in Fig. 6A, there was a rapid decrease in mean tumor volume for animals that received intratumoral injections of VV-miR-34a, VV-Smac, or both compared to those that were injected with PBS, or the VV vector alone. The average tumor volume of PBS or VV treatment reached to 3359 mm3 and 2391 mm3, respectively, on the day before the end of the experiment. In contrast, xenografts of mice tumor volume that received VV-miR-34a, VV-Smac or the combination just reached to 2058 mm3, 1198 mm3 and 247 mm3, respectively. The combination treatment was more effective than VV-miR-34a (P = 0.002) and VV-Smac alone (P = 0.002). Moreover, intratumoral injection of VV-miR-34a, VV-Smac or their combination resulted in an improved survival rate compared with PBS, or VV groups (Fig. 6B). 100% of survival rate was received in the treatment of VV-miR-34a combined with VV-Smac. 75% of survival rate was obtained in the VV-miR-34a-treated mice and 87.5% of mice were still survived in VV-Smac-treatment. In comparison, only 50% of VV-treated mice were alive in the same period. To evaluate the histopathology changes in the tumor tissues, Hematoxylin and eosin (H&E) and immunohistochemical (IHC) staining were performed on day 7 after viral injection. H&E staining demonstrated that the combined treatment of VV-miR-34a and VV-Smac caused more severe cytopathic effects including cell death, tumor necrosis and vessel growth inhibition, than individual treatment of them (Fig. 7A). IHC analysis utilizing anti-Smac antibody confirmed a strong expression of Smac in the tumor tissues followed treatment of combined or VV-Smac alone which verified successfully enhanced expression of Smac by OVVs in tumor mass (Fig. 7B). Apoptosis was further examined using a TUNEL assay. Apoptosis was significantly higher in tumors that received co-infection with the OVVs compared with VV-miR-34a or VV-Smac infection alone. No significant apoptosis was observed in VV or PBS treated tumors (Fig. 7C). Morphological changes in the tumors examined by Transmission electron microscope (TEM) analysis revealed that tumors injected with both VV-miR-34a and VV-Smac displayed classical characteristics of apoptosis including the appearance of apoptotic bodies, nuclear collapse, disintegration of the nuclear envelope, a greater nuclear-to-cytoplasm ratio, nucleus deformation, heterochromatin and chromatin condensation against the nuclear envelope (Fig. 7D). These results demonstrate that treatment with VV-miR-34a and VV-Smac has an inhibitory effect of tumor growth and prolongs the survival time of the mice. Discussion Despite great advancements in our understanding of MM, there has been little success in treatment of the disease363738. These tumors are resilient due in part to their genetic complexity. Therefore, targeting multiple pathways is essential for treatment of MM. In our study, we combined forced expression of miR-34a and Smac via OVV to treat MM aiming to amplify the antiproliferation effect synergistically by each other. Indeed, our study showed that the combined forced expression of miR-34a and Smac via OVV synergistically suppressed MM cell survival while sparing normal cells/tissues. Moreover, our results further verified that even though different target agents or pathways were involved in the treatment of combined strategy, synergic anti-tumor effect was still exerted as in predict, which suggests combination therapy targeting miR-34a and Smac provides an attractive strategy for treatment of MM. Recent findings highlight miRNA therapeutics as an attractive option for treatment of MM due to their tumor suppressive activity394041. miR-34a, a member of the miR-34 family, is one of the first miRNAs to display tumor suppressive activity. It was down-regulated in several cancer types including chronic lymphocytic leukemia42, colorectal cancer43, lung cancer44 and MM19. Moreover, recent researches have focused on its functions in modulating immune response as a putative binder of PD-L1-3′ UTR45. Downregulation of miR-34a contributes to several important oncogenic programs including tumor metastasis, cell apoptosis, cellular senescence as well as cell-cycle arrest in human leukemia cell lines, which suggests that miR-34a may be a novel target for treatment of human leukemia. Here, we explored the molecular effects induced by miR-34a expression on MM cell lines via OVV. Forced expression of miR-34a increased cell apoptosis in vitro and in vivo. And one of the underlying mechanisms is to block Bcl-2, and then release cytochrome c from mitochondria to cytoplasm which stimulates the activation of caspase pathway. Although many advantages sparked the application of miR-34a from pre-clinical research to clinical therapy, the unstable nature of its miRNA molecular structure, their rapid plasma clearance and their poor intracellular uptake have limited the use of miR-34a and other miRNAs in the clinic. Therefore, it is vital that an ideal delivery system is identified. Our study is the first to use the novel OVV as the delivery vector for miR-34a. OVV has been an attractive agent for cancer therapy due to its ability to accept large fragments of DNA, high infectious efficiency, safety and tumor affinity, and thus overcomes the major limitations of other vehicles. Various anti-tumor genes, such as p53 or GM-CSF, are engineered into OVV vector and thus exhibit profound anti-tumor effects in several cancer cells. Up to now, several OVVs armed with anti-tumor genes are in clinical trials (Table 1). Our study showed that VV-GFP infecting MM cell line exerted significant infection efficiency. Compared with transfection of synthetic mimics, both overexpression of miR-34a and downregulation of Bcl-2 by VV-miR-34a infection were significantly higher than that of synthetic mimics, which supports OVV as an ideal delivery system for miR-34a replacement in MM. Smac, as a novel tumor suppressor, exerts anti-tumor activity via the cytochrome c/Apaf-1/caspase-9 pathway. As MM is a genetically diverse malignance, identifying points of crosstalk between different anti-tumor gene regulated pathways is imperative. During the mitochondria-initiated caspase activation pathway, cytochrome c released from mitochondrial is critical. Its release is often blocked by overexpression of Bcl-2 in cancer cells. Here, we showed that increased release of cytochrome c and apoptosis by forced expression of Smac was enhanced by simultaneously forced expression of miR-34a. Co-treatment with VV-miR-34a and VV-Smac also synergistically suppressed proliferation and promoted apoptosis in MM cells in vitro and in vivo. Because miR-34a expression suppressed expression of Bcl-2 and enhanced release of cytochrome c from the mitochondrial endomembrane, miR-34a may significantly amplify the anti-tumor effect of Smac by enhancing Smac-induced activation of the Caspase pathway. It seems very important for target cancer therapy. Therefore, it is not surprising to find that VV-miR-34a has the ability to assist the release of cytochrome c in combination of VV-Smac and VV-miR-34a and synergistically increases the antiproliferation of VV-Smac on MM cell line in vitro or in vivo, which may provide a promising strategy for treatment of MM. In conclusion, we successfully constructed two novel OVVs: VV-miR-34a and VV-Smac. Data from in vitro experiments and in vivo xenograft model both demonstrated that the combination of VV-miR-34a and VV-Smac exerts powerful anti-tumor activity, apoptotic effects and gene regulation. All of these findings may lay the groundwork for future application of OVV in a clinical setting. Materials and Methods Cell lines and OVVs Human MM cell lines RPMI-8226, U266 and mouse MM cell line P3X63Ag8 were purchased from the American Type Culture Collection (ATCC, USA). Other human hematologic tumor cell lines were preserved in our lab. Human peripheral blood mononuclear cells (PBMCs) were isolated from buffy coats of healthy donors (Blood Bank, Hangzhou, Zhejiang, China) by Ficoll-Hypaque density gradient centrifugation after informed consent was signed. The study was approved by the Ethics Committee of Zhejiang University and all methods were carried out according to the approved guideline. All of the cells were cultured in RPMI-1640 culture medium supplemented with heat-inactivated fetal bovine serum (FBS, GibcoBRL, USA) at 37 °C in a humidified air atmosphere with 5% CO2. OVVs used in this study include VV vector, VV-miR-34a, VV-Smac and VV-GFP. Synthetic miR-34a mimics were produced in RiboBio co., Ltd (Guangzhou, China). Construction, identification, purification of oncolytic vaccinia viruses VV-miR-34a or VV-Smac was generated by homologous recombination of pCB-miR-34a or pCB-Smac with wild type VV in HEK293 cells. In details, the complete cDNA sequence of miR-34a or Smac gene was amplified by PCR (5′-GGAAGATCT CCTCCTGCATCC-3′[forward] and 5′-CCGCTTAAGATACCGCTCGAG-3′[reverse] for miR-34a, 5′-GGAAGATCTATGGCGGCTCTGAGA-3′[forward] and 5′- CCGCTTAAGTCAATCCTCACGCA -3′[reserve] for Smac). The synthetic DNA was digested with BglII (Takara, Japan) and ECORI (Takara, Japan) and then ligated into plasmid pCB to generate pCB-miR-34a or pCB-Smac, respectively. After sequence confirmation, pCB-miR-34a or pCB-Smac homologously recombined with wild type VV in HEK293 cells using Lipofectamine 2000 (Invitrogen, USA). After observing apparent cytopathic effect, cell culture medium was collected and repeated freezing and throwing three times. Obtained supernatant was VV-miR-34a or VV-Smac. Identification of correct VV was processed by plaque formation assay. Mycophenolic acid, dioxopurine and hypoxanthine were used to get rid of wild-type VV. Recombinant OVVs were amplified in HEK293 cells and purified by sucrose gradient ultracentrifugation. Moreover, viral titers were determined by TCID50 (median tissue culture infective dose). MTT assay For cell proliferation assays using 3-(4,5-dimethylthiazol-2-yl) -2,5-diphenyltetrazolium bromide(MTT, Sigma, USA), cells were seeded in 96-well plates at a density of 1 × 105/ml in the presence or absence of VV, VV-miR-34a, VV-Smac or VV-miR-34a combined VV-Smac in the indicated dose. After incubation for 72 hours, 20 μl of MTT solution (5 mg/ml) was added to each well and then the plates were incubated for an additional 4 hours at 37 °C. After supernatant was removed, 200 μl dimethylsulfoxide (DMSO) was added. A scientific microplate reader (Thermo, USA) was applied to detect the absorbance at a wavelength of 570 nm. Each assay was performed three times. The data was processed using Sigmaplot 6.1 package. Quantitative real-time amplification of RNAs Total RNA from MM cells treated with the indicated viruses was prepared with the TRIZOL Reagent (Invitrogen) according to manufacturer’s instructions. Oligo-dT-primed cDNA was obtained using the High Capacity cDNA Reverse Transcription Kit (TOYOBO, Japan). SYBR Premix Ex Taq (TAKARA, Japan) was used as fluorescent dye. The sequence detection system (Applied Biosystems 7500, USA) was used to detect and quantify mature miR-34a and target mRNAs according to the manufacturer’s instructions. miR-34a and mRNAs were normalized on U6 and glyceraldehyde- 3-phosphate dehydrogenase (GAPDH), respectively. Comparative RT-PCR was conducted in triplicate, including no-template controls. Relative expression was calculated using the comparative cross-threshold (Ct) method. The images were obtained using the Sigmaplot 6.1 package. The primers for miR-34a and U6 were synthesized in RiboBio co., Ltd. Other primers used in this study were followed as: 5′-CTG AAG AGT TGG CTG TC-3′ (forward) and 5′-CTA AGG GAA TGA GGC-3′ (reverse) for Smac, 5′-TGATGAACCGCTTGCTAT-3′ (forward) and 5′-TGGTCTTACTTTGAGGGA-3′ (reverse) for SIRT1. 5′-CTGGTGGACAACATCGC (forward) and 5′- GGAGAAATCAAACAGAGGC-3′ (reverse) for Bcl-2, 5′-ATG GGG AAG GTG AAG GTC G-3′ (forward) and 5′-GGG TCA TTG ATG GCA ACA ATA TC-3′ (reverse) for GAPDH. Western blot analysis RPMI-8226 cells were cultivated and treated with a MOI of 4 of VV, VV-miR-34a, VV-Smac or VV-miR-34a combined with VV-Smac, respectively, for 48 hours. Cells were washed twice with PBS and lysed in RIPA buffer (Sigma). Protein concentration of samples was measured by bicinchoninic acid (BCA) method. The protein samples were separated by SDS-polyacrylamide gel electrophoresis and then electroblotted onto PVDF membranes. The membranes were blocked in 5% non-fat milk for 2 hours and incubated with primary antibodies at 4 °C overnight. The membranes were washed and incubated with secondary antibody conjugated with horseradish peroxidase (1:5,000, CellSignaling Technology, USA). ECL detecting kit (Thermo Pierce, USA) was applied to visualize results. The images were obtained by Bio-Rad ChemiDox XRS imaging system. The primary antibodies and dilutions used were anti-caspase-3, anti-caspase-9, anti-PARP, anti-SIRT1, anti-Bcl-2, anti-XIAP, β-actin (1:1000, CST, USA), anti-GAPDH (1:1,000, Santa Cruz, USA), anti-cytochrome c (1:500, Abcam, UK). Flow cytometry analysis RPMI-8226 cells were cultured at a density of 1 × 105 cells/ml in a 6-well plate and treated with the indicated viruses at 4 MOI for 48 hours. After incubation at 37 °C, the cells were washed, resuspended in 500 μl of binding buffer and stained with 5 μl of Annexin V-FITC and 10 μl of propidiumiodide (PI) (Biouniquer, Suzhou, China) for 15 min in the dark. Then cells were examined by flow cytometry (Accuri C6, BD Biosciences, USA). The experimental data was analyzed using software of Flowjo. Confocal laser scanning microscopy study for cytochrome c staining Before immunoflorescent staining for cytochrome c, MitoTracker® Red CMXRos (Invitrogen, M7512) was chosen to staining mitochondria membrane firstly. In details, RPMI-8226 cells were infected with different viruses at an MOI of 4. After 48 hours, collecting cells and stained with prewarmed (37 °C) MitoTracker® probe (100–500 nM) for 15–45 minutes. Immunoflorescent staining for cytochrome c was performed as described previously46. In details, cells were washed with PBS, and then fixed with 4% paraformaldehyde. After permeablized by 0.3% Triton X-100 and incubated with goat serum, cells were stained with anti-cytochrome c antibody (Abcam, 1:200 dilution) overnight at 4 °C. Then, cells were incubated with a goat anti-rabbit antibody as secondary antibody (1:500) at 37 °C for 1 h and 1 μM DAPI (SouthernBiotech, USA) for 10 min. Finally, samples were examined with a Nikon confocal microscope (Nikon C1-Si, Japan), and images were processed using NIS-Elements software package. Animal experiments All animal experiments were approved by the Institutional Animal Care and Use Committee, Zhejiang University and all procedures were in according to the Guide for the Care and Use of Laboratory Animals (National Academies Press, Washington, D.C.). Female BALB/c nude mice (4-week old) were purchased from Shanghai Experimental Animal Center (Shanghai, China). For establishment of xenograft, RPMI-8226, human multiple myeloma cell lines were subcutaneously injected into the right flank of each mouse at a dose of 5 × 106 cells in 100 μl RPMI-1640. When tumors reached 100–150 mm3 in size, mice were divided randomly into five groups (six mice per group). VV, VV-Smac, VV-miR-34a and VV-Smac combined with VV-miR-34a (1 × 109 PFUs per mouse, intratumor) were injected every other day for four days. PBS treatment worked as control. Immunohistochemistry Tumors were excised and fixed in 4% paraformaldehyde, embedded in paraffin, and cut in 4-mm sections. For histopathology analysis, the paraffin sections of tumors were stained with hematoxylin and eosin (H&E). For Immunohistochemical analysis, the sections were stained with rabbit monoclonal anti-Smac antibodies at 1:500 dilutions, and then washed with PBS, incubated with the avidin-biotin-peroxidase complex reagent (Vector Laboratories, USA). The slides were detected with diaminobenzidine tetrahydrochloride solution containing 0.006% hydrogen peroxide. Hematoxylin was used as a counterstain. Tissue sections stained without primary antibodies were used as negative controls. Images were acquired using a Nikon E300 inverted fluorescence microscope equipped with a Nikon digital camera (20x). TdT-mediated dUTP-biotin nick end-labeling (TUNEL) assay The TUNEL method was used for detection of apoptotic cells. For this purpose, the in situ cell apoptosis detection kit (Roche, USA) was used. The staining was carried out according to the manufacturer’s procedures. Images were acquired using a Nikon E300 inverted fluorescence microscope equipped with a Nikon digital camera (20x).Tissue sections in PBS group were stained and served as positive controls. The TUNEL reaction preferentially labels DNA strand breaks generated during apoptosis, and allows discrimination of apoptosis from necrosis and primary DNA strand breaks induced by apoptotic agents. Transmission electron microscopy (TEM) analysis For electron microscopy analysis, tumor samples (1 mm3) were fixed in a PBS mixture containing 2.5% glutaraldehyde overnight and then incubated in 1% osmium tetroxide for 1 h. Tissues were rinsed in ddH2O, dehydrated through a graded series of ethanol and propylene oxide and finally embedded in Epon 812 resin (Shell Chemicals, USA). After examination of semithin sections, areas were selected and subjected to ultrathin sectioning. Sections collected on 200 mesh copper grids were contrasted with lead citrate and uranyl acetate, examined and photographed with a JEOL 100CX transmission electron microscope (JEOL, Akishima, Japan). Statistical analysis All data were expressed as the mean ± standard deviation and analyzed by SPSS 11.0 software (SPSS Inc., USA). Data analysis was performed using t-tests and P < 0.05 was considered statistically significant. Additional Information How to cite this article: Lei, W. et al. Combined expression of miR-34a and Smac mediated by oncolytic vaccinia virus synergistically promote anti-tumor effects in Multiple Myeloma. Sci. Rep. 6, 32174; doi: 10.1038/srep32174 (2016). Supplementary Material Supplementary Information The authors thank the scientific reviewers and the editors for kindly amending and reviewing our article. We also appreciate the assistance from the medical writers, proof-readers. This work was supported by National Natural Science Foundation of China (No. 81370645), National Natural Science Foundation of China (No. 81302036), Zhejiang Provincial Natural Science Foundation of China (No. LY13H080005), State Administration of Traditional Chinese Medicine of China (NO.JDZX2015113), Funds of Science Technology Department of Zhejiang Province (No. 2015C37035). Author Contributions All authors have made substantive intellectual contributions to this study and have given final approval of the version to be published. Each author has participated sufficiently in the work to take public responsibility for appropriate portions of the content. W.L. and S.W performed most of the experiments, analyzed data, drew the schematic image in Figure 5 and wrote the M.S draft. C.Y., X.H., Z.C. and W.H. helped to collect the materials, did parts of the experiments and give careful scrutiny for the M.S. and J.S. helped to provide with human normal peripheral blood to prepare PBMCs. X.L. supervised the study and helped in writing the M.S. and W.Q. was responsible for the overall organization of the study, planned the experiments and wrote the M.S. Figure 1 Characterization of the novel OVVs: VV-miR-34a and VV-Smac. (A) Schematic diagram of recombinant OVV structure. All viruses were constructed through homologous recombination of pCB-transgene with wild type VV in HEK293 cells. T7 promoter and gpt gene work as promoter and screen gene. (B) Infectious efficiency of OVV in MM cells was evaluated by fluorescence microscopy and FACS analysis. RPMI-8226 and U266 cells were infected with VV-GFP at multiple doses and time points and then observed under a fluorescence microscope (200×) (top panel). Two cell lines were treated with VV-GFP for 48 hours in dose-depended manner and then collected for FACS analysis (bottom panel). (C) Basal expression of miR-34a in hematologic malignancy cell lines was assessed by RT-PCR analysis (data are presented as means ± SD, n = 3). Human normal liver cell line QSG-7701 was used as the control to normalize the results. D. RT-PCR and immunoblot were carried out to analysis the enforced expression of miR-34a and Smac by OVVs. MM cell lines RPMI-8226 and U266 were infected with the indicated OVVs at a MOI of 4 for 24 hours or 48 hours. a,b. Expression of miR-34a in RPMI-8226 and U266 cells. c,d. Expression of Smac on mRNA in RPMI-8226 and U266 cells. e. Expression of Smac on protein level in RPMI-8226 cells. Full length blots were shown in Fig. S1. Figure 2 VV-miR-34a and VV-Smac synergistically induce apoptosis through activation of the caspase pathway in MM cells. (A) Three MM cell lines, QSG-7701 and human PBMCs were infected with VV, VV-miR-34a, VV-Smac and VV-miR-34a combined with VV-Smac at the MOI of 1, 2, 4, and 8. 72 hours later, cell viability rate was measured by MTT assay. The results were presented as the mean ± SD (n = 6) of three independent experiments. *represents P < 0.05, ** represents P < 0.005. (B) RPMI-8226 cells were treated with the indicated OVVs at 4 MOI. After 48 hours, whole-cell lysates were subjected to western blotting to assess the cleavage of caspase-9, -3 and PARP. (C) Annexin V/PI-staining method was used to detect the apoptosis induced by the indicated viruses at the MOI of 4 for 48 hours. (D) RPMI-8226 cells were pretreated with Z-LEHD-FMK (40uM) for 4 hours, and then infected with the indicated viruses for 48 hours at 4 MOI. Whole-cell lysates were analyzed for inhibition of activated caspase-9 by western blot (a) and apoptotic proportion by flow cytometry (b). Full length blots were shown in Fig. S1. Figure 3 VV-miR-34a and VV-Smac regulate expression of pro-survival proteins and protein inhibitors of apoptosis. (A) Expression of target genes Bcl-2 and SIRT1 of miR-34a after infected with different OVVs were evaluated on mRNA and protein level. RPMI-8226 and U266 cells were treated with the indicated viruses of 4 MOI. mRNAs of Bcl-2 and SIRT1 were extracted for RT-PCR analysis (a,b). Proteins from RPMI-8226 were harvested for Western Blot analysis using the antibodies for Bcl-2 and SIRT1(c). RPMI-8226 cells were infected by VV-miR-34a with dose-dependent manner and examined the expression of Bcl-2 (d). (B) Increased expression of miR-34a and reduction of Bcl-2 were detected on RNA levels (a,b) and protein level (c) after RPMI-8226 infected with VV-miR-34a or transfected with mimics for 24 hours. (C) Immunoblot assay evaluated the inhibition of IAPs anti-apoptosis activity by VV-Smac using antibodies for c-IAP1, c-IAP2 and XIAP. Full length blots were shown in Fig. S1. Figure 4 VV-miR-34a and VV-Smac induce mitochondria-initiated apoptosis through release of cytochrome c. (A) Up-regulation of cytochrome c was examined after infected with VV-Smac or combined with VV-miR-34a for 48 hours at 4 MOI by Western Blot. Full length blots were shown in Fig. S1. (B) Cell confocal microscopic images of cytochrome c (green) and mitochondria (red) stained by MitoTracker were collected from RPMI-8226 cells under treatment of VV-miR-34a, VV-Smac or the combined. Merged images showed the co-localization of cytochrome c and mitochondria (yellow). Figure 5 The mechanism schematic mode of viral oncolysis mediated by VV-miR-34a and VV-Smac is shown. (A) OVVs infect the tumor cells and elicit the tumor specific oncolytic effect. (B) miR-34a/Smac mediated caspase-9-dependent apoptosis pathway. Enforced expression of miR-34a by OVV blocks the function of Bcl-2, which promotes the release of cytochrome c from mitochondrial endomembrane. On the other hand, overexpression of Smac by OVV inhibits the function of inhibitors of apoptosis proteins (c-IAP, XIAP). Both of them synergistically activate the caspase-9 induced apoptosis pathway. This signal pathway image was drawn by Wen Lei. Figure 6 Combined treatment with VV-miR-34a and VV-Smac inhibits tumor growth of MM xenograft. (A) Tumor volume of various treatment groups was measured. Data are presented as means ± SD (n = 6). **P = 0.002. (B) Survival rate of mice was shown by the Kaplan–Meier survival curves. Figure 7 Histopathology analysis of tumor section in VV, VV-miR-34a, VV-Smac and the combined groups for RIMI-8226 xenograft tumor is shown. (A) Hematoxylin and eosin (HE) staining analysis. Original magnification: ×200. (B) IHC analysis for Smac expression in tumor tissues. Original magnification: ×200. (C) TUNEL assay for apoptotic cells treated with different OVVs. Original magnification: ×200. (D) Morphological observation of tumor tissues by TEM analysis. Original magnification: ×15000. Table 1 Recent clinical trials involving OVV armed antitumor genes. First author Vector Results Mastrangelo47 Vaccinia-GM-CSF Regression of injected lesions. Marshall48 Vaccinia-CEA No clinical response Mukherjee49 Vaccinia-IL-2 No clinical response Eder50 Vaccinia-PSA Stabilization of PSA levels Sanda51 Vaccinia-PSA Stabilization of PSA levels Conry52 Vaccinia-CEA No clinical response Tsang53 Vaccinia-CEA No clinical response Adams54 Vaccinia-HPV Response in cervical cancer Rochlitz55 MVA-Muc1 Responsein metastatic disease Greiner56 rV-CEA TRICOM Safe ==== Refs Lonial S. , Mitsiades C. S. & Richardson P. G. Treatment options for relapsed and refractory multiple myeloma . Clin Cancer Res 17 , 1264 –1277 (2011 ).21411442 Rajkumar S. V. Treatment of multiple myeloma . Nat Rev Clin Oncol 8 , 479 –491 (2011 ).21522124 Pichiorri F. . MicroRNAs regulate critical genes associated with multiple myeloma pathogenesis . Proc Natl Acad Sci USA 105 , 12885 –12890 (2008 ).18728182 Du J. . MicroRNA-451 regulates stemness of side population cells via PI3K/Akt/mTOR signaling pathway in multiple myeloma . Oncotarget 6 , 14993 –15007 (2015 ).25915427 Lionetti M. . Identification of microRNA expression patterns and definition of a microRNA/mRNA regulatory network in distinct molecular groups of multiple myeloma . Blood 114 , e20 –26 (2009 ).19846888 Kumar M. S. , Lu J. , Mercer K. L. , Golub T. R. & Jacks T. Impaired microRNA processing enhances cellular transformation and tumorigenesis . Nat Genet 39 , 673 –677 (2007 ).17401365 Yamamoto Y. . An integrative genomic analysis revealed the relevance of microRNA and gene expression for drug-resistance in human breast cancer cells . Mol Cancer 10 , 135 (2011 ).22051041 Misiewicz-Krzeminska I. . Restoration of microRNA-214 expression reduces growth of myeloma cells through positive regulation of P53 and inhibition of DNA replication . Haematologica 98 , 640 –648 (2013 ).23100276 Jagannathan S. , Vad N. , Vallabhapurapu S. , Anderson K. C. & Driscoll J. J. MiR-29b replacement inhibits proteasomes and disrupts aggresome+ autophagosome formation to enhance the antimyeloma benefit of bortezomib . Leukemia 29 , 727 –738 (2015 ).25234165 Morelli E. . Selective targeting of IRF4 by synthetic microRNA-125b-5p mimics induces anti-multiple myeloma activity in vitro and in vivo . Leukemia 29 , 2173 –2183 (2015 ).25987254 Di Martino M. T. . Synthetic miR-34a mimics as a novel therapeutic agent for multiple myeloma: in vitro and in vivo evidence . Clin Cancer Res 18 , 6260 –6270 (2012 ).23035210 Welch C. , Chen Y. & Stallings R. L. MicroRNA-34a functions as a potential tumor suppressor by inducing apoptosis in neuroblastoma cells . Oncogene 26 , 5017 –5022 (2007 ).17297439 Yu G. . MicroRNA-34a functions as an anti-metastatic microRNA and suppresses angiogenesis in bladder cancer by directly targeting CD44 . J Exp Clin Cancer Res 33 , 779 (2014 ).25551284 Agostini M. & Knight R. A. miR-34: from bench to bedside . Oncotarget 5 , 872 –881 (2014 ).24657911 Bader A. G. miR-34 - a microRNA replacement therapy is headed to the clinic . Front Genet 3 , 120 (2012 ).22783274 Misso G. . Mir-34: a new weapon against cancer? Mol Ther Nucleic Acids 3 , e194 (2014 ).25247240 Liu C. . The microRNA miR-34a inhibits prostate cancer stem cells and metastasis by directly repressing CD44 . Nat Med 17 , 211 –215 (2011 ).21240262 Nalls D. , Tang S. N. , Rodova M. , Srivastava R. K. & Shankar S. Targeting epigenetic regulation of miR-34a for treatment of pancreatic cancer by inhibition of pancreatic cancer stem cells . PLoS One 6 , e24099 (2011 ).21909380 Chim C. S. . Epigenetic inactivation of the miR-34a in hematological malignancies . Carcinogenesis 31 , 745 –750 (2010 ).20118199 Wang L. . A long non-coding RNA targets microRNA miR-34a to regulate colon cancer stem cell asymmetric division . Elife 5 (2016 ). Di Martino M. T. . In vivo activity of miR-34a mimics delivered by stable nucleic acid lipid particles (SNALPs) against multiple myeloma . PLoS One 9 , e90005 (2014 ).24587182 Scognamiglio I. . Transferrin-conjugated SNALPs encapsulating 2′-O-methylated miR-34a for the treatment of multiple myeloma . Biomed Res Int 2014 , 217365 (2014 ).24683542 Ghobrial I. M. , Witzig T. E. & Adjei A. A. Targeting apoptosis pathways in cancer therapy . CA Cancer J Clin 55 , 178 –194 (2005 ).15890640 Chauhan D. . Targeting mitochondrial factor Smac/DIABLO as therapy for multiple myeloma (MM) . Blood 109 , 1220 –1227 (2007 ).17032924 Chauhan D. & Anderson K. C. Mechanisms of cell death and survival in multiple myeloma (MM): Therapeutic implications . Apoptosis 8 , 337 –343 (2003 ).12815276 Feng R. . KD5170, a novel mercaptoketone-based histone deacetylase inhibitor, exerts antimyeloma effects by DNA damage and mitochondrial signaling . Mol Cancer Ther 7 , 1494 –1505 (2008 ).18566220 Ramakrishnan V. . Inhibitor of apoptosis proteins as therapeutic targets in multiple myeloma . Leukemia 28 , 1519 –1528 (2014 ).24402161 Heo J. . Randomized dose-finding clinical trial of oncolytic immunotherapeutic vaccinia JX-594 in liver cancer . Nat Med 19 , 329 –336 (2013 ).23396206 Kirn D. H. & Thorne S. H. Targeted and armed oncolytic poxviruses: a novel multi-mechanistic therapeutic class for cancer . Nat Rev Cancer 9 , 64 –71 (2009 ).19104515 Park S. H. . Phase 1b Trial of Biweekly Intravenous Pexa-Vec (JX-594), an Oncolytic and Immunotherapeutic Vaccinia Virus in Colorectal Cancer . Mol Ther 23 (2015 ). Kochneva G. . Apoptin enhances the oncolytic properties of vaccinia virus and modifies mechanisms of tumor regression . Oncotarget 5 , 11269 –11282 (2014 ).25358248 Timiryasova T. M. . Antitumor effect of vaccinia virus in glioma model . Oncol Res 11 (1999 ). Buller R. M. , Smith G. L. , Cremer K. , Notkins A. L. & Moss B. Decreased virulence of recombinant vaccinia virus expression vectors is associated with a thymidine kinase-negative phenotype . Nature 317 , 813 –815 (1985 ).4058585 Yang J. . Prevention of apoptosis by Bcl-2: release of cytochrome c from mitochondria blocked . Science 275 , 1129 –1132 (1997 ).9027314 Li L. , Peng L. & Zuo Z. Isoflurane preconditioning increases B-cell lymphoma-2 expression and reduces cytochrome c release from the mitochondria in the ischemic penumbra of rat brain . Eur J Pharmacol 586 , 106 –113 (2008 ).18355806 Richardson P. G. . Bortezomib or high-dose dexamethasone for relapsed multiple myeloma . N Engl J Med 352 , 2487 –2498 (2005 ).15958804 Kyle R. A. . Review of 1027 patients with newly diagnosed multiple myeloma . Mayo Clin Proc 78 , 21 –33 (2003 ).12528874 de la Puente P. , Muz B. , Azab F. , Luderer M. & Azab A. K. Molecularly targeted therapies in multiple myeloma . Leuk Res Treatment 2014 , 976567 (2014 ).24829804 Tagliaferri P. . Promises and challenges of MicroRNA-based treatment of multiple myeloma . Curr Cancer Drug Targets 12 , 838 –846 (2012 ).22671926 Rossi M. . From target therapy to miRNA therapeutics of human multiple myeloma: theoretical and technological issues in the evolving scenario . Curr Drug Targets 14 , 1144 –114 (2013 ).23834146 Amodio N. , Di Martino M. T. , Neri A. , Tagliaferri P. & Tassone P. Non-coding RNA: a novel opportunity for the personalized treatment of multiple myeloma . Expert Opin Biol Ther 13 Suppl 1, S125 –137 (2013 ).23692413 Blume C. J. . p53-dependent non-coding RNA networks in chronic lymphocytic leukemia . Leukemia 29 , 2015 –2023 (2015 ).25971364 Almeida A. L. . Serological under expression of microRNA-21, microRNA-34a and microRNA-126 in colorectal cancer . Acta Cir Bras 31 Suppl 1, 13 –18 (2016 ).27142899 Shi Y. , Liu C. , Liu X. , Tang D. G. & Wang J. The microRNA miR-34a inhibits non-small cell lung cancer (NSCLC) growth and the CD44hi stem-like NSCLC cells . PLoS One 9 , e90022 (2014 ).24595209 Cortez M. A. . PDL1 Regulation by p53 via miR-34 . J Natl Cancer Inst 108 (2016 ). He W. . Hsp90 inhibitor, BIIB021, induces apoptosis and autophagy by regulating mTOR-Ulk1 pathway in imatinib-sensitive and -resistant chronic myeloid leukemia cells . Int J Oncol 48 , 1710 –1720 (2016 ).26892093 Mastrangelo M. J. . Intratumoral recombinant GM-CSF-encoding virus as gene therapy in patients with cutaneous melanoma . Cancer Gene Ther 6 , 409 –422 (1999 ).10505851 Marshall J. L. . Phase I study in advanced cancer patients of a diversified prime-and-boost vaccination protocol using recombinant vaccinia virus and recombinant nonreplicating avipox virus to elicit anti-carcinoembryonic antigen immune responses . J Clin Oncol 18 , 3964 –3973 (2000 ).11099326 Mukherjee S. . Replication-restricted vaccinia as a cytokine gene therapy vector in cancer: persistent transgene expression despite antibody generation . Cancer Gene Ther 7 , 663 –670 (2000 ).10830713 Eder J. P. . A phase I trial of a recombinant vaccinia virus expressing prostate-specific antigen in advanced prostate cancer . Clin Cancer Res 6 , 1632 –1638 (2000 ).10815880 Sanda M. G. . Recombinant vaccinia-PSA (PROSTVAC) can induce a prostate-specific immune response in androgen-modulated human prostate cancer . Urology 53 , 260 –266 (1999 ).9933036 Conry R. M. . Human autoantibodies to carcinoembryonic antigen (CEA) induced by a vaccinia-CEA vaccine . Clin Cancer Res 6 , 34 –41 (2000 ).10656429 Tsang K. Y. . Generation of human cytotoxic T cells specific for human carcinoembryonic antigen epitopes from patients immunized with recombinant vaccinia-CEA vaccine . J Natl Cancer Inst 87 , 982 –990 (1995 ).7629885 Adams M. . Clinical studies of human papilloma vaccines in pre-invasive and invasive cancer . Vaccine 19 , 2549 –2556 (2001 ).11257391 Rochlitz C. . Phase I immunotherapy with a modified vaccinia virus (MVA) expressing human MUC1 as antigen-specific immunotherapy in patients with MUC1-positive advanced cancer . J Gene Med 5 , 690 –699 (2003 ).12898638 Greiner J. W. , Zeytin H. , Anver M. R. & Schlom J. Vaccine-based therapy directed against carcinoembryonic antigen demonstrates antitumor activity on spontaneous intestinal tumors in the absence of autoimmunity . Cancer Res 62 , 6944 –6951 (2002 ).12460911

PMC005xxxxxx/PMC5001437.txt

==== Front Econ Hum BiolEcon Hum BiolEconomics and Human Biology1570-677X1873-6130Elsevier Science S1570-677X(16)30010-710.1016/j.ehb.2016.03.002ArticleEarly life height and weight production functions with endogenous energy and protein inputs Puentes Esteban epuentes@fen.uchile.cla1⁎Wang Fan bBehrman Jere R. cCunha Flavio dHoddinott John eMaluccio John A. fAdair Linda S. gBorja Judith B. hMartorell Reynaldo iStein Aryeh D. ia Department of Economics, Universidad de Chile, Chileb Department of Economics, University of Houston, United Statesc Departments of Economics and Sociology and Population Studies Center, University of Pennsylvania, United Statesd Department of Economics, Rice University, United Statese Division of Nutritional Sciences and the Charles H. Dyson School of Applied Economics and Management, Cornell University and International Food Policy Research Institute, United Statesf Department of Economics, Middlebury College, United Statesg Department of Nutrition, University of North Carolina, United Statesh USC-Office of Population Studies Foundation, Inc and Department of Nutrition and Dietetics, University of San Carlos, Cebu, Philippinesi Rollins School of Public Health, Emory University, United States⁎ Corresponding author at: Diagonal Paraguay 257, of 1501, 8330015 Chile. Tel.: +56 2 29783455; fax: +56 2 29783413.Diagonal Paraguay 257, of 15018330015Chile epuentes@fen.uchile.cl1 The authors thank reviewers on previous versions for useful comments and Grand Challenges Canada (Grant 0072-03), Bill & Melinda Gates Foundation (Global Health Grant OPP1032713), and the Eunice Shriver Kennedy National Institute of Child Health and Development (Grant R01 HD070993) for financial support. The funders have no involvement in any part of the research project. 1 9 2016 9 2016 22 65 81 20 6 2015 21 2 2016 1 3 2016 © 2016 The Authors2016This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).Highlights • We estimate height and weight production functions for infants. • We focus on the role of energy and protein intake. • We use IV to control for endogeneity and estimate a number of models. • The results indicate that protein play an important role in height and weight change. We examine effects of protein and energy intakes on height and weight growth for children between 6 and 24 months old in Guatemala and the Philippines. Using instrumental variables to control for endogeneity and estimating multiple specifications, we find that protein intake plays an important and positive role in height and weight growth in the 6–24 month period. Energy from other macronutrients, however, does not have a robust relation with these two anthropometric measures. Our estimates indicate that in contexts with substantial child undernutrition, increases in protein-rich food intake in the first 24 months can have important growth effects, which previous studies indicate are related significantly to a range of outcomes over the life cycle. Keywords NutritionEarly childhoodEndogeneity of inputsGrowthProteins ==== Body 1 Introduction Inadequate child growth and weight gain are of paramount concern. Approximately 165 million children under five years old in developing countries are stunted and 100 million are underweight (Black et al. (2013)). Growing evidence indicates that early-life undernutrition is associated with, and likely in part causes, reduced education, adult cognitive skills, and wages (Grantham-McGregor et al., 2007, Engle et al., 2007, Engle et al., 2011, Victora et al., 2008, Hoddinott et al., 2008, Hoddinott et al., 2013, Behrman et al., 2009, Maluccio et al., 2009). Despite widespread concern about early-life undernutrition there is limited systematic knowledge about production technologies for key outcomes, particularly height and weight, needed to inform more-effective program and policy design. This gap is partially due to inherent difficulties in modeling these complex biological and behavioral processes—often strong assumptions are required for estimation, so that it is difficult to make definitive conclusions. A major challenge in estimating production functions for height and weight is that inputs reflect behavioral choices. Using data from the same Philippine study analyzed in this paper, Akin et al. (1992) and Liu et al. (2009) find that families allocate nutrients to compensate for prior poor health. Where allocations reflect compensatory behaviors that are not controlled for in the estimation, the estimated effect of nutrients on growth can be biased. Another challenge is measurement error in inputs. Using related data from Guatemala, Griffen (2016) finds that estimates of energy effects on height are substantially larger using instrumental variables (IV) than with ordinary least squares (OLS) probably in part due to measurement error. In this paper, we examine relations between energy intake and: (1) linear growth and (2) weight gain. We use longitudinal data from Guatemala and the Philippines that includes detailed information on anthropometric outcomes, nutrition and other inputs collected at intervals of two-three months to estimate height and weight production functions for children in the critical age range 6–24 months. In our specifications, height and weight depend on lagged height and weight, energy intakes, breastfeeding, diarrhea, and individual fixed endowments. We combine individual fixed-effects (FE) with instrumental variables (IV) to control for both endogeneity and measurement error. This paper presents three important methodological contributions. First, we estimate production functions for two countries, Guatemala and the Philippines, and for two anthropometric measures, height and weight, which allows us to compare the robustness of our findings across different settings and anthropometric outcomes. Second, we improve on previous IV literature on growth by providing details of instrument selection and an assessment of how the results are robust to changes in the instrument set. We present estimates for numerous instrument combinations, putting emphasis on those judged more reliable based on over-identification and weak instrument tests. Third, in addition to considering total energy intake, which is the nutritional input usually considered in the economics literature, we disaggregate energy intake into two components: proteins and (all) other macronutrients (which we refer to as “non-proteins”, meaning fat and carbohydrates). This emphasis on dietary quality, highlighted by Arimond and Ruel (2004), is especially relevant because it may help design interventions that better reduce stunting and underweight. We find robust and positive effects of proteins on height and weight growth. Energy from other macronutrient consumption (non-proteins), is not systematically related to these anthropometric measures, which suggests that protein-rich foods are particularly important for growth of undernourished children. 2 Specifications of height and weight production functions and identification 2.1 Input selection Our choice of inputs is guided by Black et al. (2008) who argue that inadequate diet and disease are the main immediate causes of stunting and wasting. With respect to diet, two energy sources have been identified as being especially important for child growth: proteins and non-protein energy from other macronutrients. Infants require certain minimum amounts of energy and proteins to maintain long-term good health but these requirements are heterogeneous and depend on several factors including weight and whether the child is breastfed (FAO, 2001, WHO, 2007). Children's energy requirements are partly driven by energy costs of linear growth, which has two components: (1) energy needed to synthesize growing tissues and (2) energy stored in these tissues (FAO, 2001). These comprise approximately one-third of total energy requirements during the first three months of life, but despite increasing in absolute terms they decline to only 3% by age 24 months, in part because overall energy requirements increase substantially with body size. Proteins are needed to balance nitrogen loss, maintain the body's muscle mass, and fulfill needs related to tissue deposition (WHO, 2007). There is also evidence from research on animals that protein provides anabolic drive for linear bone growth (WHO, 2007).2 To study the relative importance of protein and non-protein sources, we first examine the relationship between total energy and height and weight and then consider the potential for separate roles of the two at once in a single growth model. The comparison of proteins with non-proteins highlights the relative importance of proteins in children's diets and informs what types of interventions might have greater impact on height and weight.3 There is a limited literature focused on the distinction between total energy and protein energy. Pucilowska et al. (1993) find that high-protein supplementation in Bangladeshi children with shigellosis, a severe bacterial disease, increased weight compared to normal protein diets. A randomized evaluation for children up to 2 years of age in several European countries demonstrated that receiving baby formula with high protein content (% calories from protein) increased weight, but not height (Koletzko et al. (2009)). Both of these study populations, however, are different from the ones we examine. The Bangladeshi sample is restricted to children recovering from shigellosis while the European sample had not experienced the same nutritional deficiencies found in our samples. Using a sample more similar to ours, Moradi (2010) finds that access to high-quality protein, such as from livestock farming, better predicts height in some African countries than other energy sources. Similarly, Baten and Blum (2014), using global information for the first part of the twentieth century, that includes Guatemala and the Philippines, also find that local availability of cattle, milk and meat were an important predictor of adult height.4 A related issue is protein quality. Proteins are composed of amino acids with specific cell functions, and amino acid content defines protein quality. For instance, plant-based proteins lack essential amino acids unlike animal-based proteins (Dewey, 2013). In addition, plant-based diets have high levels of phytic acid, which might inhibit zinc absorption (Gibson, 2006), and zinc plays a key role in cellular growth and differentiation (Imdad and Bhutta, 2011). For animal-based protein, Mølgaard et al. (2011) argue that dairy intake has positive impacts on child growth. Although the mechanism is not entirely clear, this may be due to the stimulating effect on plasma insulin-like growth factor (IGF-1) (Michaelsen, 2013). Breastfeeding is another critically important source of nutrition in early life (Black et al., 2013). In this paper, we have data on breastfeeding status but not on the amount of breast milk consumed. Thus, our energy intake measures exclude energy from breastmilk requiring us to control for breastfeeding status in the models. Among diseases that affect growth, Walker et al. (2011) suggest that persistent diarrhea and other diseases can have long-lasting effects on children's physical development. Therefore, in our analyses, we incorporate diarrhea as an input, as it is considered a major contributor to stunting, wasting and child mortality (Black et al., 2013). 2.2 Height and weight production functions The main challenges for estimating height and weight production functions include the endogeneity of inputs and measurement error (Behrman and Deolalikar, 1988). To overcome these, we follow the general approach developed in recent research on production function estimation for cognitive and non-cognitive skills (Todd and Wolpin, 2003, Todd and Wolpin, 2007, Cunha and Heckman, 2007). Let hi,t denote child i height at age t, wi,t weight at age t and xi,j the input (e.g., proteins, non-proteins, or disease) at age j (For simplicity, we present the model with a single input but generalization to several inputs is straightforward.). Fairly general height and weight production functions are: (1) hi,t=αμi+∑j=1tβt−jxi,j+∈i,th (2) wi,t=σμi+∑j=1tδt−jxi,j+∈i,tw where μi is an individual fixed effect (including genetic endowments and fixed parental and household characteristics) and ∈i,th and ∈i,tw are error terms. This formulation allows the entire input history to enter into both equations up to time t. Furthermore, it allows for impacts of past inputs on current height and weight and for the possibility that such impacts differ by age. This approach also distinguishes our work from other studies using the same data. Griffen (2016) relies on the fairly strong assumption that past inputs have constant effects on height in Guatemala, so that history plays little role in growth. Similarly, height production functions estimated by de Cao (2015) in the Philippines, assume that height growth depends only on current inputs. Because they include individual fixed effects and the entire input history, Eqs. (1), (2) are difficult to estimate. For example, if inputs are treated as endogenous and an IV approach were used, it would be necessary to have at least one instrument for each period in the entire input history. Thus, instead of directly estimating these two equations, we make two further assumptions that allow less demanding specifications in terms of data and instrument requirements, while remaining more flexible than previous specifications in the literature.Assumption 1 Effects of past inputs follow a monotonic (likely decreasing) pattern at a constant rate γ for each period.5 That is: βt−j = γβt−1−j and δt−j = γδt−1−j. Assumption 2 The coefficients on inputs in the height function are the same as those in the weight function, up to a multiplicative constant δt−1−j = ((1 + σ)/α)βt−1−j. Together, these assumptions reduce the set of endogenous variables to a tractable number, thereby reducing the number of required instrumental variables. From Eq. (1) and taking first-differences in height we obtain: Δhi,t=β0xi,t+∑j=1t−1(βt−j−βt−1−j)xi,j+∈i,th−∈i,t−1h Incorporating the first assumption that βt−j = γβt−1−j, we obtain: Δhi,t=β0xi,t+(γ−1)∑j=1t−1βt−1−jxi,j+∈i,th−∈i,t−1h Next, consider the difference in Eqs. (1), (2) (after cross multiplication with σ and α): αwi,t−1−σhi,t−1=∑j=1t−1(αδt−1−j−σβt−1−j)xi,j+α∈i,t−1w−σ∈i,t−1h Under the second assumption that δt−1−j = ((1 + σ)/α)βt−1−j, we have: αwi,t−1−σhi,t−1+σ∈i,t−1h−α∈i,t−1w=∑j=1t−1βt−1−jxi,j Consequently, (3) Δhi,t=β0xi,t+α(γ−1)wi,t−1−σ(γ−1)hi,t−1+ωi,tΔh where ωi,tΔh=∈i,th+(σ(γ−1)−1)∈i,t−1h−α(γ−1)∈i,t−1w. Under these assumptions, height growth can be expressed as a function of current inputs, past height and weight, and an error involving current (t) and previous period (t − 1) shocks. Current inputs enter directly; the full history of past inputs enter indirectly through the lagged height and weight. We proceed in similar fashion for weight and obtain: (4) Δwi,t=δ0xi,t+(γ−1)(1+σ)wi,t−1−σ(γ−1)(1+σ)αhi,t−1+ωi,tΔw where ωi,tΔw=∈i,tw+σ(γ−1)(1+σ)α∈i,t−1h−[(γ−1)(1+σ)+1]∈i,t−1w. As with the change-in-height Eq. (3), the change-in-weight Eq. (4) depends on current inputs, past height and weight, and an error including current and previous period shocks.6 This framework forms the core of our approach to estimating production functions for height and weight. Estimation of Eqs. (3), (4) allow recovering β0 from Eq. (1) and δ0 from (2). 2.3 Estimation and identification Although differencing removes individual-level fixed effects and thus controls for important sources of potential bias (unobserved persistent heterogeneity including, e.g., genetic endowments and fixed parental and household characteristics), to consistently estimate the parameters in the relations for change in height (Eq. (3)) and change in weight (Eq. (4)), we still need to overcome several endogeneity problems. First, by construction previous height and weight are correlated with the error terms of Eqs. (3), (4) (see Eqs. (1), (2)). Moreover, if we assume that the household responds to past shocks as is likely and for which there is evidence for the Philippines (Akin et al., 1992, Liu et al., 2009), current inputs may be correlated with the error terms. We address potential endogeneity by using IV, which also addresses bias due to random measurement error in x under the assumption that the instruments are uncorrelated with that measurement error. The set of candidate instruments we use differs by country but draws on plausibly exogenous factors including a randomized intervention in Guatemala and prices of common foods in both countries. We treat market prices as exogenous to households (as in Liu et al. (2009)). Using prices as instruments for inputs is a well-established approach in the estimation of production functions (Todd and Wolpin (2003)). We also include past height and weight measures, hi,t−2 and wi,t−2 as instruments to help identify the effects of lagged height and weight. (Instruments are described in further detail in Section 3.3.) Using the available instruments, we endogenize protein and non-protein intakes, as well as lagged height and weight. However, we do not have access to instruments in both countries that also would allow us to control for the potential endogeneity of breastfeeding or diarrhea.7 Controlling for individual-level fixed effects is an important aspect of our approach, however, and goes part way toward addressing their potential endogeneity. For example, fixed effects control for the possibility that certain children have a pre-disposition for diarrhea, or live in particularly unsanitary households. However, if households change breastfeeding practices when health shocks affect their children's health or change sanitary conditions to reduce the diarrhea prevalence, the estimated effects of breastfeeding and diarrhea could be downward-biased. For instance, households that have increased breastfeeding could be compensating for negative health shocks, suggesting a negative relationship between growth and breastfeeding, while correcting for endogeneity could show a positive relationship (and similarly for diarrhea). Because our principal objective is to study the roles of proteins and non-proteins in the production functions, however, we do not emphasize the coefficients for diarrhea and breastfeeding but instead make clear the assumptions under which our primary coefficients of interest are consistently estimated even if breastfeeding or diarrhea are endogenous in the model. Our estimation approach is consistent provided the instruments are not correlated with the error term in the production function, conditional on breastfeeding and diarrhea as well as other covariates mentioned below. This is plausible for the same reason that the instruments are exogenous in relation to the energy inputs, e.g., that they are not correlated with individual-level time-varying health shocks.8 In principle, there also could be interactions among inputs in the production function, such as between nutrient intakes and diarrhea, or between breastfeeding and other nutrient intakes but a specification incorporating such interactions would be even more challenging to estimate, requiring additional instruments. Given that there are already four variables that we treat as endogenous in our main models (protein, non-protein, lagged height, and lagged weight), we do not estimate models with such potential interactions; instead, we studied possible interactions by splitting the sample. For instance, to examine whether diarrhea or breastfeeding interacts with diets, we estimated specifications for the sample that is breastfed and compare the results with the sample that is not breastfed. We carried out a similar exercise for diarrhea. Our results indicate that coefficients are not affected when we separate the sample by breastfeeding types. For diarrhea, there was some evidence of interaction effects, where diarrhea lowers the effects of macronutrients, but because most of the specifications suffer from problems of weak instruments, we are unable to draw strong conclusions. The estimation of the growth equations also includes an indicator for whether the child was female, number of days since the previous measurement, and age and age squared at time t. Our methods permit us to improve upon the previous literature that investigates the effects of total energy on anthropometrics. Since we do not have a single set of preferred instruments, we are able to robustly study effects of total energy on height and weight across two settings. We do this estimating the changes in height and weight, first using total energy intakes and then separating protein and energy from other macronutrient intakes to examine their relative partial effects in each model. The final estimating equations for the change in each anthropometric measure Ai,t that we estimate, adding the additional controls to Eqs. (3), (4), are: (5) ΔAi,t=λenergyAEi,t+ρ1Awi,t−1+ρ2Ahi,t−1+ρ3Adays_no_diari,t+ρ4Abfi,t+ρ5Aagei,t+ρ6Aagei,t2+ρ7Afemalei,t+ρ8Agap_msmti,t+ηi,tΔA and (6) ΔAi,t=λprotAProti,t+λnon_protANon_Proti,t+δ1Awi,t−1+δ2Ahi,t−1+δ3Adays_no_diari,t+δ4Abfi,t+δ5Aagei,t+δ6Aagei,t2+δ7Afemalei,t+δ8Agap_msmti,t+νi,tΔA where Ai,t is either weight (wi,t) or height (hi,t) of child i at age t; Ei,t, Proti,t, Non _ Proti,t correspond to the total energy intake, protein intake and non-protein intake; days_no_diari,t is the number of days without diarrhea between measurements; bfi,t is a dummy variable equal 1 if the child was breastfed during the period leading up to age t; agei,t and agei,t2 are age and age squared; femalei,t is a dummy variable equal to 1 if the child is a female; and gap_msmti,t is the number of days between measurements. Finally, the error terms in Eqs. (3), (4) exhibit serial correlation of order one by construction. We use cluster standard errors at the individual level to take into account this serial correlation, and also any possible correlation of individual error terms; using cluster standard errors is more general than a correction for serial correlation. Additionally the error terms are correlated between equations so there are possible efficiency gains of estimating a system of equations. Nonetheless given the already complex nature of the estimation, we estimate single equations. The cluster errors we calculate, therefore, can be seen as an upper bound of the standard errors. 3 Data Estimation of (5) and (6) requires high-frequency longitudinal data in early life that contain information on the outcomes (height9 and weight) and inputs (proteins and other macronutrients, breastfeeding, and diarrhea), as well as plausibly exogenous instruments. We now describe the data and contexts for two unique studies that fulfill these substantial requirements relatively well, one in Guatemala from the 1970s and the other in the Philippines from the 1980s. 3.1 Guatemala We use data from The Institute of Nutrition of Central America and Panama (INCAP) 1969–1977 nutritional supplementation trial. Four rural villages from eastern Guatemala were selected, one relatively large pair (∼900 residents) and one smaller pair (∼500 residents). At the outset, the villages were similar in terms of child nutritional status, measured as height at age three years, and were highly malnourished with over 50% of children severely stunted, i.e., with height-for-age z-score <−3. One large and one small village were randomly selected to receive a high-protein supplement (Atole); the others received an alternative supplement devoid of protein (Fresco). A 180 ml serving of Atole contained 11.5 grams of protein and 163 kcal. Fresco had no protein and a 180 ml serving had 59 kcal. The main hypothesis was that increased protein would accelerate mental development; additionally, it was expected that the high-protein nutritional supplement would affect physical growth. The nutritional supplements were distributed in centrally-located feeding centers in each village (Habicht et al., 1995). Virtually all (>98%) families participated (Martorell et al. (1995)). From 1969 to 1977, anthropometric measures (height and weight) were taken every three months for all children 24 months of age or under (including newborns entering the study) in the four villages. This yields a maximum usable sample for our analyses of 878 children measured at least twice by the age of 24 months. The amount of supplement intake was recorded daily in all villages. Home dietary information was collected every three months, including the types and amounts (except for breastmilk) of all foods and liquids consumed. These dietary histories were based on a 24-h recall period in the larger villages and a 72-h period in the smaller villages (from which we construct daily averages), and permit calculation of protein and non-protein intakes for the 24-h period by summing the nutritional content for each food item. The survey recorded the total months a child was breastfed. Nutrients from breastfeeding were not included in the nutritional intake calculations. Retrospective information on illness, specifically the length in days of episodes of diarrhea and fever, was collected semi-monthly. 3.2 The Philippines We use the Cebu Longitudinal Health and Nutritional Survey, a survey of Filipino children born between May 1983 and April 1984 in 33 rural and urban communities (barangays) in Metropolitan Cebu. The baseline survey included 3327 women sampled at a median of 30 weeks of gestation, and yielded a sample of 3080 singleton live births. This sample also exhibits high levels of undernutrition; at age 24 months, 62% of the children were stunted and 32% underweight. During the first two years of each child's life, data were collected every two months. This included anthropometric measurements, 24-h dietary recall of types and amounts (except breast milk) of all foods and liquids eaten, breastfeeding, and recent illness history. For breastfed children, the survey also collected the frequency and length of time spent breastfeeding. Total protein and energy intakes were calculated from foods consumed the previous day (24-h recall method). At each survey, mothers reported whether the child had diarrhea in the past 24 h, and if so, when the episode began, and the number of days the child had diarrhea during the previous week (Adair et al., 2011). The maximum usable sample of children between 6 and 24 months of age for the Philippines is 2713. 3.3 Variable construction Linear growth and weight gain are calculated as the difference between consecutive measurements. Although measurements were scheduled at specified intervals (every three months in Guatemala, every two in the Philippines), there were deviations including instances where a scheduled measurement did not occur. Because children experience high growth and growth spurts during the first two years of life, even differences of several days can be associated with significant differences in growth. We account for this by controlling for the exact number of days between measurements. Ideal data for this analysis would have information on protein and non-protein intakes over the entire period between measurements, but even in these uniquely comprehensive studies such detailed information is not available. Therefore, we approximate intakes over the entire period by using the average of the 24-h intakes calculated from the dietary recall information at the beginning and end of each period (which decreases measurement error relative to using only one point in time) multiplied by the exact number of days between measurements. For Guatemala, we add to this figure the intakes from the supplement (which were measured daily throughout the period) to obtain total protein and other intakes (as well as their sum, measured as total energy).10 For breastfeeding, we create a dummy indicator for whether the child was breastfed in the month previous to measurement at time t. While this does not fully exploit the detailed information available for the Philippines, it is done to have similar specifications across countries. The final input we include is diarrhea. For Guatemala, the protocol was to collect information every 15 days, so it is possible to construct the number of days experiencing diarrhea for the complete periods between anthropometric measurements.11 For the Philippines, it is only possible to construct the number of days with diarrhea during the week previous to each bimonthly anthropometric measurement. To extrapolate this to the full period between measurements, we estimate a count model for number of days with diarrhea for each two-month period with the Guatemalan data and use the estimated parameters from that model to predict number of days each Filipino child had diarrhea in each two-month period.12 As outlined in Section 2.3, in our main specifications we instrument for protein, other macronutrient intakes, and lagged height and weight. We now describe in detail the other instruments besides twice lagged height and weight. In both countries we use unit prices for various food items, selected with emphasis on foods with high protein content and/or important in the local diet. For Guatemala, prices are averages of national-level prices measured during December each year. We use lagged prices of eggs, chicken, pork, beef, dry beans, corn, and rice. Unit price variables for Guatemala are deflated and measured over the eight-year study period. For the Philippines, we use community-specific prices collected as part of the broader study. Between January 1983 and May 1986, enumerators visited two stores in each community, every other month, and collected prices (and quantity units) for a list of items. Not all items, however, were sold at each store at each visit. Consequently, there is not a complete set of prices for each item from each store (or even from each community in instances where no price was available from either store) in each measurement period. We selected as instruments the prices of dried fish, eggs, corn and tomatoes since these are the ones with the highest frequency in the sample.13 We use both current and lagged prices of those selected food items. By estimating a large set of instrument combinations, our approach does not depend on any one particular price, avoiding subjective instrument selection. For Guatemala, we also exploit the experimental variation resulting from the randomized allocation. We use a dummy variable that indicates whether the village had a feeding center that provided the high-protein supplement. We also interact this indicator with the distance from the home of the child to that feeding center. While the presence of a randomized allocation of a high-protein supplement provides an important source of exogenous variation, since there are four endogenous variables, additional instrumental variables also are used, i.e., twice lagged anthropometrics and food prices. For the Philippines we rely on price variation, which, unlike the annual Guatemalan food price data, varies both within-years and spatially, with information on these food items for the majority of measurement periods and each of the 33 communities. 3.4 Descriptive statistics Over the period from ages 6 to 24 months, each Guatemalan child is observed an average of 4.3 times and each Filipino child 9.1 times. The sample we describe includes all observations (measurements of children at different ages) with complete information for the following variables: change in height between consecutive measurement periods (linear growth), change in weight between consecutive periods (weight gain), total energy, energy from protein, energy from non-protein, breastfeeding indicator, and days with diarrhea.14 The final number of observations used in each specification varies depending on the availability of the instrumental variables used in that specification, since instruments for some observations are missing. Table 1 compares the main variables for both samples. On average and at all ages, the Filipino children in the early 1980s were taller than the Guatemalan children in the 1970s. For example, at 12 months of age, Filipino children were on average 70.7 cm tall, while their Guatemalan counterparts were 1.8 cm shorter. In terms of average weight, however, there were no significant differences between countries—at 24 months, children from both countries averaged 9.8 kg. 44% of the Guatemalan children were stunted, and 27% underweight. The corresponding levels were lower, 25% and 11%, for Filipino children. In 2011 for low- and middle-income countries, average levels of stunting were 28% and of underweight 17%, and 36% and 18% in Africa (Black et al., 2013). With broadly similar levels of stunting and underweight, thus, our historical samples remain relevant to understanding undernutrition in many countries and regions. Table 2 shows that Guatemalan children appear more likely to have been breastfed at all ages. In both countries, breastfeeding declines with age. At six months, 99% of Guatemalan children were breastfed, while at 24 months only 18% were; the proportions were 76% and 14% for Filipino children. Patterns between diarrhea and age are less clear. In Guatemala, average number of days with diarrhea (per 3-month measurement period) increases with age to 15 months, after which it declines. Levels are relatively lower in the Philippines, fluctuating between about 2 and 6 days (per 2-month period), with no clear age pattern. For Guatemala, information is complete on all of the instruments except the distance to the feeding center, which is missing for ∼5% of observations. For the Philippines, on the other hand, incomplete price availability leads to larger reductions in the sample size. The potential sample has 24,820 child-age observations; the lagged price of corn, which is the most complete, has 18,710 observations and the lagged price of tomatoes, the least complete, has 16,084 observations. 4 Results 4.1 Overview We estimate height and production functions for children 6–24 months, the period widely considered to be a critical window for post-birth nutritional investment.15 We use Generalized Method of Moments (GMM) for exactly-identified models and Limited Information Maximum Likelihood (LIML) for over-identified models because the latter allows for smaller finite-sample bias (Stock and Yogo, 2005). As noted, we cluster error terms at the individual level to take into account correlation of individual error terms and serial correlation (Baum et al., 2007).16 We first estimate height and weight production functions using only total energy (i.e., the sum of calories from protein and other sources), then we analyze separately the roles of proteins and non-proteins. In all specifications, the energy intakes, lagged height, and lagged weight are treated as endogenous, and we control for breastfeeding, number of days without diarrhea since the previous measurement, child sex, number of days since the previous measurement, and age and age squared. Because there are many potential instrument combinations, to establish general results that do not depend on one specific instrument combination, we estimated large subsets of all possible combinations. For Guatemala we first restricted the instrument sets to combinations that always had the Atole experiment indicator. Then, we systematically varied inclusion of distance interactions with Atole indicator, second lags of height, second lags of weight, and from two to four of the seven food prices (eggs, chicken, pork, beef, rice, beans and corn). For the Philippines, we systematically varied inclusion of second lags of height, second lags of weight, and from two to six of the eight (four current and four lagged) food prices (eggs, fish, tomatoes and corn). A summary of our instrument combinations is found in the Data Appendix Section 6. For Guatemala, there are 546 specifications (i.e., each with a different instrument set) for the version of the model with total energy (Eq. (5)) and 525 when proteins and non-proteins are included separately (Eq. (6)).17 The total number of specifications estimated for the Philippines is 602 for both models. For each specification, we calculate the robust versions of the Hansen-J (HJ) over-identification test, the Anderson–Rubin under-identification test (Anderson and Rubin, 1949), and the Wald F-statistic (robust Cragg–Donald or CD statistic) to detect weak instruments. Since our main models have four endogenous variables and we estimate them assuming heterokedasticity, it is not possible to compare CD statistics with critical values from Stock and Yogo (2005). The robust versions of these tests were developed in Kleibergen and Paap (2006). We also calculate for each endogenous variable Angrist and Pischke's (AP) partial F (Angrist and Pischke, 2009), which are informative about the presence of weak instruments. Finally, for all over-identified models we calculate the Hausman test of equality of OLS and IV estimates. We use the HJ over-identification and the CD statistics to focus our analysis on specifications with stronger and more exogenous instruments. In general, the Anderson–Rubin and Hausman tests strongly support our identification strategy. Based on the Anderson–Rubin test, we reject under-identification in all specifications for Guatemala, while for the Philippines we reject under-identification in 96% of the specifications. The Hausman test rejects equality of OLS and IV estimates in 99% of the specifications with total energy and 90% of the specifications with protein and non-protein separate in Guatemala and 87% and 98%, respectively, for the Philippines. Finally, we calculate the AP partial F statistic for the energy coefficient (λenergyh and λenergyw) from Eq. (5) and the protein (λproth and λprotw) and non-protein coefficients (λnon_proth and λnon_protw) from Eq. (6). These statistics are useful to make comparisons across equations and variables, but do not provide formal statistical support against weak instruments, since there are no critical values available for them. In general, the results suggest that the instruments are stronger for Guatemala: the AP partial F tends to be over 30 for the protein coefficients and over 7 for energy and non-protein coefficients. For the Philippines, the AP partial F for the total energy coefficient tends to be over 20. However, it is mostly below 5 for the protein and non-protein coefficients, which suggests that instruments are weaker in the more general specification for the Philippines.18 Despite these differences in AP statistics, results are broadly similar across countries, which suggests that we are identifying structural relationships between nutrients and anthropometrics. Since each production function is estimated multiple times, we explore distributions of estimated coefficients rather than a single or small set of “preferred” specifications, allowing us to draw more general conclusions. We do not choose or define a preferred specification because there are no obvious criteria for doing so and because of the concern that any potential preferred specification would not be robust to changes in the set of instruments. Although a priori the instruments we propose are plausibly exogenous and strong, we put relatively more confidence in those instrument sets that better satisfy over-identification and weak instrument tests. The results of each type of specification are presented in Table 3, Table 4, Table 5, Table 6 and Fig. 1, Fig. 2, Fig. 3. In Table 3, Table 5, and Fig. 1, we present the estimated overall energy coefficients. In Table 4, Table 6 (Panels A and B), and Fig. 2, we present the estimated protein coefficients, and in Table 4, Table 6 (Panels C and D), and Fig. 3, the estimated non-protein coefficients. Each table presents the 25th, 50th and 75th percentiles of the estimated coefficient distributions and, in the final two columns, the percentages of the coefficient estimates that are significantly (p < 0.05) positive or negative. For each Panel in each table, the first row reports distributions for all estimated specifications and, in subsequent rows, for specifications that are over-identified, and for those that have HJ P-values > 0.05 and CD statistics > 1, 3, or 7 (provided there are more than 10 such specifications in each case).19 These sets of specifications focus on results for which relatively strong and exogenous instruments are available. Fig. 1, Fig. 2, Fig. 3 present point estimates (and associated 95% confidence intervals) for all specifications that have HJ P-values > 0.05 and CD > 1 (corresponding to the third rows in Table 3, Table 4, Table 5, Table 6). The scale of the x-axis corresponds to the natural logarithm of CD statistics and the y-axis the coefficient values.20 To facilitate interpretation of the coefficient magnitudes, we simulate changes in height and weight when energy intakes increase ceteris paribus For this exercise, we use the most restrictive specifications with CD > 7 (or CD > 3 if there are fewer than ten specifications with CD > 7) and HJ P-values > 0.05. Within that set of specifications, we select the median coefficient and simulate effects of increasing energy intakes by 300 kcal per day, protein intakes by 10 g per day, or non-protein intakes by 250 kcal per day. Each of these is approximately one SD of respective intakes of 18-month old infants in both countries. This hypothetical daily increase is then multiplied by 90 in Guatemala and by 60 in the Philippines to approximate total intakes for a given measurement period, and then multiplied by corresponding coefficients to obtain anthropometric changes. We call this exercise median prediction. 4.2 Guatemala Table 3 summarizes for Guatemala distributions of coefficient estimates on total energy in the height and weight equations, and Fig. 1A and B show the coefficients and confidence intervals for the corresponding specifications with CD > 1. Total energy positively affects height and weight changes. These positive relationships are most evident for specifications with relatively stronger and more exogenous instruments. Our findings are consistent with previous literature that uses stronger identification assumptions estimating similar relationships from the same data sources (Habicht et al., 1995, Griffen, 2016). For height in Guatemala, estimated coefficients on total energy are positive in the vast majority of cases, positive and significant (p < 0.05) in 35% of cases, and never negative and significant. The positive relationship is more robust when we consider specifications with relatively stronger and more exogenous instruments, according to the tests. Restricting to over-identified specifications in which HJ P-values > 0.05 and CD > 3, total energy coefficient estimates are positive and significant 57% of the time. To provide further interpretation of the magnitude of the coefficients, we calculate the median prediction (Section 4.1), taking the median coefficient of the specifications with CD > 3; we calculate the effect of increasing energy per day by 300 kcal. For Guatemala, this implies a 0.62 cm predicted change in height. For weight production functions, estimated coefficients on total energy are positive and significant for 36% of specifications, and are never significantly negative. Specifications with higher CD statistics have larger proportions of positive significant coefficient estimates. Fig. 1B shows that while there are fewer specifications with higher CD statistic levels compared to the height model, for those with stronger instruments, the estimates are generally positive. The median prediction exercise indicates increasing energy intake by 300 kcal per day yields a predicted 620 g change in weight. Next, we consider the roles of protein and non-protein energy separately in the growth model. Proteins robustly and positively affect growth in height and weight in Guatemala, but the relationship of non-proteins (after controlling for protein) with these anthropometric measures is non-positive. Panel A of Table 4 (and Fig. 2A) shows that for 53% of all specifications, protein coefficient estimates are positive and significant. In specifications with CD > 3, the estimates are always positive and significant. In specifications with stronger instruments, the estimated coefficient dispersion (i.e., the distance between the 25th and 75th percentiles) decreases; for specifications with CD > 1 the ratio of the coefficients in the 75th and 25th percentiles is 1.3, while for the specifications with CD > 3 the ratio is 1.06. Our median prediction exercise indicates that if protein were to increase by 10 g per day, the predicted change in height is 0.39 cm. For weight change (Panel B of Table 4 and Fig. 2B), we find an even more robust pattern for proteins. In nearly all specifications (92%), protein coefficient estimates are positive and significant, and for specifications with CD > 1, they are always positive and significant. For all specifications, the estimate at the 75th percentile is only 1.2 times larger than that at the 25th percentile. This pattern of stability and significance of coefficient estimates also can be seen in Fig. 2B where the dispersion of the estimated coefficients is small, and there is a clear pattern of positive and significant effects of protein intake on weight growth. An increment in protein intake of 10 g per day results in a predicted 195 g change in weight. By contrast, there is little evidence that energy from non-proteins affects changes in height and weight. Panel C in Table 4 and Fig. 3A show that for Guatemala, in nearly all cases (98%) the estimated coefficient is insignificant in the height model. For the weight production function (Panel D of Table 4 and Fig. 3B), the point estimates are never significant. 4.3 Philippines Table 5 shows the distribution of the total energy coefficient estimates for the Philippines and Fig. 1A and B the corresponding coefficients and confidence intervals for specifications with CD > 1. As in Guatemala, positive relations are most evident for specifications with relatively stronger and more exogenous instruments. The positive impacts of total energy on height and weight are consistent with those found under somewhat stronger identification assumptions and using the same data, by Liu et al. (2009) and de Cao (2015). Across all specifications summarized in the Panel A of Table 5, 13% have positive and significant coefficient estimates (p < 0.05), while none have negative and statistically significant estimates. Restricting results to the 45 specifications with HJ test P-values > 0.05 and CD > 7, 64% of estimated total energy coefficients are positive and significant. Specifications with higher CD statistics tend to have more concentrated coefficient estimate distributions. If daily energy intake increases by 300 kcal the predicted change in height is 0.18 cm. For weight, evidence is similar regarding the role of total energy. The bottom panel of Table 5 indicates that for 15% of all the specifications in the Philippines, the estimated coefficient on total energy is positive and significant and never negative and significant. Specifications with the highest CD statistics tend to have larger shares of positive and significant coefficient estimates. Our median prediction results in a predicted change in weight of 37 g. Panel A of Table 6 (and Fig. 2A) shows that for 39% of all specifications, protein coefficient estimates are positive and significant. While there are fewer specifications with strong instruments than in Guatemala, for specifications with CD > 3, 100% of the coefficient estimates are positive and significant. In specifications with stronger instruments, the estimated coefficients dispersion decreases. Increasing protein consumption by 10 g per day is predicted to result in a 2.24 cm change in height. For all specifications (Panel B of Table 6 and Fig. 2B), 48% of estimated coefficients on protein for weight are positive and significant – 100% in specifications with CD > 3. Similar to Guatemala, coefficient estimate dispersion decreases with stronger instruments. Increasing protein consumption by 10 g per day results in a predicted 703 g change in weight. Somewhat surprisingly, non-protein intakes are generally negatively related to both height and weight gain. For height, Panel C of Table 6 reports that 88% of the specifications with the strongest instruments (CD > 3) yield negative and significant estimated coefficients. For weight, 100% of estimates in specifications with the strongest instruments are negative and significant. These findings for non-protein energy for the Philippines are somewhat counter-intuitive, because they suggest that such energy intakes are detrimental to growth. Most individual foods (including those consumed in these regions during the study periods), however, include both proteins and non-proteins and virtually all diets do. Consequently, it is unlikely that actual intakes would change in a fashion that increased energy from non-proteins while simultaneously holding proteins constant. Since Filipino children's diets included both intakes, on net any negative effects of other macronutrient sources would have been partly or fully offset by protein effects. For example, not including breastmilk, at age 6 months, 93% of children had some protein consumption and from ages 14 to 24 months, all did. Moreover, at age 6 months 75% of children are breastfed, which also provides protein intakes. In Section 4.5, we show that the model predicts that a dietary change (relatively rich in proteins but with some energy from other sources) indeed has positive effects on height and weight, despite negative coefficient estimates on non-proteins. There are several potential explanations for the finding that non-proteins are less robustly related to anthropometrics than proteins. First, it is possible that energy from macronutrients other than proteins do not affect height and weight, at least aggregating the other macronutrients as we do. Second, it may be that non-linearities are not captured. For instance, it could happen that carbohydrates and fat need some proteins to have an effect on anthropometrics—if protein intakes are zero or very low, other intakes would not affect height and weight. Third, dietary changes after children stop breastfeeding can result in poorer quality diets, especially poor quality of carbohydrates and low micronutrient density, weakening any potential link to anthropometrics. Fourth, the available instruments simply may not be powerful enough to detect effects of other macronutrients; protein and non-protein intakes are highly correlated (even before instrumentation), making it difficult econometrically to identify their distinct effects; in that sense, Guatemala greatly benefits from the experimental Atole intervention, which provides a clear and strong exogenous variation for protein, though it is less powerful for other macronutrients. 4.4 Effects of other inputs and controls In addition to the different nutrition intakes, our analysis provides estimates of the coefficients on lagged height, lagged weight, breastfeeding, and diarrhea. The results clearly indicate some catch-up height and weight growth. The lagged height coefficient is consistently negative and mostly significant in the change-in-height equation, indicating that shorter children at the end of one period tend to grow more in the next period. Similarly, the lagged weight coefficient is consistently negative and mostly significant in the weight equation so that lighter children at the end of one period gain more weight in the following period. With the caveat that the estimates for breastfeeding and diarrhea are potentially biased due to endogeneity, our coefficient estimates for number of days without diarrhea are consistently positive and significant for weight in both samples, suggesting that diarrhea has detrimental effects on weight gain as generally found in the literature. The coefficient estimates for breastfeeding are positive and mostly significant for Guatemala. In the Philippines, the coefficient estimates generally show a positive association between breastfeeding and height while the associations between breastfeeding and weight show no consistent pattern, similar to findings from Adair and Popkin (1996).21 4.5 Counterfactual exercise: increasing nutritional intakes We next simulate the full effects of additional protein and non-protein intakes on child height and weight for the Philippines, complementing the simpler median predictions we used when interpreting individual coefficients. From the set of specifications with HJ P-values > 0.05, we select the specification with the highest CD. The simulation is based on adding one egg per week to a child's diet, assuming no other changes in diet and no change in diarrhea. Eggs are good for such simulations. They were widely available in the localities where these studies are situated and are easily consumed by infants. They not only contain highly bioavailable protein, but also contain energy from other macronutrients, similar to many other naturally protein-rich foods. A medium (44 g), whole raw egg contains on average 5.5 g of protein and 40.9 calories from non-protein.22,23 Based on our parameter estimates, a child who consumed an additional egg per week on top of existing diet, for 18 months – from 6 to 24 months of age – would gain an additional 0.72 cm in height and 265 grams in weight. 5 Conclusions Arimond and Ruel (2004) described associations between children's dietary diversity and their height. We build on their insights, examining effects of diet and particularly diet composition on height and weight growth for children between ages 6 and 24 months, giving special attention to differences between diets rich and poor in proteins. We improve upon previous literature by making weaker identifying assumptions, considering two important anthropometric measures—height and weight, investigating the robustness of our results to the use of a number of different instruments, and separately investigating the effects of energy from proteins and from non-proteins while controlling for breastfeeding and diarrhea. We take advantage of two rich databases, one for Guatemala and the other for the Philippines, which have longitudinal information on height, weight, and protein and energy intakes with high frequencies of observations. IV estimation strategies are used to overcome endogeneity and measurement error problems, using food prices and, in the case of Guatemala, a randomized nutritional intervention, as instruments. Because there are many instruments and instrument combinations available, we present results that comprehensively summarize these combinations rather than selecting only a single set of instruments. Our findings indicate that increasing energy intake increases both height and weight in both countries. But the source of that energy, protein versus non-protein, matters. In these poor populations characterized by high levels of chronic undernutrition, increases in protein intake drive increases in child height and weight. These results provide evidence on an important puzzle in the literature while pointing to possible modifications to interventions designed to improve children's nutritional status. A systematic review by Manley et al. (2013) using meta-analysis techniques shows that while the average impact of income transfers from social protection programs on height-for-age is positive, effect sizes are small and not statistically significant. If households use these transfers largely to increase the quantity of calories consumed, if the increases in protein consumption is small in magnitude, or if these proteins are not allocated to children, then our results suggest that such transfers will have little impact on child height—precisely what Manley et al. (2013) find. Headey and Hoddinott (2015) examine impacts of Green Revolution-induced increases in rice productivity on children's anthropometric status. They find no impact of these on child height, results also consistent with what we observe here. Our findings, in conjunction with these other studies, suggest that interventions designed to increase household incomes may only improve children's nutritional status when they are linked to mechanisms that also improve the quality of children's diets. Such interventions, e.g., linking nutritional behavior change communication to social protection interventions or “nutrition-sensitive agriculture” await further study. Funding The authors thank Grand Challenges Canada (Grant 0072-03), Bill and Melinda Gates Foundation (Global Health Grant OPP1032713), and the Eunice Shriver Kennedy National Institute of Child Health and Development (Grant R01 HD070993) for Financial Support. The funders have no involvement in the analysis and interpretation of the data, writing of the paper, or the decision to submit the paper for publication. Conflict of interest There are no conflicts of interest. Appendix A Supplementary data The following are the supplementary data to this article: Acknowledgements This version of the paper has benefited with comments made by two anonymous referees and participants of the seminars at LACEA, PAA and University of Pennsylvania. 2 Micronutrients also play important roles in tissue building (WHO, 2007), but there is limited information about them in our data. Hence our focus on protein and non-protein energy. 3 While other individual macronutrients may have different relationships with growth (WHO, 2007), separating them into their components while still treating them as endogenous was empirically infeasible. 4 Relatedly, and using the same data from the Philippines that we use, Bhargava (2016) studies the association of macronutrients (proteins) and micronutrients (calcium) with anthropometrics, finding that both, protein and calcium are strongly associated with height and weight in the first 24 months of life and also on adolescence. However, Bhargava (2016) only controls for individual effects, assuming several time varying variables as exogenous. 5 While it seems most likely that nutritional inputs would have a larger impact during the 6–24 month age window we model, assuming it is decreasing is not strictly necessary. The rate can be different for the height and weight equations; we assume that is similar only for illustration purposes. 6 Specifications of the change-in-height equation that exclude lagged weight, and the change-in-weight equation that exclude lagged height were also estimated. Results were similar to the more general specification (available on request). 7 Previous work using the Philippine data has used rainfall as an instrument for diarrhea (Akin et al., 1992). We attempted to endogenize diarrhea using spatial and temporal variation in rainfall and temperature as instruments in Guatemala, but they had minimal predictive power. To keep the structure parallel across the countries, we do not use rainfall to endogenize diarrhea in either country. 8 For instance, if some other disease is important in the production function, and we are not including it, our results hold if the instrumental variables are orthogonal to this other disease. 9 In both settings, children under 24 months were measured lying down, per standard anthropometric measurement practice. This measurement is sometimes referred to as length, rather than height. 10 For Guatemala we use an individual-level fixed-effects model to impute nutrient intakes for approximately 5% of missing observations. See Data Appendix Section 1. 11 Approximately 45% of such 15-day visits were missed. In those instances, we assume the child had similar diarrhea patterns across all 15-day intervals during that growth period and scale-up the observed number of days accordingly. 12 See Data Appendix Section 2 for details of the estimation of the count model for diarrhea. 13 See Data Appendix Section 3 for further details on prices. 14 For the Philippines, the number of available observations is constant across variables, but decreases with child age due to attrition. For Guatemala, the number of children with available information on intakes and diarrhea is smaller than the number with anthropometric measures because the dietary and morbidity information for infants under 12 months was not collected until 1973. 15 There are additional substantive, as well as practical, reasons for the 6–24 month window. First, during the first six months most infants are breastfed; indeed WHO recommends exclusive breastfeeding from birth to age six months. Therefore, before that age proteins and non-proteins in the diet reflect non-exclusive breastfeeding that could be detrimental to growth. Second, it is not possible to study the production function at earlier ages because our final specification models growth and the candidate instrumental variables include second lags of height and weight (Section 2.2). Because we model growth and use these second lags, however, the analysis does incorporate information on individuals prior to six months of age. Third, while the frequency of measurements differs, both samples have measurements at ages six and 24 months, facilitating comparability. 16 The specifications also include predicted days of diarrhea. We do not explicitly account this in calculating the standard errors, instead relying on the general correction provided by clustered standard error calculations. 17 The reduction in specifications arises because 21 specifications that include both protein and non-protein are exactly-identified with three instruments. 18 Results available on request. 19 Restricting the sample to those with HJ p-values > 0.10 generates similar results; see Data Appendix Section 5. 20 The number of observations used varies for each specification. In the Data Appendix Section 4, we show that the results do not depend on the number of observations used. 21 All results available on request. 22 Agricultural Research Service of the United States Department of Agriculture. http://ndb.nal.usda.gov/ndb/foods/show/112 accessed on 17th September 2014. 23 If households were to purchase the eggs, the cost would have been ∼0.37% of the annual average income. Appendix A Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.ehb.2016.03.002. Fig. 1 Total energy coefficients. (A) Change in height: total energy coefficients. (B) Change in weight: total energy coefficients. Fig. 2 Protein coefficients. (A) Change in height: protein coefficients. (B) Change in weight: protein coefficients. Fig. 3 Non-protein coefficients. (A) Change in height: non-protein coefficients. (B) Change in weight: non-protein coefficients. Table 1 Guatemala, nutritional outcomes and inputs. Table 1 Height (cm) Change in height Weight (grams) Change in weight Total energy (kcal) Non-protein (kcal) Protein (grams) Mean (sd) Mean (sd) Mean (sd) Mean (sd) Mean (sd) Mean (sd) Mean (sd) 6 months 62.97 5.19 6871.99 1424.26 131.86 113.82 4.51 (2.38) (1.34) (959.14) (470.80) (149.30) (131.73) (5.45) 9 months 66.21 3.46 7516.29 698.85 218.16 191.06 6.77 (2.69) (1.55) (1085.65) (469.85) (193.81) (170.68) (6.85) 12 months 68.91 2.96 7979.84 500.85 340.90 301.12 9.95 (3.00) (1.47) (1147.19) (463.02) (232.77) (206.26) (7.78) 15 months 71.01 2.40 8292.93 461.96 511.06 451.65 14.85 (3.21) (1.35) (1117.15) (432.51) (245.47) (218.69) (8.39) 18 months 73.25 2.29 8712.95 431.83 656.70 581.27 18.86 (3.36) (1.41) (1118.61) (495.17) (271.83) (241.61) (9.29) 21 months 75.47 2.33 9186.83 505.42 767.85 678.13 22.43 (3.47) (1.39) (1129.93) (481.52) (293.42) (261.17) (10.08) 24 months 77.53 2.23 9752.69 604.67 847.75 747.65 25.03 (3.55) (1.44) (1168.04) (523.06) (303.51) (271.84) (10.37) Observations 3802 3802 3802 3802 3802 3802 3802 Philippines, nutritional outcomes and inputs 6 months 64.27 3.26 6856.72 736.13 204.93 182.57 5.59 (2.57) (1.66) (903.05) (415.00) (249.68) (221.73) (7.52) 8 months 66.80 2.54 7302.63 440.11 285.67 254.65 7.76 (2.71) (1.42) (964.47) (383.29) (279.54) (246.70) (8.99) 10 months 68.92 2.13 7642.79 338.95 349.93 312.18 9.44 (2.80) (1.39) (1028.15) (402.86) (300.60) (264.36) (10.12) 12 months 70.72 1.82 7948.05 300.68 407.33 362.27 11.27 (2.96) (1.29) (1079.27) (391.39) (310.79) (273.21) (10.56) 14 months 72.29 1.58 8225.85 278.16 477.29 423.20 13.52 (3.07) (1.22) (1115.39) (377.09) (325.60) (284.74) (11.56) 16 months 73.73 1.45 8512.16 283.81 540.50 479.13 15.34 (3.24) (1.19) (1111.83) (389.74) (328.22) (285.88) (12.22) 18 months 75.12 1.43 8797.30 286.79 589.04 521.87 16.79 (3.38) (1.23) (1143.54) (392.57) (334.33) (291.48) (12.43) 20 months 76.50 1.42 9104.64 316.95 640.35 567.20 18.29 (3.51) (1.29) (1177.77) (397.24) (347.83) (303.38) (12.74) 22 months 77.73 1.30 9436.95 338.55 681.66 603.35 19.58 (3.61) (1.33) (1210.32) (413.78) (355.37) (310.89) (12.75) 24 months 79.13 1.43 9782.39 349.09 710.41 627.28 20.78 (3.68) (1.19) (1233.11) (418.55) (354.38) (309.68) (12.99) Observations 24,820 24,820 24,820 24,820 24,820 24,820 24,820 Table 2 Guatemala, other inputs. Table 2 Breastfed Days with diarrhea Female Time between measurement (days) Age (days) Mean (sd) Mean (sd) Mean (sd) Mean (sd) Mean (sd) 6 months 0.99 6.52 0.51 91.95 182.62 (0.12) (12.44) (0.50) (5.09) (3.66) 9 months 0.97 9.43 0.51 95.15 273.37 (0.17) (15.80) (0.50) (20.52) (4.17) 12 months 0.92 12.18 0.50 96.82 364.59 (0.28) (16.08) (0.50) (24.77) (4.88) 15 months 0.81 12.60 0.54 94.26 456.95 (0.40) (15.63) (0.50) (16.29) (4.08) 18 months 0.59 11.43 0.53 94.48 547.98 (0.49) (15.77) (0.50) (18.90) (3.51) 21 months 0.34 9.97 0.53 95.75 638.72 (0.47) (14.90) (0.50) (26.05) (3.29) 24 months 0.18 7.57 0.53 100.14 730.64 (0.38) (13.65) (0.50) (34.32) (3.21) Observations 3802 3802 3802 3802 3802 Philippines, other inputs 6 months 0.76 1.54 0.53 61.77 186.41 (0.43) (2.76) (0.50) (8.92) (6.03) 8 months 0.72 4.38 0.53 60.23 246.59 (0.45) (4.23) (0.50) (5.84) (5.57) 10 months 0.68 3.11 0.53 62.03 307.98 (0.47) (2.35) (0.50) (8.61) (6.03) 12 months 0.62 2.38 0.53 61.72 369.10 (0.49) (2.37) (0.50) (8.93) (6.36) 14 months 0.53 5.25 0.53 61.48 430.07 (0.50) (5.15) (0.50) (8.55) (6.46) 16 months 0.44 4.98 0.53 61.36 490.90 (0.50) (5.06) (0.50) (8.92) (6.47) 18 months 0.34 1.99 0.53 61.54 551.72 (0.47) (2.35) (0.50) (9.56) (6.16) 20 months 0.26 6.22 0.53 61.45 612.72 (0.44) (6.25) (0.50) (8.86) (6.48) 22 months 0.19 2.78 0.53 60.83 673.14 (0.39) (3.51) (0.50) (8.55) (6.11) 24 months 0.14 1.83 0.53 61.59 734.06 (0.34) (2.73) (0.50) (9.03) (6.33) Observations 24,820 24,820 24,820 24,820 24,820 Table 3 Impact of total energy intake on change in heights and weights, Guatemala. Table 3 Total energy Distribution of total energy coefficient sig > 0 sig < 0 # of sp. p25 p50 p75 %-Sig %-Sig Panel A: Height All IV 546 −0.0182 0.0099 0.0288 35 0 All over-identified IV 525 −0.0179 0.0107 0.0289 36 0 CD > 1 P-val HJ > 5 137 −0.0090 0.0034 0.0170 15 0 CD > 3 P-val HJ > 5 21 0.0009 0.0231 0.0438 57 0 Panel B: Weight All IV 546 −0.0061 0.0059 0.0159 36 0 All over-identified IV 525 −0.0036 0.0060 0.0159 38 0 CD > 1 P-val HJ > 5 129 −0.0024 0.0050 0.0233 32 0 CD > 3 P-val HJ > 5 36 0.0142 0.0230 0.0239 83 0 CD = Robust Kleibergen-Paap F statistic, P-value, J = P-value of Hansen J stat × 100. 1st column: # of specifications that meet criteria; 2nd–4th col: percentile of distribution of estimated coefficients. 5th (6th) column: percent of estimated coefficients that are positive (negative) and significant at 5% significance level. 1st row: all specifications; 2nd row: all over-identified specifications for which # of IVs># of endogenous variables. Other rows include all specifications satisfying the indicated criteria based on the CD and HJ tests. All specifications include breastfeeding, diarrhea, sex, age, and age squared as covariates and a seasonal dummy for the Philippines, and lagged height and lagged weight, both of which are treated as endogenous. Height coefficients are divided by 1000 for presentation purposes. Table 4 Impact of protein and non-protein energy on change in heights and weights, Guatemala. Table 4 Protein Distribution of protein coefficient sig>0 sig<0 # of esp. p25 p50 p75 %-Sig %-Sig Panel A: Height (protein) All IV 525 0.0666 0.1047 0.1293 53 0 All over-identified IV 448 0.0774 0.1044 0.1268 58 0 CD > 1 P-val HJ > 5 163 0.0931 0.1067 0.1232 77 0 CD > 3 P-val HJ > 5 48 0.1043 0.1079 0.1106 100 0 Panel B: Weight (protein) All IV 525 0.0541 0.0588 0.0632 92 0 All over-identified IV 448 0.0543 0.0586 0.0627 97 0 CD > 1 P-val HJ > 5 347 0.0540 0.0571 0.0614 100 0 CD > 3 P-val HJ > 5 132 0.0534 0.0542 0.0567 100 0 Non-protein Distribution of non-protein coefficient sig>0 sig<0 # of esp. p25 p50 p75 %-Sig %-Sig Panel C: Height (non-protein) All IV 525 −0.0170 −0.0045 0.0039 0 2 All over-identified IV 448 −0.0161 −0.0042 0.0033 0 1 CD > 1 P-val HJ > 5 163 −0.0136 −0.0053 0.0018 0 3 CD > 3 P-val HJ > 5 48 −0.0059 −0.0028 0.0016 0 0 Panel D: Weight (non-protein) All IV 525 −0.0019 −0.0012 −0.0005 0 0 All over-identified IV 448 −0.0018 −0.0012 −0.0006 0 0 CD > 1 P-val HJ > 5 347 −0.0016 −0.0012 −0.0006 0 0 CD > 3 P-val HJ > 5 132 −0.0013 −0.0011 −0.0006 0 0 See Table 3 notes. Table 5 Impact of total energy intake on change in heights and weights, Philippines. Table 5 Total energy Distribution of total energy coefficient sig>0 sig<0 # of esp. p25 p50 p75 %-Sig %-Sig Panel A: Height (SeeFig. 1A) All IV 602 −0.0039 0.0069 0.0166 13 0 All over-identified IV 602 −0.0039 0.0069 0.0166 13 0 CD > 1 P-val HJ > 5 313 −0.0174 0.0067 0.0147 18 0 CD > 3 P-val HJ > 5 118 0.0035 0.0087 0.0140 37 0 CD > 7 P-val HJ > 5 45 0.0076 0.0098 0.0123 64 0 Panel B: Weight (SeeFig. 1B) All IV 602 0.0013 0.0044 0.0220 15 0 All over-identified IV 602 0.0013 0.0044 0.0220 15 0 CD > 1 P-val HJ > 5 284 0.0013 0.0058 0.0229 7 0 CD > 3 P-val HJ > 5 65 0.0024 0.0063 0.0152 15 0 CD > 7 P-val HJ > 5 15 0.0013 0.0020 0.0031 33 0 See Table 3 notes. Table 6 Impact of protein and non-protein energy on change in heights and weights, Philippines. Table 6 Protein Distribution of protein coefficient sig>0 sig<0 # of esp. p25 p50 p75 %-Sig %-Sig Panel A: Height (SeeFig. 2A) All IV 602 0.6826 1.0868 1.6848 39 0 All over-identified IV 448 0.7758 1.1194 1.7353 46 0 CD > 1 P-val HJ > 5 248 0.8633 1.1247 1.4188 77 0 CD > 3 P-val HJ > 5 16 0.6947 0.9324 1.0274 100 0 Panel B: Weight (SeeFig. 2B) All IV 602 0.2972 0.3887 0.4818 48 0 All Over-Identified IV 448 0.3145 0.3991 0.4813 56 0 CD > 1 P-val HJ > 5 242 0.3185 0.3766 0.4406 90 0 CD > 3 P-val HJ > 5 16 0.2631 0.2929 0.3110 100 0 Non-protein Distribution of non-protein coefficient sig>0 sig<0 # of esp. p25 p50 p75 %-Sig %-Sig Panel C: Height (SeeFig. 3A) All IV 602 −0.2792 −0.1739 −0.0943 0 32 All over-identified IV 448 −0.2795 −0.1789 −0.1110 0 39 CD > 1 P-val HJ > 5 248 −0.2362 −0.1777 −0.1269 0 67 CD > 3 P-val HJ > 5 16 −0.1564 −0.1283 −0.0912 0 88 Panel D: Weight (SeeFig. 3B) All IV 602 −0.0754 −0.0592 −0.0433 0 38 All over-identified IV 448 −0.0748 −0.0606 −0.0463 0 47 CD > 1 P-val HJ > 5 242 −0.0676 −0.0577 −0.0475 0 80 CD > 3 P-val HJ > 5 16 −0.0460 −0.0433 −0.0379 0 100 See Table 3 notes. ==== Refs References Angrist J. Pischke J.-S. Mostly Harmless Econometrics. An Empiricist Companion 2009 Princeton University Press Adair L. Popkin B. Low birth weight reduces the likelihood of breast-feeding among Filipino infants J. Nutr. 126 1 1996 103 112 8558289 Adair L. Popkin B. Akin J. Guilkey D. Gultiano S. Borja J. Perez L. Kuzawa C. McDade T. Hindin M. Cohort profile: the Cebu longitudinal health and nutrition survey Int. J. Epidemiol. 40 3 2011 619 625 20507864 Akin J. Guilkey D. Popkin B. Flieger W. Briscoe J. Black R.E. Adair Linda S. A child health production function estimated from longitudinal data J. Dev. Econ. 38 2 1992 323 351 12285368 Anderson T.W. Rubin H. Estimation of the parameters of a single equation in a complete system of stochastic equations Ann. Math. Stat. 20 1949 46 63 Arimond M. Ruel M. Dietary diversity is associated with child nutritional status: evidence from 11 demographic and health surveys J. Nutr. 134 10 2004 2579 2585 15465751 Bhargava A. Protein and micronutrient intakes are associated with child growth and morbidity from infancy to adulthood in the Philippines J. Nutr. 146 1 2016 133 141 26661837 Baum C. Schaffer M. Stillman S. Enhanced routines for instrumental variables/generalized method of moments estimation and testing Stata J. 7 4 2007 465 506 Baten J. Blum M. Why are you tall while others are short? Agricultural production and other proximate determinants of global heights Eur. Rev. Econ. Hist. 18 2 2014 144 165 Behrman J.R. Deolalikar A.B. Health and Nutrition. Handbook of Development Economics vol. 1 1988 631 711 Behrman J.R. Calderon M.C. Preston S.H. Hoddinott J.F. Martorell R. Stein A.D. Nutritional supplementation of girls influences the growth of their children: prospective study in Guatemala Am. J. Clin. Nutr. 90 5 2009 1372 1379 19793851 Black R.E. Allen L.H. Bhutta Z.A. Caulfield L.E. de Onis M. Ezzati M. Mathers C. Rivera J. for the Maternal Child Undernutrition Study Group Maternal and child undernutrition: global and regional exposures and health consequences Lancet 371 9608 2008 243 260 18207566 Black R.E. Victora C.G. Walker S.P. Butha Z.A. Crhistian P. de Onis M. Ezzati M. Grantham-McGregor S. Katz J. Martorell R. Uauy R. the Maternal Child Nutrition Study Group Maternal and child undernutrition and overweight in low-income and middle-income countries Lancet 382 9890 2013 427 451 23746772 Cunha F. Heckman J.J. The technology of skill formation Am. Econ. Rev. 97 2 2007 31 47 de Cao E. The height production function from birth to age two J. Hum. Cap. 2015 (forthcoming) Dewey K. Meeting protein needs at 6 to 25 months of age Food Nutr. Bull. 34 2 2013 The United Nations University Engle P.L. Black M.M. Behrman J.R. Cabral de Mello M. Gertler P.J. Kapiriri L. Martorell R. Young M.E. the International Child Development Steering G Strategies to avoid the loss of developmental potential in more than 200 million children in the developing world Lancet 369 9557 2007 229 242 17240290 Engle P.L. Fernald L.C.H. Alderman H. Behrman J.R. O’Gara C. Yousafzai A. de Mello M.C. Hidrobo M. Ulkuer N. Ertem I. Global Child Development Steering Group Strategies for reducing inequalities and improving developmental outcomes for young children in low and middle income countries Lancet 378 9799 2011 1339 1353 21944378 FAO Human Energy Requirements. Report of a Joint FAO/WHO/UNU Expert Consultation, Rome, 17–24 October 2001 2001 Food and Agricultural Organization Gibson R.S. Zinc: the missing link in combating micronutrient malnutrition in developing countries Proc. Nutr. Soc. 65 1 2006 51 60 16441944 Grantham-McGregor S. Cheung Y.B. Cueto S. Glewwe P. Ritcher L. Strupp B. Developmental potential in the first 5 years for children in developing countries Lancet 369 9555 2007 60 70 17208643 Griffen A.S. Height and calories in early childhood Econ. Hum. Biol. 20 2016 55 69 26656205 Habicht J.P. Martorell R. Rivera J.A. Nutritional impact of supplementation in the INCAP longitudinal study: analytic strategies and inferences J. Nutr. 125 1995 1042S 1050S 7722706 Headey D. Hoddinott J. The Nutritional Impacts of Bangladesh's Green Revolution 2015 Mimeo, International Food Policy Research Institute Washington, DC Hoddinott J. Behrman J.R. Maluccio J.A. Melgar P. Quisumbing A.R. Ramirez-Zea M. Stein A.D. Yount K.M. Martorell R. Adult consequences of growth failure in early childhood Am. J. Clin. Nutr. 98 5 2013 1170 1178 24004889 Hoddinott J.F. Maluccio J.A. Behrman J.R. Flores R. Martorell R. Effect of a nutrition intervention during early childhood on economic productivity in Guatemalan adults Lancet 371 9610 2008 411 416 18242415 Imdad A. Bhutta Z. Effect of preventive zinc supplementation on linear growth in children under 5 years of age in developing countries: a meta-analysis of studies for input to the lives saved tool BMC Public Health 11 Suppl. 3 2011 S22 21501440 Kleibergen F. Paap R. Generalized reduced rank tests using the singular value decomposition J. Econ. 127 1 2006 97 126 Koletzko B. von Kries R. Closa R. Escribano J. Scaglioni S. Giovannini M. Beyer J. Demmelmair H. Gruszfeld D. Dobrzanska A. Sengier A. Langhendries J.P. Rolland Cachera M.-F. Veit Grote for the European Childhood Obesity Trial Study Group Lower protein in infant formula is associated with lower weight up to age 2 y: a randomized clinical trial Am. J. Clin. Nutr. 89 2009 1836 1845 19386747 Liu H. Mroz T. Adair L. Parental compensatory behaviors and early child health outcomes in Cebu, Philippines J. Dev. Econ. 90 2 2009 209 230 22505790 Maluccio J.A. Hoddinott J.F. Behrman J.R. Quisumbing A.R. Martorell R. Stein A.D. The impact of nutrition during early childhood on education among Guatemalan adults Econ. J. 119 537 2009 734 763 Manley J. Gitter S. Slavchevska V. How effective are cash transfers at improving nutritional status? World Dev. 48 2013 133 155 Martorell R. Habicht J.P. Rivera J.A. History and design of the INCAP longitudinal study (1969–77) and its follow-up (1988–89) J. Nutr. 125 1995 1027S 1041S 7536830 Michaelsen K. Cow's milk in the prevention and treatment of stunting and wasting Food Nutr. Bull. 34 2 2013 The United Nations University Moradi A. Nutritional status and economic development in sub-Saharan Africa, 1950–1980 Econ. Hum. Biol. 8 2010 16 29 20071250 Mølgaard C. Larnkjær A. Arnberg K. Michaelsen K. Milk and growth in children: effects of whey and casein. Nestle Nutrition Workshop Series Pediatrics 67 2011 67 78 Pucilowska J. Davenport M. Kabir I. Clemmons D. Thissen J.P. Butler T. Underwood L. The effect of dietary protein supplementation on insulin-like growth factors (Igfs) and IGF-binding proteins in children with shigellosis J. Clin. Endocrinol. Metab. 77 6 1993 1516 1521 7505287 Stock J.H. Yogo M. Testing for weak instruments in linear IV regression Andrews D.W.K. Stock J.H. Identification and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg 2005 Cambridge University Press Cambridge 80 108 Todd P.E. Wolpin K.I. On the specification and estimation of the production function for cognitive achievement Econ. J. 113 2003 F3 F33 Todd P.E. Wolpin K.I. The production of cognitive achievement in children: home, school, and racial test score gaps J. Hum. Cap. 1 1 2007 91 136 Victora C.G. Adair L. Fall C. Hallal P.C. Martorell R. Rithcer L. Sachdev H.S. for the Maternal and Child Undernutrition Study Group Maternal and child undernutrition: consequences for adult health and human capital Lancet 371 9609 2008 340 357 18206223 Walker S.P. Wachs T.D. Grantham-McGregor S. Black M.M. Nelson C.A. Huffman S.L. Baker-Henningham H. Chang S.M. Hamadani J.D. Lozoff B. Meeks Gardner J.M. Powell C.A. Rahman A. Richter L. Inequality in early childhood: risk and protective factors for early child development Lancet 378 9799 2011 1325 1338 21944375 WHO Protein And Amino Acid Requirements In Human Nutrition. Report of a Joint WHO/FAO/UNU Expert Consultation. World Health Organization technical report series, No. 935 2007

PMC005xxxxxx/PMC5001507.txt

==== Front Expert Rev Mol MedExpert Rev Mol MedERMExpert Reviews in Molecular Medicine1462-3994Cambridge University Press Cambridge, UK 10.1017/erm.2016.13S146239941600013200013ReviewThe immunological landscape in necrotising enterocolitis THE IMMUNOLOGICAL LANDSCAPE IN NECROTISING ENTEROCOLITISTHE IMMUNOLOGICAL LANDSCAPE IN NECROTISING ENTEROCOLITISCho Steven X. 12Berger Philip J. 12Nold-Petry Claudia A. 12†Nold Marcel F. 12*†1 Ritchie Centre, Hudson Institute of Medical Research, Melbourne, Australia2 Department of Paediatrics, Monash University, Melbourne, Australia* Corresponding author: Marcel F. Nold, Hudson Institute of Medical Research, 27-31 Wright St., Melbourne, Victoria, Australia. E-mail: marcel.nold@monash.edu† These authors contributed equally to this work. 2016 24 6 2016 18 e12© Cambridge University Press 20162016Cambridge University PressThis is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution, and reproduction in any medium, provided the original work is properly cited.Necrotising enterocolitis (NEC) is an uncommon, but devastating intestinal inflammatory disease that predominantly affects preterm infants. NEC is sometimes dubbed the spectre of neonatal intensive care units, as its onset is insidiously non-specific, and once the disease manifests, the damage inflicted on the baby's intestine is already disastrous. Subsequent sepsis and multi-organ failure entail a mortality of up to 65%. Development of effective treatments for NEC has stagnated, largely because of our lack of understanding of NEC pathogenesis. It is clear, however, that NEC is driven by a profoundly dysregulated immune system. NEC is associated with local increases in pro-inflammatory mediators, e.g. Toll-like receptor (TLR) 4, nuclear factor-κB, tumour necrosis factor, platelet-activating factor (PAF), interleukin (IL)-18, interferon-gamma, IL-6, IL-8 and IL-1β. Deficiencies in counter-regulatory mechanisms, including IL-1 receptor antagonist (IL-1Ra), TLR9, PAF-acetylhydrolase, transforming growth factor beta (TGF-β)1&2, IL-10 and regulatory T cells likely facilitate a pro-inflammatory milieu in the NEC-afflicted intestine. There is insufficient evidence to conclude a predominance of an adaptive Th1-, Th2- or Th17-response in the disease. Our understanding of the accompanying regulation of systemic immunity remains poor; however, IL-1Ra, IL-6, IL-8 and TGF-β1 show promise as biomarkers. Here, we chart the emerging immunological landscape that underpins NEC by reviewing the involvement and potential clinical implications of innate and adaptive immune mediators and their regulation in NEC. ==== Body Introduction Necrotising enterocolitis (NEC) is a serious gastrointestinal disease that most commonly afflicts infants born prematurely. Although infrequent, NEC is a major cause of morbidity and mortality in neonatal intensive care units (NICUs). In older children, NEC occurs most commonly in association with cyanotic heart disease or major cardiac surgery (Ref. 1). NEC is a multifactorial disease whose pathogenesis remains poorly understood despite decades of research. However, risk factors for NEC have been identified, namely prematurity, formula feeding, hypoxic–ischaemic injury and abnormal bacterial colonisation. Yet, no single risk factor is essential, and the mechanisms by which each precipitates NEC are largely unknown. Nonetheless, evidence is mounting that formula feeding, hypoxia–ischemia, and dysbiosis lead to inflammation, and that immaturity of the immune system in preterm babies – although itself poorly characterised – is one of the pivotal pathogenic factors in NEC. Here, we review current knowledge on inflammation and immunity in NEC and highlight frontiers emerging in this field. Epidemiology, staging criteria and disease outcomes Death of extremely premature infants from most causes has decreased across the period from 2000 to 2011, whereas the incidence of death from NEC has increased (Ref. 2). Thus, NEC is now the most common cause of death between days 15 and 60 (Ref. 2). The overall incidence of NEC is 1–3 per 1000 live births (Ref. 3), but reaches 11% in very low birth weight infants (VLBW, <1500 g) (Ref. 4). NEC-associated mortality has changed little over the past 50 years, ranging from 20 to 30% in confirmed cases (Ref. 5). Approximately 20–50% of NEC infants require surgery; mortality then rises to about 65% (Refs 4, 6, 7). Treatment options for NEC infants are limited to bowel rest, antibiotics and supportive therapy, e.g. blood pressure management (Ref. 8). Decisions on such treatment or escalation to surgery are aided by Bell's staging criteria (Refs 9, 10) (Fig. 1). The clinical presentation of stage I NEC is largely non-specific, which explains why diagnosing NEC early is difficult. It is for this reason, and because NEC often manifests rapidly and quickly wreaks intestinal and systemic havoc that many neonatologists perceive NEC as an ever-looming spectre in NICUs. Figure 1. Modified Bell's staging criteria for necrotising enterocolitis, adapted from (Ref. 10). Short-term consequences of NEC include severe multisystem morbidity, leading to extended hospitalisation with all its financial and social burdens (Ref. 11). The cost of surgically managed NEC is enormous at approximately US$200,000 per survivor in excess of the per-baby cost of routine neonatal intensive care (Refs 11, 12). In childhood, prior history of NEC is an independent risk factor for bowel-related chronic conditions such as diarrhoea and constipation (Ref. 13). Similarly, neurodevelopmental issues often persist into later life and may include epilepsy, attention deficit hyperactivity disorder, cerebral palsy, deafness, blindness and compromised mental and psychomotor functions (Refs 13, 14, 15). Half of all surgically managed NEC infants develop some degree of short-bowel syndrome/intestinal failure (Ref. 16), and poor growth is common, particularly in extremely low birth weight (ELBW, <1000 g) NEC infants (Ref. 15). NEC pathogenesis and risk factors Prematurity NEC incidence and severity are most strongly associated with prematurity, quantified either as low gestational age (GA) or low weight at birth (Refs 17, 18, 19). Briefly, NEC may arise on the basis of the interactions between two poorly developed systems, namely the intestine and the immune system (Refs 20, 21, 22) (Fig. 2). Immaturity of intestinal motility and mucosal/barrier functions facilitates a potentially harmful composition of the microbiome and bacterial translocation (Fig. 2a). Thus confronted with bacteria, the premature immune system responds by unleashing a violent inflammatory storm (Fig. 2e) that overwhelms the extant endogenous counter-regulatory mechanisms (Fig. 2f), leading to cell death and subsequent release of intracellular components such as stored cytokines termed alarmins (Fig. 2i) (Ref. 23), thus perpetuating the inflammatory storm (Fig. 2g). As described below in detail, a poorly controlled, excessive inflammatory response is one of the major factors that not only triggers the cascade that ultimately leads to NEC, but also maintains disease activity as part of a vicious cycle (Fig. 2g). Figure 2. Model of NEC pathogenesis in the preterm intestine. (a) Multiple factors are involved in the precipitation of NEC, including dysbiosis, formula feeding, and ischaemic/hypoxic assaults. (b) Inappropriate increases in abundance of, and signalling by, pro-inflammatory pattern recognition receptors (PRRs) such as TLR4 contribute to the initiation of a cascade that involves (c) antigen processing by antigen-presenting cells such as dendritic cells (DCs) and (d) activation of other immune cells such as T cells, monocytes, macrophages and regulatory T cells (Tregs), leading to (e) an inappropriate and excessive increase of pro-inflammatory cytokines, chemokines and transcription factors. (f) A deficiency in counter-regulatory mediators contributes to this pro-inflammatory milieu to self-perpetuate and spiral out of control – (g) a vicious cycle is formed. (h) Inflammation-, ischaemia/reperfusion- and hypoxia-associated injury compromises the endothelial integrity of the local blood vessels, which also feeds the vicious cycle. (i) Necrotic cell death of the intestinal epithelium ensues, further exacerbating tissue injury and inflammation. (j) In line with the clinical stages (see Fig. 1), NEC severity can range from mild intestinal injury to segmental or even complete destruction of the intestinal epithelium. (k) Disintegration of the intestinal epithelium compromises its barrier functions, ultimately leading to rampant bacterial translocation into the lamina propria and the systemic circulation. Sepsis, multi-organ failure and death ensue. *, systemic data. #, strong evidence to be harmful only from one paper. Formula feeding Formula feeding is a well-established risk factor for NEC (Fig. 2a), and the incidence of NEC in infants fed their own mother's milk is reduced compared with formula-fed infants (Ref. 24). Exclusive feeding with their own mother's milk was also associated with fewer episodes of late-onset sepsis and/or NEC (OR 0.18; 95% CI 0.04–0.79, P = 0.02) and shorter duration of hospital stay compared with formula- or donor breast milk-fed infants (Ref. 25). A meta-analysis of studies comparing formula with donor breast milk in preterm or LBW infants revealed that formula triples the risk of NEC (Ref. 26). Infant formula contains components such as unbound free fatty acids (Ref. 27) that may facilitate NEC, and is deficient in potentially protective factors such as anti-inflammatory cytokines, immunoglobulins, growth factors, and microbiota, which are present in breast milk (Refs 28, 29). Further details are discussed in the relevant sections below. Hypoxia–ischaemia Historically, intestinal hypoxic–ischaemic injury was considered the single most important factor initiating and perpetuating NEC, a view consistent with the predominant pathologic finding being coagulative necrosis, a common sequela of prior ischaemia (Ref. 30). In addition, term neonates with NEC often have conditions such as chronic heart disease that favour hypoxic or ischaemic states (Fig. 2a) (Refs 31, 32). However, no primary hypoxic–ischaemic event can be identified in most preterm infants presenting with NEC. The appearance of NEC at 2–3 weeks of age (Ref. 33) (when pronounced or prolonged hypoxia/ischaemia is uncommon) rather points to a role of intestinal bacterial colonisation, which is usually nearly complete by this time. Microbial colonisation The gut microflora plays an important role in regulating gut immune homeostasis, e.g. by dampening excessive inflammatory responses and establishing an environment ‘tolerogenic’ for commensal bacteria (Ref. 34). This dampening process may be disrupted in NEC because of lower microflora diversity compared with preterm controls (Ref. 35). It is currently unclear whether the dysbiosis (Fig. 2a) that often accompanies NEC is a consequence or one of the causes of abnormal immune interactions between gut bacteria and the preterm intestine. Nevertheless, the role of the initial microbial colonisation in NEC is probably important as experimental NEC does not develop in the absence of bacteria, i.e. in germ-free piglets (Ref. 36) or mice treated with antibiotics (Ref. 37). Of note, animal studies have implicated Clostridium butyricum in NEC (Refs 38, 39, 40), and a recent study in human infants found this bacterium in the stool of 80% of NEC infants compared with 12% of controls (Ref. 41). Although these findings are promising, it is too early to conclude that C. butyricum is a bacterial cause of NEC. Animal models of NEC Much of our understanding of NEC pathogenesis stems from animal models of the disease, with the majority using rats, mice or piglets [reviewed in (Ref. 42)]. Most published models employ one or several of the known risk factors that induce NEC-like intestinal injury. The earliest NEC model, dating to 1974, subjected newborn rats to formula feeding and hypoxic stress (Ref. 43). This model is still used today, the most common variant being to subject caesarean-born preterm rats to formula feeding, hypoxia and hypothermia. Other variants of the hypoxia–hypothermia model include using caesarean-born E18.5 mice (Ref. 44), naturally delivered newborn mice (Ref. 45) and 7–10-day-old mice (Ref. 46). As newborn mice are more difficult to feed and handle than rats, the variant using 10-day-old mice is widely used today. Less commonly, murine NEC is induced using 2,4,6-trinitrobenzene sulphonic acid by gavage or enema in 10-day-old mice (Ref. 47), ablation of Paneth cells in combination with gavage feeding of Klebsiella in 14–16-day-old mice (Ref. 48), and by oral administration of Cronobacter sakazakii in 3-day-old mice (Ref. 49). Rabbit and hamster NEC models are occasionally employed, but most large animal-work on NEC is conducted in piglets. The gastrointestinal tract of newborn piglets closely resembles that of human babies in terms of anatomy, physiology, development and function. Piglet NEC models commonly comprise preterm birth, parenteral nutrition and formula feeding, but no exposure to hypoxia or hypothermia. NEC can also be modelled in primates, but such research is rarely undertaken as it requires preterm delivery and care for weeks in a NICU-like setting (Ref. 50). Another rare model is the gnotobiotic quail, primarily used for investigation of the role of clostridia in NEC (Refs 39, 51). NEC and Immunity The relationship between the immature preterm immune system and NEC is complex. A number of innate and adaptive immune mediators have been implicated in NEC, as summarised in Figure 3; note the distinction between local and systemic events. It is also important to keep in mind that evidence from human resection specimens is virtually always obtained from advanced NEC stages; therefore, knowledge on the intestinal events occurring in early human NEC is all but non-existent. Figure 3. Summary of the regulation and role of immune mediators in NEC. Green, protective; Grey, inconclusive; and Red, harmful. White text, animal data; purple text, human data; yellow text, animal and human data; black text outline, functional and/or genetic data. Ig, immunoglobulin; IFNγ, interferon gamma; IL, interleukin; IL-1Ra, interleukin-1 receptor antagonist; IL-1R8, IL-1 receptor 8; IL-17R, IL-17 receptor; MyD88, myeloid differentiation factor 88; NF-κB, nuclear factor-κB; NOD-2, nucleotide-binding oligomerisation domain-containing protein 2; PAF, platelet-activating factor; PAF-AH, PAF-acetylhydrolase; RORC, RAR-related orphan receptor C; TNF, tumour necrosis factor; TRIF, toll/IL-1R domain containing adaptor inducing IFNβ; TLR, toll-like receptor; TGF-β, transforming growth factor beta; *, may be protective in NEC, but Ig supplementation has not proven effective; #, strong evidence to be harmful only from one paper. The immune system in preterm neonates Detailed discussion of this topic is beyond the scope of this review, but briefly: The immune system is divided into two arms, innate and adaptive immunity. The newborn relies predominantly on innate immunity during early life as maturation of adaptive immunity lags behind that of innate immunity (Ref. 52). Within the adaptive arm, type 2 T-cell polarisation predominates in mother and foetus, thus protecting both from graft-versus-host-type rejections, which are mediated by type 1-polarised responses (Ref. 53). Compared with term infants, other differences include lower immune cell counts (Ref. 54), lower expression of major histocompatibility class II molecules (Ref. 55), and reduced phagocytic ability of monocytes and neutrophils (Ref. 56). Innate immunity The innate arm of immunity is phylogenetically older than adaptive immunity and functions as the first line of defence against potential pathogens. Innate immunity has two key components; a static component that consists of epithelial surfaces such as the skin and the gastrointestinal epithelium, which serve as physical barriers against microbial entry, and a reactive component, which involves tissue-resident and patrolling immune cells that are poised to respond rapidly to potential threats. Pattern recognition receptors (PRRs) PRRs play a central role in innate immunity, as they recognise pathogen-associated molecular patterns of invading pathogens and initiate signalling cascades that lead to target-independent inflammatory responses. As they are expressed by most cell types, PRRs perform a key function in frontline surveillance (Ref. 57). Two families of PRRs, Toll-like receptors (TLRs) and Nod-like receptors (NLRs), have been implicated in NEC. Toll-like receptors In the intestine, TLRs are expressed by immune cells and intestinal epithelial cells (IECs) (Ref. 58). A fine balance is required between preventing tissue invasion by gut bacteria on the one hand and establishing tolerance of a luminal commensal, symbiotic gut flora on the other. Therefore, the function of TLRs must be tightly controlled, particularly during the transition of the newborn gut from a germ-free intrauterine environment to postnatal exposure to colonising bacteria. Of note, much of our knowledge on TLRs in NEC stems from animal experiments, and it should be kept in mind that animal and human data are not always congruent. TLR4 Among the TLRs, TLR4 has received by far the most attention in the context of NEC. TLR4 is activated by the Gram-negative bacterial cell wall component lipopolysaccharide (LPS), a prototypical trigger of inflammation. Abundance and function of TLR4 is tightly regulated: Late in murine pregnancy (up to day 18; normal duration 21 days), Tlr4 mRNA expression increases, but rapidly decreases immediately following birth, thus adapting innate responses to the new environment (Ref. 59). Functionally, murine foetal IECs are significantly more responsive to LPS than IECs isolated on postnatal days 1 and 6 (Ref. 60). Xenografts from more immature human foetal ileum also express 3-fold more TLR4 than more mature grafts when transplanted into SCID (severe combined immunodeficiency) mice (Ref. 61). TLR4 gene and protein expression are elevated in the small intestinal mucosa of both human and mouse NEC compared with healthy controls (Fig. 2b) (Refs 59, 62, 63). This important signalling node is also target of mediators in breast milk such as soluble CD14, lactadherin, lactoferrin and 2′-fucosyllactose (Ref. 64). In a study in which lactating mice were milked under anaesthesia, mouse breast milk attenuated murine NEC by reducing TLR4 signalling, and overexpression of TLR4 in the intestinal epithelium reverses these protective effects (Ref. 65). In mice, excessive TLR4 expression was moreover linked to inhibition of intestinal repair, via activation of the p53-up-regulated modulator of apoptosis (Ref. 66) as well as induction of endoplasmic reticulum (ER) stress in intestinal stem cells (Ref. 67). Increased ER stress and apoptosis have been observed in the intestinal crypts of human NEC patients (Ref. 67). A pathogenic role of TLR4 in NEC appears likely, as TLR4-deficient mice (Ref. 37) and mice with non-functional TLR4 (Ref. 63) were protected against NEC-associated tissue damage, and a small molecule TLR4 inhibitor (C34) administered by oral gavage reduced ileal NEC injury (Ref. 68). Interestingly, enterocyte-specific deletion of TLR4 also efficiently protected from NEC, suggesting that the epithelium participates in this aspect of the disease (Ref. 37). Indeed, there is evidence that TLR4 expression in the intestinal epithelium may influence the recruitment and polarisation of T cells in the intestinal mucosa (Ref. 69). TLR9 Interestingly, TLR9, which recognises the characteristically CpG-rich bacterial DNA, acts as a counter-regulator of the disease-promoting effects of TLR4 in NEC (Ref. 59). Regulation of Tlr9 gene expression in the murine ileum is opposite to that of TLR4, so that Tlr9 decreases during late pregnancy, but increases at birth (Ref. 59). Mouse pups receiving two injections of 1 mg/kg CpG-DNA per day (Ref. 59) or once-daily oral CpG-DNA (Ref. 46) exhibited reduced NEC severity compared to vehicle-treated pups, demonstrating a functional relevance for TLR9 in NEC. Conversely, a mutation rendering TLR9 unresponsive to CpG-DNA causes increased NEC severity in mice (Ref. 59). Similarly, Lactobacillus rhamnosus-mediated protection in murine NEC is also dependent on TLR9 activation, as protection was abolished upon selective lentiviral knockdown of intestinal epithelial TLR9 (Ref. 46). A small human study showed that TLR9 protein abundance was reduced in NEC patients compared with controls (Fig. 2b) (Ref. 59); however, a protective function of TLR9 has not been confirmed in humans. TLR5 and other TLRs Gene expression of Tlr1, -2, -3, -6 and -7 was increased in ileal tissue of NEC rats compared with dam-fed controls, with only Tlr5 decreased (Refs 70, 71). A NEC-associated decrease in Tlr5 is consistent with TLR5 knockout mice developing spontaneous colitis (Refs 72, 73). The underlying mechanism between decreased TLR5 and chronic intestinal inflammation remains unknown, but it was speculated that absence of epithelial TLR5 may reduce epithelial barrier functions and thus increase bacterial translocation (Ref. 72). Alternatively, the decrease in Tlr5 mRNA may be secondary to increased TLR2 and -4 activation (Ref. 74); notably, Tlr2 and -4 are elevated in resected intestinal tissue from infants with stage III NEC (Ref. 62). In summary, aberrantly elevated TLR4 signalling has a pathogenic role in NEC, whereas TLR9 and possibly TLR5 act as counter-regulators of TLR4. The functional relevance of other TLRs in the disease remains poorly defined. Nucleotide-binding oligomerisation domain (NOD)-like receptors NLRs are intracellular PRRs and are critical mediators of the assembly of the inflammasome, which converts the pro-forms of the pro-inflammatory cytokines interleukin (IL)-1β and IL-18 into their mature, active forms. Data on NLRs in NEC are scant. NOD-2 NOD-2 is a sensor of bacterial cell-wall fragments, specifically muramyl dipeptide (MDP). NOD-2 mediates production of anti-bacterial defensins in epithelial Paneth cells (Ref. 75) and elicits immune responses through the nuclear factor (NF)-κB pathway (Ref. 76). NOD-2 activity may exert protective effects in NEC as daily injections of MDP almost completely abolished NEC-associated intestinal tissue damage in mice (Ref. 77). Similarly, in humans, NOD-2 loss-of-function mutations has been associated with Crohn's disease (CD) (Refs 78, 79) and VLBW infant carriers of two or more NOD-2 loss-of-function alleles had an increased risk for NEC requiring surgery (OR 3.57; 95% CI 1.3–10.0, P = 0.03) (Ref. 80). Mediators of innate immunity IL-1 IL-1 is the prototypical pro-inflammatory cytokine, and is induced in numerous cell types by a wide variety of triggers. Active at picogram concentrations, IL-1 induces a plethora of inflammatory effects, including the production of other pro-inflammatory mediators, tissue damage and fever (Ref. 81). The two isoforms, IL-1α and IL-1β, bind to the same heterodimeric cell surface receptor (Ref. 81). Activation and release of IL-1β are tightly controlled by post-translational mechanisms such as processing by caspase-1, which in turn is regulated by the inflammasome. Therefore, data on IL1B mRNA not accompanied by protein measurements may not be indicative of biological activity and should be interpreted with great caution. IL-1 binding to its receptor triggers a signalling cascade that results in activation of pro-inflammatory transcription factors such as NF-κB and AP-1, which in turn induce pro-inflammatory cytokines such as IL-6, tumour necrosis factor (TNF) and IL-1 itself (Ref. 81). Studies on IL-1α in NEC are rare. In caesarean-delivered preterm piglets with NEC, lysates of the small intestine exhibited increased IL1A mRNA abundance compared to colostrum-fed controls (Refs 82, 83). This increase in IL1A expression was rapid, occurring at 8 h and persisting for up to 34 h post-NEC induction (Ref. 82). IL-1β protein was elevated systemically (Ref. 84) and in intestinal tissue in animal models of NEC (Fig. 2e) (Refs 70, 85). In newborn rats, 48 h of formula feeding alone increased IL-1β protein in the terminal ileum 3-fold compared with dam-fed controls (Ref. 70). Induction of NEC increased IL-1β up to 6-fold compared with dam-fed controls (Refs 70, 85). Importantly, the authors highlighted that increases in IL-1β preceded tissue injury, which did not occur before 72 h (Ref. 70). In one of the few human studies on IL-1β in NEC, ileal IL1B mRNA in surgical NEC infants was more than 10-fold higher compared with GA-matched non-NEC controls (Ref. 86). Similarly, in situ hybridisation experiments showed a more than 2-fold increase in IL1B mRNA in full-thickness sections of stage III NEC infants compared with surgical controls (Ref. 87). Systemically, there was no difference between the pre-operative serum IL-1β abundance in NEC babies and non-NEC controls (Ref. 88). Similarly, limited time course experiments in human NEC infants beginning at NEC onset (defined by a combination of clinical and laboratory findings) and covering 8, 24, 48 and 72 h showed no significant change in serum IL-1β (Ref. 89). However, there was a trend towards higher IL-1β abundance in stage III infants compared with stage I and II infants (Ref. 89). Overall, the available data indicate that increased IL-1 precedes NEC injury, suggesting that IL-1 aggravates tissue damage and contributes to NEC initiation and perpetuation of the vicious cycle (Fig. 2g). IL-1 receptor antagonist (IL-1Ra) IL-1Ra is an anti-inflammatory cytokine that functions by competitively inhibiting the binding of the two pro-inflammatory ligands IL-1α and IL-1β to their receptor. IL-1Ra is in clinical use as reviewed in (Ref. 81), though at present not in NEC. As IL-1Ra is one of the endogenous counter-regulatory mechanisms induced by inflammation, its abundance is often associated with disease severity in inflammatory diseases. However, the considerable increases in IL-1Ra observed in NEC (Ref. 89) clearly do not curtail the overwhelming inflammation that underpins NEC; perhaps IL-1Ra concentrations are insufficiently elevated in the gut where the inflammatory damage is occurring. Interestingly, IL-1Ra was decreased 2–3 weeks prior to NEC onset in buccal swabs from at-risk infants (Ref. 90), suggesting a causative connection between NEC and IL-1Ra deficiency (Fig. 2f). Indeed, IL-1Ra shows promise as a NEC biomarker as described below. Tumour necrosis factor TNF, like IL-1, is a key pro-inflammatory cytokine that activates inflammatory mediators such as NF-κB in virtually any cell type. TNF was increased systemically (Ref. 91) and in intestinal tissue (Ref. 92) of NEC patients compared with non-NEC controls (Fig. 2e), but was not indicative of disease severity (Refs 88, 93, 94, 95). Ileal and systemic TNF were also increased in rat models of NEC (Refs 96, 97, 98), with the mRNA rising as early as 1.5 h after the first feed (Ref. 99). Although others did not observe such increases in TNF (Ref. 100), functional data indicate a disease-promoting role for TNF. Inhibition of TNF via administration of a monoclonal anti-TNF antibody (Refs 98, 101), pentoxiphylline (Ref. 102), etanercept (Ref. 103) or infliximab (Ref. 104) significantly reduced intestinal inflammation and tissue injury in neonatal NEC rats. However, others have reported no significant improvement with pentoxiphylline in hypoxia/reperfusion-induced rabbit NEC (Ref. 105). These observations suggest that TNF contributes to NEC progression, likely with a major role in the early stages of the disease. The usefulness of TNF as a biomarker in NEC appears limited. IL-6 IL-6 is an important acute phase immune mediator; for example, it stimulates hepatocytes to produce acute-phase proteins such as C-reactive protein (CRP). In fact, both CRP and IL-6 are in clinical use as biomarkers of acute inflammation (Ref. 106). It is likely that excessive IL-6 plays a pathogenic role in NEC. Genetic analysis of IL-6 single nucleotide polymorphisms (SNPs) in neonates of 32 weeks gestation or less revealed that Caucasians with IL-6 rs1800795, an SNP that is associated with increased plasma IL-6 in neonates (Ref. 107), were six times more likely to develop NEC and seven times more likely to progress to stage III disease (Ref. 108). These observations agree with studies that demonstrated elevated IL-6 protein (Ref. 109) and mRNA expression (Refs 62, 110) in resected intestinal tissue of stage III NEC patients compared with controls (Fig. 2e). IL-6 may thus be useful as a biomarker in NEC; see the Biomarkers section. IL-10 IL-10 is an important dampener of immune responses in the intestine, and loss of IL-10 or its receptor (IL-10R) results in early-onset inflammatory bowel disease in humans (Ref. 111) and mice (Ref. 112). Although the interaction between the intestinal microbiome and immunity is not part of this review, it is interesting to note that the intestinal inflammation of IL-10-deficient mice does not develop in a pathogen-free environment (Ref. 112). IL-10 functionality in macrophages curtails intestinal inflammation, as specific knockout of IL-10R signalling in intestinal lamina propria-resident macrophages results in severe spontaneous colitis in mice (Ref. 113). The number of regulatory T cells (Treg), an important source of intestinal IL-10 (Ref. 114), was reduced in the ileum of NEC rats compared to dam-fed controls (Fig. 2f) (Ref. 115). Similarly, in humans, the total number of CD4+Foxp3+ Treg and the Treg/T effector ratio was reduced in the lamina propria of surgical NEC infants compared to surgical controls (Ref. 86). Mice deficient in IL-10 exhibited more severe epithelial damage and overall NEC injury than wild-type controls (Fig. 2f) (Ref. 116). Moreover, administration of exogenous IL-10 to IL-10-deficient mice prior to NEC induction prevented mucosal injury (Ref. 116). IL-10 as a protective factor in NEC is supported by the observation that human breast milk contains high concentrations of bioactive IL-10 (Ref. 117) and lower IL-10 abundance in breast milk correlates with increased human NEC incidence (Ref. 118). However, a deficiency in IL-10 is not observed in human NEC; indeed, both serum and ileal IL-10 were markedly increased in infants diagnosed with NEC, particularly in those with advanced NEC (Refs 86, 88, 89), which, as with IL-1Ra, is likely part of the immune system's inadequate attempt at countering the excessive inflammation. As NEC predominantly affects preterm infants, it should also be noted that prematurity does not predispose to IL-10 deficiency (Refs 119, 120) or inducibility by TLR agonists (Refs 119, 121, 122). It thus appears likely that IL-10 contributes to dampening inflammation in NEC, but its precise role in NEC pathogenesis remains unclear. Mediators of innate immune signalling Nuclear factor-κB NF-κB is the prototypical pro-inflammatory transcription factor, with many pathways converging at this central node of inflammatory signalling. TLR-, IL-1 receptor (IL-1R)-, and TNFR-activation trigger a cascade that leads to release of cytoplasmic NF-κB from its inhibitory protein, the inhibitor of κB (IκB), allowing NF-κB to translocate to the nucleus and to actuate the transcription of pro-inflammatory mediators, including cytokines, chemokines and leukocyte adhesion molecules (Ref. 123). Developmental regulation of NF-κB pathway components may favour NEC, e.g. a reduced abundance of IκB in foetal primary IEC compared with mature adult enterocytes (called T84 cells) (Ref. 124). In animals, vaginal birth may trigger a transient, low-grade increase in NF-κB activation in the small intestine, possibly allowing a tolerogenic immune surveillance of the early stages of bacterial colonisation (Ref. 60): NF-κB was activated in murine IECs as early as 60 min after natural birth in the absence of inflammatory stimuli (Ref. 60) before its activation returned to baseline by 24 h (Ref. 99). Conversely, NF-κB activity was nearly undetectable in the small intestine of newborn rats delivered by caesarean section (Ref. 125). These findings may contribute to the unexpected observation that vaginal birth is a risk factor for early onset NEC (defined as <14 days, stage II or higher) in human preterm infants of <33 weeks GA (Ref. 126). However, the association between vaginal birth and intestinal NF-κB activation has not been demonstrated in human infants. On the other hand, there is clear evidence for an involvement of NF-κB in NEC. First, NEC severity was correlated with increased NF-κB activity in the epithelial cells of caesarean-born pups (Fig. 2e) (Refs 71, 99, 125), and second, specific inhibition of NF-κB (using a NEMO-binding domain peptide) in NEC rats markedly reduces disease incidence and severity (Ref. 125). Furthermore, in a human study, 100% of NEC infants were carriers of the NFKB1 variant –94delATTG, which leads to more pronounced inflammatory responses to LPS (Ref. 127), compared to 65% of the non-NEC infants (Ref. 128). MyD88 (myeloid differentiation factor 88), TRIF [Toll/IL-1R domain containing adaptor inducing interferon (IFN)β] and IL-1R8 (IL-1 receptor 8, previously called SIGIRR) The first step in the TLR- and IL-1R signalling cascades is recruitment of adapter molecules to the intracellular domains of the receptors. For example, TLR4 activates two signalling pathways, one via the adapter MyD88 and one via TRIF (Ref. 129). In concordance with the finding that TLR4-deficient mice were protected from NEC injury (Ref. 37), deficiency in MyD88 (Ref. 130) and TRIF (Ref. 37) also attenuated the disease (Fig. 2b). Unexpectedly, the protection conferred by the absence of MyD88 was not as complete as that observed in mice deficient in TLR4 and TRIF, indicating an important role for TRIF-dependent signalling in NEC (Ref. 37). Similarly, a deficiency in IL-1R8, which is a negative regulator of TLR- and IL-1R signalling (Refs 131, 132), may also be important as a small study associated NEC infants with stop-, missense- or splice region-IL-1R8 variants (Fig. 2b) (Ref. 133). Adaptive immunity The immune system's adaptive arm responds to highly specific antigens, which must be processed and presented, again in a highly specific fashion, by antigen-presenting cells (APC). The prototypic APC are dendritic cells (DC), which present antigens to T and B cells, the major effector cells of adaptive immunity. Such presentation results in the polarisation of naïve CD4+ T helper (Th) cells into different subsets, including Th1, Th2, Th17 and Treg, with the subset determination depending on the state of the APC, the antigen, its presentation, and the local cytokine milieu. Each subset is characterised by predominance of a transcription factor (T-bet, GATA-3, Ror-γt and Foxp3, respectively) and signature cytokines (IFNγ, IL-4, IL-17A and IL-10, respectively). Generally, the subsets antagonise each other, e.g. Th1 cytokines inhibit Th2 polarisation. There are conflicting data on the lymphocyte fraction of the inflammatory tissue infiltrate in NEC: Whereas a lamina propria CD4+ T cell component of 30–40% in NEC mouse pups and human infants was reported (Ref. 69), others observed a paucity of lymphocytes in the inflammatory infiltrate in human NEC infants (Refs 47, 134). Thus, the data discussed below need to be interpreted with caution. Nevertheless, some animal studies provide evidence to support a role for CD4+ T cell influx as an important pathogenic event in NEC. For example, recombination activating gene-deficient (Rag1−/−) mice, which are deficient in functional T and B cells, exhibit significantly reduced NEC-associated intestinal injury and Il1b expression compared with wild-type controls (Ref. 69). In addition, adoptive transfer of naïve CD4+ T cells to Rag1−/− mice prior to NEC induction restored susceptibility to severe NEC (Ref. 69). Furthermore, transfer and repopulation of Rag1−/− mice with CD4+ T cells from wild-type mice with NEC led to intestinal damage and increased Il1b expression after 48 h (Ref. 69). RNA sequencing of ileal samples from surgical NEC infants also revealed strongly altered T and B cell signalling in NEC compared with non-NEC preterm controls (Ref. 135). Although surprisingly little information is available on the role of Th subsets in initiation and/or perpetuation of NEC, some of the signature cytokines have been investigated. Th1 Cytokines IFNγ IFNγ is the signature cytokine of Th1 immune responses. It contributes to the differentiation of Th1 cells and exerts pro-inflammatory actions by inducing Th1 chemokines, activating macrophages and facilitating phagocytosis (Ref. 136). The combined effects of IFNγ are critical to clearance of intracellular pathogens. Of note, prematurity is associated with a reduced capacity to mount Th1 responses and produce IFNγ (Ref. 137). Whereas one human study reported no difference between peri-operative serum IFNγ in NEC infants and non-NEC controls (Ref. 88), others found a 4-fold higher frequency of cells spontaneously secreting IFNγ in peripheral blood mononuclear cells (PBMCs) isolated from stage II and III NEC infants at diagnosis compared with age-matched healthy controls (Ref. 138). Similarly, contradictory observations were made on IFNG mRNA in intestinal resection specimens (Refs 139, 140). In rats and mice, the data more clearly point to a disease-promoting role for IFNγ, as ileal IFNγ protein abundance dramatically increased after induction of experimental NEC compared with dam-fed controls (Fig. 2e) (Refs 70, 141). Mechanistically, excessive IFNγ interferes with epithelial barrier integrity and regeneration, including function of intercellular gap junctions and IEC migration (two processes impaired in wild-type NEC mice but unaffected in IFNγ-deficient NEC mice) (Ref. 141). Abrogation of these detrimental effects of IFNγ is likely to contribute to the observation that 10-day-old IFNγ-deficient mice are completely protected from NEC-associated ileal tissue damage (Ref. 141). IL-12 The principal function of IL-12 is to promote and maintain Th1 polarisation, for example by induction of IFNγ. Animal studies of NEC are inconclusive about IL-12, one reporting lower (Ref. 100), others higher (Refs 142, 143), expression. Interestingly, in human infants, reduced IL-12 abundance might be a risk factor for NEC: Preterm infants with a low bioactivity IL-12p40 promoter polymorphism exhibited a higher risk of NEC (CTCTAA allele, OR 2.9, 95% CI 1.4–6.0, P = 0.004) compared with infants with homozygous IL-12 CTCTGC alleles (Ref. 144). IL-18 IL-18 is a pleiotropic cytokine with functions in innate and adaptive immunity. In concert with IL-12, IL-18 enhances IFNγ production and promotes Th1 differentiation (Ref. 145). In experimental NEC, IL-18 appears to aggravate the disease process. Ileal IL-18 protein abundance increased progressively with severity of NEC injury in rats (Fig. 2e) (Refs 143, 146). Furthermore, IL-18-deficient mice were partially protected from NEC injury (Ref. 147), and the protection of anti-TNF treatment was associated with reduced intestinal IL-18 protein (Ref. 101). However, the available human evidence disagrees with the animal findings. Ileal IL18 mRNA was decreased in NEC infants compared with controls (Ref. 86). Similarly, a low-expression polymorphism (IL-18 A-607) was more frequent in infants with stage III NEC than in those with stage I/II (Ref. 148), and plasma IL-18 was moderately reduced in ELBW infants who subsequently developed NEC compared with infants that did not (Ref. 149). Th2 Cytokines Th2 cytokines studied in NEC include IL-4, IL-5, and IL-13. IL-4 is the signature cytokine of the Th2 subset as it promotes Th2 polarisation, suppresses Th1 responses, and induces B cell immunoglobulin class switching to IgE. The functions of IL-13 are similar to those of IL-4, including IgE class switching and activation of mast cells and eosinophils. IL-5 acts on eosinophils, promoting their activation, survival, and adhesion (Ref. 145). The intrauterine environment favours Th2 polarisation (Ref. 53). Increased ileal IL-4 and IL-5 accompanies NEC progression in rats (Ref. 70). Similarly, in a small human study, PBMCs isolated from stage II and III NEC infants at diagnosis exhibited 3-fold more cells spontaneously secreting IL-4 than GA-matched healthy controls (Ref. 138). However, comparing pre-operative NEC infants and GA-matched controls, serum IL-4 was not different, while IL-5 was 50% lower (Ref. 88), a surprising finding as onset of NEC coincides with eosinophilia (Ref. 150). Moreover, infants affected by NEC less frequently carried a high-bioactivity variant of the IL-4Rα chain (Ref. 151). A marked increase in ileal IL-13 in NEC rats occurred after onset of tissue injury (Ref. 70). Others have proposed that IL-13 protects the gut by curbing excessive IL-17 and limiting its colitogenic effects (Ref. 152). However, IL-13 also causes epithelial dysfunction such as goblet cell hyperplasia and mucus hypersecretion. Th17 Cytokines The Th17 signature cytokine, IL-17A, has several pro-inflammatory effects that are important for host protection against extracellular bacteria, including induction of chemokines (CXCL1, CXCL6 and CXCL10) and neutrophil recruitment and activation (Ref. 145). IL-23 induces Th17 polarisation, stimulates IL-17A in effector T cells, and is necessary for differentiation and effector functions of Th17 cells. Dysregulation of the Th17 pathway has been linked to inflammatory bowel diseases such as CD and ulcerative colitis (Ref. 153). Th17 responses likely also play a pathogenic role in NEC. For example, RNA sequencing has revealed remarkable similarities in the signalling pathways affected by NEC, CD and paediatric CD (Ref. 135). Lamina propria CD4+ Th17 cells were more than 2-fold more abundant in NEC mice compared with controls (Ref. 69), and intestinal IL-17A and IL-17 receptor A (IL-17RA) was increased in mouse and human NEC (Ref. 69). These observations are in agreement with formula-fed preterm NEC baboons who exhibit a 5-fold increase in ileal IL17A gene expression compared with GA-matched non-NEC preterm controls (Ref. 50), and with ileal Il23 mRNA being 6-fold higher in NEC rats than in dam-fed controls (Refs 142, 154). Moreover, intraperitoneal injection of recombinant IL-17A in newborn mice led to loss of intercellular tight junctions in the villi, reduced enterocyte proliferation and increased crypt apoptosis (Ref. 69). The detrimental effects of IL-17A in murine NEC were mediated by IL-17R, as these effects were abrogated by blockade of IL-17R with an antibody (Ref. 69). Similarly, inhibition of STAT3, a critical mediator of T cell differentiation towards a Th17 phenotype, using the compound WP1066 was also protective against murine NEC; WP1066 reduced Th17 cells and increased Tregs (Ref. 69). In fact, the balance between Tregs and Th17 cells may be critical in facilitating NEC, as one of the consequences of TLR4 deficiency was restoration of the Treg/Th17 ratio and near complete prevention of the NEC-associated intestinal infiltration of CD4+ T cells (Ref. 69). By contrast, systemic IL-17 was reduced in 21-day-old babies that subsequently developed NEC compared with infants that did not (Ref. 149). Likewise, there were 50% fewer of the Th17-associated intestinal intraepithelial γδ-T cells in the ileum of acute surgical NEC infants than in non-NEC controls (Ref. 155). Furthermore, expression of the Th17 transcription factor RAR-related orphan receptor C (RORC) was 10-fold less in the ileal mucosa of NEC infants compared to non-NEC controls (Ref. 155). In summary, although a disease-promoting role for Th17 polarisation may be emerging (Ref. 69), the data from humans frequently contradict those from animals in the field of adaptive immunity in NEC, and there are only few mechanistic studies. Moreover, the possibility that different Th subsets may dominate during different NEC stages remains poorly studied; thus, current evidence does not allow a conclusion on the relevance of Th polarisation in NEC. Immunoglobulins Immunoglobulins (Ig) are produced by B cells in five isotypes (IgA, IgD, IgE, IgG and IgM) and function as antibodies or receptors that target foreign invaders such as bacteria, viruses, fungi, parasites and toxins, assisting in their neutralisation in cooperation with other immune cells. Ig-mediated host defence in the gut is immature even in infants born at term and is thus temporarily dependent on Ig transfer from the mother (Ref. 156). Breast milk is a major source of Ig for the newborn infant and has been proposed as one of the major factors by which breast milk protects against NEC. The Ig content in infant formula is low or absent (Ref. 28). Ig supplementation was suggested as a prophylactic for NEC, with two small human studies reporting successful reduction of NEC incidence in infants orally administered either IgG alone (Ref. 157) or a mixture of IgA and IgG (Ref. 158). However, in a larger study, oral supplementation of IgG alone did not reduce NEC incidence (Ref. 159). A systematic review of these studies concluded that IgG or IgG+IgA demonstrated no significant reduction in incidence of definite NEC, suspected NEC, need for surgery or death from NEC in preterm and LBW infants (Ref. 160). Similarly, a systematic review of intravenous immunoglobulin administration to preterm or LBW infants or both also did not find any statistically significant difference in the incidence of NEC (Ref. 161). Thus, current evidence does not support the administration of oral or intravenous Ig for the prevention of NEC. Chemokines IL-8 IL-8 is a member of the CXC chemokine family that is produced by a variety of immune and non-immune cells (Ref. 145). Its main effector role is to recruit neutrophils to the site of inflammation. The premature human gut readily produces IL-8 (Ref. 162) and unlike in the adult immune system, IL-8 production is also a major T-cell effector function in preterm infants (Ref. 163). IL8 mRNA expression was increased in intestinal resection specimens from NEC infants compared with non-NEC controls (Fig. 2e) (Refs 86, 164) and serum IL-8 holds substantial potential as a diagnostic marker for NEC (see the ‘Biomarkers’ section). Other mediators Transforming growth factor beta (TGF-β) The biological activities of TGF-β are pleiotropic and strongly dependent on the target cell/organ and the local cytokine milieu. In the context of adaptive immunity, TGF-β may support anti- and pro-inflammatory responses, for example suppressing Th1- and Th2-polarisation and promoting Treg functions, but also inducing Th17 cell differentiation (Ref. 165). However, there is good evidence that TGF-β-deficiency promotes NEC. Disruption of TGF-β-signalling via depletion of TGF-βRII significantly increased the severity of platelet-activating factor (PAF) + LPS-induced NEC injury in 10–12-day-old mice compared with controls (Ref. 166). Moreover, tissue damage was ameliorated by enteral TGF-β2-supplementation in the PAF + LPS model and in formula-, hypoxia-, and cold stress-triggered mouse NEC (Ref. 166). Likewise, oral administration of TGF-β1 to NEC rats resulted in moderate suppression of NF-κB activation in ileal IECs and was associated with a 20% overall reduction in NEC incidence compared with vehicle-fed controls (Ref. 167). Intestinal TGF-β2 bioactivity, protein abundance, and gene expression were markedly reduced in NEC patients compared to GA-matched non-NEC controls (Fig. 2f) (Ref. 166) and preterm versus term infants (Refs 50, 166). A similar TGF-β2 deficiency was observed in the intestine of formula-fed preterm baboons, and was even more pronounced in preterm baboons with NEC (Ref. 50). In fact, the protective properties of human breast milk may in part be mediated by TGF-β2, which it contains in high quantities (Ref. 28). Mechanistic studies showed that human adult PBMC-derived macrophages develop increasing LPS tolerance when exposed to media conditioned by increasingly mature intestines, an effect mediated primarily through TGF-β2 and to a lesser extent by TGF-β1 (Ref. 166). In the developing intestine, macrophage production of pro-inflammatory cytokines may thus be suppressed by TGF-β, promoting tolerogenicity to commensal bacteria (Ref. 166). This function combines with TGF-β2-mediated cytoprotection (Refs 168, 169), rendering TGF-β2 a key protective player in NEC. Platelet-activating factor PAF is a pro-inflammatory phospholipid mediator that activates pathways such as protein-kinase C (PKC), mitogen-activated protein kinases (MAPK), phosphatidylinositol 3-kinase (PI3K) and NF-κB (Refs 170, 171). Intravenous administration of PAF in adult rats causes NEC-like ischaemic necrosis in the small intestine (Ref. 172). Pre-treatment with a low dose of LPS further aggravates these lesions, suggesting a synergistic effect between TLR signalling and PAF in intestinal disease (Ref. 172). Whereas a moderate elevation of circulating and stool PAF is physiological upon commencement of enteral feeding in newborn babies (Refs 173, 174), this increase is more pronounced in formula-fed infants. Even higher PAF concentrations were observed in NEC patients compared with non-NEC controls (Fig. 2e) (Refs 91, 174). The human neonate has a reduced capability to control substantial increases in PAF as the activity of its degrading enzyme, PAF-acetylhydrolase (PAF-AH) remains low in the first few weeks of life (Fig. 2f) (Ref. 175). Unlike formula, breast milk contains PAF-AH, which likely contributes to breast milk-mediated protection from NEC (Ref. 176). Although postnatally the circulating PAF-AH concentrations are similar in term and preterm infants (Ref. 175), in the setting of NEC, PAF-AH activity is reduced by more than 50% compared with non-NEC controls (Ref. 91). PAF-AH-deficient mice were more than twice as likely to develop experimental NEC than wild-type mice and exhibited a significantly higher abundance of inflammatory mediators such as CXCL1 and inducible nitric oxide synthase (Ref. 177). A beneficial role of PAF-AH is supported by the demonstration that intravenously administered recombinant PAF-AH protected against experimental NEC injury (Ref. 178). Furthermore, blockade of the PAF receptor ameliorated NEC-associated tissue damage in rats (Ref. 179) and piglets (Ref. 180). Biomarkers Among the major challenges clinicians face in caring for infants who may have NEC, unequivocal identification of the disease in its early stages, differentiating it from sepsis or spontaneous intestinal perforation (SIP), and deciding if and when to proceed with surgery, stand out. Identifying and validating biomarkers to guide clinical decision making would represent a major advance in neonatal medicine. A recent review summarised potential biomarkers in NEC (Ref. 181); here, we focus on promising candidates with a relevance to immunology. The acute phase reactant CRP is widely used as a marker of inflammation in NEC and many other diseases. Whereas CRP is non-specific and cannot be used to differentiate NEC from sepsis, it may be useful to determine disease progression; for example, a persistently elevated CRP may be indicative of treatment failure, whereas normalisation may indicate success. IL-6 may have greater sensitivity and specificity than CRP for charting NEC disease. In surgical NEC infants, serum IL-6 was up to 60-fold higher than in controls (Refs 88, 109, 110), and was correlated with disease severity. In small studies on blood samples obtained within 48 h of NEC diagnosis (Refs 94, 95), IL-6 was undetectable in stage I, whereas the mean concentrations were 127 pg/ml in stage II and 3127 pg/ml in stage III patients. Post-operatively, stage III patients exhibited a decline in serum IL-6 to stage II levels, and importantly, mean IL-6 was 3-fold higher in infants that subsequently died than in survivors (Ref. 95). Furthermore, pre-operative IL-6 concentrations were markedly lower in SIP, a condition that is sometimes difficult to differentiate from NEC (Refs 62, 182). Similarly, a mathematical model employing the sequential use of IL-10, IL-6 and RANTES plasma measurements predicted the development of disseminated intravascular coagulation in VLBW infants with severe sepsis and NEC (Ref. 183). Thus, IL-6 may assist clinicians in assessing NEC disease severity and progression, and in distinguishing between NEC and SIP, but not sepsis. Several other biomarkers of immune function and intestinal injury have been suggested to predict the progress of NEC. Higher plasma and urinary abundance of I-FABP is correlated with more severe intestinal damage, and predicted the need for surgery (Refs 184, 185, 186). Similarly, increased serum IL-1Ra (>130 000 pg/ml) at NEC onset was 92% specific in identifying infants whose disease subsequently progressed to stage III (Ref. 89). Moreover, serum IL-1Ra (>60 000 pg/ml) at NEC onset was 100% specific and 68% sensitive in classifying patients as suspected (stage I) or definite (stages II and III) NEC (Ref. 89). NEC patients exhibited higher serum IL-8 than healthy infants and babies with sepsis and non-inflammatory intestinal conditions (Refs 88, 109, 149, 187). Increased serum IL-8 at diagnosis of NEC predicted the need for surgery and correlated with 60-day mortality (Ref. 188). Pre-operative serum IL-8 moreover predicted subsequent NEC severity, with 20-fold more IL-8 in pan-intestinal than in focal NEC cases (2750 versus 171 pg/ml) (Ref. 189). Furthermore, compared with pre-operative abundance, serum IL-8 dropped by 60% in focal NEC, by 92% in multifocal NEC, and by 96% in pan-intestinal NEC by post-operative day 3 in infants that survived the disease; note there were no postoperative data on non-survivors, as most of them died within 24 h (Ref. 189). Like IL-6, IL-8 also differentiated NEC from SIP (Ref. 182). Deficiencies in anti-inflammatory mediators such as TGF-β1 and inter-alpha inhibitor protein (IaIp) may serve as predictive biomarkers for NEC onset. ELBW infants who subsequently developed NEC exhibited low circulating TGF-β1 from the first day of life, a concentration <1380 pg/ml predicted 64% of NEC cases (Ref. 149). Similarly, low plasma IaIp differentiated between NEC and non-specific abdominal disorders (Ref. 190). Stool measurement of calprotectin also has potential for predicting NEC onset and severity (Refs 191, 192, 193). However, larger studies are required to resolve the wide variations in calprotectin concentrations in faecal matter and to establish universal thresholds for NEC diagnosis (Ref. 181). It is common practice to combine several biomarkers into scores to achieve maximal diagnostic and predictive power. An example in NEC is the combined use of serum amyloid A and apoplipoprotein CII (ApoSAA score) that can guide the decision to initiate antibiotic treatment, as the ApoSAA score stratifies infants into low- and high-risk groups for sepsis/NEC (Ref. 194). The LIT [liver-fatty acid binding protein, intestinal-fatty acid binding protein (I-FABP) and trefoil factor-3] score can be used to determine NEC from sepsis and, more importantly, to differentiate the need for surgery and predict chance of survival (Ref. 195). Clinical trials No immune mediator or inhibitor has been tested for efficacy in treatment of NEC. The clinical trials landscape is dominated by probiotics, which are thought to restore the gut flora to its healthy, diverse state, thus indirectly modulating the preterm gut immune system towards its tolerogenic poise. Conclusion and Outlook Although the current data paints a complex picture of the vicious disease cycle in NEC (Fig. 2), two features stand out. First, there is a clear link between marked increases in certain pro-inflammatory mediators, including TLR4, TNF, IL-18, IFNγ, PAF, IL-6, IL-8, IL-1β, NF-κB and possibly IL-17A, in intestinal tissue on the one hand, and increased NEC severity on the other. Thus, beyond confirming that NEC occurs in the setting of excessive inflammation, research should focus on the aforementioned mediators in order to identify potential therapeutic targets. Second, it is likely that deficiencies in protective mediators such as TLR9, IL-1R8, IL-1Ra, TGFβ2, PAF-AH, and IL-10, as well as in Tregs, permit development of excessive inflammation in NEC, and thereby predispose infants to the disease. On the biomarker front, immune mediators such as IL-1Ra, IL-6, IL-8, and TGF-β1 have emerged as promising candidates. Measurement of gut-specific markers such as I-FABP also demonstrates potential for management of NEC. Before clinical implementation, these observations need to be confirmed in larger multicentre trials. At present we have insufficient evidence to draw a conclusion on the involvement of adaptive immunity in NEC initiation and perpetuation; however, recent data on a pathogenic role of Th17 responses and a Th17/Treg imbalance invite further exploration. On a cautionary note, findings in animal models and in the human have frequently proven contradictory, pointing to the danger inherent to relying too heavily on animal work. Furthermore, increasing the use of ‘omics’-approaches in NEC research will identify yet unknown mediators that contribute to NEC pathogenesis, and studies addressing the contribution of different cell types to the disease (e.g. IEC versus macrophages versus lymphocytes) are needed. The large datasets that are becoming increasingly available should also be mined in search for abnormalities such as SNPs or other genetic variants relevant to NEC. While we have in recent years made progress in understanding some aspects of NEC, it is clear that a major research effort is required if an immunology-based treatment for NEC is to emerge. Only such an effort can banish the spectre of NEC, which looms over NICUs and continues to kill preterm infants. Acknowledgements and funding We thank Sue Panckridge for both her artistic advice on the design of the figures as well as her contributions towards the graphic elements. This work was supported by two Project Grants by the National Health and Medical Research Council (M.F.N. and C.A.N., grant numbers 1012353 and 1043845); the Marian and E.H. Flack Trust (M.F.N. and C.A.N.); the Future Leader Fellowship from the National Heart Foundation of Australia (C.A.N., CF14/3517), the ANZ-Trustees (C.A.N., CT-20681); Monash University's Larkins Fellowship (M.F.N.); the Hudson Institute's Star Recruitment Fellowship (M.F.N.); the Monash University's Faculty Postgraduate Research Scholarship (S.X.C.) and the Victorian Government's Operational Infrastructure Support Program. Conflict of interest None. Further reading, resources and contacts http://www.morgansfund.org/ The Morgan Leary Vaughan Fund for Necrotizing Enterocolitis (NEC) is an all-volunteer, public charity dedicated to promoting public awareness about NEC and the potentially devastating effects it can have on preemies and their families, and to advancing research to prevent, diagnose, treat, and ultimately, cure NEC. http://necsociety.org/ The NEC society is a non-profit organisation that seeks to raise awareness of the risk factors of NEC in the wider community, advocate for better policies and practices to best protect preterm infants from NEC and assist in encouraging future NEC research. http://www.preemieparentalliance.wildapricot.org/ The Preemie Parent Alliance represents a diverse set of organisations that provide information, support and resources to parents of preterm infants. ==== Refs References 1. Al Tawil K. (2013 ) Risk factors, characteristics and outcomes of necrotizing enterocolitis in late preterm and term infants . Journal of Neonatal-Perinatal Medicine 6 , 125 -130 24246514 2. Patel R.M. (2015 ) Causes and timing of death in extremely premature infants from 2000 through 2011 . New England Journal of Medicine 372 , 331 -340 25607427 3. Henry M.C. and Moss R.L. (2009 ) Necrotizing enterocolitis . Annual Review of Medicine 60 , 111 -124 4. Stoll B.J. (2010 ) Neonatal outcomes of extremely preterm infants from the NICHD Neonatal Research Network . Pediatrics 126 , 443 -456 20732945 5. Fitzgibbons S.C. (2009 ) Mortality of necrotizing enterocolitis expressed by birth weight categories . Journal of Pediatric Surgery 44 , 1072 -1075 ; discussion 75-619524719 6. Blakely M.L. (2005 ) Postoperative outcomes of extremely low birth-weight infants with necrotizing enterocolitis or isolated intestinal perforation: a prospective cohort study by the NICHD Neonatal Research Network . Annals of Surgery 241 , 984 -989 ; discussion 89-9415912048 7. ANZNN (Australian and New Zealand Neonatal Network) . (2012 ) Report of the Australian and New Zealand Neonatal Network 2008 and 2009. Sydney : ANZNN . 8. Hall N.J. , Eaton S. and Pierro A. (2013 ) Royal Australasia of Surgeons Guest Lecture. Necrotizing enterocolitis: prevention, treatment, and outcome . Journal of Pediatric Surgery 48 , 2359 -2367 24314171 9. Bell M.J. (1978 ) Neonatal necrotizing enterocolitis. Therapeutic decisions based upon clinical staging . Annals of Surgery 187 , 1 -7 413500 10. Walsh M.C. and Kliegman R.M. (1986 ) Necrotizing enterocolitis: treatment based on staging criteria . Pediatric Clinics of North America 33 , 179 -201 3081865 11. Bisquera J.A. , Cooper T.R. and Berseth C.L. (2002 ) Impact of necrotizing enterocolitis on length of stay and hospital charges in very low birth weight infants . Pediatrics 109 , 423 -428 11875136 12. Ganapathy V. , Hay J.W. and Kim J.H. (2012 ) Costs of necrotizing enterocolitis and cost-effectiveness of exclusively human milk-based products in feeding extremely premature infants . Breastfeeding Medicine: the Official Journal of the Academy of Breastfeeding Medicine 7 , 29 -37 21718117 13. Pike K. (2012 ) Outcomes at 7 years for babies who developed neonatal necrotising enterocolitis: the ORACLE Children Study. Archives of disease in childhood . Fetal and Neonatal Edition 97 , F318 -F322 22933088 14. Schulzke S.M. , Deshpande G.C. and Patole S.K. (2007 ) Neurodevelopmental outcomes of very low-birth-weight infants with necrotizing enterocolitis: a systematic review of observational studies . Archives of Pediatrics & Adolescent Medicine 161 , 583 -590 17548764 15. Hintz S.R. (2005 ) Neurodevelopmental and growth outcomes of extremely low birth weight infants after necrotizing enterocolitis . Pediatrics 115 , 696 -703 15741374 16. Murthy K. (2014 ) Short-term outcomes for preterm infants with surgical necrotizing enterocolitis . Journal of Perinatology: Official Journal of the California Perinatal Association 34 , 736 -740 25144157 17. Luig M. and Lui K. (2005 ) Epidemiology of necrotizing enterocolitis. Part II. Risks and susceptibility of premature infants during the surfactant era: a regional study . Journal of Paediatrics and Child Health 41 , 174 -179 15813870 18. Buch N.A. (2001 ) An epidemiological study of neonatal necrotizing enterocolitis . Saudi Medical Journal 22 , 231 -237 11307109 19. Guthrie S.O. (2003 ) Necrotizing enterocolitis among neonates in the United States . Journal of Perinatology: Official Journal of the California Perinatal Association 23 , 278 -285 12774133 20. McElroy S.J. and Weitkamp J.H. (2011 ) Innate immunity in the small intestine of the preterm infant . Neoreviews 12 , e517 -e526 22639551 21. Battersby A.J. and Gibbons D.L. (2013 ) The gut mucosal immune system in the neonatal period . Pediatric Allergy and Immunology 24 , 414 -421 23682966 22. Neu J. (2007 ) Gastrointestinal development and meeting the nutritional needs of premature infants . American Journal of Clinical Nutrition 85 , 629S -634S 17284768 23. Bianchi M.E. (2007 ) DAMPs, PAMPs and alarmins: all we need to know about danger . Journal of Leukocyte Biology 81 , 1 -5 17032697 24. Schanler R.J. , Shulman R.J. and Lau C. (1999 ) Feeding strategies for premature infants: beneficial outcomes of feeding fortified human milk versus preterm formula . Pediatrics 103 (6 Pt 1), 1150 -1157 10353922 25. Schanler R.J. (2005 ) Randomized trial of donor human milk versus preterm formula as substitutes for mothers’ own milk in the feeding of extremely premature infants . Pediatrics 116 , 400 -406 16061595 26. Quigley M. and McGuire W. (2014 ) Formula versus donor breast milk for feeding preterm or low birth weight infants . Cochrane Database of Systematic Reviews 4 , CD002971 24752468 27. Penn A.H. (2012 ) Digested formula but not digested fresh human milk causes death of intestinal cells in vitro: implications for necrotizing enterocolitis . Pediatric Research 72 , 560 -567 23007028 28. Chatterton D.E. (2013 ) Anti-inflammatory mechanisms of bioactive milk proteins in the intestine of newborns . International Journal of Biochemistry & Cell Biology 45 , 1730 -1747 23660296 29. Castanys-Munoz E. , Martin M.J. and Vazquez E. (2016 ) Building a beneficial microbiome from birth . Advances in Nutrition 7 , 323 -330 26980815 30. Ballance W.A. (1990 ) Pathology of neonatal necrotizing enterocolitis: a ten-year experience . Journal of pediatrics 117 (1 Pt 2), S6 -S13 2362230 31. Lambert D.K. (2007 ) Necrotizing enterocolitis in term neonates: data from a multihospital health-care system . Journal of Perinatology: Official Journal of the California Perinatal Association 27 , 437 -443 17392837 32. Fisher J.G. (2014 ) Serious congenital heart disease and necrotizing enterocolitis in very low birth weight neonates . Journal of the American College of Surgeons 220 , 1018 -1026 25868405 33. Uauy R.D. (1991 ) Necrotizing enterocolitis in very low birth weight infants: biodemographic and clinical correlates . National Institute of Child Health and Human Development Neonatal Research Network. The Journal of Pediatrics 119 , 630 -638 1919897 34. Maynard C.L. (2012 ) Reciprocal interactions of the intestinal microbiota and immune system . Nature 489 , 231 -241 22972296 35. Wang Y. (2009 ) 16S rRNA gene-based analysis of fecal microbiota from preterm infants with and without necrotizing enterocolitis . ISME Journal 3 , 944 -954 19369970 36. Sangild P.T. (2006 ) Diet- and colonization-dependent intestinal dysfunction predisposes to necrotizing enterocolitis in preterm pigs . Gastroenterology 130 , 1776 -1792 16697741 37. Sodhi C.P. (2012 ) Intestinal epithelial Toll-like receptor 4 regulates goblet cell development and is required for necrotizing enterocolitis in mice . Gastroenterology 143 , 708 -718 e1-522796522 38. Bousseboua H. (1989 ) Experimental cecitis in gnotobiotic quails monoassociated with Clostridium butyricum strains isolated from patients with neonatal necrotizing enterocolitis and from healthy newborns . Infection and Immunity 57 , 932 -936 2917793 39. Waligora-Dupriet A.J. (2005 ) Evidence for clostridial implication in necrotizing enterocolitis through bacterial fermentation in a gnotobiotic quail model . Pediatric Research 58 , 629 -635 16189185 40. Azcarate-Peril M.A. (2011 ) Acute necrotizing enterocolitis of preterm piglets is characterized by dysbiosis of ileal mucosa-associated bacteria . Gut Microbes 2 , 234 -243 21983069 41. Cassir N. (2015 ) Clostridium butyricum strains and Dysbiosis linked to Necrotizing Enterocolitis in preterm Neonates . Clinical Infectious Diseases: an Official Publication of the Infectious Diseases Society of America 61 , 1107 -1526084844 42. Lu P. (2014 ) Animal models of gastrointestinal and liver diseases . Animal Models of Necrotizing Enterocolitis: Pathophysiology, Translational Relevance, and Challenges. American Journal of Physiology. Gastrointestinal and Liver Physiology 306 , G917 -G928 24763555 43. Barlow B. (1974 ) An experimental study of acute neonatal enterocolitis--the importance of breast milk . Journal of Pediatric Surgery 9 , 587 -595 4138917 44. Premkumar M.H. (2014 ) Argininosuccinate lyase in enterocytes protects from development of necrotizing enterocolitis . American Journal of Physiology – Gastrointestinal and Liver Physiology 307 , G347 -G354 24904080 45. Tian R. (2010 ) Characterization of a necrotizing enterocolitis model in newborn mice . International Journal of Clinical and Experimental Medicine 3 , 293 -302 21072263 46. Good M. (2014 ) Lactobacillus rhamnosus HN001 decreases the severity of necrotizing enterocolitis in neonatal mice and preterm piglets: evidence in mice for a role of TLR9 . American Journal of Physiology – Gastrointestinal and Liver Physiology 306 , G1021 -G1032 24742987 47. MohanKumar K. (2012 ) Gut mucosal injury in neonates is marked by macrophage infiltration in contrast to pleomorphic infiltrates in adult: evidence from an animal model . American Journal of Physiology – Gastrointestinal and Liver Physiology 303 , G93 -102 22538401 48. Zhang C. (2012 ) Paneth cell ablation in the presence of Klebsiella pneumoniae induces necrotizing enterocolitis (NEC)-like injury in the small intestine of immature mice . Disease Models & Mechanisms 5 , 522 -532 22328592 49. Emami C.N. (2012 ) Role of neutrophils and macrophages in the pathogenesis of necrotizing enterocolitis caused by Cronobacter sakazakii . Journal of Surgical Research 172 , 18 -28 21601887 50. Namachivayam K. (2013 ) Smad7 inhibits autocrine expression of TGF-beta2 in intestinal epithelial cells in baboon necrotizing enterocolitis . American Journal of Physiology – Gastrointestinal and Liver Physiology 304 , G167 -G180 23154975 51. Butel M.J. (1998 ) Clostridial pathogenicity in experimental necrotising enterocolitis in gnotobiotic quails and protective role of bifidobacteria . Journal of Medical Microbiology 47 , 391 -399 9879939 52. Dowling D.J. and Levy O. (2014 ) Ontogeny of early life immunity . Trends in Immunology 35 , 299 -310 24880460 53. Levy O. (2007 ) Innate immunity of the newborn: basic mechanisms and clinical correlates . Nature Reviews Immunology 7 , 379 -390 54. Sharma A.A. (2012 ) The developing human preterm neonatal immune system: a case for more research in this area . Clinical Immunology 145 , 61 -68 22926079 55. Jones C.A. , Holloway J.A. and Warner J.O. (2002 ) Phenotype of fetal monocytes and B lymphocytes during the third trimester of pregnancy . Journal of Reproductive Immunology 56 , 45 -60 12106883 56. Strunk T. (2004 ) Differential maturation of the innate immune response in human fetuses . Pediatric Research 56 , 219 -226 15181184 57. Kumar H. , Kawai T. and Akira S. (2011 ) Pathogen recognition by the innate immune system . International Review of Immunology 30 , 16 -34 58. Abreu M.T. (2010 ) Toll-like receptor signalling in the intestinal epithelium: how bacterial recognition shapes intestinal function . Nature Reviews Immunology 10 , 131 -144 59. Gribar S.C. (2009 ) Reciprocal expression and signaling of TLR4 and TLR9 in the pathogenesis and treatment of necrotizing enterocolitis . Journal of Immunology 182 , 636 -646 60. Lotz M. (2006 ) Postnatal acquisition of endotoxin tolerance in intestinal epithelial cells . Journal of Experimental Medicine 203 , 973 -984 16606665 61. Nanthakumar N. (2011 ) The mechanism of excessive intestinal inflammation in necrotizing enterocolitis: an immature innate immune response . PLoS ONE 6 , e12 62. Chan K.Y. (2014 ) Genome-wide expression profiles of Necrotizing Enterocolitis versus spontaneous intestinal perforation in human intestinal tissues: dysregulation of functional pathways . Annals of Surgery 260 , 1128 -1137 24368664 63. Leaphart C.L. (2007 ) A critical role for TLR4 in the pathogenesis of necrotizing enterocolitis by modulating intestinal injury and repair . Journal of Immunology 179 , 4808 -4820 64. He Y. , Lawlor N.T. and Newburg D.S. (2016 ) Human milk components modulate toll-like receptor-mediated inflammation . Advances in Nutrition 7 , 102 -111 26773018 65. Good M. (2015 ) Breast milk protects against the development of necrotizing enterocolitis through inhibition of Toll-like receptor 4 in the intestinal epithelium via activation of the epidermal growth factor receptor . Mucosal Immunology 8 , 1166 -1179 25899687 66. Neal M.D. (2012 ) Toll-like receptor 4 is expressed on intestinal stem cells and regulates their proliferation and apoptosis via the p53 up-regulated modulator of apoptosis . Journal of Biological Chemistry 287 , 37296 -37308 22955282 67. Afrazi A. (2014 ) Toll-like receptor 4-mediated endoplasmic reticulum stress in intestinal crypts induces necrotizing enterocolitis . Journal of Biological Chemistry 289 , 9584 -9599 24519940 68. Neal M.D. (2013 ) Discovery and validation of a new class of small molecule Toll-like receptor 4 (TLR4) inhibitors . PLoS ONE 8 , e12 69. Egan C.E. (2015 ) Toll-like receptor 4-mediated lymphocyte influx induces neonatal necrotizing enterocolitis . Journal of Clinical Investigation 70. Liu Y. (2009 ) Changes in intestinal Toll-like receptors and cytokines precede histological injury in a rat model of necrotizing enterocolitis . American Journal of Physiology – Gastrointestinal and Liver Physiology 297 , G442 -G450 19608731 71. Le Mandat Schultz A. (2007 ) Expression of TLR-2, TLR-4, NOD2 and pNF-kappaB in a neonatal rat model of necrotizing enterocolitis . PLoS ONE 2 , e12 72. Vijay-Kumar M. (2007 ) Deletion of TLR5 results in spontaneous colitis in mice . Journal of Clinical Investigation 117 , 3909 -3921 18008007 73. Carvalho F.A. (2012 ) Interleukin-1beta (IL-1beta) promotes susceptibility of Toll-like receptor 5 (TLR5) deficient mice to colitis . Gut 61 , 373 -384 21646247 74. van Aubel R.A. (2007 ) Ligand-induced differential cross-regulation of Toll-like receptors 2, 4 and 5 in intestinal epithelial cells . Molecular Immunology 44 , 3702 -3714 17493681 75. Voss E. (2006 ) NOD2/CARD15 mediates induction of the antimicrobial peptide human beta-defensin-2 . Journal of Biological Chemistry 281 , 2005 -2011 16319062 76. Philpott D.J. (2014 ) NOD proteins: regulators of inflammation in health and disease . Nature Reviews Immunology 14 , 9 -23 77. Richardson W.M. (2010 ) Nucleotide-binding oligomerization domain-2 inhibits toll-like receptor-4 signaling in the intestinal epithelium . Gastroenterology 139 , 904 -917 , 17 e1-620580721 78. Hugot J.P. (2001 ) Association of NOD2 leucine-rich repeat variants with susceptibility to Crohn's disease . Nature 411 , 599 -603 11385576 79. Ogura Y. (2001 ) A frameshift mutation in NOD2 associated with susceptibility to Crohn's disease . Nature 411 , 603 -606 11385577 80. Hartel C. (2016 ) NOD2 loss-of-function mutations and risks of necrotizing Enterocolitis or focal intestinal perforation in very low-birth-weight Infants . Inflammatory Bowel Diseases 22 , 249 -256 26752461 81. Dinarello C.A. , Simon A. and van der Meer J.W. (2012 ) Treating inflammation by blocking interleukin-1 in a broad spectrum of diseases . Nature Reviews Drug Discovery 11 , 633 -652 22850787 82. Siggers J. (2011 ) Transition from parenteral to enteral nutrition induces immediate diet-dependent gut histological and immunological responses in preterm neonates . American Journal of Physiology – Gastrointestinal and Liver Physiology 301 , G435 -G445 21700903 83. Siggers J. (2013 ) Postnatal amniotic fluid intake reduces gut inflammatory responses and necrotizing enterocolitis in preterm neonates . American Journal of Physiology – Gastrointestinal and Liver Physiology 304 , G864 -G875 23518680 84. Guven A. (2011 ) Melatonin ameliorates necrotizing enterocolitis in a neonatal rat model . Journal of Pediatric Surgery 46 , 2101 -2107 22075339 85. Yurttutan S. (2013 ) Protective effects of colchicine in an experimental model of necrotizing enterocolitis in neonatal rats . Journal of Surgical Research 183 , 156 -162 23465391 86. Weitkamp J.H. (2013 ) Necrotising enterocolitis is characterised by disrupted immune regulation and diminished mucosal regulatory (FOXP3)/effector (CD4, CD8) T cell ratios . Gut 62 , 73 -82 22267598 87. Viscardi R.M. (1997 ) Inflammatory cytokine mRNAs in surgical specimens of necrotizing enterocolitis and normal newborn intestine . Pediatric Pathology & Laboratory Medicine: Journal of the Society for Pediatric Pathology, Affiliated with the International Paediatric Pathology Association 17 , 547 -559 88. Benkoe T. (2013 ) Comprehensive evaluation of 11 cytokines in premature infants with surgical necrotizing enterocolitis . PLoS ONE 8 , e12 89. Edelson M.B. , Bagwell C.E. and Rozycki H.J. (1999 ) Circulating pro- and counterinflammatory cytokine levels and severity in necrotizing enterocolitis . Pediatrics 103 (4 Pt 1), 766 -771 10103300 90. Murgas Torrazza R. (2013 ) Pilot study using proteomics to identify predictive biomarkers of necrotizing enterocolitis from buccal swabs in very low birth weight infants . Neonatology 104 , 234 -242 24030199 91. Caplan M.S. (1990 ) Role of platelet activating factor and tumor necrosis factor-alpha in neonatal necrotizing enterocolitis . Journal of Pediatrics 116 , 960 -964 2348301 92. Baregamian N. (2009 ) Tumor necrosis factor-alpha and apoptosis signal-regulating kinase 1 control reactive oxygen species release, mitochondrial autophagy, and c-Jun N-terminal kinase/p38 phosphorylation during necrotizing enterocolitis . Oxidative Medicine and Cellular Longevity 2 , 297 -306 20716917 93. Harris M.C. (1994 ) Cytokine elevations in critically ill infants with sepsis and necrotizing enterocolitis . Journal of Pediatrics 124 , 105 -111 8283358 94. Morecroft J.A. (1994 ) Plasma cytokine levels in necrotizing enterocolitis . Acta Paediatrica 396 , 18 -20 8086674 95. Morecroft J.A. (1994 ) Plasma interleukin-6 and tumour necrosis factor levels as predictors of disease severity and outcome in necrotizing enterocolitis . Journal of Pediatric Surgery 29 , 798 -800 8078025 96. Sheng Q. (2014 ) Human beta-defensin-3 promotes intestinal epithelial cell migration and reduces the development of necrotizing enterocolitis in a neonatal rat model . Pediatric Research 76 , 269 -279 24956228 97. Cetinkaya M. (2013 ) CDP-choline reduces severity of intestinal injury in a neonatal rat model of necrotizing enterocolitis . Journal of Surgical Research 183 , 119 -128 23228325 98. Seitz G. (2005 ) Protective effect of tumor necrosis factor alpha antibody on experimental necrotizing enterocolitis in the rat . Journal of Pediatric Surgery 40 , 1440 -1445 16150346 99. Rentea R.M. (2013 ) Early enteral stressors in newborns increase inflammatory cytokine expression in a neonatal necrotizing enterocolitis rat model . European Journal of Pediatric Surgery:Official Journal of Austrian Association of Pediatric Surgery = Zeitschrift fur Kinderchirurgie 23 , 39 -47 100. Nadler E.P. (2000 ) Expression of inducible nitric oxide synthase and interleukin-12 in experimental necrotizing enterocolitis . Journal of Surgical Research 92 , 71 -77 10864485 101. Halpern M.D. (2006 ) Reduction of experimental necrotizing enterocolitis with anti-TNF-alpha . American Journal of Physiology – Gastrointestinal and Liver Physiology 290 , G757 -G764 16269520 102. Travadi J. (2006 ) Pentoxifylline reduces the incidence and severity of necrotizing enterocolitis in a neonatal rat model . Pediatric Research 60 , 185 -189 16864701 103. Yurttutan S. (2014 ) Beneficial effects of Etanercept on experimental necrotizing enterocolitis . Pediatric Surgery International 30 , 71 -77 24072202 104. Tayman C. (2016 ) TNF-alpha blockade efficiently reduced severe intestinal damage in necrotizing enterocolitis . Journal of Investigative Surgery 18 , 1 -9 105. Erdener D. (2004 ) Pentoxifylline does not prevent hypoxia/reoxygenation-induced necrotizing enterocolitis. An experimental study . Biology of the Neonate 86 , 29 -33 15017117 106. Tanaka T. and Kishimoto T. (2014 ) The biology and medical implications of interleukin-6 . Cancer Immunology Research 2 , 288 -294 24764575 107. Kilpinen S. (2001 ) The promoter polymorphism of the interleukin-6 gene regulates interleukin-6 production in neonates but not in adults . European Cytokine Network 12 , 62 -68 11282548 108. Franklin A.L. (2015 ) Are immune modulating single nucleotide polymorphisms associated with necrotizing enterocolitis? Scientific Reports 5 , 18369 26670709 109. Ren Y. (2005 ) Up-regulation of macrophage migration inhibitory factor in infants with acute neonatal necrotizing enterocolitis . Histopathology 46 , 659 -667 15910597 110. Chan K.Y. (2012 ) Immunoregulatory protein profiles of necrotizing enterocolitis versus spontaneous intestinal perforation in preterm infants . PLoS ONE 7 , e12 111. Kotlarz D. (2012 ) Loss of interleukin-10 signaling and infantile inflammatory bowel disease: implications for diagnosis and therapy . Gastroenterology 143 , 347 -355 22549091 112. Kuhn R. (1993 ) Interleukin-10-deficient mice develop chronic enterocolitis . Cell 75 , 263 -274 8402911 113. Zigmond E. (2014 ) Macrophage-restricted interleukin-10 receptor deficiency, but not IL-10 deficiency, causes severe spontaneous colitis . Immunity 40 , 720 -733 24792913 114. Uhlig H.H. (2006 ) Characterization of Foxp3+CD4+CD25+ and IL-10-secreting CD4+CD25+ T cells during cure of colitis . Journal of Immunology 177 , 5852 -5860 115. Liu Y. (2013 ) Lactobacillus reuteri DSM 17938 changes the frequency of Foxp3+ regulatory T cells in the intestine and mesenteric lymph node in experimental necrotizing enterocolitis . PLoS ONE 8 , e12 116. Emami C.N. (2012 ) Role of interleukin-10 in the pathogenesis of necrotizing enterocolitis . American Journal of Surgery 203 , 428 -435 22450025 117. Garofalo R. (1995 ) Interleukin-10 in human milk . Pediatric Research 37 (4 Pt 1), 444 -449 7596683 118. Fituch C.C. (2004 ) Concentrations of IL-10 in preterm human milk and in milk from mothers of infants with necrotizing enterocolitis . Acta Paediatrica 93 , 1496 -1500 15513579 119. Dembinski J. (2003 ) Modulation of pro- and anti-inflammatory cytokine production in very preterm infants . Cytokine 21 , 200 -206 12788309 120. Blanco-Quiros A. (2000 ) Cord blood interleukin-10 levels are increased in preterm newborns . European Journal of Pediatrics 159 , 420 -423 10867846 121. Lavoie P.M. (2010 ) Profound lack of interleukin (IL)-12/IL-23p40 in neonates born early in gestation is associated with an increased risk of sepsis . Journal of Infectious Diseases 202 , 1754 -1763 20977341 122. Dirix V. , Vermeulen F. and Mascart F. (2013 ) Maturation of CD4+ regulatory T lymphocytes and of cytokine secretions in infants born prematurely . Journal of Clinical Immunology 33 , 1126 -1133 23793781 123. Li Q. and Verma I.M. (2002 ) NF-kappaB regulation in the immune system . Nature Reviews Immunology 2 , 725 -734 124. Claud E.C. (2004 ) Developmentally regulated IkappaB expression in intestinal epithelium and susceptibility to flagellin-induced inflammation . Proceedings of the National Academy of Sciences of the United States of America 101 , 7404 -7408 15123821 125. De Plaen I.G. (2007 ) Inhibition of nuclear factor-kappaB ameliorates bowel injury and prolongs survival in a neonatal rat model of necrotizing enterocolitis . Pediatric Research 61 , 716 -721 17426653 126. Yee W.H. (2012 ) Incidence and timing of presentation of necrotizing enterocolitis in preterm infants . Pediatrics 129 , e298 -e304 22271701 127. Adamzik M. (2013 ) The NFKB1 promoter polymorphism (-94ins/delATTG) alters nuclear translocation of NF-kappaB1 in monocytes after lipopolysaccharide stimulation and is associated with increased mortality in sepsis . Anesthesiology 118 , 123 -133 23249930 128. Sampath V. (2011 ) The NFKB1 (g.-24519delATTG) variant is associated with necrotizing enterocolitis (NEC) in premature infants . Journal of Surgical Research 169 , e51 -e57 21529841 129. O'Neill L.A. and Bowie A.G. (2007 ) The family of five: TIR-domain-containing adaptors in Toll-like receptor signalling . Nature Reviews Immunology 7 , 353 -364 130. Yang G. (2014 ) Critical role of myeloid differentiation factor 88 in necrotizing enterocolitis . Pediatric Research 75 , 707 -715 24614801 131. Garlanda C. (2013 ) Negative regulatory receptors of the IL-1 family . Seminars in Immunology 25 , 408 -415 24239046 132. Nold-Petry C.A. (2015 ) IL-37 requires the receptors IL-18Ralpha and IL-1R8 (SIGIRR) to carry out its multifaceted anti-inflammatory program upon innate signal transduction . Nature Immunology 16 , 354 -365 25729923 133. Sampath V. (2015 ) SIGIRR genetic variants in premature infants with necrotizing enterocolitis . Pediatrics 135 , e1530 -4 25963006 134. Remon J.I. (2015 ) Depth of bacterial invasion in resected intestinal tissue predicts mortality in surgical necrotizing enterocolitis . Journal of Perinatology 35 , 755 -762 25950918 135. Tremblay E. (2016 ) Gene expression profiling in necrotizing enterocolitis reveals pathways common to those reported in Crohn's disease . BMC Medical Genomics 9 , 6 26801768 136. Schoenborn J.R. and Wilson C.B. (2007 ) Regulation of interferon-gamma during innate and adaptive immune responses . Advances in Immunology 96 , 41 -101 17981204 137. Basha S. , Surendran N. and Pichichero M. (2014 ) Immune responses in neonates . Expert Review of Clinical Immunology 10 , 1171 -1184 25088080 138. Abdelhamid A.E. (2011 ) In vitro cow's milk protein-specific inflammatory and regulatory cytokine responses in preterm infants with necrotizing enterocolitis and sepsis . Pediatric Research 69 , 165 -169 20975616 139. Ford H. (1997 ) The role of inflammatory cytokines and nitric oxide in the pathogenesis of necrotizing enterocolitis . Journal of Pediatric Surgery 32 , 275 -282 9044137 140. Pender S.L. (2003 ) Matrix metalloproteinases in necrotising enterocolitis . Pediatric Research 54 , 160 -164 12736398 141. Leaphart C.L. (2007 ) Interferon-gamma inhibits intestinal restitution by preventing gap junction communication between enterocytes . Gastroenterology 132 , 2395 -2411 17570214 142. Coursodon-Boyiddle C.F. (2012 ) Pomegranate seed oil reduces intestinal damage in a rat model of necrotizing enterocolitis . American Journal of Physiology – Gastrointestinal and Liver Physiology 303 , G744 -G751 22821948 143. Halpern M.D. (2002 ) Up-regulation of IL-18 and IL-12 in the ileum of neonatal rats with necrotizing enterocolitis . Pediatric Research 51 , 733 -739 12032269 144. Bokodi G. (2007 ) Association of interferon gamma T+874A and interleukin 12 p40 promoter CTCTAA/GC polymorphism with the need for respiratory support and perinatal complications in low birthweight neonates . Archives of Disease in Childhood. Fetal and Neonatal Edition 92 , F25 -F29 16754651 145. Akdis M. (2011 ) Interleukins, from 1 to 37, and interferon-gamma: receptors, functions, and roles in diseases . Journal of Allergy and Clinical Immunology 127 , 701 -721 e1-7021377040 146. Halpern M.D. (2003 ) Ileal cytokine dysregulation in experimental necrotizing enterocolitis is reduced by epidermal growth factor . Journal of Pediatric Gastroenterology and Nutrition 36 , 126 -133 12500008 147. Halpern M.D. (2008 ) Decreased development of necrotizing enterocolitis in IL-18-deficient mice . American Journal of Physiology – Gastrointestinal and Liver Physiology 294 , G20 -G26 17947451 148. Heninger E. (2002 ) Genetic variants of the interleukin-18 promoter region (-607) influence the course of necrotising enterocolitis in very low birth weight neonates . European Journal of Pediatrics 161 , 410 -411 12174825 149. Maheshwari A. (2014 ) Cytokines associated with necrotizing enterocolitis in extremely-low-birth-weight infants . Pediatric Research 76 , 100 -108 24732104 150. Calhoun D.A. (2000 ) Granulocyte-macrophage colony-stimulating factor and interleukin-5 concentrations in premature neonates with eosinophilia . Journal of Perinatology: Official Journal of the California Perinatal Association 20 , 166 -171 10802841 151. Treszl A. (2003 ) Lower prevalence of IL-4 receptor alpha-chain gene G variant in very-low-birth-weight infants with necrotizing enterocolitis . Journal of Pediatric Surgery 38 , 1374 -1378 14523823 152. Mannon P. and Reinisch W. (2012 ) Interleukin 13 and its role in gut defence and inflammation . Gut 61 , 1765 -1773 22942239 153. Khor B. , Gardet A. and Xavier R.J. (2011 ) Genetics and pathogenesis of inflammatory bowel disease . Nature 474 , 307 -317 21677747 154. Underwood M.A. (2014 ) Bifidobacterium longum subsp. infantis in experimental necrotizing enterocolitis: alterations in inflammation, innate immune response, and the microbiota . Pediatric Research 76 , 326 -333 25000347 155. Weitkamp J.H. (2014 ) Small intestinal intraepithelial TCRgammadelta+ T lymphocytes are present in the premature intestine but selectively reduced in surgical necrotizing enterocolitis . PLoS ONE 9 , e12 156. Brandtzaeg P. (2003 ) Mucosal immunity: integration between mother and the breast-fed infant . Vaccine 21 , 3382 -3388 12850345 157. Rubaltelli F.F. , Benini F. and Sala M. (1991 ) Prevention of necrotizing enterocolitis in neonates at risk by oral administration of monomeric IgG . Developmental Pharmacology & Therapeutics 17 , 138 -143 1841829 158. Eibl M.M. (1988 ) Prevention of necrotizing enterocolitis in low-birth-weight infants by IgA–IgG feeding . New England Journal of Medicine 319 , 1 -7 3288866 159. Lawrence G. (2001 ) Enteral human IgG for prevention of necrotising enterocolitis: a placebo-controlled, randomised trial . Lancet 357 , 2090 -2094 11445103 160. Foster J.P. , Cole M.J. and Seth R. (2016 ) Oral immunoglobulin for preventing necrotizing enterocolitis in preterm and low birth weight neonates . Cochrane Database of Systematic Reviews 4 , CD001816 27040323 161. Ohlsson A. and Lacy J.B. (2013 ) Intravenous immunoglobulin for preventing infection in preterm and/or low birth weight infants . Cochrane Database of Systematic Reviews 7 , CD000361 162. Nanthakumar N.N. (2000 ) Inflammation in the developing human intestine: a possible pathophysiologic contribution to necrotizing enterocolitis . Proceedings of the National Academy of Sciences of the United States of America 97 , 6043 -6048 10823949 163. Gibbons D. (2014 ) Interleukin-8 (CXCL8) production is a signatory T cell effector function of human newborn infants . Nature Medicine 20 , 1206 -1210 164. Nadler E.P. (2001 ) Intestinal cytokine gene expression in infants with acute necrotizing enterocolitis: interleukin-11 mRNA expression inversely correlates with extent of disease . Journal of Pediatric Surgery 36 , 1122 -1129 11479840 165. Travis M.A. and Sheppard D. (2014 ) TGF-beta activation and function in immunity . Annual Review of Immunology 32 , 51 -82 166. Maheshwari A. (2011 ) TGF-beta2 suppresses macrophage cytokine production and mucosal inflammatory responses in the developing intestine . Gastroenterology 140 , 242 -253 20875417 167. Shiou S.R. (2013 ) Oral administration of transforming growth factor-beta1 (TGF-beta1) protects the immature gut from injury via Smad protein-dependent suppression of epithelial nuclear factor kappaB (NF-kappaB) signaling and proinflammatory cytokine production . Journal of Biological Chemistry 288 , 34757 -34766 24129565 168. Nguyen D.N. (2014 ) Transforming growth factor-beta2 and endotoxin interact to regulate homeostasis via interleukin-8 levels in the immature intestine . American Journal of Physiology – Gastrointestinal and Liver Physiology 307 , G689 -G699 25147235 169. Nguyen D.N. (2015 ) Protective effects of transforming growth factor beta2 in intestinal epithelial cells by regulation of proteins associated with stress and endotoxin responses . PLoS ONE 10 , e12 170. Izumi T. and Shimizu T. (1995 ) Platelet-activating factor receptor: gene expression and signal transduction . Biochimica et Biophysica Acta 1259 , 317 -333 8541341 171. De Plaen I.G. (2000 ) Lipopolysaccharide activates nuclear factor kappaB in rat intestine: role of endogenous platelet-activating factor and tumour necrosis factor . British Journal of Pharmacology 129 , 307 -314 10694237 172. Hsueh W. (2003 ) Neonatal necrotizing enterocolitis: clinical considerations and pathogenetic concepts . Pediatric and Developmental Pathology 6 , 6 -23 12424605 173. MacKendrick W. (1993 ) Increase in plasma platelet-activating factor levels in enterally fed preterm infants . Biology of the Neonate 64 , 89 -95 8260551 174. Amer M.D. (2004 ) Platelet-activating factor concentration in the stool of human newborns: effects of enteral feeding and neonatal necrotizing enterocolitis . Biology of the Neonate 85 , 159 -166 14646336 175. Caplan M. (1990 ) Serum PAF acetylhydrolase increases during neonatal maturation . Prostaglandins 39 , 705 -714 2371412 176. Moya F.R. (1994 ) Platelet-activating factor acetylhydrolase in term and preterm human milk: a preliminary report . Journal of Pediatric Gastroenterology and Nutrition 19 , 236 -239 7815247 177. Lu J. (2010 ) Dual roles of endogenous platelet-activating factor acetylhydrolase in a murine model of necrotizing enterocolitis . Pediatric Research 68 , 225 -230 20531249 178. Caplan M.S. (1997 ) The role of recombinant platelet-activating factor acetylhydrolase in a neonatal rat model of necrotizing enterocolitis . Pediatric Research 42 , 779 -783 9396557 179. Caplan M.S. (1997 ) The platelet-activating factor receptor antagonist WEB 2170 prevents neonatal necrotizing enterocolitis in rats . Journal of Pediatric Gastroenterology and Nutrition 24 , 296 -301 9138176 180. Ewer A.K. (2004 ) The role of platelet activating factor in a neonatal piglet model of necrotising enterocolitis . Gut 53 , 207 -213 14724152 181. Ng P.C. , Ma T.P. and Lam H.S. (2015 ) The use of laboratory biomarkers for surveillance, diagnosis and prediction of clinical outcomes in neonatal sepsis and necrotising enterocolitis . Archives of Disease in Childhood. Fetal and Neonatal Edition 100 , F448 -F452 25555389 182. Bhatia A.M. (2014 ) Cytokine levels in the preterm infant with neonatal intestinal injury . American Journal of Perinatology 31 , 489 -496 23966125 183. Ng P.C. (2006 ) Early prediction of sepsis-induced disseminated intravascular coagulation with interleukin-10, interleukin-6, and RANTES in preterm infants . Clinical Chemistry 52 , 1181 -1189 16613997 184. Gregory K.E. (2014 ) Urinary intestinal fatty acid binding protein predicts necrotizing enterocolitis . Journal of Pediatrics 164 , 1486 -1488 24636854 185. Edelson M.B. (1999 ) Plasma intestinal fatty acid binding protein in neonates with necrotizing enterocolitis: a pilot study . Journal of Pediatric Surgery 34 , 1453 -1457 10549746 186. Schurink M. (2015 ) Intestinal fatty acid-binding protein as a diagnostic marker for complicated and uncomplicated necrotizing enterocolitis: a prospective cohort study . PLoS ONE 10 , e12 187. Harris M.C. (2005 ) Cytokine elaboration in critically ill infants with bacterial sepsis, necrotizing entercolitis, or sepsis syndrome: correlation with clinical parameters of inflammation and mortality . Journal of Pediatrics 147 , 462 -468 16227031 188. Benkoe T. (2014 ) Interleukin-8 predicts 60-day mortality in premature infants with necrotizing enterocolitis . Journal of Pediatric Surgery 49 , 385 -389 24650462 189. Benkoe T. (2012 ) Interleukin 8 correlates with intestinal involvement in surgically treated infants with necrotizing enterocolitis . Journal of Pediatric Surgery 47 , 1548 -1554 22901915 190. Chaaban H. (2010 ) Inter-alpha inhibitor protein level in neonates predicts necrotizing enterocolitis . Journal of Pediatrics 157 , 757 -761 20955849 191. Yoon J.M. (2014 ) Fecal calprotectin concentration in neonatal necrotizing enterocolitis . Korean Journal of Pediatrics 57 , 351 -356 25210522 192. Albanna E.A. , Ahmed H.S. and Awad H.A. (2014 ) Stool calprotectin in necrotizing enterocolitis . Journal of Clinical Neonatoloy 3 , 16 -19 193. Bin-Nun A. (2015 ) Rapid fecal calprotectin (FC) analysis: point of care testing for diagnosing early necrotizing enterocolitis . American Journal of Perinatology 32 , 337 -342 25111039 194. Ng P.C. (2010 ) Host-response biomarkers for diagnosis of late-onset septicemia and necrotizing enterocolitis in preterm infants . Journal of Clinical Investigation 120 , 2989 -3000 20592468 195. Ng E.W. (2013 ) Gut-associated biomarkers L-FABP, I-FABP, and TFF3 and LIT score for diagnosis of surgical necrotizing enterocolitis in preterm infants . Annals of Surgery 258 , 1111 -1118 23470582

PMC005xxxxxx/PMC5001535.txt

==== Front Sci RepSci RepScientific Reports2045-2322Nature Publishing Group srep3198610.1038/srep3198627558955ArticleMolecular hydrogen suppresses activated Wnt/β-catenin signaling Lin Yingni 1Ohkawara Bisei 1Ito Mikako 1Misawa Nobuaki 2Miyamoto Kentaro 1Takegami Yasuhiko 1Masuda Akio 1Toyokuni Shinya 2Ohno Kinji a11 Division of Neurogenetics, Center for Neurological Diseases and Cancer, Nagoya University Graduate School of Medicine, Nagoya, Japan2 Department of Pathology and Biological Responses, Graduate school of Medicine, Nagoya University Graduate School of Medicine, Nagoya, Japana ohnok@med.nagoya-u.ac.jp25 08 2016 2016 6 3198616 03 2016 01 08 2016 Copyright © 2016, The Author(s)2016The Author(s)This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/Molecular hydrogen (H2) is effective for many diseases. However, molecular bases of H2 have not been fully elucidated. Cumulative evidence indicates that H2 acts as a gaseous signal modulator. We found that H2 suppresses activated Wnt/β-catenin signaling by promoting phosphorylation and degradation οf β-catenin. Either complete inhibition of GSK3 or mutations at CK1- and GSK3-phosphorylation sites of β-catenin abolished the suppressive effect of H2. H2 did not increase GSK3-mediated phosphorylation of glycogen synthase, indicating that H2 has no direct effect on GSK3 itself. Knock-down of adenomatous polyposis coli (APC) or Axin1, which form the β-catenin degradation complex, minimized the suppressive effect of H2 on β-catenin accumulation. Accordingly, the effect of H2 requires CK1/GSK3-phosphorylation sites of β-catenin, as well as the β-catenin degradation complex comprised of CK1, GSK3, APC, and Axin1. We additionally found that H2 reduces the activation of Wnt/β-catenin signaling in human osteoarthritis chondrocytes. Oral intake of H2 water tended to ameliorate cartilage degradation in a surgery-induced rat osteoarthritis model through attenuating β-catenin accumulation. We first demonstrate that H2 suppresses abnormally activated Wnt/β-catenin signaling, which accounts for the protective roles of H2 in a fraction of diseases. ==== Body The effects of H2 have been reported in 166 disease models and human diseases1. Prominent effects have been observed especially in oxidative stress-mediated diseases and inflammatory diseases. H2 was first reported to be a selective scavenger of •OH and peroxynitrite2. Cumulative evidence, however, suggests that H2 functions as a signaling modulator345. In this study, we dissected the effects of H2 on Wnt/β-catenin signaling. Wnt/β-catenin signaling controls cell proliferation and differentiation by regulating expression of target genes. In the absence of Wnt ligands, β-catenin is steadily phosphorylated by casein kinase 1 (CK1) at Ser45 and glycogen synthase kinase 3 (GSK3) at Ser33/Ser37/Thr41 at its N-terminus in a degradation complex assembled by Axin1 and adenomatous polyposis coli (APC), and is subsequently degraded through the β-transducin repeat-containing protein (β-TrCP)-mediated ubiquitin/proteasome pathway6. Wnt ligands or GSK3 inhibitors [lithium chloride (LiCl) and 6-bromoindirubin-3′-oxime (BIO)] suppress phosphorylation and degradation of β-catenin. Consequently, β-catenin accumulates in the cytoplasm and then translocates into the nucleus to interact with T-cell factor/lymphoid enhancing factor (TCF/LEF) to activate transcription of the Wnt/β-catenin target genes. Aberrant activation of Wnt/β-catenin signaling is associated with a number of diseases including cancers and degenerative diseases7. Osteoarthritis (OA) is characterized by degradation of extracellular matrix (ECM) molecules, loss of articular cartilages, and formation of osteophytes. Development and aggravation of OA are associated with abnormal activation of Wnt/β-catenin signaling8910. H2 is beneficial for musculoskeletal diseases including inflammatory and mitochondrial myopathies11, microgravity-induced bone loss12, post-ovariectomy osteopenia13, rheumatoid arthritis (RA)1415, and psoriasis-associated arthritis16. However, no study has demonstrated the effect of H2 on OA to the best of our knowledge. In this study, we observed that H2 inhibited Wnt/β-catenin signaling activated by Wnt3a, LiCl, or BIO in L and HeLa cells. H2 promoted phosphorylation, ubiquitination, and subsequent degradation of β-catenin without directly affecting mRNA level of β-catenin. The effect of H2 required CK1/GSK3-phosphorylation sites on β-catenin, the CK1/GSK3 activities, as well as APC and Axin1 activities. We confirmed the suppressive effect of H2 on Wnt/β-catenin signaling in chondrocytes and observed a protective effect of H2 on OA progression. We report that H2 is an inhibitor for activated Wnt/β-catenin signaling, which provides additional evidence that H2 is a gaseous signal modulator. Results H2 suppresses activated Wnt/β-catenin signaling In order to examine whether H2 affects Wnt/β-catenin signaling, we first conducted Topflash luciferase reporter assay in L cells with 10% H2 or 10% nitrogen (N2) gas. Topflash luciferase reporter plasmid carries 8 copies of TCF-binding sites in the promoter region and the firefly luciferase cDNA to quantify activation of Wnt/β-catenin signaling. Addition of Wnt3a or a GSK3 inhibitor, LiCl or BIO, to the culture medium for 24 h increased Topflash reporter activity, which, however, was attenuated by H2 (Fig. 1a). Similar effects of H2 on the Wnt/β-catenin signaling were also observed in HeLa cells (Supplementary Fig. S1a), suggesting that H2 suppressed activation of Wnt/β-catenin signaling in different cell lines. We also examined the expression of an endogenous target gene of Wnt/β-catenin signaling, Axin2, and found the suppressive effect of H2 on Wnt3a-, LiCl-, or BIO-induced upregulation of Axin2 mRNA in L cells (Fig. 1b). Then, we examined whether H2 decreases the level of β-catenin, the transcriptional co-activator, by Western blotting. H2 reduced accumulation of endogenous β-catenin induced by Wnt3a, LiCl, or BIO (Fig. 1c–e), as well as accumulation of exogenous myc-β-catenin (Supplementary Fig. S1c) in L cells. Consistently, the nuclear accumulation of β-catenin induced by Wnt3a, LiCl, or BIO was also decreased by H2 in L cells (Supplementary Fig. S1d). H2, however, failed to suppress basal expression level of β-catenin in HeLa cells (Supplementary Fig. S1e). Time course analysis revealed that the suppressive effect of H2 on β-catenin accumulation was prominent in the first 6 h in L cells (Fig. 1f and supplementary Fig. S1f). In all the experiments stated above, we used 10% N2 gas as a control for 10% H2 gas to match O2 concentrations. We observed that minimal reduction of O2 concentration by 10% N2 gas had no effect on the β-catenin level (Supplementary Fig. S1g). Therefore, we consistently used 10% N2 gas in a control group in the following experiments. We previously reported a prominent protective effect of administration of H2 water and intermittent inhalation of H2 gas, but not continuous inhalation of H2 gas, in a rat model of Parkinson’s disease17. Therefore, we examined whether intermittent exposure to H2 gas has a more suppressive effect on Wnt/β-catenin signaling. We then followed the protocol of intermittent administration of H2 gas as described previously17. Briefly, L cells were exposed to 10% H2 gas for 15 min followed by air for 45 min using a time controller, and the cycle was repeated for 24 h. We added 5% CO2 throughout the cycle. We found that intermittent H2 treatment was less effective than continuous H2 treatment on suppressing β-catenin level (Fig. 1g). H2 has been reported to inhibit mitogen-activated protein (MAP) kinase signaling in cell lines and rodent disease models345. A previous report shows that inhibition of JNK, but not of ERK or p38 MAP kinase, decreases Wnt3a-induced β-catenin accumulation18. As we found that H2 suppressed Wnt/β-catenin signaling, we next asked whether H2 suppresses Wnt/β-catenin signaling through inhibition of JNK signaling. We pretreated L cells with 40 μM SP600125, which efficiently inhibited JNK-mediated phosphorylation of c-Jun (Fig. 1h). L cells were then added with Wnt3a or BIO with 10% H2 or 10% N2 gas for 1 h. SP600125 decreased Wnt3a- or BIO-induced β-catenin accumulation, but did not abolish the suppressive effect of H2 on β-catenin accumulation (Fig. 1h), suggesting that the effect of H2 on Wnt/β-catenin signaling is independent of JNK signaling. Then, we asked whether the suppressive effect of H2 on Wnt/β-catenin signaling is operational in vivo. A previous study shows that starvation induces nuclear accumulation of β-catenin in the liver in mice19. Consistently, we detected accumulation of β-catenin in the liver of starved mice. Ad libitum oral intake of H2 water suppressed β-catenin accumulation in the liver in starved but not fed mice (Supplementary Fig. S1h). All these data point to the notion that H2 suppresses activation of Wnt/β-catenin signaling by reducing β-catenin accumulation. H2 promotes β-catenin degradation We found that mRNA levels of β-catenin remained unchanged by H2 treatment in the absence and presence of Wnt/β-catenin signaling activators (Wnt3a, BIO, or LiCl) in L cells (Fig. 2a) and HeLa cells (Supplementary Fig. S2a), indicating that H2 acts on β-catenin protein synthesis/degradation rather than its gene expression. To further dissect this hypothesis, we used cycloheximide (CHX) to block protein biosynthesis of β-catenin and conducted the CHX chase assay to determine whether H2 accelerates the degradation rate of β-catenin. Because endogenous β-catenin is hard to be detected in L cells, we expressed myc-β-catenin in L cells and found that H2 accelerated degradation of myc-β-catenin (Fig. 2b). We additionally performed the CHX chase assay in HeLa cells and found that H2 modestly accelerated degradation of endogenous β-catenin (Supplementary Fig. S2b,c). Intracellular β-catenin is degraded by the ubiquitin-proteasome system. Consequently, we used MG132 for proteasome inhibition to determine whether the ubiquitin–proteasome system is involved in H2-induced down-regulation of β-catenin. We found that MG132 minimized the suppressive effect of H2 on BIO-induced β-catenin accumulation (Fig. 2c). We also observed that H2 facilitated ubiquitination of β-catenin (Fig. 2d). These results suggest that H2 enhances proteasome-mediated β-catenin degradation. H2 enhances β-catenin phosphorylation In canonical Wnt/β-catenin signaling, β-catenin is sequentially phosphorylated at Ser45 by CK1 and at Ser33/Ser37/Thr41 by GSK3 in a complex with Axin1 and APC. Phosphorylation of these residues is required to trigger proteasome-mediated β-catenin degradation20. We then tested whether the effect of H2 was on either phosphorylation or degradation of β-catenin. Since the level of phospho-β-catenin is dependent on the total amount of β-catenin, we added MG132 to block β-catenin degradation in the proteasome pathway. H2 increased phosphorylation of β-catenin at Ser45 and at Ser33/Ser37/Thr41 with the treatment of MG132 (Fig. 3a). The concentration of BIO (2 μM) used in these studies was likely to partially inhibit GSK3 activity in L cells, because the phosphorylations on Ser33/Ser37/Thr41 of β-catenin were detected to some extent even in the presence of 2 μM BIO (Fig. 3a). We thus increased BIO concentrations to examine whether the effect of H2 requires GSK3 activity. We found that H2 was able to decrease β-catenin accumulation induced by low BIO concentrations (1 μM and 2 μM) but not by high BIO concentrations (4 μM and 6 μM) (Fig. 3b). These results indicate that the inhibitory effect of H2 on activated Wnt/β-catenin signaling requires GSK3 activity. Then, we examined the effect of H2 in two cell lines, HCT-116 colon cancer cells and HepG2 liver carcinoma cells, which carry mutations in CTNNB1 encoding β-catenin at its phosphorylation sites. In accordance with the guidelines by the Human Genome Organization (HUGO) (http://www.genenames.org/), CTNNB1 and Ctnnb1 are used in this communication to indicate the genes for β-catenin in human and mouse, respectively. HCT-116 cells are heterozygous for a deletion mutation at Ser45, the CK1 regulatory site, in the β-catenin gene21. HepG2 cells are heterozygous for an inframe-deletion lacking a potential GSK3-phosphorylation site in the β-catenin gene22. We found that the inhibitory effect of H2 on Wnt3a- or BIO-induced β-catenin accumulation in HCT-116 cells (Fig. 3c) was less than those observed in L cells (Fig. 1c,e) and HeLa cells (Supplementary Fig. S1e). The reduction of the H2 effect in HCT-116 cells was likely due to the presence of a deletion at Ser45. In HepG2 cells, Wnt3a or BIO failed to increase the truncated β-catenin and H2 showed no suppressive effect on the truncated β-catenin (Fig. 3d). In contrast, wild-type β-catenin in HepG2 cells was increased by Wnt3a or BIO, and was suppressed by H2 (Fig. 3d). To further confirm these results, we transfected L cells with HA-wild-type-β-catenin (HA-WT-β-catenin) or two β-catenin mutants (HA-ΔN-β-catenin and HA-S37A-β-catenin). HA-ΔN-β-catenin lacks the whole N-terminal region of β-catenin phosphorylated by CK1 and GSK323. Another mutant HA-S37A-β-catenin has a Ser-to-Ala mutation at one of the GSK3-phosphorylated sites23. Enhanced Wnt/β-catenin signaling activity in L cells expressing HA-WT-β-catenin, but not HA-ΔN-β-catenin or HA-S37A-β-catenin, was decreased by H2 (Fig. 3e). These results indicate that CK1/GSK3-phosphorylation sites of β-catenin and the dual CK1/GSK3 activity are required for H2-medicated enhanced β-catenin degradation. Next, we analyzed whether H2 regulates phosphorylation of glycogen synthase (GS), which is another substrate of GSK3. We found that H2 did not affect GSK3-mediated phosphorylation of GS in L cells (Supplementary Fig. S3a). Thus, H2 specifically enhances GSK3-mediated phosphorylation of β-catenin, but not GS. In addition to kinase activities, phosphorylation status of β-catenin also depends on phosphatase activities. Okadaic acid, an inhibitor for protein phosphatase 2A (PP2A), promotes β-catenin hyperphosphorylation on serine-threonine residues24. To check whether PP2A mediates the effect of H2 on β-catenin degradation, L cells were pretreated with 30 nM okadaic acid and were added with BIO with 10% H2 or 10% N2 gas for 12 h. We observed that okadaic acid could not abrogate the suppressive effect of H2 on β-catenin accumulation in BIO-treated cells (Fig. 3f), indicating that H2 promoted β-catenin degradation independent of PP2A. In canonical Wnt/β-catenin signaling, two scaffold proteins, APC and Axin1, bind to β-catenin to form a degradation complex, and facilitate phosphorylation and ubiquitination of β-catenin. We then investigated whether APC, Axin1, or both are involved in H2-medicated β-catenin degradation. First, in HT-29 human colon cancer cells carrying truncated APC21, H2 failed to decrease β-catenin levels (Fig. 4a). Second, siRNA-mediated knock-down of APC in L cells abrogated the suppressive effect of H2 on β-catenin accumulation induced by BIO (Fig. 4b, Supplementary Fig. S3b,c). Third, similar to APC knock-down, knock-down of Axin1 also attenuated the effect of H2 (Fig. 4c and Supplementary Fig. S3d). In addition, H2 did not affect mRNA levels of APC and Axin1 (Supplementary Fig. S3e,f). Fourth, we conducted co-immunoprecipitation assay to examine the effect of H2 on the β-catenin degradation complex. To directly observe the effect of H2 on the degradation complex, we added MG132 to block β-catenin degradation in the proteasome pathway. We detected that H2 enhanced phosphorylation and ubiquitination of β-catenin in Axin1-immunoprecipitates, but had no effect on the interactions of Axin1 with β-catenin, APC, and GSK3 (Supplementary Fig. S3g,h). Taken together, H2 enhances phosphorylation of β-catenin mediated by CK1/GSK3 in the degradation complex formed by APC and Axin1, which subsequently leads to enhanced ubiquitination and degradation of β-catenin. H2 suppresses Wnt/β-catenin signaling in chondrocytes and protects against cartilage degradation in osteoarthritis (OA) Abnormal activation of Wnt/β-catenin signaling has been reported to be involved in the development and aggravation of OA8, which is characterized by degradation of ECM molecules, loss of articular cartilages, and formation of osteophytes. We then examined whether H2 inhibits Wnt/β-catenin signaling in HCS-2/8 human chondrosarcoma cells and human osteoarthritic chondrocyte (OAC) cells. We first confirmed that H2 downregulated Topflash reporter activity in HCS-2/8 cells stimulated by Wnt3a or BIO (Fig. 5a). Similarly, in human OAC cells stimulated by Wnt3a or BIO, H2 decreased β-catenin accumulation (Fig. 5b, Supplementary Fig. S4a,b) and suppressed mRNA levels of Axin2 (Fig. 5c, Supplementary Fig. S4a,b), but not of β-catenin (Supplementary Fig. S4c). Chondrocytes produce and maintain ECM, which mostly consists of collagens and proteoglycans. Dysfunction of articular chondrocytes in OA disturbs synthesis of ECM and enhances degradation of ECM. SRY-box 9 (Sox9), as a transcriptional factor, activates a number of cartilage ECM genes including COL2A1 and ACAN25 and plays an essential role in chondrogenic differentiation. Physiological interaction between Sox9 and β-catenin results in their mutual degradation by the ubiquitin/proteasome system26. Increased level of β-catenin protein promotes degradation of Sox9 protein and vice versa26. We found that, in human OAC cells, stimulation of Wnt/β-catenin signaling by Wnt3a or BIO downregulated SOX9 expression, which was partly rescued by H2 treatment (Fig. 5d, Supplementary Fig. S4a,b). H2 is thus expected to increase the level of Sox9 protein by suppressing Wnt/β-catenin signaling. This mechanism is likely to account for the H2-mediated up-regulation of Sox9 transcript, because Sox9 protein upregulates SOX9 mRNA expression via forming a positive feedback loop27. We also compared gene expressions of other chondrogenic markers including MMP3 encoding catabolic metalloproteinase 3, COL2A1 encoding collagen type II α1, and ACAN encoding aggrecan in 3 clones of OAC cells derived from 3 patients (clone 1 in Fig. 5e–g, clone 2 in Supplementary Fig. S4a, and clone 3 in Supplementary Fig. S4b). In the absence of H2, we observed that Wnt3a and BIO induced expressions of the MMP3, COL2A1, and ACAN genes in variable directions from clone to clone. For example, BIO increased MMP3 in clones 1 and 2, and decreased MMP3 in clone 3. H2 changed expressions of these genes in favorable (downregulation of MMP3, and upregulation of COL2A1 and ACAN) or unfavorable directions from clone to clone. To summarize, H2 suppresses Wnt/β-catenin signaling in human OAC cells, but the effect of H2 on ECM production cannot be evaluated due to variable responses of human OAC cells to Wtn3a and BIO. We also conducted Alcian blue staining in differentiated ATDC5 mouse chondrogenic cells, and found that both Wnt3a- and BIO-induced loss of proteoglycans were marginally reversed by H2 (Fig. 5h) without affecting proliferation of ATDC5 cells (Supplementary Fig. S4d). We next examined the effects of H2 on destabilization of the medial meniscus (DMM)-induced OA cartilage in rats, where Wnt/β-catenin signaling is abnormally activated8. Rats drank degassed water (control) or supersaturated H2 water (7 ppm) ad libitum from days 0 to 56 after DMM surgery. No difference in body weights was noted between the control and H2 groups (Supplementary Fig. S4e). We observed a tendency that 8-week administration of H2 water after surgery partially improved Safranin O-staining on the articular surface and minimally preserved the structure of articular cartilage (Fig. 6a). We also found that H2 decreased the percentage of β-catenin-positive cells and inhibited accumulation of β-catenin in cartilage chondrocytes in the DMM group without affecting β-catenin expression in sham groups (Fig. 6b). Additionally, we found that the expression of Sox9 was decreased in cartilage chondrocytes in the DMM group, which was partially rescued by H2 (Fig. 6c). These data indicate that H2 suppresses Wnt/β-catenin signaling in articular chondrocytes and partially ameliorates cartilage degradation and OA progression in a rat OA model. Discussion Although more than 321 studies have demonstrated beneficial effects of H2 on both animal models and human diseases1, molecular target(s) of H2 have not been fully elucidated. H2 was initially reported as a selective scavenger of hydroxyl radical (•OH)2. However, H2 is a stable gas and the reaction rate constant of H2 and •OH is in the order of 107 M−1·s−1, which is much lower than the reaction rate constants of •OH with other molecules (109 to 1010 M−1·s−1)28. Additionally, the breath H2 concentration comes to the baseline level in 30 min after taking 200 ml saturated H2 water in healthy individuals29. As the reaction rate constant of H2 is low and the dwell time of H2 in our body is short, H2 is unlikely to efficiently remove •OH. Thus, yet unidentified mechanisms should underlie the therapeutic effect of H2. A previous report showed that oral intake of H2 water increases gastric secretion of ghrelin in mice30. In addition, intermittent, but not continuous, inhalation of H2 gas ameliorates a rat model of Parkinson’s disease17. We have also shown that H2 alters signaling activities in mast cells and macrophages without directly scavenging reactive oxygen/nitrogen species34. In this study, we first showed that H2 inhibited endogenous β-catenin accumulation induced by Wnt3a and GSK3 inhibitors (LiCl and BIO), as well as exogenous β-catenin accumulation induced by a transgene, but had no effect on the basal endogenous β-catenin level. The effect of H2 was thus likely to be observed when Wnt/β-catenin signaling was abnormally activated. Then, we found that H2 increased β-catenin phosphorylation without attenuating the activity of PP2A, and accelerated β-catenin degradation without decreasing its mRNA level. Given that complete GSK3 inhibition and mutations at the CK1- or GSK3-phosphorylation sites of β-catenin nullified the H2 effect, the dual CK1 and GSK3 activities were required for H2-medicated β-catenin degradation. As H2 had no effect on GSK3-mediated phosphorylation of GS, which is another substrate of GSK3, H2 was likely to enhance GSK3 activity only in the degradation complex with APC and Axin1, but had no direct effect on GSK3 itself. Enhancement of CK1-mediated β-catenin phosphorylation by H2 is also in accordance with the assumption that the H2 works on the degradation complex, not on CK1 or GSK3. Additionally, as both CK1 and GSK3 are constitutively active in the resting state and as no allosteric activators are known, H2 is unlikely to be able to allosterically upregulate activities of CK1 and GSK3. To further confirm that the effect of H2 is on phosphorylation and not ubiquitination of β-catenin, we examined the effect of H2 on HT-29 human colon cancer cells harboring truncated APC. The truncated APC retains all three 15-aa repeats and three of seven 20-aa repeats, and lacks all three Axin1-binding sites and four of seven 20-aa repeats. A previous study shows that the three 20-aa repeats that are retained in the truncated APC are sufficient for β-catenin ubiquitination, and the interaction between the truncated APC and E3-ligase β-TrCP is not affected in HT-29 cells21. The Axin1-binding sites, which are deficient in HT-29 cells, markedly facilitate β-catenin phosphorylation, but are not required for β-catenin ubiquitination2131. Thus, the low degradation rate of β-catenin in HT-29 cells is due to reduced phosphorylation of β-catenin because of lack of Axin1-binding sites21. Lack of the effect of H2 in HT-29 cells thus indicates that H2 enhances phosphorylation but not ubiquitination of β-catenin. In addition, our observation that knock-down of either APC or Axin1 minimizes the suppressive effect of H2 on β-catenin accumulation indicates that both APC and Axin1 are required for the H2 effect. We next examined the protein-protein interactions within the degradation complex using co-immunoprecipitation assay, and found that H2 had no effect on the interactions among Axin1, APC, β-catenin, and GSK3, but still could enhance CK1/GSK3-mediated phosphorylation of β-catenin. Water structure plays an active role in protein folding, enzyme catalysis, and cell signaling32. Alteration of the water structure will thus have a direct effect on biological systems. The H2O-H2 interaction accompanied by charge transfer is stronger than predicted33, which might affect the water structure. Additionally, as H2 can easily diffuse into every cellular compartment, solubilized H2 interfacing with other biomolecules possibly alters their hydration structures and subsequently their activities. In Wnt/β-catenin signaling, H2 may dynamically enhance protein-protein interactions during the process of β-catenin phosphorylation, which, however, could not be detected in our co-immunoprecipitation assays because of a rapid decrease of H2 concentration in a test tube34. Alternatively, H2 may modulate interaction of one of molecules constituting the degradation complex with a yet unidentified molecule, which leads to enhancement of β-catenin phosphorylation. We have previously shown that H2 prevents degranulation of mast cells not by a radical scavenging effect, but by downregulating NADPH oxidase activity via attenuating the phosphorylation of the FcεRI-associated Lyn and its downstream signal transduction molecules (ERK1/2, JNK, p38 MAP kinase, and Akt) without affecting other signaling molecules [apoptosis signal-regulating kinase 1 (ASK1) and nuclear factor of kappa light polypeptide gene enhancer in B-cells inhibitor, alpha (IκB-α)]3. Similarly, in RAW264 macrophage cells, H2 reduces lipopolysaccharide/interferon γ (LPS/IFNγ)-induced nitric oxide (NO) release by suppressing the activity of inducible nitric oxide synthase (iNOS) not by affecting NADPH oxidase activity but by suppressing phosphorylation of ASK1 and its downstream signaling molecules (p38 MAP kinase, JNK, and IκB-α)4. We here demonstrate that H2 suppresses Wnt/β-catenin signaling without scavenging hydroxyl radicals or peroxynitrite. Inhibition of JNK, but not of ERK or p38 MAP kinase, suppresses Wnt3a-induced β-catenin accumulation18. However, the effect of H2 was unlikely to be dependent on JNK because inhibition of JNK did not attenuate the effect of H2. Akt phosphorylates GSK3 on Ser9 to inactivate it and inhibition of Akt activates GSK335. Since we found that H2 had no effect on GSK3-mediated GS phosphorylation, it is unlikely that H2 activates GSK3 via suppression of Akt. Accordingly, H2 is able to specifically modulate signaling pathways in cell- and disease-specific manners. In this study, we showed that H2 could not decrease abnormally elevated β-catenin in HCT-116, HepG2, and HT-29 cancer cells. Thus, H2 is unlikely to suppress proliferation of these cells. Previous studies have shown that H2 inhibited cell proliferation of human tongue carcinoma cells HSC-436 and human fibrosarcoma cells HT-108036. Similarly, a combination of H2 and 5-fluorouracil induced apoptosis of colon 26 cells37. Additionally, H2 suppressed the expression of vascular endothelial growth factor (VEGF) in human lung adenocarcinoma cells A54938. Although the effect of H2 on Wnt/β-catenin signaling has not been dissected in these cells, H2 might have achieved the tumor-suppressing effects by suppressing Wnt/β-catenin signaling. Alternatively, H2 might have suppressed cell proliferation by modulating other signaling pathway(s) and/or molecules. As stated in introduction, abnormal activation of Wnt/β-catenin signaling deteriorates OA and thus Wnt/β-catenin signaling can be a therapeutic target for OA. In this study, we first show that H2 suppresses Wnt/β-catenin signaling in human OAC cells and also rescues Wnt3a- or BIO-induced loss of proteoglycan in differentiated ATDC5 chondrogenic cells. Additionally, H2 tended to protect against cartilage degradation in the DMM-induced OA model in rats, although statistical significance was not observed. The insignificant effect of H2 on the OA model may be partly accounted for by a limited effect of H2 on Wnt/β-catenin signaling in cartilage chondrocytes. Alternatively, the DMM-induced OA progression is accelerated by other signaling pathways that are insensitive to H2. In contrast to the deleterious effect of aberrantly activated Wnt/β-catenin signaling in OA, excessive inhibition of Wnt/β-catenin signaling also worsens OA39, possibly due to an essential role of Wnt/β signaling in cartilage development and homeostasis40. As H2 did not change the levels of endogenous β-catenin and Sox9 on the sham-operated side, H2 is expected to have no effect on cartilage development and homeostasis. Our results suggest that H2 may be able to attenuate OA progression in humans. However, other signaling pathways and/or molecules, including transforming growth factor β (TGF-β) signaling and inflammatory responses, are also involved in OA development and progression41. In this study, we could not exclude the possibility that H2 modulated other signaling pathways or molecules to ameliorate OA. We present that H2 suppresses abnormally activated Wnt/β-catenin signaling, which plays pivotal roles in diverse pathophysiologic processes, by enhancing β-catenin phosphorylation in the degradation complex (Fig. 7). For a considerable fraction of the 166 disease models and human diseases, for which the effects of H2 have been documented1, H2 is likely to have exerted beneficial effects by suppressing Wnt/β-catenin signaling. Materials and Methods Cell culture with H2 gas Cells were cultured in a culture dish in a 560-ml closed plastic box that was covered with aluminum and humidified with water at the base of the box. The box was put in a convection incubator (SLI-221, EYELA). We adjusted the temperature of the incubator to make the temperature inside the box 37 °C using an electronic thermometer. In the H2 group, H2 gas (3 ml/min or 6 ml/min) was mixed with CO2-added air (5% CO2 and 95% air, 60 ml/min) to make 5% or 10% H2 gas, which was delivered into the box via an afferent tube and out of the box via an efferent tube connected to a draft chamber. In the control group, 3 ml/min or 6 ml/min of N2 gas was used instead of H2 gas to control for O2 concentrations. We measured the hydrogen concentration in the medium by equilibrating 1 ml medium with 100 ml of 100% N2 gas and by injecting 1 ml equilibrated gas into a gas chromatography connected to a semiconductor gas detector (EAGanalyzer GS-23, SensorTec). As shown in Supplemental Fig. S1i, H2 concentrations were detectable in culture medium in 2 min after administration of H2 gas. The H2 concentration in the culture medium stayed stable after 20 min of H2 administration. The amount of H2 dissolved in the culture medium doubled by increasing H2 concentration from 5% to 10%. The pH of the culture medium after culturing L cells for 24 h with 10% N2 or 10% H2 gas administration was 7.71 ± 0.03 and 7.67 ± 0.02 (mean and SEM, n = 3), respectively, with no statistical difference by Student’s t-test. Cell culture, chemicals, and reporter assay L, L Wnt3a, HeLa, HCT-116, and HT-29 cells were obtained from ATCC. HepG2 and ATDC5 cells were from RIKEN BioResource Center. HCS-2/8 cells were kindly provided by Dr. Masaharu Takigawa at Okayama University. Studies using human OAC cells were approved by the Ethical Review Committee of Nagoya University Graduate School of Medicine, and were performed in accordance with the relevant guidelines by MHLW, Japan. After a written informed consent was given, OAC cells were obtained from patients who underwent total joint replacement for severe knee OA. L, L Wnt3a, HeLa, HCT-116, HT-29, HepG2, HCS-2/6, and human OAC cells were cultured in the Dulbecco’s Modified Eagle’s medium (DMEM, Gibco) supplemented with 10% fetal bovine serum (FBS, Thermo Scientific). ATDC5 cells were cultured in DMEM/F12 (a mixture of Dulbecco’s modified Eagle’s medium and Ham’s F12 medium, Sigma–Aldrich) supplemented with 5% FBS. L and L Wnt3a cells were cultured for 4 d to make control conditioned medium (CM) and Wnt3a CM, respectively. Lithium chloride (LiCl), cycloheximide (CHX), N-benzyoloxycarbonyl (Z)-Leu-Leu-leucinal (MG132), dimethyl sulfoxide (DMSO), and SP600125 were purchased from Wako. Okadaic acid was bought from Research Biochemicals International (RBI), and 6-bromoindirubin-3′-oxime (BIO) was from Sigma-Aldrich. To quantify the canonical Wnt/β-catenin signaling activity, cells were transfected with the Topflash luciferase reporter plasmid (M50 Super 8 × Topflash plasmid, Addgene) and the Renilla luciferase plasmid (phRL-TK, Promega). L cells were transfected using Lipofectamine 2000 (Invitrogen), whereas HeLa and HCS-2/8 cells were transfected using FuGENE 6 (Roche). Twenty-four hours later, cells were treated with either 50% control CM, 50% Wnt3a CM, 30 mM LiCl, or 2–4 μM BIO in 10% H2 or 10% N2 gas for 24 h. Luciferase activity was measured in triplicate by the Dual Luciferase Reporter Assay System (Promega). Plasmids, siRNAs, and transfection A plasmid carrying myc-β-catenin (XE28 XBC 40) was kindly gifted from Dr. Takamasa Yamamoto at National Institute for Basic Biology. Plasmid carrying HA-WT-β -catenin, HA-ΔN-β-catenin, and HA-S37A-β-catenin were kindly provided by Dr. Eisuke Nishida at Kyoto University. L cells were transfected with plasmids using Lipofectamine 2000 (Invitrogen). Sequences of siRNAs against APC and Axin1 were adopted from previous studies4243 and were synthesized by Sigma-Aldrich. To knock-down APC, L cells were transfected with 280 pmol APC-targeting siRNA (APC-siRNA#1 or APC-siRNA#2) or control siRNA (Cont-siRNA) by Lipofectamine RNAiMax (Invitrogen). Forty-eight hours later, cells were treated with 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. To knock-down Axin1, we performed siRNA transfection in two consecutive days as described previously43. Briefly, L cells were transfected with 140 pmol Axin1-targeting siRNA (Axin1-siRNA) or Cont-siRNA by Lipofectamine RNAiMax, and on the second day we repeated the same steps. Twenty-four hours later, cells were treated with 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. Preparation of cell lysates and Western blotting Cells were washed with PBS twice and were harvested with PLC buffer containing 50 mM HEPES (pH 7.0), 150 mM NaCl, 10% glycerol, 1% TritonX-100, 1.5 mM MgCl2, 1 mM EGTA, 100 mM NaF, 10 mM sodium pyrophosphate, 1 μg/μl aprotinin, 1 μg/μl leupeptin, 1 μg/μl pepstatin A, 1 mM PMSF, 1 mM sodium orthovanadate, and the Phosphatase Inhibitor Cocktail (PhosSTOP, Roche). The lysates were incubated on ice for 15 min, sonicated for 10 sec, and then centrifuged at 20,600 × g at 4 °C for 15 min. Preparation of nuclear fraction is described in details in the Supplementary information. The total protein concentrations of the lysates were measured by Pierce 660 nm Protein Assay Reagent. Cell lysates were boiled for 5 min in 2 × Laemmli buffer, separated on a 7.5% or 10% SDS-polyacrylamide gel, and transferred to a polyvinylidene fluoride membrane (Immobilon-P, Millipore). Membranes were washed in Tris-buffered saline containing 0.05% Tween 20 (TBS-T) and blocked for 1 h at room temperature in TBS-T with 5% skim milk or 5% bovine serum albumin (BSA). The membranes were then incubated overnight at 4 °C with specific antibodies listed in Supplementary Table S1. The membranes were washed with TBS-T and incubated with secondary goat anti-mouse IgG (1: 5000, LNA931V/AG, GE Healthcare) or anti-rabbit IgG (1: 5000, LNA934V/AE, GE Healthcare) antibody conjugated to horseradish peroxidase (HRP) for 1 h at room temperature. The bound antibodies were detected with Amersham ECL Western blotting detection reagents (GE Healthcare), and the signal intensities were quantified with the ImageJ program. In vivo ubiquitination assay Proteins were immunoprecipitated from L cell lysates as shown in the Supplementary information. Gene expression analysis Total RNAs from cells were extracted by RNeasy Mini Kit (Qiagen) according to the manufacturer’s instructions and then reverse-transcribed into complementary DNA (cDNA) using an oligo-dT primer (Thermo Fisher Scientific) and ReverTraAce (Toyobo). Real-time quantitative PCR (qPCR) was performed with the LightCycler 480 (Roche Applied Science) using the SYBR Premix ExTaq (Takara Bio). The expression level of a specific gene was normalized by the levels of β2-microglobulin (β2-MG). PCR primers are shown in Supplementary Table S2. All real-time qPCR experiments were performed in triplicate. Co-immunoprecipitation Details are shown in the Supplementary information. Alcian blue staining Alcian blue staining in differentiated ATDC5 cells was performed as described previously with slight modifications8. Details are shown in the Supplementary information. Cell proliferation assay Cell proliferation of ATDC5 cells were estimated by a BrdU cell proliferation ELISA kit (Roche). Details are shown in the Supplementary information. Animal experiments and administration of supersaturated hydrogen water All animal studies were approved by the Animal Care and Use Committee of the Nagoya University Graduate School of Medicine, and were performed in accordance with the relevant guidelines by MEXT, Japan. Details of experiments with fed and starved mice are shown in the Supplemental information. For generation of osteoarthritis (OA) model, eight-week-old male Sprague-Dawley rats were purchased from Japan SLC. Rats were anesthetized with isopropentane. The OA model was generated by resection of the menisco-tibial ligament to destabilize medial meniscus (DMM) in the right knee. On the sham-operated left side, the skin and joint capsule was incised and sutured. Rats were randomly divided into 2 groups: the degassed water group (control) and the H2 water group with unlimited access to degassed water and H2 water after surgery, respectively. H2 water was freshly prepared every evening using Hydrogen Water 7.0 (Ecomo International), which was kindly provided by MiZ Co. Ltd. The concentration of dissolved H2 was 5 to 7 ppm, whereas the concentration of saturated H2 under the standard ambient temperature and pressure (SATP) is 1.6 ppm. The H2 concentrations in the glass vessel for rodents exponentially decreased with a half-life of 1.09 h34. As rodents drink water every hour at night, the average H2 concentration in the water that the rodents drank was predicted to be 1.66 ppm34. Rats were sacrificed 8 weeks after surgery. Tissues around the knees were fixed overnight in 4% paraformaldehyde at 4 °C, dehydrated, and embedded in paraffin. Safranin O and fast green stainings were performed on the sagittal sections. The modified Mankin histologic scores on both tibial and femoral sides of articular cartilages to estimate severity of OA were graded by 2 blinded investigators as described previously44. One well-cut and well-stained sagittal section in the medial region of the joint per joint was scored. Details of immunofluorescence staining of β-catenin and Sox9 are shown in the Supplementary information. Statistical analysis All values were presented as the mean and SEM. For in cellulo studies, values are normalized to those of cells treated with 50% control CM and 10% N2 gas, unless otherwise indicated. Statistical significance was estimated either by Student’s t-test or two-way repeated measures ANOVA test. Bonferroni correction was applied to Student’s t-test for multiple comparisons. P-values less than 0.05 were considered significant. Additional Information How to cite this article: Lin, Y. et al. Molecular hydrogen suppresses activated Wnt/β-catenin signaling. Sci. Rep. 6, 31986; doi: 10.1038/srep31986 (2016). Supplementary Material Supplementary Information We would like to thank Dr. Masaharu Takigawa at Okayama University for providing HCS-2/8 cells, Dr. Takamasa Yamamoto at National Institute for Basic Biology for providing the XE28 XBC 40 plasmid, and Dr. Eisuke Nishida at Kyoto University for providing HA-tagged wild-type β-catenin, ΔN-β-catenin and S37A-β-catenin plasmids. We also would like to thank MiZ Co. Ltd. for providing Hydrogen Water 7.0. Author Contributions K.O. and B.O. conceived the study. Y.L. and B.O. designed experiments. Y.L. performed most experiments with the help of B.O., M.I., A.M., Y.T., K.M. and N.M. conducted the Safranin O and fast green staining. Y.T. and K.M. blindly scored severity of OA. Y.L., B.O. and K.O. wrote the manuscript. All authors reviewed the manuscript. Figure 1 H2 suppresses Wnt/β-catenin signaling in L cells. (a,b) Cells were treated with control CM (Cont.), Wnt3a CM, 30 mM LiCl, or 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. The Wnt/β-catenin signaling activity was measured by Topflash luciferase reporter assay (n = 4) (a) and expression of Axin2 was quantified by qRT-PCR (n = 3–4) (b). *P < 0.05 and **P < 0.01 by Student’s t-test. (c–e) Cells were treated with pairs of CM (Cont.) and Wnt3a CM (c); 30 mM NaCl and 30 mM LiCl (d); and 0.02% DMSO and 2 μM BIO/0.02% DMSO (e) with 10% H2 or 10% N2 gas for 24 h. Representative Western blots are shown. (f) Cells were treated with 2 μM BIO with 10% H2 or 10% N2 gas for indicated periods of time. The upper panel shows representative Western blots and the lower panel shows densitometry of β-catenin/β-actin (n = 4). #P < 0.05 by two-way repeated measures ANOVA. **P < 0.01 by Student’s t-test with Bonferroni correction for each pair of H2 and N2. (g) Cells were treated with control CM (Cont.), Wnt3a CM, 30 mM LiCl, or 2 μM BIO with either 10% N2 gas, intermittent 10% H2 gas, or continuous 10% H2 gas for 24 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 3). *P < 0.05 and **P < 0.01 by Student’s t-test with Bonferroni correction. (h) Cells were pretreated with 40 μM SP600125 for 30 min. Cells were then added with control CM (Cont.), Wnt3a CM, or 2 μM BIO with 10% H2 or 10% N2 gas for 1 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 4). *P < 0.05 and **P < 0.01 by Student’s t-test. Figure 2 H2 promotes β-catenin degradation in L cells. (a) Cells were treated with control CM (Cont.), Wnt3a CM, 30 mM LiCl, or 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. Expression of Ctnnb1 encoding β-catenin was quantified by qRT-PCR (n = 3). (b) Cells transfected with myc-β-catenin (XE28 XBC plasmid) were exposed to a combination of 10 μg/ml cycloheximide (CHX) and 2 μM BIO with 10% H2 or 10% N2 gas for indicated periods of time. Representative Western blots are shown with densitometry of myc-β-catenin/β-actin (n = 4). Two groups were not statistically different by two-way repeated measures ANOVA. *P < 0.05 and **P < 0.01 by Student’s t-test with Bonferroni correction for each pair of H2 and N2. (c) Cells were treated with 2 μM BIO, 10 μM MG132, or a combination of both with 10% H2 or 10% N2 gas for 12 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 6). Note that mono-ubiquitinated β-catenin is visible in MG132-treated cells. *P < 0.05 by Student’s t-test. (d) Cells were treated with 10 μM MG132 in the presence or absence of 2 μM BIO with 10% H2 or 10% N2 gas for 12 h. Cells were then immunoprecipitated (IP) with antibodies against β-catenin or mouse control IgG (IgG). Blots were immunostained with antibodies against ubiquitin and β-catenin, and representative blots are shown. Figure 3 H2 facilitates β-catenin phosphorylation. (a) L cells were treated with 10 μM MG132 in the presence or absence of 2 μM BIO with 10% H2 or 10% N2 gas for 6 h. Cell lysates were then immunoprecipitated (IP) with an antibody against β-catenin or mouse control IgG (IgG). Blots were immunostained with an antibody against phospho-β-catenin (at Ser45), phospho-β-catenin (at Ser33/Ser37/Thr41) and β-catenin. Representative Western blots of IP samples and input samples were shown in the upper panels and lower panels, respectively. Arrows indicate mono-ubiquitinated phospho-β-catenin. (b) L cells were exposed to indicated concentrations of BIO with 10% H2 or 10% N2 gas for 24 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 4). *P < 0.05 by Student’s t-test with Bonferroni correction. (c,d) HCT-116 cells (c) and HepG2 cells (d) were incubated with control CM (Cont.), Wnt3a CM, or 2 μM BIO with 10% H2 or 10% N2 for 24 h. HCT-116 cells carry Ser45-deleted β-catenin in one allele. HepG2 cells carry truncated β-catenin in one allele. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 3). Wild-type and Ser45-deleted β-catenin were analyzed together for HCT-116 cells (c), whereas wild-type and truncated β-catenin were individually analyzed for HepG2 cells (d). *P < 0.05 by Student’s t-test. (e) L cells were transfected with wild-type (HA-WT-β-catenin) or mutant forms of HA-β-catenin (HA-ΔN-β-catenin and HA-S37A-β-catenin). Cell were then treated with 10% H2 or 10% N2 gas for 24 h. Wnt/β-catenin signaling activity was measured by Topflash luciferase reporter assay (n = 4). *P < 0.05 by Student’s t-test. (f) L cells were pretreated with 30 nM okadaic acid for 30 min. Cells were then added with 2 μM BIO with 10% H2 or 10% N2 gas for 12 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 4). *P < 0.05 by Student’s t-test. Figure 4 H2-mediated β-catenin degradation requires activities of APC and Axin1. (a) HT-29 cells carrying truncated APC were incubated with control CM (Cont.), Wnt3a CM, or 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 3). No statistical difference by Student’s t-test. (b) L cells were transfected with Cont-siRNA or APC-siRNA#2, and treated with 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 3). *P < 0.05 by Student’s t-test. (c) L cells were transfected with Cont-siRNA or Axin1-siRNA, and treated with 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 3). *P < 0.05 by Student’s t-test. Figure 5 H2 inhibits Wnt/β-catenin signaling in chondrocytes. (a) HCS-2/6 human chondrosarcoma cells were treated with control CM (Cont.), Wnt3a CM, or 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. The Wnt/β-catenin signaling activity was measured by Topflash luciferase reporter assay (n = 9). *P < 0.05 and **P < 0.01 by Student’s t-test. (b) Human OAC cells were treated with control CM (Cont.), Wnt3a CM, or 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. Representative Western blots are shown with densitometry of β-catenin/β-actin (n = 4). *P < 0.05 by Student’s t-test. (c–g) Human OAC cells (clone 1) were treated with control CM (Cont.), Wnt3a CM, or 2 μM BIO with 10% H2 or 10% N2 gas for 24 h. Expression of AXIN2 (n = 3) (c), SOX9 (n = 6) (d), MMP3 (n = 6) (e), COL2A1 (n = 3) (f), and ACAN (n = 3) (g) were quantified by qRT-PCR. *P < 0.05 by Student’s t-test. (h) Differentiated ATDC5 cells were treated with control CM (Cont.), Wnt3a CM, or 2 μM BIO with 10% H2 or 10% N2 gas for 48 h. Proteoglycans were stained with Alcian blue (upper panel) and quantified by measuring the optical intensity at 630 nm of the cell lysates (n = 9 for Cont.-treated and BIO-treated cells, n = 6 for Wnt3a-treated cells). *P < 0.05 by Student’s t-test. Non-significant P values less than 0.10 are indicated above each pair. Figure 6 H2 modestly ameliorates cartilage degradation and improves expressions of β–catenin and Sox9 in a rat osteoarthritis (OA) model. (a) OA phenotype was generated by DMM surgery in Wistar/ST rats. Degassed water (control) or H2 water was administered ad libitum for 8 weeks after surgery. Each rat had DMM and sham surgeries in the right and left knees, respectively. Knee joints were stained with Safranin O and fast green, and representative images are shown (scale bar = 100 μm). Upper and lower sides of the image are femoral and tibial bones, respectively. Each image was blindly graded by modified Mankin score to evaluate severity of OA (sham groups, n = 5; DMM groups, n = 6). Mean and SEM are plotted. No statistical difference by Student’s t-test. (b,c) Immunofluorescence staining with antibodies against β-catenin (b) and Sox9 (c) in rat articular cartilage. Representative immunostaining images (scale bar = 50 μm) are shown with the ratio of β-catenin-positive cells (n = 9) (b) and Sox9-positive cells (n = 9) (c) to DAPI-positive cells, and average fluorescence intensity of β-catenin (n = 9) (b) and Sox9 (n = 9) (c). Boxed regions in the merged images are magnified in the rightmost column to show nuclear translocation of β-catenin (scale bar = 50 μm) (b). Three fields of the tibial cartilage area per section and 3 different sections per group were analyzed. Mean and SEM are plotted (n = 9). Intensity of β-catenin and Sox9 is normalized to the mean of the sham operated side of rats taking degassed water. *P < 0.05 and **P < 0.01by Student’s t test. Figure 7 Schematic diagram showing that H2 promotes phosphorylation and degradation of β-catenin in activated Wnt/β-catenin signaling. β-catenin is phosphorylated at one site by CK1 and at three sites by GSK3 in the degradation complex composed of CK1, GSK3, APC, and Axin1. Wnt3a or BIO inhibits phosphorylation and ubiquitination of β-catenin, and increases the intracellular level of β-catenin, which drives Wnt/β-catenin signaling from the “off-state” to the “on-state”. H2 works on the degradation complex to enhance phosphorylation and degradation of β-catenin when Wnt/β-catenin signaling is activated by Wnt3a or BIO, and moves the “on-state” toward the “off-state”. H2 has no direct effect on the protein phosphatase 2A (not shown) or the ubiquitination process. ==== Refs Ichihara M. . Beneficial biological effects and the underlying mechanisms of molecular hydrogen - comprehensive review of 321 original articles . Med Gas Res 5 , 12 (2015 ).26483953 Ohsawa I. . Hydrogen acts as a therapeutic antioxidant by selectively reducing cytotoxic oxygen radicals . Nat Med 13 , 688 –694 (2007 ).17486089 Itoh T. . Molecular hydrogen suppresses FcεRI-mediated signal transduction and prevents degranulation of mast cells . Biochem Biophys Res Commun 389 , 651 –656 (2009 ).19766097 Itoh T. . Molecular hydrogen inhibits lipopolysaccharide/interferon γ-induced nitric oxide production through modulation of signal transduction in macrophages . Biochem Biophys Res Commun 411 , 143 –149 (2011 ).21723254 Sobue S. . Simultaneous oral and inhalational intake of molecular hydrogen additively suppresses signaling pathways in rodents . Mol Cell Biochem 403 , 231 –241 (2015 ).25707580 MacDonald B. T. , Tamai K. & He X. Wnt/β-Catenin Signaling: Components, Mechanisms, and Diseases . Dev Cell 17 , 9 –26 (2009 ).19619488 Luo J. . Wnt signaling and human diseases: what are the therapeutic implications? Lab Invest 87 , 97 –103 (2007 ).17211410 Takamatsu A. . Verapamil Protects against Cartilage Degradation in Osteoarthritis by Inhibiting Wnt/β-Catenin Signaling . PLoS One 9 , e92699 (2014 ).24658359 Zhu M. . Activation of β-Catenin Signaling in Articular Chondrocytes Leads to Osteoarthritis-Like Phenotype in Adult β-Catenin Conditional Activation Mice . J Bone Miner Res 24 , 12 –21 (2009 ).18767925 van den Bosch M. H. . Induction of Canonical Wnt Signaling by Synovial Overexpression of Selected Wnts Leads to Protease Activity and Early Osteoarthritis-Like Cartilage Damage . The American Journal of Pathology 185 , 1970 –1980 (2015 ).25976248 Ito M. . Open-label trial and randomized, double-blind, placebo-controlled, crossover trial of hydrogen-enriched water for mitochondrial and inflammatory myopathies . Med Gas Res 1 , 24 –24 (2011 ).22146674 Sun Y. , Shuang F. , Chen D. M. & Zhou R. B. Treatment of hydrogen molecule abates oxidative stress and alleviates bone loss induced by modeled microgravity in rats . Osteoporosis Int 24 , 969 –978 (2013 ). Guo J.-D. , Li L. , Shi Y.-M. , Wang H.-D. & Hou S.-X. Hydrogen water consumption prevents osteopenia in ovariectomized rats . Br J Pharmacol 168 , 1412 –1420 (2013 ).23121335 Ishibashi T. . Consumption of water containing a high concentration of molecular hydrogen reduces oxidative stress and disease activity in patients with rheumatoid arthritis: an open-label pilot study . Med Gas Res 2 , 1 –8 (2012 ).22273079 Ishibashi T. . Therapeutic efficacy of infused molecular hydrogen in saline on rheumatoid arthritis: A randomized, double-blind, placebo-controlled pilot study . Int Immunopharmacol 21 , 468 –473 (2014 ).24929023 Ishibashi T. . Improvement of psoriasis-associated arthritis and skin lesions by treatment with molecular hydrogen: A report of three cases . Mol Med Rep 12 , 2757 –2764 (2015 ).25936373 Ito M. . Drinking hydrogen water and intermittent hydrogen gas exposure, but not lactulose or continuous hydrogen gas exposure, prevent 6-hydorxydopamine-induced Parkinson’s disease in rats . Med Gas Res 2 , 15 –15 (2012 ).22608009 Vikram A. . Canonical Wnt signaling induces vascular endothelial dysfunction via p66Shc-regulated reactive oxygen species . Atertio Thromb Vasc Biol 34 , 2301 –2309 (2014 ). Liu H. . Wnt signaling regulates hepatic metabolism . Sci Signal 4 , ra6 (2011 ). Liu C. . Control of β-Catenin Phosphorylation/Degradation by a Dual-Kinase Mechanism . Cell 108 , 837 –847 (2002 ).11955436 Yang J. . Adenomatous Polyposis Coli (APC) Differentially Regulates β-Catenin Phosphorylation and Ubiquitination in Colon Cancer Cells . J Biol Chem 281 , 17751 –17757 (2006 ).16798748 Coste A. d. L. . Somatic mutations of the β-catenin gene are frequent in mouse and human hepatocellular carcinomas . Proc Natl Acad Sci USA 95 , 8847 –8851 (1998 ).9671767 Imajo M. , Miyatake K. , Iimura A. , Miyamoto A. & Nishida E. A molecular mechanism that links Hippo signalling to the inhibition of Wnt/β-catenin signalling . The EMBO Journal 31 , 1109 –1122 (2012 ).22234184 Serres M. . Hyperphosphorylation of β-Catenin on Serine-Threonine Residues and Loss of Cell–Cell Contacts Induced by Calyculin A and Okadaic Acid in Human Epidermal Cells . Exp Cell Res 231 , 163 –172 (1997 ).9056423 Oh C.-d. . SOX9 Regulates Multiple Genes in Chondrocytes, Including Genes Encoding ECM Proteins, ECM Modification Enzymes, Receptors, and Transporters . PLoS One 9 , e107577 (2014 ).25229425 Akiyama H. . Interactions between Sox9 and β-catenin control chondrocyte differentiation . Genes Dev 18 , 1072 –1087 (2004 ).15132997 Kumar D. & Lassar A. B. The Transcriptional Activity of Sox9 in Chondrocytes Is Regulated by RhoA Signaling and Actin Polymerization . Mol Cell Biol 29 , 4262 –4273 (2009 ).19470758 Ohno K. , Ito M. , Ichihara M. & Ito M. Molecular Hydrogen as an Emerging Therapeutic Medical Gas for Neurodegenerative and Other Diseases . Oxid Med Cell Longev 2012 , 11 (2012 ). Ito M. . Drinking hydrogen water and intermittent hydrogen gas exposure, but not lactulose or continuous hydrogen gas exposure, prevent 6-hydorxydopamine-induced Parkinson’s disease in rats . Med Gas Res 2 , 2045 –9912 (2012 ). Matsumoto A. . Oral ‘hydrogen water’ induces neuroprotective ghrelin secretion in mice . Sci Rep 3 , 3273 (2013 ).24253616 Hinoi T. . Complex formation of adenomatous polyposis coli gene product and axin facilitates glycogen synthase kinase-3 beta-dependent phosphorylation of beta-catenin and down-regulates beta-catenin . J Biol Chem 275 , 34399 –34406 (2000 ).10906131 Chaplin M. Do we underestimate the importance of water in cell biology? Nat Rev Mol Cell Biol 7 , 861 –866 (2006 ).16955076 Belpassi L. . Charge-transfer energy in the water-hydrogen molecular aggregate revealed by molecular-beam scattering experiments, charge displacement analysis, and ab initio calculations . J Am Chem Soc 132 , 13046 –13058 (2010 ).20804191 Hasegawa S. . Molecular hydrogen alleviates motor deficits and muscle degeneration in mdx mice . Redox Rep 15 , 1 –9 (2016 ).26866650 Cross D. A. E. , Alessi D. R. , Cohen P. , Andjelkovich M. & Hemmings B. A. Inhibition of glycogen synthase kinase-3 by insulin mediated by protein kinase B . Nature 378 , 785 –789 (1995 ).8524413 Saitoh Y. , Okayasu H. , Xiao L. , Harata Y. & Miwa N. Neutral pH hydrogen-enriched electrolyzed water achieves tumor-preferential clonal growth inhibition over normal cells and tumor invasion inhibition concurrently with intracellular oxidant repression . Oncol Res 17 , 247 –255 (2008 ).19192719 Runtuwene J. . Hydrogen-water enhances 5-fluorouracil-induced inhibition of colon cancer . PeerJ 7 (2015 ). Ye J. . Inhibitory effect of electrolyzed reduced water on tumor angiogenesis . Biol Pharm Bull 31 , 19 –26 (2008 ).18175936 Zhu M. . Inhibition of β-catenin signaling in articular chondrocytes results in articular cartilage destruction . Arthritis Rheum 58 , 2053 –2064 (2008 ).18576323 Chun J. S. , Oh H. , Yang S. & Park M. Wnt signaling in cartilage development and degeneration . BMB Rep 41 , 485 –494 (2008 ).18682032 Pulsatelli L. , Addimanda O. , Brusi V. , Pavloska B. & Meliconi R. New findings in osteoarthritis pathogenesis: therapeutic implications . Ther Adv Chronic Dis 4 , 23 –43 (2013 ).23342245 Li S. . Chemoprevention of Familial Adenomatous Polyposis by Bromo-noscapine (EM011) in the Apc(Min/+) Mouse Model . International Journal of Cancer Journal International du Cancer 131 , 1435 –1444 (2012 ).22052467 Rada P. . WNT-3A Regulates an Axin1/NRF2 Complex That Regulates Antioxidant Metabolism in Hepatocytes . Antioxidants & Redox Signaling 22 , 555 –571 (2015 ).25336178 Furman B. D. . Joint degeneration following closed intraarticular fracture in the mouse knee: a model of posttraumatic arthritis . J Orth Res 25 , 578 –592 (2007 ).

PMC005xxxxxx/PMC5001550.txt

==== Front Database (Oxford)Database (Oxford)databadatabaDatabase: The Journal of Biological Databases and Curation1758-0463Oxford University Press 2754284510.1093/database/baw120baw120Original ArticleThe Markyt visualisation, prediction and benchmark platform for chemical and gene entity recognition at BioCreative/CHEMDNER challenge Pérez-Pérez Martin 1Pérez-Rodríguez Gael 1Rabal Obdulia 2Vazquez Miguel 3Oyarzabal Julen 2Fdez-Riverola Florentino 1Valencia Alfonso 3Krallinger Martin 3Lourenço Anália 12*1 ESEI - Department of Computer Science, University of Vigo, Ourense, Spain2 Small Molecule Discovery Platform, Molecular Therapeutics Program, Center for Applied Medical Research (CIMA), University of Navarra, Pamplona, Spain3 Structural Computational Biology Group, Structural Biology and BioComputing Programme, Spanish National Cancer Research Centre, Madrid, Spain4 CEB - Centre of Biological Engineering, University of Minho, Braga, Portugal*Corresponding author: Tel: +34 988 387 013; Fax: +34 988 387 001; E-mail: analia@uvigo.esCitation details: Pérez-Pérez,M., Pérez-Rodríguez,G., Rabal,O. et al. The Markyt visualisation, prediction and benchmark platform for chemical and gene entity recognition at BioCreative/CHEMDNER challeng. Database (2016) Vol. 2016: article ID baw120; doi:10.1093/database/baw120 2016 19 8 2016 19 8 2016 2016 baw12025 4 2016 19 7 2016 02 8 2016 © The Author(s) 2016. Published by Oxford University Press.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Biomedical text mining methods and technologies have improved significantly in the last decade. Considerable efforts have been invested in understanding the main challenges of biomedical literature retrieval and extraction and proposing solutions to problems of practical interest. Most notably, community-oriented initiatives such as the BioCreative challenge have enabled controlled environments for the comparison of automatic systems while pursuing practical biomedical tasks. Under this scenario, the present work describes the Markyt Web-based document curation platform, which has been implemented to support the visualisation, prediction and benchmark of chemical and gene mention annotations at BioCreative/CHEMDNER challenge. Creating this platform is an important step for the systematic and public evaluation of automatic prediction systems and the reusability of the knowledge compiled for the challenge. Markyt was not only critical to support the manual annotation and annotation revision process but also facilitated the comparative visualisation of automated results against the manually generated Gold Standard annotations and comparative assessment of generated results. We expect that future biomedical text mining challenges and the text mining community may benefit from the Markyt platform to better explore and interpret annotations and improve automatic system predictions. Database URL: http://www.markyt.org, https://github.com/sing-group/Markyt ==== Body Introduction A contemporary, well-recognized challenge of Bioinformatics is to develop specialized methods and tools that enable the systematic and large-scale integration of scientific literature, biological databases and experimental data (1–3). These tools have the potential of considerably reducing the time of database curation and enabling on demand and highly specialized access to literature and database contents (4–6). Initiatives such as BioCreative have been brewing the development of such tools, by providing annotated literature corpora (7, 8) and enabling the controlled comparison of systems performing the automated recognition of biomedical entities of practical interest (9, 10). Under this scenario, the latest BioCreative V CHEMDNER patents challenge (11, 12), which addressed the automatic extraction of chemical and biological data from medicinal chemistry patents, aimed to go a step forward and integrated new computational means to optimize the efforts of both annotators and participants. Specifically, a new Web-based visualisation, prediction and benchmark platform was devised in support of the chemical and gene entity recognition tasks (Figure 1). The CHEMDNER challenge organizers used this platform, named Markyt, to prepare the annotated document sets and to evaluate the predictions of the participating systems. The platform provided a user-friendly document visualisation environment, where human annotators could manage annotation sets and project administrators could evaluate the quality of the annotations throughout the annotation process. On the other hand, Markyt offered participants the possibility of evaluating their predictions on different annotated document sets, so that they could explore prediction–annotation mismatches and acquire insights on possible system improvements. Also, it was used for the final submission of task predictions and the automated scoring of the teams. Currently, the platform is supporting post-workshop prediction evaluation, i.e. any developer can now test their software against CHEMDNER corpora and compare their results with those obtained in the competition. Figure 1. Main use cases of Markyt at BioCreative V CHEMDNER patents challenge. The system helped the organizers and annotators to prepare the annotated document sets, supported the work of text miners while tuning up their systems, and enabled the evaluation and ranking of final predictions. The aim of this paper is to describe the operation of Markyt platform for chemical and gene entity recognition at BioCreative/CHEMDNER challenge and its support to the broader use of the challenge’s resources by text mining developers. The next sections present the architectural design of the platform and show how the platform was utilized by the different users throughout the challenge. Materials and Methods This section describes the main features of the Markyt platform for the visualisation, prediction and benchmark of chemical and gene entity recognition in medicinal chemistry patents under the scope of the BioCreative V CHEMDNER patents challenge (11, 12). As illustrated in Figure 2, Markyt platform was used to revise the manual labelling of the datasets for the chemical entity mention in patents (CEMP) and gene and protein related object (GPRO) tasks, which entailed the detection of chemical named entity mentions and mentions of gene and protein related objects in patent titles and abstracts, respectively. The annotated document sets used for training and development were produced with the intent of supporting the improvement of the automatic prediction tools enrolled in the challenge. Conversely, the test sets were used in the controlled comparison of the performance of the participating systems. Markyt enabled both the analysis of automatic predictions by participants and the controlled comparison of the performance of the various systems. Figure 2. The Markyt platform for chemical and gene entity recognition at BioCreative/CHEMDNER challenge. Markyt was used in CEMP and GPRO tasks, supporting the preparation of training, development and test sets and enabling controlled prediction evaluation and benchmarking. Next, we detail the main aspects of the platform, in terms of software architecture, challenge requirements and user–system interaction. General architecture Markyt is a Web-based platform that was initially devised for the management of multi-user iterative annotation projects (13). Supported by open-source consolidated technologies and presents a modular design, which enabled the development of new modules in response to CHEMDNER’s requirements. The platform follows the Model-View-Controller architectural pattern and was developed using the open source CakePHP Web framework (14). At the core of its architecture are consolidated Web technologies. PHP programming language (version 5.5) and the MySQL database engine (version 5.1.73) support the server side operations. HTML5 (http://www.w3.org/TR/html5/) and CSS3 technologies (http://www.css3.info/) provide for common interface features. Browser-independent implementation of common DOM range and selection tasks is achieved using the Rangy library (http://code.google.com/p/rangy/). Finally, Ajax and JQuery (http://jquery.com/) technologies help in user–system interaction, such as document manipulation, event handling, animation and efficient use of the network layer. Annotation environment The annotation environment of Markyt supports the customized deployment of multi-user and multi-round annotation projects. It provides an intuitive interface to visualize and edit annotated document sets, keeps track of multiple rounds of annotation and allows the comparison of annotation quality across rounds and among annotators. At the technical level, this tool manipulates documents in HTML format and encoded using UTF-8. Annotation classes are represented by HTML class labels and customized to meet the specifications of the project. Moreover, Markyt allows three main types of annotations: manual, i.e. performed by a human annotator; automatic, i.e. originating from an automatic recognition system; and semi-automatic, i.e. automatic annotation of text fragments that are similar to a manual annotation. Two annotation projects were created for the CHEMDNER challenge, one for the CEMP task and another one for the GPRO task, so that annotators could manage each task independently and adequately. Furthermore, the CEMP and GPRO annotation projects were configured according to the following annotation guidelines: Annotations could include single or multiple words, or even partial fragments of a word. Annotation class is unique and exclusive so that the annotations can only belong to one annotation class. Nested annotations are not allowed, i.e. the offsets of one annotation cannot be between the offsets of another annotation in the same document. CEMP entity mentions were divided into eight classes: SYSTEMATIC, TRIVIAL, FAMILY, FORMULA, ABBREVIATIONS, IDENTIFIERS, MULTIPLE and NO CLASS (details can be found at http://www.biocreative.org/media/store/files/2015/cemp_patent_guidelines_v1.pdf). GPRO entity mentions were classified into four classes: NESTED MENTIONS, IDENTIFIER, FULL NAME and ABBREVIATION (details can be found at http://www.biocreative.org/media/store/files/2015/gpro_patent_guidelines_v1.pdf). Project administrators and human annotators were instructed on how to operate Markyt annotation environment and were assisted throughout the annotation process. Markyt was used primarily as a visualisation and editing tool, which helped project administrators and annotators to discuss the quality of the annotations in the different document sets. Under this scenario, the automatic annotation recommendation module of Markyt played an important part in helping reduce the number of false negative mentions. More specifically, this annotation module was utilized to produce annotation recommendations based on the annotation history of the project. That is, the tool detected unlabelled text mentions that are exact, case insensitive matches of manual expert annotations and prompted recommendations to be manually revised. Manual annotations and automatic annotations were made visually distinguishable to make human inspection easier. Likewise, the human curator was only asked to remove the automatic recommendations that were incorrect. At the end of the manual revision, all remaining automatic annotations were accepted. Prediction analysis CHEMDNER released training and development annotated document sets that participants could use to improve the performance of their automatic prediction tools and a blinded test set for which participants had to submit predictions to be evaluated against manual annotations. Markyt platform provided participants with an analytical environment for evaluating their predictions for the different gold standards. In previous editions of the challenge, the organizers made available an evaluation script to score the predictions (http://www.biocreative.org/resources/biocreative-ii5/evaluation-library/). However, it was not straightforward to identify which terms were often missed or particular scenarios where the algorithm would output false positive predictions. Such exploration was either conducted in a manual way, which was time consuming, or supported by in-house software, which implied additional programming effort for most participants. Markyt analytical environment aimed to bridge this gap and equip the teams with means to calculate prediction scores and explore the most important prediction–annotation mismatches without any additional programming costs. Therefore, it provides the calculation of micro- and macro-average standard performance statistics, such as precision, recall and F-score (15, 16). Furthermore, it enables the examination of annotation mismatches, i.e. false positive (FP) and false negative (FN) annotations. Three main statistics are examined: false negative (FN) results corresponding to incorrect negative predictions (i.e. cases that were part of the gold standard, but missed by the automatic system), false positive (FP) results being cases of incorrect positive predictions (i.e. wrong results predicted by the automatic system that had no corresponding annotation in the gold standard) and true positive (TP) results consisting of correct positive predictions (i.e. correct predictions matching exactly with the gold standard annotations). Correspondingly, recall is the percentage of correctly labelled positive results over all positive cases, i.e. it is a measure of the ability of a system to identify positive cases. (1) recall= TPTP+FN Precision is the percentage of correctly labelled positive results over all positive labelled results, i.e. it is a measure of the reproducibility of a classifier of the positive results. (2) precision= TPTP+FP And, the F-score is the harmonic mean between precision and recall. (3) F-score=2*precision×recallprecision+recall Partial hits, i.e. predictions that only in part overlapped with the manually defined gold standard annotations, were not taken into account in the analyses. Micro-average statistics were calculated globally by counting the total true positives, false negatives and false positives. Conversely, macro-average statistics were calculated by counting the true positives, false negatives and false positives on a per-document basis and then averaged across documents. Benchmarking CHEMDNER participants could submit a total of five runs per task for final evaluation. The micro-averaged recall, precision and F-score statistics were used for final prediction scoring, and F-score was used as main evaluation metric. Furthermore, Markyt analytical environment supported the examination of the statistical significance of each prediction with respect to the other final submissions by means of bootstrap resampling simulation, in a similar way to what was done in the previous CHEMDNER challenge (9, 17). This statistical analysis was done for both the CEMP and GPRO tasks by taking 2500 bootstrapped samples from all the documents in the test sets (a total of 7000 documents in each set) that had annotations. The micro-average F-scores for each team on each sample were calculated and these 2500 resampled results were then used to calculate the standard deviation of the F-score of each team (SDs). Teams were grouped based on statistically significant difference (at two SD) between results. Results Management of annotation projects and gold standard preparation The CHEMDNER challenge involved the annotation of a total of 21 000 medicinal chemistry patents (11, 12). The patents came from the following agencies: the World Intellectual Property Organization (WIPO), the European Patent Office (EPO), the United States Patent and Trademark Office (USPTO), Canadian Intellectual Property Office (CIPO), the German Patent and Trade Mark Office (DPMA) and the State Intellectual Property Office of the People's Republic of China (SIPO). CEMP and GPRO tasks were supported by the same document sets, but annotation was independent and complied with the guidelines established by the organizers for each task. Details on task guidelines can be found at http://www.biocreative.org/media/store/files/2015/cemp_patent_guidelines_v1.pdf. As Figure 3 illustrates, Markyt annotation environment was used to manage the manual annotations in the document sets. Annotators were able to create, edit or delete annotations, navigate to specific documents or search for matches of a particular annotation. The task-dependent annotation types were defined at project configuration and colour tagged for immediate visual perception. For example, annotations of systematic chemical names in CEMP sets were tagged in yellow. Figure 3. Some features of the Markyt document manual annotation environment. Examples were taken from one of the CEMP document sets and illustrate the visualisation of document contents and existing annotations, distribution of annotation per class, editing features, document search and annotation search. Besides providing basic means of text annotation, Markyt helped in minimising typical errors in repetitive tasks. For example, the annotator could apply the same operation to multiple inter-document occurrences of the marked text fragment. Markyt also allowed the search of documents containing a given annotation, and navigation to specific documents, which are operations suggested by the annotators as means to expedite annotation revision. Later, CEMP and GPRO test sets were submitted to an additional semi-automatic process of annotation (Figure 4). The manual annotations were used as ground truth by the automatic recommendation module of Markyt. That is, any text fragment matching one of these annotations and without an annotation was treated as a potential miss. Figure 4. Semi-automated revision workflow of the CHEMDNER test set. The manually annotated document set is enriched with automatic annotation recommendations to be revised by the experts. Recommendations are based on unlabelled text mentions that match manual annotations. Annotators were required only to edit or eliminate non-qualifying recommendations. The annotation environment presented an integrated view of these recommended annotations together with the manual annotations to simplify manual expert revision. Automatic annotations were visually differentiated by the use of bordered marks. Annotators took advantage of the search and navigation features to inspect recommendations. Specially, annotators were instructed to edit or eliminate recommendations as considered appropriate. No manual action was required for accepting annotation recommendations, as this operation was performed automatically at the end of the revision round. Prediction analysis Markyt measured the performance of the competing systems by comparing their predictions for gold standards (Figure 5). During the challenge, analysis was performed on demand by the participants. At the end of the challenge, this analysis was executed by the organizers as part of the automatic comparison and scoring of all competing teams. Figure 5. Markyt environment for prediction analysis. Teams could upload an unlimited number of predictions against the competition data sets. Markyt provided common performance metrics as well as details on false positives and false negatives. Final submissions were ranked based on micro averaged F-score. Access to challenge participants was made simple. Using challenge credentials, the participant could submit prediction files (compliant with the CHEMDNER predictions file format, which is exemplified in the text data set available at http://www.biocreative.org/media/store/files/2015/CHEMDNER_TEST_TEXT.tar.gz). Markyt would perform the analysis and send the results encrypted via email (to avoid unnecessary wait when processing a large number of predictions). Then, the participant could access the system and visualize the prediction results privately. The analysis report consisted of precision, recall and F-score statistics, a table with the distribution of true positive and false negative results per annotation class, and tables listing the top false positive and top false negative predictions (Figure 6). The micro-average weights each annotation class equally whereas the macro-average weights each document equally, regardless of how many annotations are found in the document. Thus, macro-averaged results provided a straightforward way to compute statistical significance. Figure 6. Snippets of the information displayed in a CEMP prediction analysis report. The report shows the distribution of true positives and false negatives per annotation type (A), and performance is described in terms of macro- and micro-averaged precision, recall and F-score statistics (B and C). The tool enables the exploration of prediction matches and mismatches in individual documents (D). Also, it enables the inspection of the most frequent misclassifications (E) and the documents with more mismatches (F). The lists of false negative results (i.e. gold standard annotations missed by the automatic system) and false positive predictions (i.e. predictions that did not have a match in the gold standard) provided the team with contextualized examples of the most frequent misclassifications made by their system. For example, in the prediction report shown in Figure 6, the term ‘alkyl’ was at the top of the false negatives of a prediction run for the CEMP task, with a total of 66 missed occurrences. Conversely, the document EP2610263A1 was the document with more incorrect positive predictions, with a total of 34 false positives. So, participants could use Markyt prediction reports to keep track of the mistakes being committed by their automatic systems, gain an understanding about what motivated the incorrect predictions and work on possible solutions. There were no restrictions to the number of prediction analyses that a participant could run in Markyt during the challenge. In the final submission, each participant could submit a maximum of five runs per task and micro-averaged statistics were used to determine the top-scoring run for each team (the best F-score results). Furthermore, Markyt analytical environment enabled an overall view of misclassifications by all the systems for the CHEMDNER organizers. Specifically, Markyt showed the false positives and false negatives common to most systems and exposed the most ‘difficult’ documents, i.e. enabled further understanding of where/when the predictive ability of the automatic systems is more limited (Table 1). This information is considered valuable to better assess the difficulties that the automatic processing of chemical patents presents to those currently developing chemical recognition systems, and the capabilities of existing systems. Such insight will help prepare future editions of the CHEMDNER challenge and is believed to be of added value to the participating team (a macroscopic observation of each system’s prediction scores). Further details on this analysis can be found at the challenge overview paper (11, 12). Table 1. Top document mismatches and term mismatches for final CEMP team submissions Top false negatives Avg(FN) per run Top false positives Avg(FP) per run Documents EP2033959A1 19 EP2033959A1 53 EP2610263A1 13 EP2546253A1 49 EP2546253A1 13 EP2610263A1 47 DE102004060041A1 8 DE102004060041A1 36 US20100184776 4 US20090170813 21 Terms Acid 33 Sodium 42 Soy isoflavone 10 Alkyl 29 Fibroblast 7 Ester 29 Docetaxel 6 Opioid 28 Aromatic or 4 Calcium 25 heteroaromatic 4 Post-workshop software benchmarking After the challenge, the results of the top-scoring run for each team, including recall, precision, and F-score for the best (micro-averaged) run of each system, were made public (11, 12). In particular, Markyt prediction analysis environment is now open to anyone who wishes to compare the performance of his system to CHEMDNER results. To further evaluate the significance of the difference between system performances, results were submitted to a bootstrap resampling. As exemplified in Figure 7, the evaluation tables of the CEMP and GPRO tasks depict the precision, recall and F-score of the best run of each team as well as illustrate the position of each team are covered within two standard deviations (SDs). When running a new prediction analysis at Markyt, the tool produces the usual performance statistics and compares these to the performance statistics of the systems that originally participated in the challenge. Therefore, developers receive in-depth information of their system’s predictive abilities against CHEMDNER gold standards and existing systems. Figure 7. Example of CEMP post-workshop benchmarking in Markyt. New predictions on CEMP test set can be compared against the best predictions obtained during the competition (A) and the system is ranked accordingly (C). Also, general prediction statistics (B) can be explored per class type (E), looking into document matches and mismatches (D and E). Benefits of using Markyt Table 2 summarizes the functionalities made available to the three user types of BioCreative/CHEMDNER, i.e. challenge organizer, text miner and expert annotator, in support of the main usage scenarios presented during the competition and after the workshop. Most notably, Markyt supports post-workshop system evaluation against the best performing systems in the competition. Table 2. Main features of Markyt available to BioCreative/CHEMDNER and post-workshop users User type Role Features Challenge organizer Competition management Management of multi-user and iterative annotation projects Annotation quality control and consensus analysis Exportation of annotation sets (inline and stand-off formats) Publication of competition tasks Evaluation and ranking of final team submissions CHEMDNER participant or text miner System tuning and challenge submission Prediction evaluation against various gold standards Validation of submission compliance with CHEMDNER format Identification of top prediction-annotation mismatches by document and type Visualisation of annotation mismatches Submission of final predictions Post-workshop benchmarking Post-workshop evaluation against CHEMDNER resources (corpora and challenge scores) Expert annotator Manual curation Highlight all occurrences of selected text fragment in document (or page of documents) Automatic annotation of a text fragment across all documents Remember last annotation type for new annotations Remember last annotation metadata for new annotations (e.g. last database id) Search by annotation Auto saving to prevent data loss Annotation revision Visual differentiation of manual and automatic annotations to facilitate curation Download suggested terms to review, indexed by document Implicit acceptance of automatic annotations to speed up curation Batch change of annotation type for all annotations of a given term and type Delete all annotations of a given term (type-specific or not) Conclusion Creating gold standards and enabling the controlled comparison of automatic prediction systems are key steps to keep improving the performance of automatic prediction systems in practical biomedical scenarios. Here, we described the Markyt Web-based platform for the visualisation, prediction and benchmark of chemical and gene entity recognition at the BioCreative/CHEMDNER challenge. This platform supported the preparation of the annotated document sets and, in particular, provides a semi-automatic curation workflow to improve the quality of the post-workshop test set (11, 12). Furthermore, Markyt allowed developers to test their predictions against the different CHEMDNER corpora and, more recently, to compare the performance of their systems with CHEMDNER final system ranking. The ultimate purpose was to provide computational support to challenge organizers and participants, and to make the resources and evaluation methods of BioCreative/CHEMDNER challenge readily available to the text mining community. Previous BioCreative tasks did not address the visualisation aspect sufficiently. Markyt bridged this gap by helping to prepare the annotated sets (targeting the needs manifested by human annotators) and enabling the analysis of prediction-annotation mismatches (helping text miners understand where/when the automatic systems tend to fail). Likewise, and although BioCreative related resources and tools are publicly available, no platform provided support to post-workshop benchmarking, namely, the development of new systems and the development of the participating systems. Markyt has an open and general purpose architectural design that allows the integration of new subsystems or the modification of existing subsystems by third-parties. Markyt is also domain/application agnostic, notably it handles main document formats (TXT, HTML and XML), allows the customisation of annotation types and annotation metadata (e.g. database identifiers), and enables the customisation of quality monitoring (e.g. evaluation over one or several rounds of annotation, and looking into different metrics). Hence, Markyt has the potential of being adapted to other text mining tasks, including challenges or benchmark projects, annotation projects and database curation. Finally, and thanks to the feedback received through the challenge, Markyt subsystems are being improved and they will be part of a new community-geared metaserver evaluation system. This innovative system will support the work of participants in upcoming editions of BioCreative/CHEMDNER as well as provide broader benchmarking and annotation services for text miners, database annotators and anyone that wishes to perform system benchmarking. Furthermore, the development of such novel community-geared metaserver will be under the umbrella of the OpenMinTeD project (http://openminted.eu/), the European open mining infrastructure for text and data, which will certainly help to identify and accommodate the requirements and concerns of different user communities as well as promote the use of the system worldwide. Acknowledgements This document reflects only the author’s views and the European Union is not liable for any use that may be made of the information contained herein. Funding This work was partially funded by the [14VI05] Contract-Programme from the University of Vigo and the Agrupamento INBIOMED from DXPCTSUG-FEDER unha maneira de facer Europa (2012/273) as well as by the Foundation for Applied Medical Research, University of Navarra (Pamplona, Spain). The research leading to these results has received funding from the European Union's Seventh Framework Programme FP7/REGPOT-2012-2013.1 under grant agreement n° 316265, BIOCAPS. Conflicts of interest: The authors have no competing interests. ==== Refs References 1 Fluck J. Hofmann-Apitius M. (2014 ) Text mining for systems biology . Drug Discov. Today , 19 , 140 –144 .24070668 2 Cohen K.B. Hunter L.E. (2013 ) Chapter 16: text mining for translational bioinformatics . PLoS Comput. Biol ., 9 , e1003044. 23633944 3 Krallinger M. Leitner F. Valencia A. (2010 ) Analysis of biological processes and diseases using text mining approaches . Methods Mol. Biol ., 593 , 341 –382 .19957157 4 Fleuren W.W.M. Alkema W. (2015 ) Application of text mining in the biomedical domain . Methods , 74 , 97 –106 .25641519 5 Lu Z. Hirschman L. (2012 ) Biocuration workflows and text mining: overview of the BioCreative 2012 Workshop Track II . Database (Oxford) , 2012 , bas043 .23160416 6 Krallinger M. Valencia A. Hirschman L. (2008 ) Linking genes to literature: text mining, information extraction, and retrieval applications for biology . Genome Biol ., 9 Suppl 2 , S8. 18834499 7 Krallinger M. Rabal O. Leitner F. (2015 ) The CHEMDNER corpus of chemicals and drugs and its annotation principles . J. Cheminform , 7 , S2. 25810773 8 Leitner F. Krallinger M. Cesareni G. Valencia A. (2010 ) The FEBS Letters SDA corpus: a collection of protein interaction articles with high quality annotations for the BioCreative II.5 online challenge and the text mining community . FEBS Lett ., 584 , 4129 –4130 .20728446 9 Krallinger M. Leitner F. Rabal O. (2015 ) CHEMDNER: The drugs and chemical names extraction challenge . J. Cheminform , 7 , S1. 25810766 10 Krallinger M. Vazquez M. Leitner F. (2011 ) The Protein-Protein Interaction tasks of BioCreative III: classification/ranking of articles and linking bio-ontology concepts to full text . BMC Bioinformatics , 12 Suppl 8 , S3. 22151929 11 Krallinger M. Rabal O. Lourenço A. (2016 ) Evaluation, corpora and analysis of chemical and gene/protein name recognition in patents: the CHEMDNER patents text mining task at BioCreative V . Database. 12 Krallinger M. Rabal O. Lourenço A. (2015 ) Overview of the CHEMDNER patents task. In: Proceedings of the Fifth BioCreative Challenge Evaluation Workshop, Sevilla, Spain. pp. 63 –75 . 13 Pérez-Pérez M. Glez-Peña D. Fdez-Riverola F. Lourenço A. (2014 ) Marky: a tool supporting annotation consistency in multi-user and iterative document annotation projects . Comput. Methods Programs Biomed ., 118 , 242 –251 .25480679 14 Iglesias M. (2011 ) CakePHP 1.3 Application Development Cookbook. Packt Publishing, Birmingham. 15 Sokolova M. Lapalme G. (2009 ) A systematic analysis of performance measures for classification tasks . Inf. Process. Manag ., 45 , 427 –437 . 16 Blair D.C. (1979 ) Information Retrieval, 2nd ed. C.J. Van Rijsbergen. London: Butterworths; 1979: 208 pp. Price: $32.50 . J. Am. Soc. Inf. Sci ., 30 , 374 –375 . 17 Smith L. Tanabe L.K. Ando R.J. (2008 ) Overview of BioCreative II gene mention recognition . Genome Biol ., 9 Suppl 2 , S2 .18834493

PMC005xxxxxx/PMC5001580.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2706013710.1093/nar/gkw139Survey and SummaryTranscriptional gene silencing in humans Weinberg Marc S. 123Morris Kevin V. 145*1 Department of Molecular and Experimental Medicine, The Scripps Research Institute, La Jolla, CA 92037, USA2 Wits/SAMRC Antiviral Gene Therapy Research Unit, School of Pathology, University of the Witwatersrand, WITS 2050, South Africa3 HIV Pathogenesis Research Unit, Department of Molecular Medicine and Haematology, School of Pathology, University of the Witwatersrand, WITS 2050, South Africa4 Center for Gene Therapy, City of Hope – BeckmanResearch Institute; Duarte, CA 91010, USA5 School of Biotechnology and Biomedical Sciences, University of New South Wales, Kensington, NSW, 2033 Australia* To whom correspondence should be addressed. Tel: +626 256 4673 (Ext. 82839); Fax: +858 784 2131; Email: kmorris@scripps.edu19 8 2016 07 4 2016 07 4 2016 44 14 6505 6517 23 2 2016 22 2 2016 22 1 2016 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.It has been over a decade since the first observation that small non-coding RNAs can functionally modulate epigenetic states in human cells to achieve functional transcriptional gene silencing (TGS). TGS is mechanistically distinct from the RNA interference (RNAi) gene-silencing pathway. TGS can result in long-term stable epigenetic modifications to gene expression that can be passed on to daughter cells during cell division, whereas RNAi does not. Early studies of TGS have been largely overlooked, overshadowed by subsequent discoveries of small RNA-directed post-TGS and RNAi. A reappraisal of early work has been brought about by recent findings in human cells where endogenous long non-coding RNAs function to regulate the epigenome. There are distinct and common overlaps between the proteins involved in small and long non-coding RNA transcriptional regulatory mechanisms, suggesting that the early studies using small non-coding RNAs to modulate transcription were making use of a previously unrecognized endogenous mechanism of RNA-directed gene regulation. Here we review how non-coding RNA plays a role in regulation of transcription and epigenetic gene silencing in human cells by revisiting these earlier studies and the mechanistic insights gained to date. We also provide a list of mammalian genes that have been shown to be transcriptionally regulated by non-coding RNAs. Lastly, we explore how TGS may serve as the basis for development of future therapeutic agents. cover-date19 August 2016 ==== Body INTRODUCTION The history of RNA-directed transcriptional gene silencing (TGS) Almost three decades ago, Marjorie Matzke et al. observed that over-expression of a transgene led to DNA hypermethylation and transcriptional silencing in doubly transformed tobacco plants (1) (Figure 1). Mechanistically, this type of silencing in plants was found to be the result of small non-coding RNAs directing epigenetic changes, specifically DNA methylation, to those loci containing sequences homologous to the small RNA. The phenomenon was termed small RNA-directed transcriptional gene silencing (TGS). TGS was later shown in Arabidopsis to require the action of RNA-dependent DNA methylation (2,3) and members of the Argonaute protein family (4). A few years later RNA interference (RNAi), mediated by double-stranded RNAs, was discovered as a powerful post-TGS (PTGS) system against messenger RNAs (mRNAs) in plants (5), and a few months later in Caenorhabditis elegans (6). Figure 1. Regulatory non-coding RNA timeline. A timeline of some important observations in RNA biology are shown leading up to our collective understanding of non-coding RNA-directed transcriptional gene silencing (TGS). (5,71,170–177). Transcriptional gene silencing in humans The study of small non-coding RNA-directed TGS has been carried out in various model organisms such as plants (Arabidopsis thaliana), yeast (Saccharomyces pombe), flies (Drosophila melanogaster) and worms (C. elegans) (reviewed extensively in (7,8)). A decade ago, the first report of RNA-directed TGS in human cells was observed when exogenous siRNAs were used to silence a transgenic elongation factor 1 α promoter driving a Green Fluorescent Protein (GFP) reporter gene (9) (Figure 1). Importantly, the observed silencing was clearly at the transcriptional level, as indicated by nuclear run-on analysis. Moreover, silencing was also epigenetic: inhibition was abrogated by 5′ Aza-cytadine (5′ AzaC) and Trichostatin A (TSA), compounds involved in inhibiting DNA methylation and histone deacetylalion, respectively (9). This early observation was soon followed by other studies (10,11), all of which confirmed that small non-coding RNAs could functionally control gene transcription and epigenetic states in human cells. But the underlying mechanism of action remained unknown. Mechanisms of small non-coding RNA-directed TGS TGS is mechanistically distinct from the abundantly studied PTGS pathway of RNAi. One notable difference is that TGS results in long-term stable epigenetic modifications to gene expression that can be passed on to daughter cells during cellular division (reviewed in (12)). Early observations postulated that siRNA-directed TGS functioned through an epigenetic nuclear mechanisms distinct from RNAi-mediated PTGS in the cytoplasm (13). For instance 5′ AzaC and TSA were functional in reverting the siRNA targeted TGS, indicating epigenetic modes of gene regulation were at play in siRNA-directed TGS, and not via a PTGS-based mechanism (9). Indeed, recent studies have observed that two different siRNAs, one targeted to the promoter and one targeted to exon 1 of the coding transcript, can functionally repress the targeted gene in a TGS or PTGS based manner (14). A lot has been gleaned over the last decade regarding the mechanism of action for RNA-directed TGS in human cells. Studies carried out to determine the underlying mechanism of siRNA-directed TGS revealed that RNA-mediated TGS is operative through RNA-directed methylation of histone 3 lysine's 9 and 27 (H3K9 and H3K27, respectively) and DNA methylation at the targeted promoter (9,11,15–21) (Figure 2). These promoter-directed siRNAs interact with a low level expressed (∼1–2%) promoter associated RNA, which is essentially the 5′ UTR of the protein coding gene (16,22) (Figure 2). It is worth noting that most genes and gene promoters appear to be transcribed to some extent (23,24) and experimental observations suggest that non-coding RNAs interact with target loci via Watson–Crick-based RNA:RNA hybridization (16,22) and not by double-stranded DNA invasion. Temporal studies have determined that exogenously introduced siRNAs targeted to a promoter region interact first with Argonautes 1 and 2 (AGO1 and AGO2) (17,25,26). SiRNA and AGO interactions is found within the first 24 h, at the siRNA targeted promoter and is followed shortly thereafter with the recruitment of the H3K9me2 and H3K27me3 silent state epigenetic marks (17), and later by the recruitment of DNA methyltransferase and DNA methylation at 72–96 h for some genes (14). It should be noted, however, that the role of DNA methylation in TGS in human cells is not as clearly understood as in plants; DNA methylation at the targeted promoter is not always observed in human TGS applications (Table 1). These effects may be explained by the duration of RNA targeting to the promoter, the presence of robust siRNA targeting (e.g. delivery to the nucleus), the presence and abundance of promoter-occupied RNAs and/or the dynamic interplay of proteins interacting with the promoter. Despite differences in the various experimental observations, a key consistent feature has been the observations that promoter-directed small RNAs can modulate gene transcription and that some level of epigenetic based silencing is ongoing in the observed silenced genes. Figure 2. Small non-coding RNA pathways in human cells. Small non-coding RNAs can be generated as priRNAs where they are (A) processed by Drosha and DGCR8 into miRNAs which are (B) exported from the nucleus and (C) loaded into RISC where they can affect mRNA expression by (D) binding and blocking mRNA translation or (E) cutting the target mRNA. Some miRNAs may also be retained in the nucleus (F) where they can interact with epigenetic remodeling proteins and (G) recruit the complexes to target loci in the genome resulting in (H) localized chromatin compaction and epigenetic silencing. Table 1. Mammalian genes transcriptionally regulated by non-coding RNAs Gene(s) Gene symbol Effector RNA Cell line Therapeutic relevance References Eukaryotic translation elongation factor 1 α EEF1A1 siRNA HEK293T (9,16,18) HIV-1/SIV 5′ LTR siRNA, sasRNA Jurkat, Tzmb, T-cells, in vivo (mouse) Regulation of HIV-1 (18–21,26,79,99,100,104,140–144) nitric-oxide synthase NOS siRNA HAEC Cardiac disease (145) E-cadherin CDH1 siRNA HCT116; MCF7 Cancer, tumor suppressor (10) BCL-2 (oncogene) BCL-2 sasRNA HeLa, 293 Cancer, oncogene (146) Fibronectin FN1 siRNA HeLa (129,147) Huntingtin gene HTT siRNA Glioblastoma Monogenetic diseases (148) Non-sense codon-containing immunoglobulin minigenes (Ig)-mu and Ig-gamma siRNA HeLa Immunologic diseases (149) INK4B/Cyclin-dependent kinase inhibitor 2B/p15 + ARF + INK4A/ Cyclin-dependen kinase inhibitor 2A isoform 3/p16 CDKN2B+ CDKN2A siRNA HEK293T Cancer, tumor suppressor (150) INK4A/Cyclin-dependen kinase inhibitor 2A isoform 3/p16 CDKN2A siRNA HEK293T Cancer, tumor suppressor (151) Plasminogen activator, urokinase PLAU siRNA PC3 and in vivo Cancer (152) Chemokine receptor 5 CCR5 siRNA HEK293T HIV-1 co-receptor (16,17) Breast cancer-associated gene 1 BRCA1 siRNA T47D Cancer, oncogene (153) Progesterone receptor PGR siRNA T47D Cancer (11,25,119,153,154) Huntingtin HD siRNA T47D Monogenetic diseases (25) Androgen receptor AR siRNA T47D Cancer, spinal bulbar muscular atrophy (25,155) v-myc avian myelocytomatosis viral oncogene homolog c-MYC siRNA/sasRNA PC3, HCT113, 293, Hela, MCF7 Cancer, oncogene (22,97,156,157) Papillomavirus-16 oncogenes HPV-16 siRNA HeLa HPV (158) v-akt murine thymoma viral oncogene homolog 1 AKT-1 siRNA 293HEK Cancer, oncogene (156) Kirsten rat sarcoma viral oncogene KRAS siRNA 293HEK Cancer, oncogene (156) Dual specificity phosphatase 6 DUSP6 siRNA CFPAC Cancer, tumor suppressor (156) Myostatin MSTN siRNA C2C12 mouse myoblasts muscle hypertrophy (159) Runt-related transcription factor 3 RUNX3 siRNA stomach carcinoma cell line SGC7901 osteoarthritis (160) Small nuclear 7sk (RNAi functional in the nucleus of human cells) 7SK siRNA Hela, 293HEK cells 7SK/P-TEFb control of HIV-1 (126) met proto-oncogene (hepatocyte growth factor receptor) c-Met siRNA/asRNA SKHep1C3 cells Cancer, oncogene (161) Periostin POSTN siRNA/sasRNA PC3 Cancer and metastasis (101) Heparanase (endo-h-D-glucuronidase) HPA siRNA PC3, EJ and SGC-7901 cells Cancer, angiogenesis (162) phosphoglycerate kinase 1 promoter driving GFP pgk-1 siRNA 293, HeLa (98) Interleukin 2 IL2 shRNA Jurkat Immunologic (163) Ubiquitin C UBC siRNA; shRNA HEK293GT (14) Transforming growth factor-β receptor II TGFβII shRNA rat SBC10 Cancer, angiogenesis (102) Vascular endothelial growth factor VEGF-A shRNA mouse C166 and in vivo Cancer, angiogenesis (107,122) Ras association domain family 1 RASSF1A shRNA HeLa Cancer, oncogene (15,17) Tubulin folding cofactor E-like TBCEL/ LRRC35 miRNA HCT116 Kenny-Caffey syndrome (KCS) (150) Ras p21 protein activator 2 RASA2 miRNA HCT116 Cancer, tumor suppressor (150) Rhophilin, Rho GTPase binding protein 2 RHPN2 miRNA HCT116 (150) Wolf-Hirschhorn syndrome candidate 1 WHSC1 miRNA HCT116 Wolf-Hirschhorn syndrome (150) Homeobox D4 HOXD4 miRNA MCF7; MDA-MB-231 (164) HIV-1 LTR miRNA Jurkat, T-cells HIV-1 infection (28) HIV-1 TAR miRNA Tzmb, Jurkat, T-cells HIV-1 infection (29,165) OCT4 and Nanog (pluripotent factor) OCT4 and Nanog lncRNA (antisense, pseudogene) MCF7 Cancer, pluripotency (54) PTENpg1 asRNA alpha PTEN lncRNA (Trans-antisense, pseudogene) 293, Hela, Jurkat Cancer, tumor suppressor (53) P21 tumor suppressor P21 lncRNA (antisense) MCF7 Cancer, tumor suppressor (39) P15 tumor suppressor P15 lncRNA (antisense) HL-60, KG-1, Kasumi-1, DG-75, Raji and Ramos Cancer, tumor suppressor (38) alpha-globin gene HBA2 lncRNA Embryonic stem cells Alpha thalassemia (40) DM1 insulator DM1/SIX5 lncRNA Primary fibroblasts (166) Herpes LAT lncRNA in vivo (mice) HPV (167) P53, lncRNA-p21 hnRNP-K lncRNA MEF, in vivo (mice) Cancer, oncogene (57) lncRNA HOTAIR HOX lncRNA MDA-MB-231, SK-BR-3, MCF-10A, MCF-7, HCC1954, T47D and MDA-MB-453 cell lines. Human tissue samples Cancer (86) HIV-1 HIV-1 lncRNA (antisense) Jurkat, T-cells, Tzmb, 293HEK HIV-1 infection (49–52) v-myc avian myelocytomatosis viral oncogene neuroblastoma derived homolog MYCN MYCNOS lncRNA (Cis-antisense) Lan6 Cancer, oncogene, neuroblastoma (43) neuroblastoma associated transcript 1 NBAT-1 lncRNA Neuroblastoma primary tumors Cancer, neuroblastoma (48) brain-derived neurotrophic factor BDNF lncRNA (antisense) Human brain, mouse in vivo Huntington disease, Alzheimer disease and Parkinson disease (41,42) short-chain dehydrogenase/reductase family member 4 DHRS4 lncRNA (antisense) HepG2 and HL7702 cell (44) potassium voltage-gated channel, KQT-like subfamily, member 1 Kcnq1ot1 lncRNA (antisense) Human placenta-derived JEG-3 cells Romano-Ward syndrome, Jervell and Lange-Nielsen syndrome and familial atrial fibrillation (168) moesin at 5p14.1 in Autism Moesin lncRNA (antisense) Human brain (postmortem cerebral cortex) Autism spectrum (113) C-terminal binding protein 1 CTBP1 lncRNA (antisense) LNCaP, VCaP, DU145 Prostate cancer (169) Reports from 2004 to present using exogenously administered small interfering RNAs (siRNAs), small hairpin RNAs, small antisense RNAs (sasRNA) and endogenously expressed microRNAs (miRNAs) or long non-coding RNAs (lncRNAs) effector transcripts to modulate gene transcription in mammalian cells are shown. Those genes targeted and their therapeutic relevant disease is also shown. The endogenous pathway of TGS in human cells; rise of long non-coding RNAs While small RNAs were observed early on to regulate gene transcription in human cells by the targeting of epigenetic silencing complexes to those loci containing complementarity to the small RNAs (Figure 1), the endogenous mechanism(s) driving this form of gene regulation in the context of human cells remained largely unknown. MicroRNAs (miRNAs) have been shown to be endogenous drivers of TGS with some genes in human cells (27–30)(Table 1). In 2005, through the efforts of the FANTOM and ENCODE consortia, it started to become apparent that a large fraction of the human genome was generating long non-coding RNAs (lncRNAs) and that many of these transcripts were antisense to protein-coding counterparts (31,32). Several of these sense/antisense or bidirectionally-transcribed genes are evolutionarily conserved, suggesting some functional cues for retention of these elements (33,34). Indeed, studies with imprinted genes and X-inactivation found that cis acting long non-coding RNAs (lncRNAs) were actively involved in epigenetic regulation of these, dosage-dependent, regulated loci (35). In 2008, a number of important studies confirmed the role of lncRNAs as endogenous drivers of TGS in human cells, in particular those, which were antisense to their protein-coding counterparts (reviewed in (36,37)). Antisense lncRNAs were shown to regulate the p15 (38) and p21 (39) tumor suppressor genes (Table 1). The over-expression of these antisense lncRNAs resulted in TGS of their protein-coding counterpart while their repression resulted in the de-repression/transcriptional activation (38,39). Support for the role of antisense lncRNAs as active endogenous regulators of gene expression was evident in an earlier understated study, which indicated that antisense transcripts might also be involved in CpG methylation in thalassemia (40). Antisense lncRNAs are now known to affect TGS for genes such as BDNF (41,42), MYCN (43), DHRS4 (44), KCNQ1 (45–47), NBAT (48) and HIV-1 (49–52) (Table 1). Interestingly, non-coding transcripts derived from pseudogenes of Phosphatase and tensin homolog (PTEN) (53) and OCT4 (54), which contain significant homology to their protein-coding counterparts, have also been observed to be involved in directing TGS and subsequent PTEN and OCT4 suppression (Table 1). Also, the PTEN antisense pseudogene-directed TGS of PTEN is one of the first bona fide examples of a trans-functional lncRNA (53). LncRNAs direct TGS and CpG methylation (55). Cis regulation and antisense lncRNAs have also been observed to epigenetically regulate ribosomal genes (56) as well as p21 (57) and MYCN (43), genes involved in cell regulation and cancer. The plethora of lncRNA functions are extensive and the range includes: protein modifiers (58,59), scaffolds for tethering proteins (43,60,61), miRNAs (62–64), splicing modifiers (65), cellular body transformation (66–69), enhancer function and gene activation (70–73), and epigenetic modifiers (53,74,75), (reviewed in (76)). Collectively, the observations to date suggest that we are only now just beginning to realize the complexity and pervasiveness of lncRNA functional regulation in epigenetic and transcriptional states. Mechanisms of small and long non-coding RNA-directed TGS To date there are ∼55 reports of small RNA-directed TGS and ∼10 of antisense lncRNA-directed TGS (Table 1). Mechanistically, much of what we know about how small non-coding RNAs, such as siRNAs, miRNAs and small antisense RNAs (sasRNAs)-directed TGS, have been determined from cell culture studies. The promoter-targeted small RNAs interact with various proteins to guide TGS, beginning in the first 24 h, with direct interactions with AGO1 and AGO2 (17,18,25) followed shortly thereafter by interactions at the targeted promoter with DNMT3a (14,18,77,78), HDAC1 (14,20) and resulting ultimately in histone 3 lysine 9 di-methylation and histone 3 lysine 27 tri-methylation (H3K9me2 and H3K27me3, respectively)(14,16–18,20,79,80) (Figure 3). SiRNA-directed TGS has also been observed to occur in the absence of DNA methylation, suggesting that alternative routes may be present for RNA-mediated transcriptional and epigenetic silencing (10). Small RNA-directed TGS appears to require a template or target transcript at the corresponding targeted promoter (16,22), similar to the method by which plants utilize RNA Polymerase V transcribed and processed siRNAs to regulate DNA methylation and TGS (reviewed in (81)). Notably, in plants there is a requirement for RNA-dependent RNA polymerase (RdRP) activity to amplify RNA polymerase V transcript-directed TGS (81), whereas humans lack such a polymerase, which opens up a methodology for specific RNA-directed epigenetic modes of regulation. Curiously, this is exactly what lncRNAs appear to be doing in human cells via cis and trans-specific targeting of epigenetic complexes to particular loci (Figure 3 and Table 1), similar to what is also observed in Saccharomyces cerevisiae, which also lacks RdRP activity (82,83). Figure 3. Antisense RNA-directed TGS. Small antisense non-coding RNAs can be (A) introduced into the nucleus and (B) interact with and recruit epigenetic silencing complexes consisting of DNMT3a, Ago1, EZH2 and HDAC1 to homology containing targeted loci by interactions with low copy promoter-associated transcripts resulting in (C) epigenetic silencing consisting of histone and DNA methylation and ultimately chromatin compaction of the targeted locus. (D) Long antisense non-coding RNAs have also been observed to interact with similar epigenetic silencing complexes (53,54) and (E) localize with these complexes at targeted loci resulting in (C) epigenetic silencing of the lncRNA targeted locus. Early studies carried out with S. cerevisiae indicated that antisense non-coding RNAs function endogenously to direct epigenetic gene silencing in place of RdRP-mediated mechanisms (82,83). The parallels between S. cerevisiae and previous observations of small antisense RNA-directed TGS in human cells (18) have emerged, suggesting that antisense transcripts also function to direct TGS (Figure 3). Most notable are the observations that particular antisense lncRNAs, first observed in tumor suppressor genes, p15 (38) and p21(39), function to epigenetically modulate their protein-coding counterparts (Figure 1, Table 1). One interesting, and surprisingly overlooked, early study found antisense transcription was involved in DNA methylation in Thalassemia (40), and even early work linked antisense transcripts and DNA methylation in regulating HIV (51) and MYC (84,85). Mechanistically, far less is known about how antisense lncRNAs direct epigenetic silencing in human cells. Studies carried out with the lncRNA, HOTAIR, indicate that bimodal chromatin modifying complexes can be localized to the HOX locus via the action of this lncRNA (86). A common theme is also evident with Kcnq1ot1 (45) and the p53 regulatory lincRNA p21, which indicates that the entire p53 expressed pathway is controlled by the action of this lncRNA at the p53 locus (57). Indeed, many lncRNAs have been observed to be associated with chromatin (87), but mechanistic insights into the process of lncRNA-directed gene regulation remain less clear. Interesting insights into the mechanism of action of lncRNA-directed TGS came from a recent study looking at the PTEN pseudogene. It had been reported previously that the PTEN pseudogene functions as a miRNA ‘sponge’ (64), similar to the CEBPA lncRNA that acts to sponge DNMT1 away from the CEBPA promoter (88). Studies to interrogate the PTEN pseudogene in greater detailed determined that this pseudogene also expressed an antisense lncRNA in trans which functions to direct TGS to the PTEN promoter and control PTEN expression epigenetically (53). Mechanistically, the PTEN pseudogene expressed antisense lncRNA modulated PTEN transcription by recruiting DNMT3a and EZH2 to the PTEN promoter. The parallels between the functions of the PTEN pseudogene and previous observations with small antisense ncRNA-directed TGS are notable, as both involved the action of DNMT3a (Figure 3). It is noteworthy that DNMT3a is the only known de novo DNA methyltransferase in human cells (89) and has been observed previously to be the only DNA methylatransferase to bind non-coding RNAs including small ncRNAs, both antisense and double stranded RNAs (18,77,78,90), and lncRNAs (53,91). There is an interesting connection between DNMT3a and epigenetic silencing, which including studies indicating DNTM3a co-immunoprecipitates with HDAC1 (92,93) and EZH2 (94), as well as early predictions that DNA methylation is an active participant in X-inactivation (95), one of the first bona fide lncRNA regulatory pathways described. Collectively, a paradigm is emerging in human cells, which proposes that non-coding RNAs, both small and long forms (Figure 3), function through the action of DNMT3a to modulate chromatin and epigenetic states of gene expression. While there are several other mechanisms of action described for lncRNAs in human cells, the interactions with DNMT3a and targeting of transcriptional and epigenetic states is of particular interest, as this mode of gene regulation has the potential to be long-lasting, heritable and may be of significant relevance to the development of targeted therapeutics (reviewed in (96)). Therapeutic applications of RNA-directed epigenetic regulation of gene expression The utility of small RNA-induced TGS as a therapeutic has been largely ignored, mainly due to the pervasiveness of using RNAi targeted approaches to degrade mRNAs. The main concern with RNAi and post-transcriptional mechanisms of gene silencing (Figure 2) is the duration of their therapeutic effect. The effector siRNAs required to drive RNAi must be administered continuously to repress a therapeutic target gene. This is not the case with RNA-induced TGS, where stable, long-term, silencing can be achieved following a relatively short duration of promoter targeting with the siRNAs (19,20,97–100) or small antisense RNA (14,101). This is because the mode of action for the observed gene silencing is transcriptional and driven ultimately by epigenetic silencing (79,102) and not ‘slicing’ of the genes messenger RNA as is the case with RNAi. One universal hurdle that both RNAi and RNA-induced TGS face with is the targeted delivery of the effector RNAs to those cells requiring treatment. One approach is to utilize synthetic antisense oligonucleotides targeted to promoters of interest. This approach has worked with regards to blocking transcription (103) but was not found to induce robust epigenetic silencing, unless the particular oligonucleotides were RNA based (104). However, it may be that better interrogation of each non-coding RNA targeted promoter is required to delineate the best promoter-associated transcripts to target and that many of the earlier studies may have neglected this notion. Indeed, establishing TGS in the absence of a target promoter RNA has not been reported and attempts by some groups, including ours, have proven fruitless. Another approach might be to deliver the effector RNAs using receptor targeted aptamers, which has shown promise for targeting HIV infected cells (105,106). While delivery remains an important concern, the notion that one needs to only target a particular gene for 2–4 days to instill stable epigenetic silencing is promising with regards to minimizing the need for sustained delivery. Recent studies suggest that small RNA-directed TGS is feasible and that stable epigenetic marks can be imposed at small RNA target loci in vivo (99,107). Another area of therapeutic utility can be found in the plethora of lncRNAs that are appearing to be involved in various diseases. Emerging evidence suggest that non-coding RNAs play a wide role (108) in various disease states in humans. Genome-wide observations of diseased states, such as heart failure (109), indicate significant differential and discordant expression between protein-coding and non-coding antisense and pseudogenes is prevalent (110). To date the list of those lncRNAs involved in human diseases is expanding at an unprecedented rate. LncRNAs have been observed in disease ranging from Cancer (57,86,105,106), to HIV (111,112), to autism (113), to pluripotency and differentiation (114–116). It is worth underscoring that many of the disease relevant lncRNAs have been observed to be antisense to particular protein coding genes. A significant obstacle to using RNAi and other post-transcriptional effectors for targeting antisense lncRNAs is the fact that double stranded siRNAs have an ability to target both sense and antisense transcripts (117). The use of RNA-directed TGS avoids this issue by targeting the lncRNA promoter with single stranded antisense transcripts (52). The targeting of endogenous effector antisense lncRNAs can result in the de-repression and subsequent transcriptional activation of the lncRNA targeted locus (Figure 4). Using this mode of action, it becomes feasible to activate gene expression to affect those protein-coding genes under sustained lncRNA-directed TGS (Figure 4). This has proven an effective approach to inducing genes both in vitro (39,52,118–121) and in vivo (42,107,122), but presupposes that there are known antisense lncRNAs regulating the therapeutic target gene. Collectively, the advantages to using RNA-directed TGS as a therapeutic are many and include: (i) strand specific targeting of a gene, (ii) stable long-term epigenetic based silencing can be established to particular genes of therapeutic interest and (iii) antisense RNA-based approaches work as well, if not better than double stranded RNAs, as the endogenous pathway of RNA-directed TGS appears to contain significant overlap with small antisense RNAs and antisense lncRNAs (Figure 3). Figure 4. LncRNA pathways of transcriptional silencing and de-repression. LncRNAs can be expressed in (A) Cis or trans and can (B) interact with those proteins involved in epigenetic silencing. The lncRNAs act to (C) target and tether the epigenetic silencing complexes to homology containing loci resulting in (D) chromatin compaction and transcrtiptinal gene silencing of the targeted locus. These endogenous regulatory lncRNAs can be targeted with (E) antisense oligonucleotides or (F) siRNAs, which results in the loss of the lncRNA and activation/de-repression of those loci actively under lncRNA regulation. CONCLUSION It has been roughly 10 years since the first observation that promoter-directed RNAs can affect gene transcription (Figure 1 and Table 1). This seminal observation in 2004 (9) was indicative of a role for RNA in regulating gene expression, a notion proposed ∼5 decades ago but largely overlooked (123,124). Possible reasons for the poor early adoption of RNA-directed TGS (Table 1) are varied but may include (i) the unfortunate retraction of a similar paper published in Nature (125), and/or (ii) the overwhelmingly positive response to PTGS and the rejection of any RNAi-related phenomena occurring in the nucleus, despite the fact that RNAi was shown to be functional in the human nucleus in 2005 (126) and confirmed in many subsequent studies (25–27,127–131). The notion that RNA may function as the master gene regulator in the cell was something proposed by Britten and Davidson in 1969 (123), which at the time was largely neglected by the broad scientific community. With the advent of high-throughput technologies and the findings from ENCODE, that most of the human genome is transcribed and likely plays a functional role (132–139), it is becoming apparent that Britten and Davidson's theory should be reappraised. Certainly, lncRNAs are abundantly active in the nucleus, and many of them are active modulators of transcriptional and epigenetic modes of gene expression (reviewed in (37,76) and appear to share many of the mechanistic characteristics observed in small RNA-directed TGS (Figure 3). Collectively, the mounting observations that antisense non-coding RNAs, both small and long RNAs, directed to gene promoters can affect transcription by the recruitment of silent state epigenetic complexes suggests that a pervasive and underappreciated role for non-coding RNAs is part of the basic fabric of life. Knowledge of this molecular pathway may prove incredibly insightful with regards to the development of disease, including epigenetic silencing of gene expression and the development of new-targeted therapeutics aimed at specifically affecting gene expression. The next decade could prove an exciting time for our understanding of non-coding RNAs in the transcriptional gene expression and their application as novel therapeutics. FUNDING KVM acknowledges support from NIH [PO1 AI099783-01, RO1 CA151574-01, R01 DK104681-01] and the Australian Research Council Future Fellowship [FT130100572]. Funding for open access charge: [PO1 AI099783-01, RO1 CA151574-01, RO1 DK104681-01]; ARC Future Fellowship [T130100572] to KVM. MSW acknowledges support from the Strategic Health Innovation Partnerships (SHIP) Unit of South African Medical Research Council (SAMRC), with funds received from the South African Department of Science and Technology (DST). Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Matzke M.A. Primig M. Trnovsky J. Matzke A.J.M. Reversible methylation and inactivation of marker genes in sequentially transformed tobacco plants EMBO J. 1989 8 643 649 16453872 2. Mette M.F. Aufsatz W. Van der Winden J. Matzke A.J.M. Matzke M.A. Transcriptional silencing and promoter methylation triggered by double-stranded RNA EMBO J. 2000 19 5194 5201 11013221 3. Wassenegger M. Heimes S. Riedel L. Sänger H.L. RNA-directed de novo methylation of genomic sequences in plants Cell 1994 76 567 576 8313476 4. Lippman Z. May B. Yordan C. Singer T. Martienssen R. Distinct mechanisms determine transposon inheritance and methylation via small interfering RNA and histone modification PLoS Biol. 2003 1 E67 14691539 5. Waterhouse P.M. Graham M.W. Wang M.B. Virus resistance and gene silencing in plants can be induced by simultaneous expression of sense and antisense RNA Proc. Natl. Acad. Sci. U.S.A. 1998 95 13959 13964 9811908 6. Fire A. Xu S. Montgomery M.K. Kostas S.A. Driver S.E. Mello C.C. Potent and specific genetic interference by double-stranded RNA in Caenorhabditis elegans Nature 1998 391 806 811 9486653 7. Green V.A. Weinberg M.S. Small RNA-induced transcriptional gene regulation in mammals mechanisms, therapeutic applications, and scope within the genome Prog. Mol. Biol. Transl. Sci. 2011 102 11 46 21846568 8. Morris K.V. RNA-directed transcriptional gene silencing and activation in human cells Oligonucleotides 2009 19 299 306 19943804 9. Morris K.V. Chan S.W. Jacobsen S.E. Looney D.J. Small interfering RNA-induced transcriptional gene silencing in human cells Science 2004 305 1289 1292 15297624 10. Ting A.H. Schuebel K.E. Herman J.G. Baylin S.B. Short double-stranded RNA induces transcriptional gene silencing in human cancer cells in the absence of DNA methylation Nat. Genet. 2005 37 906 910 16025112 11. Janowski B.A. Huffman K.E. Schwartz J.C. Ram R. Hardy D. Shames D.S. Minna J.D. Corey D.R. Inhibiting gene expression at transcription start sites in chromosomal DNA with antigene RNAs Nat. Chem. Biol. 2005 1 210 215 16408037 12. Morris K.V. RNA-mediated transcriptional gene silencing in human cells Curr. Top. Microbiol. Immunol. 2008 320 211 224 18268846 13. Morris K.V. siRNA-mediated transcriptional gene silencing: the potential mechanism and a possible role in the histone code Cell Mol. Life Sci. 2005 62 3057 3066 16314933 14. Hawkins P.G. Santoso S. Adams C. Anest V. Morris K.V. Promoter targeted small RNAs induce long-term transcriptional gene silencing in human cells Nucleic Acids Res. 2009 2984 2995 19304753 15. Castanotto D. Tommasi S. Li M. Li H. Yanow S. Pfeifer G.P. Rossi J.J. Short hairpin RNA-directed cytosine (CpG) methylation of the RASSF1A gene promoter in HeLa cells Mol. Ther. 2005 12 179 183 15963934 16. Han J. Kim D. Morris K.V. Promoter-associated RNA is required for RNA-directed transcriptional gene silencing in human cells Proc. Natl. Acad. Sci. U.S.A. 2007 104 12422 12427 17640892 17. Kim D.H. Villeneuve L.M. Morris K.V. Rossi J.J. Argonaute-1 directs siRNA-mediated transcriptional gene silencing in human cells Nat. Struct. Mol. Biol. 2006 13 793 797 16936726 18. Weinberg M.S. Villeneuve L.M. Ehsani A. Amarzguioui M. Aagaard L. Chen Z.X. Riggs A.D. Rossi J.J. Morris K.V. The antisense strand of small interfering RNAs directs histone methylation and transcriptional gene silencing in human cells RNA 2006 12 256 262 16373483 19. Suzuki K. Ishida T. Yamagishi M. Ahlenstiel C. Swaminathan S. Marks K. Murray D. McCartney E.M. Beard M.R. Alexander M. Transcriptional gene silencing of HIV-1 through promoter targeted RNA is highly specific RNA Biol. 2011 8 1035 1046 21955498 20. Suzuki K. Juelich T. Lim H. Ishida T. Watanebe T. Cooper D.A. Rao S. Kelleher A.D. Closed chromatin architecture is induced by an RNA duplex targeting the HIV-1 promoter region J. Biol. Chem. 2008 283 23353 23363 18519571 21. Yamagishi M. Ishida T. Miyake A. Cooper D.A. Kelleher A.D. Suzuki K. Watanabe T. Retroviral delivery of promoter-targeted shRNA induces long-term silencing of HIV-1 transcription Microbes Infect. 2009 11 500 508 19233310 22. Napoli S. Pastori C. Magistri M. Carbone G.M. Catapano C.V. Promoter-specific transcriptional interference and c-myc gene silencing by siRNAs in human cells EMBO J. 2009 28 1708 1719 19461583 23. Seila A.C. Calabrese J.M. Levine S.S. Yeo G.W. Rahl P.B. Flynn R.A. Young R.A. Sharp P.A. Divergent transcription from active promoters Science 2008 322 1849 1851 19056940 24. Core L.J. Waterfall J.J. Lis J.T. Nascent RNA sequencing reveals widespread pausing and divergent initiation at human promoters Science 2008 322 1845 1848 19056941 25. Janowski B.A. Huffman K.E. Schwartz J.C. Ram R. Nordsell R. Shames D.S. Minna J.D. Corey D.R. Involvement of AGO1 and AGO2 in mammalian transcriptional silencing Nat. Struct. Mol. Biol. 2006 13 787 792 16936728 26. Ahlenstiel C.L. Lim H.G. Cooper D.A. Ishida T. Kelleher A.D. Suzuki K. Direct evidence of nuclear Argonaute distribution during transcriptional silencing links the actin cytoskeleton to nuclear RNAi machinery in human cells Nucleic Acids Res. 2012 40 1579 1595 22064859 27. Kim D.H. Saetrom P. Snove O. Jr Rossi J.J. MicroRNA-directed transcriptional gene silencing in mammalian cells Proc. Natl. Acad. Sci. U.S.A. 2008 105 16230 16235 18852463 28. Omoto S. Fujii Y.R. Regulation of human immunodeficiency virus 1 transcription by nef microRNA J. Gen. Virol. 2005 86 751 755 15722536 29. Klase Z. Kale P. Winograd R. Gupta M.V. Heydarian M. Berro R. McCaffrey T. Kashanchi F. HIV-1 TAR element is processed by Dicer to yield a viral micro-RNA involved in chromatin remodeling of the viral LTR BMC Mol. Biol. 2007 8 63 17663774 30. Tan Y. Zhang B. Wu T. Skogerbo G. Zhu X. Guo X. He S. Chen R. Transcriptional inhibition of Hoxd4 expression by noncoding RNAs in human breast cancer cells BMC Mol. Biol. 2009 10 12 19232136 31. Carninci P. Kasukawa T. Katayama S. Gough J. Frith M.C. Maeda N. Oyama R. Ravasi T. Lenhard B. Wells C. The transcriptional landscape of the mammalian genome Science 2005 309 1559 1563 16141072 32. Katayama S. Tomaru Y. Kasukawa T. Waki K. Nakanishi M. Nakamura M. Nishida H. Yap C.C. Suzuki M. Kawai J. Antisense transcription in the mammalian transcriptome Science 2005 309 1564 1566 16141073 33. Dahary D. Elroy-Stein O. Sorek R. Naturally occurring antisense: transcriptional leakage or real overlap Genome Res. 2005 15 364 368 15710751 34. Lavorgna G. Dahary D. Lehner B. Sorek R. Sanderson C.M. Casari G. In search of antisense Trends Biochem. Sci. 2004 29 88 94 15102435 35. Latos P.A. Barlow D.P. Regulation of imprinted expression by macro non-coding RNAs RNA Biol. 2009 6 100 106 19229135 36. Malecova B. Morris K.V. Transcriptional gene silencing through epigenetic changes mediated by non-coding RNAs Curr. Opin. Mol. Ther. 2010 12 214 222 20373265 37. Morris K.V. Long antisense non-coding RNAs function to direct epigenetic complexes that regulate transcription in human cells Epigenetics 2009 4 296 301 19633414 38. Yu W. Gius D. Onyango P. Muldoon-Jacobs K. Karp J. Feinberg A.P. Cui H. Epigenetic silencing of tumour suppressor gene p15 by its antisense RNA Nature 2008 451 202 206 18185590 39. Morris K.V. Santoso S. Turner A.M. Pastori C. Hawkins P.G. Bidirectional transcription directs both transcriptional gene activation and suppression in human cells PLoS Genet. 2008 4 e1000258 19008947 40. Tufarelli C. Stanley J.A. Garrick D. Sharpe J.A. Ayyub H. Wood W.G. Higgs D.R. Transcription of antisense RNA leading to gene silencing and methylation as a novel cause of human genetic disease Nat. Genet. 2003 34 157 165 12730694 41. Lipovich L. Dachet F. Cai J. Bagla S. Balan K. Jia H. Loeb J.A. Activity-dependent human brain coding/noncoding gene regulatory networks Genetics 2012 192 1133 1148 22960213 42. Modarresi F. Faghihi M.A. Lopez-Toledano M.A. Fatemi R.P. Magistri M. Brothers S.P. van der Brug M.P. Wahlestedt C. Inhibition of natural antisense transcripts in vivo results in gene-specific transcriptional upregulation Nat. Biotechnol. 2012 30 453 459 22446693 43. Vadie N. Saayman S. Lenox A. Ackley A. Clemson M. Burdach J. Hart J. Vogt P.K. Morris K.V. MYCNOS functions as an antisense RNA regulating MYCN RNA Biol. 2015 12 893 899 26156430 44. Li Q. Su Z. Xu X. Liu G. Song X. Wang R. Sui X. Liu T. Chang X. Huang D. AS1DHRS4, a head-to-head natural antisense transcript, silences the DHRS4 gene cluster in cis and trans Proc. Natl. Acad. Sci. U.S.A. 2012 109 14110 14115 22891334 45. Kanduri C. Kcnq1ot1: a chromatin regulatory RNA Semin. Cell Dev. Biol. 2011 22 343 350 21345374 46. Mohammad F. Pandey R.R. Nagano T. Chakalova L. Mondal T. Fraser P. Kanduri C. Kcnq1ot1/Lit1 noncoding RNA mediates transcriptional silencing by targeting to the perinucleolar region Mol. Cell Biol. 2008 28 3713 3728 18299392 47. Pandey R.R. Mondal T. Mohammad F. Enroth S. Redrup L. Komorowski J. Nagano T. Mancini-Dinardo D. Kanduri C. Kcnq1ot1 antisense noncoding RNA mediates lineage-specific transcriptional silencing through chromatin-level regulation Mol. Cell 2008 32 232 246 18951091 48. Pandey G.K. Mitra S. Subhash S. Hertwig F. Kanduri M. Mishra K. Fransson S. Ganeshram A. Mondal T. Bandaru S. The risk-associated long noncoding RNA NBAT-1 controls neuroblastoma progression by regulating cell proliferation and neuronal differentiation Cancer Cell 2014 26 722 737 25517750 49. Landry S. Halin M. Lefort S. Audet B. Vaquero C. Mesnard J.M. Barbeau B. Detection, characterization and regulation of antisense transcripts in HIV-1 Retrovirology 2007 4 71 17910760 50. Ludwig L.B. Ambrus J.L. Jr Krawczyk K.A. Sharma S. Brooks S. Hsiao C.B. Schwartz S.A. Human immunodeficiency virus-Type 1 LTR DNA contains an intrinsic gene producing antisense RNA and protein products Retrovirology 2006 3 80 17090330 51. Miller R.H. Human immunodeficiency virus may encode a novel protein on the genomic DNA plus strand Science 1988 239 1420 1422 3347840 52. Saayman S. Ackley A. Turner A.M. Famiglietti M. Bosque A. Clemson M. Planelles V. Morris K.V. An HIV-encoded antisense long noncoding RNA epigenetically regulates viral transcription Mol. Ther. 2014 22 1164 1175 24576854 53. Johnsson P. Ackley A. Vidarsdottir L. Lui W.O. Corcoran M. Grander D. Morris K.V. A pseudogene long-noncoding-RNA network regulates PTEN transcription and translation in human cells Nat. Struct. Mol. Biol. 2013 20 440 446 23435381 54. Hawkins P.G. Morris K.V. Transcriptional regulation of Oct4 by a long non-coding RNA antisense to Oct4-pseudogene 5 Transcription 2010 1 165 175 21151833 55. Mohammad F. Pandey G.K. Mondal T. Enroth S. Redrup L. Gyllensten U. Kanduri C. Long noncoding RNA-mediated maintenance of DNA methylation and transcriptional gene silencing Development 2012 139 2792 2803 22721776 56. Bierhoff H. Schmitz K. Maass F. Ye J. Grummt I. Noncoding transcripts in sense and antisense orientation regulate the epigenetic state of ribosomal RNA genes Cold Spring Harb. Symp. Quant. Biol. 2010 75 357 364 21502405 57. Huarte M. Guttman M. Feldser D. Garber M. Koziol M.J. Kenzelmann-Broz D. Khalil A.M. Zuk O. Amit I. Rabani M. A large intergenic noncoding RNA induced by p53 mediates global gene repression in the p53 response Cell 2010 142 409 419 20673990 58. Gong C. Li Z. Ramanujan K. Clay I. Zhang Y. Lemire-Brachat S. Glass D.J. A long non-coding RNA, LncMyoD, regulates skeletal muscle differentiation by blocking IMP2-mediated mRNA translation Dev. Cell 2015 34 181 191 26143994 59. Wang P. Xue Y. Han Y. Lin L. Wu C. Xu S. Jiang Z. Xu J. Liu Q. Cao X. The STAT3-binding long noncoding RNA lnc-DC controls human dendritic cell differentiation Science 2014 344 310 313 24744378 60. Blewitt M.E. Gendrel A.V. Pang Z. Sparrow D.B. Whitelaw N. Craig J.M. Apedaile A. Hilton D.J. Dunwoodie S.L. Brockdorff N. SmcHD1, containing a structural-maintenance-of-chromosomes hinge domain, has a critical role in X inactivation Nat. Genet. 2008 40 663 669 18425126 61. Tsai M.C. Manor O. Wan Y. Mosammaparast N. Wang J.K. Lan F. Shi Y. Segal E. Chang H.Y. Long noncoding RNA as modular scaffold of histone modification complexes Science 2010 329 689 693 20616235 62. Wang L. Guo Z.Y. Zhang R. Xin B. Chen R. Zhao J. Wang T. Wen W.H. Jia L.T. Yao L.B. Pseudogene OCT4-pg4 functions as a natural micro RNA sponge to regulate OCT4 expression by competing for miR-145 in hepatocellular carcinoma Carcinogenesis 2013 34 1773 1781 23615404 63. Wang Y. Xu Z. Jiang J. Xu C. Kang J. Xiao L. Wu M. Xiong J. Guo X. Liu H. Endogenous miRNA sponge lincRNA-RoR regulates Oct4, Nanog, and Sox2 in human embryonic stem cell self-renewal Dev. Cell 2013 25 69 80 23541921 64. Poliseno L. Salmena L. Zhang J. Carver B. Haveman W.J. Pandolfi P.P. A coding-independent function of gene and pseudogene mRNAs regulates tumour biology Nature 2010 465 1033 1038 20577206 65. Zhou H.L. Luo G. Wise J.A. Lou H. Regulation of alternative splicing by local histone modifications: potential roles for RNA-guided mechanisms Nucleic Acids Res. 2014 42 701 713 24081581 66. Liu Y. Luo F. Xu Y. Wang B. Zhao Y. Xu W. Shi L. Lu X. Liu Q. Epithelial-mesenchymal transition and cancer stem cells, mediated by a long non-coding RNA, HOTAIR, are involved in cell malignant transformation induced by cigarette smoke extract Toxicol. Appl. Pharmacol. 2015 282 9 19 25447409 67. Xue M. Li X. Li Z. Chen W. Urothelial carcinoma associated 1 is a hypoxia-inducible factor-1alpha-targeted long noncoding RNA that enhances hypoxic bladder cancer cell proliferation, migration, and invasion Tumour Biol. 2014 35 6901 6912 24737584 68. Gutschner T. Hammerle M. Eissmann M. Hsu J. Kim Y. Hung G. Revenko A. Arun G. Stentrup M. Gross M. The noncoding RNA MALAT1 is a critical regulator of the metastasis phenotype of lung cancer cells Cancer Res. 2013 73 1180 1189 23243023 69. Ramos A.D. Andersen R.E. Liu S.J. Nowakowski T.J. Hong S.J. Gertz C.C. Salinas R.D. Zarabi H. Kriegstein A.R. Lim D.A. The long noncoding RNA Pnky regulates neuronal differentiation of embryonic and postnatal neural stem cells Cell Stem Cell 2015 16 439 447 25800779 70. Orom U.A. Derrien T. Beringer M. Gumireddy K. Gardini A. Bussotti G. Lai F. Zytnicki M. Notredame C. Huang Q. Long noncoding RNAs with enhancer-like function in human cells Cell 2010 143 46 58 20887892 71. Kim T.K. Hemberg M. Gray J.M. Costa A.M. Bear D.M. Wu J. Harmin D.A. Laptewicz M. Barbara-Haley K. Kuersten S. Widespread transcription at neuronal activity-regulated enhancers Nature 2010 465 182 187 20393465 72. Schaukowitch K. Joo J.Y. Liu X. Watts J.K. Martinez C. Kim T.K. Enhancer RNA facilitates NELF release from immediate early genes Mol. Cell 2014 56 29 42 25263592 73. Pulakanti K. Pinello L. Stelloh C. Blinka S. Allred J. Milanovich S. Kiblawi S. Peterson J. Wang A. Yuan G.C. Enhancer transcribed RNAs arise from hypomethylated, Tet-occupied genomic regions Epigenetics 2013 8 1303 1320 24135681 74. Beckedorff F.C. Ayupe A.C. Crocci-Souza R. Amaral M.S. Nakaya H.I. Soltys D.T. Menck C.F. Reis E.M. Verjovski-Almeida S. The intronic long noncoding RNA ANRASSF1 recruits PRC2 to the RASSF1A promoter, reducing the expression of RASSF1A and increasing cell proliferation PLoS Genet. 2013 9 e1003705 23990798 75. Kotake Y. Nakagawa T. Kitagawa K. Suzuki S. Liu N. Kitagawa M. Xiong Y. Long non-coding RNA ANRIL is required for the PRC2 recruitment to and silencing of p15(INK4B) tumor suppressor gene Oncogene 2011 30 1956 1962 21151178 76. Morris K.V. Mattick J.S. The rise of regulatory RNA Nat. Rev. Genet. 2014 15 423 437 24776770 77. Holz-Schietinger C. Reich N.O. RNA modulation of the human DNA methyltransferase 3A Nucleic Acids Res. 2012 40 8550 8557 22730298 78. Ross J.P. Suetake I. Tajima S. Molloy P.L. Recombinant mammalian DNA methyltransferase activity on model transcriptional gene silencing short RNA-DNA heteroduplex substrates Biochem. J. 2010 432 323 332 20846120 79. Suzuki K. Shijuuku T. Fukamachi T. Zaunders J. Guillemin G. Cooper D. Kelleher A. Prolonged transcriptional silencing and CpG methylation induced by siRNAs targeted to the HIV-1 promoter region J. RNAi Gene Silencing 2005 1 66 78 19771207 80. Turner A.M. Ackley A.M. Matrone M.A. Morris K.V. Characterization of an HIV-targeted transcriptional gene-silencing RNA in primary cells Hum. Gene Ther. 2012 23 473 483 22122263 81. Matzke M.A. Mosher R.A. RNA-directed DNA methylation: an epigenetic pathway of increasing complexity Nat. Rev. Genet. 2014 15 394 408 24805120 82. Camblong J. Iglesias N. Fickentscher C. Dieppois G. Stutz F. Antisense RNA stabilization induces transcriptional gene silencing via histone deacetylation in S. cerevisiae Cell 2007 131 706 717 18022365 83. Camblong J. Beyrouthy N. Guffanti E. Schlaepfer G. Steinmetz L.M. Stutz F. Trans-acting antisense RNAs mediate transcriptional gene cosuppression in S. cerevisiae Genes Dev. 2009 23 1534 1545 19571181 84. Celano P. Berchtold C.M. Kizer D.L. Weeraratna A. Nelkin B.D. Baylin S.B. Casero R.A. Jr Characterization of an endogenous RNA transcript with homology to the antisense strand of the human c-myc gene J. Biol. Chem. 1992 267 15092 15096 1378845 85. Spicer D.B. Sonenshein G.E. An antisense promoter of the murine c-myc gene is localized within intron 2 Mol. Cell. Biol. 1992 12 1324 1329 1545813 86. Gupta R.A. Shah N. Wang K.C. Kim J. Horlings H.M. Wong D.J. Tsai M.C. Hung T. Argani P. Rinn J.L. Long non-coding RNA HOTAIR reprograms chromatin state to promote cancer metastasis Nature 2010 464 1071 1076 20393566 87. Khalil A.M. Guttman M. Huarte M. Garber M. Raj A. Rivea Morales D. Thomas K. Presser A. Bernstein B.E. van Oudenaarden A. Many human large intergenic noncoding RNAs associate with chromatin-modifying complexes and affect gene expression Proc. Natl. Acad. Sci. U.S.A. 2009 106 11667 11672 19571010 88. Di Ruscio A. Ebralidze A.K. Benoukraf T. Amabile G. Goff L.A. Terragni J. Figueroa M.E. De Figureido Pontes L.L. Alberich-Jorda M. Zhang P. DNMT1-interacting RNAs block gene-specific DNA methylation Nature 2013 503 371 376 24107992 89. Bestor T.H. The DNA methyltransferases of mammals Hum. Mol. Genet. 2000 9 2395 2402 11005794 90. Jeffery L. Nakielny S. Components of the DNA methylation system of chromatin control are RNA-binding proteins J. Biol. Chem. 2004 279 49479 49487 15342650 91. Wang L. Zhao Y. Bao X. Zhu X. Kwok Y.K. Sun K. Chen X. Huang Y. Jauch R. Esteban M.A. LncRNA Dum interacts with Dnmts to regulate Dppa2 expression during myogenic differentiation and muscle regeneration Cell Res. 2015 25 335 350 25686699 92. Fuks F. Hurd P.J. Deplus R. Kouzarides T. The DNA methyltransferases associate with HP1 and the SUV39H1 histone methyltransferase Nucleic Acids Res. 2003 31 2305 2312 12711675 93. Fuks F. Burgers W.A. Godin N. Kasai M. Kouzarides T. Dnmt3a binds deacetylases and is recruited by a sequence-specific repressor to silence transcription EMBO J. 2001 20 2536 2544 11350943 94. Vire E. Brenner C. Deplus R. Blanchon L. Fraga M. Didelot C. Morey L. Van Eynde A. Bernard D. Vanderwinden J.M. The Polycomb group protein EZH2 directly controls DNA methylation Nature 2006 439 871 874 16357870 95. Riggs A.D. X inactivation, differentiation, and DNA methylation Cytogenet. Cell Genet. 1975 14 9 25 1093816 96. Morris K.V. Non-coding RNAs, epigenetic memory, and the passage of information to progeny RNA Biol. 2009 6 242 247 19305164 97. Mehndiratta M. Palanichamy J.K. Pal A. Bhagat M. Singh A. Sinha S. Chattopadhyay P. CpG hypermethylation of the C-myc promoter by dsRNA results in growth suppression Mol. Pharm. 2011 8 2302 2309 21879731 98. Esmaeili F. Bamdad T. Ghasemi S. Stable suppression of gene expression by short interfering RNAs targeted to promoter in a mouse embryonal carcinoma stem cell line In Vitro Cell. Dev. Biol. Anim. 2010 46 834 840 20872183 99. Suzuki K. Hattori S. Marks K. Ahlenstiel C. Maeda Y. Ishida T. Millington M. Boyd M. Symonds G. Cooper D.A. Promoter targeting shRNA suppresses HIV-1 infection in vivo through transcriptional gene silencing Mol. Ther. Nucleic acids 2013 2 e137 24301868 100. Lim H.G. Suzuki K. Cooper D.A. Kelleher A.D. Promoter-targeted siRNAs induce gene silencing of simian immunodeficiency virus (SIV) infection in vitro Mol. Ther. 2008 16 565 570 18227841 101. Lister N.C. Clemson M. Morris K.V. RNA-directed epigenetic silencing of periostin inhibits cell motility R. Soc. Open Sci. 2015 2 140545 26543579 102. Kim J.W. Zhang Y.H. Zern M.A. Rossi J.J. Wu J. Short hairpin RNA causes the methylation of transforming growth factor-beta receptor II promoter and silencing of the target gene in rat hepatic stellate cells Biochem. Biophys. Res. Commun. 2007 359 292 297 17533113 103. Watts J.K. Yu D. Charisse K. Montaillier C. Potier P. Manoharan M. Corey D.R. Effect of chemical modifications on modulation of gene expression by duplex antigene RNAs that are complementary to non-coding transcripts at gene promoters Nucleic Acids Res. 2010 38 5242 5259 20403811 104. Knowling S.S.K. Turner A.M. Uhlmann E. Lehmann T. Vollmer J. Morris K.V. Chemically modified oligonucleotides modulate an epigenetically varied and transient form of transcription silencing of HIV-1 in human cells Mol. Ther. Nucleic Acids 2012 1 e16 23343927 105. Zhou J. Satheesan S. Li H. Weinberg M.S. Morris K.V. Burnett J.C. Rossi J.J. Cell-specific RNA aptamer against human CCR5 specifically targets HIV-1 susceptible cells and inhibits HIV-1 infectivity Chem. Biol. 2015 22 379 390 25754473 106. Neff C.P. Zhou J. Remling L. Kuruvilla J. Zhang J. Li H. Smith D.D. Swiderski P. Rossi J.J. Akkina R. An aptamer-siRNA chimera suppresses HIV-1 viral loads and protects from helper CD4(+) T cell decline in humanized mice Sci. Transl. Med. 2011 3 66ra66 107. Turunen M.P. Lehtola T. Heinonen S.E. Assefa G.S. Korpisalo P. Girnary R. Glass C.K. Vaisanen S. Yla-Herttuala S. Efficient regulation of VEGF expression by promoter-targeted lentiviral shRNAs based on epigenetic mechanism: a novel example of epigenetherapy Circ. Res. 2009 105 604 609 19696410 108. Hart J.R. Roberts T.C. Weinberg M.S. Morris K.V. Vogt P.K. MYC regulates the non-coding transcriptome Oncotarget 2014 5 12543 12554 25587025 109. Di Salvo T.G. Guo Y. Su Y.R. Clark T. Brittain E. Absi T. Maltais S. Hemnes A. Right ventricular long noncoding RNA expression in human heart failure Pulm. Circ. 2015 5 135 161 25992278 110. Johnsson P. Lipovich L. Grander D. Morris K.V. Evolutionary conservation of long noncoding RNAs; sequence, structure, function Biochim. Biophys. Acta 2013 1840 1063 1071 24184936 111. Saayman S. Ali S.A. Morris K.V. Weinberg M.S. The therapeutic application of CRISPR/Cas9 technologies for HIV Expert Opin. Biol. Ther. 2015 15 819 830 25865334 112. Kobayashi-Ishihara M. Yamagishi M. Hara T. Matsuda Y. Takahashi R. Miyake A. Nakano K. Yamochi T. Ishida T. Watanabe T. HIV-1-encoded antisense RNA suppresses viral replication for a prolonged period Retrovirology 2012 9 38 22569184 113. Kerin T. Ramanathan A. Rivas K. Grepo N. Coetzee G.A. Campbell D.B. A noncoding RNA antisense to moesin at 5p14.1 in autism Sci. Transl. Med. 2012 4 128ra140 114. Guttman M. Donaghey J. Carey B.W. Garber M. Grenier J.K. Munson G. Young G. Lucas A.B. Ach R. Bruhn L. lincRNAs act in the circuitry controlling pluripotency and differentiation Nature 2011 477 295 300 21874018 115. Loewer S. Cabili M.N. Guttman M. Loh Y.H. Thomas K. Park I.H. Garber M. Curran M. Onder T. Agarwal S. Large intergenic non-coding RNA-RoR modulates reprogramming of human induced pluripotent stem cells Nat. Genet. 2010 42 1113 1117 21057500 116. Kretz M. Siprashvili Z. Chu C. Webster D.E. Zehnder A. Qu K. Lee C.S. Flockhart R.J. Groff A.F. Chow J. Control of somatic tissue differentiation by the long non-coding RNA TINCR Nature 2013 493 231 235 23201690 117. Wei J.X. Yang J. Sun J.F. Jia L.T. Zhang Y. Zhang H.Z. Li X. Meng Y.L. Yao L.B. Yang A.G. Both strands of siRNA have potential to guide posttranscriptional gene silencing in mammalian cells PLoS One 2009 4 e5382 19401777 118. Schwartz J.C. Younger S.T. Nguyen N.B. Hardy D.B. Monia B.P. Corey D.R. Janowski B.A. Antisense transcripts are targets for activating small RNAs Nat. Struct. Mol. Biol. 2008 15 842 848 18604220 119. Janowski B.A. Younger S.T. Hardy D.B. Ram R. Huffman K.E. Corey D.R. Activating gene expression in mammalian cells with promoter-targeted duplex RNAs Nat. Chem. Biol. 2007 3 166 173 17259978 120. Li L.C. Okino S.T. Zhao H. Pookot D. Place R.F. Urakami S. Enokida H. Dahiya R. Small dsRNAs induce transcriptional activation in human cells Proc. Natl. Acad. Sci. U.S.A. 2006 103 17337 17342 17085592 121. Place R.F. Li L.C. Pookot D. Noonan E.J. Dahiya R. MicroRNA-373 induces expression of genes with complementary promoter sequences Proc. Natl. Acad. Sci. U.S.A. 2008 105 1608 1613 18227514 122. Turunen M.P. Husso T. Musthafa H. Laidinen S. Dragneva G. Laham-Karam N. Honkanen S. Paakinaho A. Laakkonen J.P. Gao E. Epigenetic upregulation of endogenous VEGF-A reduces myocardial infarct size in mice PLoS One 2014 9 e89979 24587164 123. Britten R.J. Davidson E.H. Gene regulation for higher cells: a theory Science 1969 165 349 357 5789433 124. Britten R.J. Davidson E.H. Repetitive and non-repetitive DNA sequences and a speculation on the origins of evolutionary novelty Q. Rev. Biol. 1971 66 111 138 5160087 125. Taira K. Induction of DNA methylation and gene silencing by short interfering RNAs in human cells Nature 2006 441 1176 16810259 126. Robb G.B. Brown K.M. Khurana J. Rana T.M. Specific and potent RNAi in the nucleus of human cells Nat. Struct. Mol. Biol. 2005 12 133 137 15643423 127. Rudel S. Flatley A. Weinmann L. Kremmer E. Meister G. A multifunctional human Argonaute2-specific monoclonal antibody RNA 2008 14 1244 1253 18430891 128. Ameyar-Zazoua M. Rachez C. Souidi M. Robin P. Fritsch L. Young R. Morozova N. Fenouil R. Descostes N. Andrau J.C. Argonaute proteins couple chromatin silencing to alternative splicing Nat. Struct. Mol. Biol. 2012 19 998 1004 22961379 129. Allo M. Agirre E. Bessonov S. Bertucci P. Gomez Acuna L. Buggiano V. Bellora N. Singh B. Petrillo E. Blaustein M. Argonaute-1 binds transcriptional enhancers and controls constitutive and alternative splicing in human cells Proc. Natl. Acad. Sci. U.S.A. 2014 111 15622 15629 25313066 130. Gagnon K.T. Li L. Janowski B.A. Corey D.R. Analysis of nuclear RNA interference in human cells by subcellular fractionation and Argonaute loading Nat. Protoc. 2014 9 2045 2060 25079428 131. Gagnon K.T. Li L. Chu Y. Janowski B.A. Corey D.R. RNAi factors are present and active in human cell nuclei Cell Rep. 2014 6 211 221 24388755 132. Rosenbloom K.R. Dreszer T.R. Long J.C. Malladi V.S. Sloan C.A. Raney B.J. Cline M.S. Karolchik D. Barber G.P. Clawson H. ENCODE whole-genome data in the UCSC Genome Browser: update 2012 Nucleic Acids Res. 2012 40 D912 D917 22075998 133. Pennisi E. Genomics. ENCODE project writes eulogy for junk DNA Science 2012 337 1159 1161 22955811 134. Pei B. Sisu C. Frankish A. Howald C. Habegger L. Mu X.J. Harte R. Balasubramanian S. Tanzer A. Diekhans M. The GENCODE pseudogene resource Genome Biol. 2012 13 R51 22951037 135. Djebali S. Davis C.A. Merkel A. Dobin A. Lassmann T. Mortazavi A. Tanzer A. Lagarde J. Lin W. Schlesinger F. Landscape of transcription in human cells Nature 2012 489 101 108 22955620 136. Derrien T. Johnson R. Bussotti G. Tanzer A. Djebali S. Tilgner H. Guernec G. Martin D. Merkel A. Knowles D.G. The GENCODE v7 catalog of human long noncoding RNAs: analysis of their gene structure, evolution, and expression Genome Res. 2012 22 1775 1789 22955988 137. Bernstein B.E. Birney E. Dunham I. Green E.D. Gunter C. Snyder M. An integrated encyclopedia of DNA elements in the human genome Nature 2012 489 57 74 22955616 138. Consortium E.P. The ENCODE (ENCyclopedia Of DNA Elements) Project Science 2004 306 636 640 15499007 139. Rinn J.L. Chang H.Y. Genome regulation by long noncoding RNAs Annu. Rev. Biochem. 2012 81 145 166 22663078 140. Suzuki K. Suter S. Ward R. Cooper D.A. Kelleher A.D. Silencing gene expression by short interfering RNA targeting NF-kB binding site within 5′ LTR 10th Conference on Retroviruses and Opportunistic Infections 2003 Boston 141. Turner A.M. De La Cruz J. Morris K.V. Mobilization-competent Lentiviral Vector-mediated Sustained Transcriptional Modulation of HIV-1 Expression Mol. Ther. 2009 17 360 368 19066594 142. Turner A.M. Morris K.V. Controlling transcription with noncoding RNAs in mammalian cells Biotechniques 2010 48 ix–xvi 143. Singh A. Palanichamy J.K. Ramalingam P. Kassab M.A. Bhagat M. Andrabi R. Luthra K. Sinha S. Chattopadhyay P. Long-term suppression of HIV-1C virus production in human peripheral blood mononuclear cells by LTR heterochromatization with a short double-stranded RNA J. Antimicrob. Chemother. 2013 69 404 415 24022068 144. Ahlenstiel C. Mendez C. Lim S.T. Marks K. Turville S. Cooper D.A. Kelleher A.D. Suzuki K. Novel RNA duplex locks HIV-1 in a latent state via chromatin-mediated transcriptional silencing Mol. Ther. Nucleic Acids 2015 4 e261 26506039 145. Zhang M.X. Zhang C. Shen Y.H. Wang J. Li X.N. Chen L. Zhang Y. Coselli J.S. Wang X.L. Effect of 27nt small RNA on endothelial nitric-oxide synthase expression Mol. Biol. Cell 2008 19 3997 4005 18614799 146. Hoffman A.R. Hu J.F. Directing DNA methylation to inhibit gene expression Cell. Mol. Neurobiol. 2006 26 425 438 16710755 147. Allo M. Buggiano V. Fededa J.P. Petrillo E. Schor I. de la Mata M. Agirre E. Plass M. Eyras E. Elela S.A. Control of alternative splicing through siRNA-mediated transcriptional gene silencing Nat. Struct. Mol. Biol. 2009 16 717 724 19543290 148. Park C.W. Chen Z. Kren B.T. Steer C.J. Double-stranded siRNA targeted to the huntingtin gene does not induce DNA methylation Biochem. Biophys. Res. Commun. 2004 323 275 280 15351733 149. Buhler M. Mohn F. Stalder L. Muhlemann O. Transcriptional silencing of nonsense codon-containing immunoglobulin minigenes Mol. Cell 2005 18 307 317 15866173 150. Gonzalez S. Pisano D.G. Serrano M. Mechanistic principles of chromatin remodeling guided by siRNAs and miRNAs Cell Cycle 2008 7 2601 2608 18719372 151. Wang X. Feng Y. Pan L. Wang Y. Xu X. Lu J. Huang B. The proximal GC-rich region of p16(INK4a) gene promoter plays a role in its transcriptional regulation Mol. Cell. Biochem. 2007 301 259 266 17333389 152. Pulukuri S.M. Rao J.S. Small interfering RNA directed reversal of urokinase plasminogen activator demethylation inhibits prostate tumor growth and metastasis Cancer Res. 2007 67 6637 6646 17638874 153. Yue X. Schwartz J.C. Chu Y. Younger S.T. Gagnon K.T. Elbashir S. Janowski B.A. Corey D.R. Transcriptional regulation by small RNAs at sequences downstream from 3′ gene termini Nat. Chem. Biol. 2010 6 621 629 20581822 154. Chu Y. Yue X. Younger S.T. Janowski B.A. Corey D.R. Involvement of argonaute proteins in gene silencing and activation by RNAs complementary to a non-coding transcript at the progesterone receptor promoter Nucleic Acids Res. 2010 38 7736 7748 20675357 155. Cho S. Park J.S. Kang Y.K. AGO2 and SETDB1 cooperate in promoter-targeted transcriptional silencing of the androgen receptor gene Nucleic Acids Res. 2014 42 13545 13556 25183519 156. Ackley A. Lenox A. Stapleton K. Knowling S. Lu T. Sabir K.S. Vogt P.K. Morris K.V. An algorithm for generating small RNAs capable of epigenetically modulating transcriptional gene silencing and activation in human cells Mol. Ther. Nucleic Acids 2013 2 e104 23839098 157. Zakaria M.K. Khan I. Mani P. Chattopadhyay P. Sarkar D.P. Sinha S. Combination of hepatocyte specific delivery and transformation dependent expression of shRNA inducing transcriptional gene silencing of c-Myc promoter in hepatocellular carcinoma cells BMC Cancer 2014 14 582 25108398 158. Palanichamy J.K. Mehndiratta M. Bhagat M. Ramalingam P. Das B. Das P. Sinha S. Chattopadhyay P. Silencing of integrated human papillomavirus-16 oncogenes by small interfering RNA-mediated heterochromatization Mol. Cancer Ther. 2010 9 2114 2122 20587668 159. Roberts T.C. Andaloussi S.E. Morris K.V. McClorey G. Wood M.J. Small RNA-mediated epigenetic myostatin silencing Mol. Ther. Nucleic Acids 2012 1 e23 23344003 160. Feng X.Z. He X.S. Zhuang Y.Z. Luo Q. Jiang J.H. Yang S. Tang X.F. Liu J.L. Chen T. Investigation of transcriptional gene silencing and mechanism induced by shRNAs targeted to RUNX3 in vitro World J. Gastroenterol. 2008 14 3006 3014 18494051 161. Salvi A. Arici B. Portolani N. Giulini S.M. De Petro G. Barlati S. In vitro c-met inhibition by antisense RNA and plasmid-based RNAi down-modulates migration and invasion of hepatocellular carcinoma cells Int. J. Oncol. 2007 31 451 460 17611703 162. Jiang G. Zheng L. Pu J. Mei H. Zhao J. Huang K. Zeng F. Tong Q. Small RNAs targeting transcription start site induce heparanase silencing through interference with transcription initiation in human cancer cells PLoS One 2012 7 e31379 22363633 163. Murayama A. Sakura K. Nakama M. Yasuzawa-Tanaka K. Fujita E. Tateishi Y. Wang Y. Ushijima T. Baba T. Shibuya K. A specific CpG site demethylation in the human interleukin 2 gene promoter is an epigenetic memory EMBO J. 2006 25 1081 1092 16498406 164. Tan Y. Zhang B. Wu T. Skogerbo G. Zhu X. Guo X. He S. Chen R. Transcriptional inhibition of Hoxd4 expression by miRNA-10a in human breast cancer cells BMC Mol. Biol. 2009 10 12 19232136 165. Ouellet D.L. Plante I. Landry P. Barat C. Janelle M.E. Flamand L. Tremblay M.J. Provost P. Identification of functional microRNAs released through asymmetrical processing of HIV-1 TAR element Nucleic Acids Res. 2008 36 2353 2365 18299284 166. Cho D.H. Thienes C.P. Mahoney S.E. Analau E. Filippova G.N. Tapscott S.J. Antisense transcription and heterochromatin at the DM1 CTG repeats are constrained by CTCF Mol. Cell 2005 20 483 489 16285929 167. Wang Q.Y. Zhou C. Johnson K.E. Colgrove R.C. Coen D.M. Knipe D.M. Herpesviral latency-associated transcript gene promotes assembly of heterochromatin on viral lytic-gene promoters in latent infection Proc. Natl. Acad. Sci. U.S.A. 2005 102 16055 16059 16247011 168. Mohammad F. Mondal T. Guseva N. Pandey G.K. Kanduri C. Kcnq1ot1 noncoding RNA mediates transcriptional gene silencing by interacting with Dnmt1 Development 2010 137 2493 2499 20573698 169. Takayama K. Horie-Inoue K. Katayama S. Suzuki T. Tsutsumi S. Ikeda K. Urano T. Fujimura T. Takagi K. Takahashi S. Androgen-responsive long noncoding RNA CTBP1-AS promotes prostate cancer EMBO J. 2013 32 1665 1680 23644382 170. Brenner S. Jacob F. Meselson M. An unstable intermediate carrying information from genes to ribosomes for protein synthesis Nature 1961 190 576 581 20446365 171. Berget S.M. Moore C. Sharp P.A. Spliced segments at the 5′ terminus of adenovirus 2 late mRNA Proc. Natl. Acad. Sci. U.S.A. 1977 74 3171 3175 269380 172. Chow L.T. Gelinas R.E. Broker T.R. Roberts R.J. An amazing sequence arrangement at the 5′ ends of adenovirus 2 messenger RNA Cell 1977 12 1 8 902310 173. Brannan C.I. Dees E.C. Ingram R.S. Tilghman S.M. The product of the H19 gene may function as an RNA Mol. Cell. Biol. 1990 10 28 36 1688465 174. Brockdorff N. Ashworth A. Kay G.F. McCabe V.M. Norris D.P. Cooper P.J. Swift S. Rastan S. The product of the mouse Xist gene is a 15 kb inactive X-specific transcript containing no conserved ORF and located in the nucleus Cell 1992 71 515 526 1423610 175. Lee R.C. Feinbaum R.L. Ambros V. The C. elegans heterochronic gene lin-4 encodes small RNAs with antisense complementarity to lin-14 Cell 1993 75 843 854 8252621 176. Wightman B. Ha I. Ruvkun G. Posttranscriptional regulation of the heterochronic gene lin-14 by lin- 4 mediates temporal pattern formation in C. elegans Cell 1993 75 855 862 8252622 177. Fire A. Xu S. Montgomery M.K. Kostas S.A. Driver S.E. Mello C.C. Potent and specific genetic interference by double stranded RNA in Caenorhabditis elegans Nature 1998 391 806 811 9486653

PMC005xxxxxx/PMC5001582.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2708493710.1093/nar/gkw252Gene regulation, Chromatin and EpigeneticsTranscription facilitates sister chromatid cohesion on chromosomal arms Bhardwaj Shweta †Schlackow Margarita †Rabajdova Miroslava Gullerova Monika *Sir William Dunn School of Pathology, South Parks Road, Oxford OX1 3RE, UK* To whom correspondence should be addressed. Tel: +44 1865 285658; Email: monika.gullerova@path.ox.ac.uk† These authors contributed equally to this paper as first authors.19 8 2016 15 4 2016 15 4 2016 44 14 6676 6692 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Cohesin is a multi-subunit protein complex essential for sister chromatid cohesion, gene expression and DNA damage repair. Although structurally well studied, the underlying determinant of cohesion establishment on chromosomal arms remains enigmatic. Here, we show two populations of functionally distinct cohesin on chromosomal arms using a combination of genomics and single-locus specific DNA-FISH analysis. Chromatin bound cohesin at the loading sites co-localizes with Pds5 and Eso1 resulting in stable cohesion. In contrast, cohesin independent of its loader is unable to maintain cohesion and associates with chromatin in a dynamic manner. Cohesive sites coincide with highly expressed genes and transcription inhibition leads to destabilization of cohesin on chromatin. Furthermore, induction of transcription results in de novo recruitment of cohesive cohesin. Our data suggest that transcription facilitates cohesin loading onto chromosomal arms and is a key determinant of cohesive sites in fission yeast. cover-date19 August 2016 ==== Body INTRODUCTION Cohesin is a conserved multi-subunit protein complex that plays an essential role in sister chromatid cohesion and proper chromosome segregation. Cohesin is also involved in other fundamental processes such as gene expression regulation and DNA damage repair. The core cohesin complex consists of structural maintenance of chromosomes (SMC) proteins, Psm3 and Psm1, and the kleisin subunit Rad21 (Table 1) (1,2). SMC proteins are characterized by a globular hinge domain surrounded by two α-helices that fold back onto themselves at the hinge, thereby bringing the N- and C-termini together to form an ABC-type nucleotide binding domain (NBD) (3). The SMC proteins form V-shaped Psm3-Psm1 heterodimers that interact with the N- and C-terminal domains of Rad21, thus forming the tripartite cohesin ring (4). The ring is further stabilized by the essential subunit Psc3 which is recruited by Rad21 (5). Experimental evidence strongly suggests that cohesion is maintained via topological entrapment of sister chromatids by the cohesin ring (6,7). Table 1. Nomenclature of the cohesin subunits in various organisms Mammals D. melanogaster S. cerevisiae S. pombe Function SMC1A Smc1 Smc1 Psm1 Core cohesin (mitosis) SMC1B Core cohesin (meiosis) SMC3 Smc3 Smc3 Psm3 Core cohesin RAD21 Rad21/Vtd Mcd1/Scc1 Rad21 Core cohesin (mitosis) REC8 C(2)M Rec8 Rec8 Core cohesin (meiosis) SA1/STAG1 SA (Stromalin) Scc3 Psc3 Core cohesin (mitosis) SA2/STAG2 SA2 (Stromalin-2) SA3/STAG3 Core cohesin (meiosis) The association of cohesin with chromatin is dependent on a loading complex, which consists of two essential subunits, Mis4 and Ssl3 in fission yeast (8,9). It has been speculated that the loading complex stimulates the ATPase activity of cohesin (10–13) and creates a DNA ‘entry gate’ via the transient opening of the Psm3-Psm1 hinge interface in yeast (14,15). In budding yeast, the activity of a loading complex is counteracted by the ‘anti-establishment’ activity of Wpl1 that destabilizes the Smc3-Scc1 interface (16) and forms a DNA ‘exit gate’ (17). In fission yeast, chromatin bound Psm3 is acetylated by the Eso1 acetyltransferase. Acetylated Wpl1 resistant cohesin exhibits increased dwelling times on chromatin and is believed to be the topologically bound cohesin that entraps sister chromatids until mitosis (Table 2) (18,19). Table 2. Nomenclature of the regulatory proteins involved in the cohesion cycle Mammals D. melanogaster S. cerevisiae S. pombe Function NIPBL/SCC2 Nipped-B Scc2 Mis4 Cohesin loading MAU2/SCC4 Mau2 Scc4 Ssl3 Cohesin loading ESCO1 Eco/Deco Eco1/Ctf7 Eso1 Cohesion establishment ESCO2 San PDS5A Pds5 Pds5 Pds5 Cohesion maintenance PDS5B/APRIN WAPL/WAPAL Wapl Rad61/Wpl1 Wpl1 Cohesion maintenance SORONIN/CDCA5 Dmt (Dalmatian) – – Cohesion maintenance HDAC8 – Hos1 – Cohesin deacetylase Shugosin1 Sse1 Esp1 Protection of centromeric cohesion Separase Sse1 Esp1 Separase Cohesin removal Polo like Kinase 1 (PLK1) Polo Cdc5 Plk1 Cohesin removal Surprisingly, the underlying determinant of cohesin loading sites remains unclear. Considerable differences have been documented between different eukaryotes. In Xenopus, cohesin is recruited to chromatin via Dbf4/Drf1 Dependent Kinase (DDK), a component of the pre-replicative complex (20,21). In contrast, in S. pombe, cohesin enrichment at centromeric and peri-centromeric sites is attributed to heterochromatin protein Swi6 (22,23), whereas in S. cerevisiae the kinetochore proteins play a critical role in the recruitment of the cohesin complex (24,25). While centromeres and peri-centromeric regions constitute the strongest cohesin binding sites in all eukaryotes, cohesin also binds to ‘Cohesin Associated Regions’ or CARs on chromosomal arms (26–30). In metazoans, cohesin overlaps with the mediator complex, CTCF (CCCTC-binding factor) and tissue-specific transcription factors (27,31–32). In budding yeast, cohesin is actively moved by ongoing transcription away from its loading sites and accumulates between convergent genes (26,28). Interestingly, in S. pombe cohesin appears to be a combination of both: it associates with its loader at sites of strong transcriptional activity (29), while a sub-set re-locates and accumulates between convergent genes (33). Despite such diversity, cohesin exhibits an ordered and highly reproducible chromatin association pattern, suggesting the presence of a hitherto uncharacterized determinant. We have successfully applied a combination of single-locus DNA fluorescent in situ hybridization assay and bioinformatics analysis of genome wide ChIP-chip data of cohesin proteins (29) to present a functional analysis of cohesin dynamics across the fission yeast genome. We identified that transcription mediates cohesion establishment on chromosomal arms at the sites of cohesin loading. MATERIALS AND METHODS Cell culture Most experiments were performed in S. pombe cycling cells (∼80% G2), unless indicated. All strains were cultured in rich or minimal medium supplemented with essential amino acids. Wild type (972), mis4-367ts and rad21-K1ts cultures were synchronized in G1 by nitrogen starvation in EMM (minus NH4Cl) for 16 h. Subsequently, cells were resuspended in rich medium at an OD600 = 0.2 and shifted to 37°C, for a total of 8 h. Aliquots were taken at 1 h intervals for FACS. For heat shock experiments, log phase cultures were shifted from 32 to 42°C for 30 min and processed for FISH or ChIP-qPCR. Cdc25-22 or cdc10-129 strains were grown at 25°C, followed by a shift to 37°C to synchronize the cells in G2 or G1 respectively. All strains used in this study are listed in Supplementary Table S3. Mammalian cell line HEK293T was cultured in DMEM supplemented with 10% FBS + antibiotics. For α-amanitin treatment, cells were grown to 70% confluency and treated with α-amanitin at 2 μg/ml for 36 h (recommended incubation time is 12–48 h) (34) and processed for ChIP-qPCR or western blotting. Chromatin immunoprecipitation (ChIP) ChIP was performed using exponential cultures (OD600 < 0.5) according to previously published protocols. Cells were crosslinked with 3% para-formaldehyde for 30 min; glycine quenched and lysed in buffer I (50 mM HEPES (pH 7.5), 140 mM NaCl, 1 mM EDTA (pH 7.5), 1% Triton X-100, 0.1% sodium deoxycholate, protease inhibitiors) using the MagnaLyzer. Whole cell lysates were sonicated with a diagenode Bioruptor at high intensity for 15 min, with 30 s ON/OFF intervals. DNA was visualized on an agarose gel showing fragments of 200–500 bp. 100 μg of sonicated, precleared chromatin was incubated per antibody at 4°C overnight. In experiments with RNase treatment, mix of RNase A/T was added 30 min prior to incubation with antibodies. Crosslinked immuno-complexes were captured with Protein A agarose beads (Millipore) and washed once each with Buffer I, Buffer II (50 mM HEPES (pH 7.5), 500 mM NaCl, 1 mM EDTA (pH 7.5),1% Triton-X-100, 0.1% sodium deoxycholate), Buffer III (10 mM Tris-Cl (pH 8.0), 250 mM LiCl, 1 mM EDTA (pH 7.5), 0.5% NP-40, 0.5% sodium deoxycholate) and TE (10 mM Tris–Cl and 1 mM EDTA). Immunoprecipitated DNA was eluted from the beads in TE + 1% SDS. Samples were reverse crosslinked with 3 μg/ml of RNaseA (Roche) at 65°C overnight, treated with 20 μg Proteinase K (Roche) at 45°C for 2 h and purified using Qiagen PCR purification columns. DNA was eluted twice with 35 μl MilliQ water. 2 μl of each sample was used for qPCR (SensiMix, Bioline). In experiments with chemical treatment, cells were treated with 1,10-phenananthroline (300 μg/ml in ethanol, 30 min) and ChIP-qPCR analysis was performed as above. Antibodies were obtained from Abcam; Anti-GFP (ab290), RNAPII (ab817) and histone H3 (ab1791). All primer pairs used for ChIP analysis are listed in Supplementary Table S5. FACS To determine cell cycle progression the DNA content was measured by propidium iodide staining of ethanol-fixed cells. 1 ml aliquots were pelleted and fixed in cold 70% ethanol. Pellets were resuspended in 500 μl of 50 mM sodium citrate and treated with 1 mg/ml RNaseA, 2 h at 37°C. Pellets were washed with 50 mM sodium citrate, resuspended in 50 mM sodium citrate + propidium iodide (8 μg/ml) and incubated at 4°C for 2 h (to overnight). Subsequently, pellets were washed with 50 mM sodium citrate and DNA content analyzed using the CellQuest pro software for the FACS Calibur machine. Data analysis ChIP-chip data (GEO accession number GEO:GSE13517) were reanalyzed. All data for Pk9 tagged Rad21, Psc3, Pds5, Mis4, Ssl3 and untagged control were normalized to control whole cell extract (WCE) fraction values. The original data generated on the Affymetrix S_pombea520106F platform was adjusted to the current S. pombe annotation (EF2 annotation from the iGenomes database; this annotation is equivalent to the current ASM294v2 annotation). In particular, we observed a shift of 80,031 bp in the genome annotation on chromosome 2 since 2004 and two regions of 1000 ‘N's were replaced by 100 ‘N's. Data was plotted in MATLAB. Swi6 data was obtained from the first biological repeat performed previously (35) (GEO:GSE3186, sample GSM71577). Data were generated with the NCI Pombe 44K v1.0 platform from 2005 and readjusted to the current S. pombe annotation similarly to above. Peak-caller The MATLAB peakfinder algorithm was employed to log2 normalized signal data (http://www.mathworks.co.uk/matlabcentral/fileexchange/25500-peakfinder). The algorithm employs a minimal threshold and a cut-off value, by a factor of which the peak must be higher than the average surrounding data. Based on data noise and distribution this threshold was set to 0.3 for Rad21, 0.4 for Swi6, 0.5 for Mis4 and Sfc6, and 0.7 for all other datasets. The cut-off was set to 0.2 for all datasets. Peaks were extended to both sides until log2 reached 0 (corresponding to the value where WCE normalization data becomes higher than the specific signal). The identified peaks were further subjected to a custom written Perl code. The criteria employed were a minimal peak-width of 1000 bp and an average signal throughout the peak. According to data distribution the average signal value throughout a peak was chosen as 0.2 for Rad21, 0.3 for Mis4, Sfc6 and Swi6, and 0.4 for all other datasets. The average peak signal was computed as the average of the log2 values of each probe within the peak. Gene expression analysis Gene expression in reads per kilobase per million (RPKM) values were computed as the average from previously published RNA-Seq data (36) generated from two biological replicates. The RPKM value of the most highly expressed overlapping gene was assigned to the peak. This was applied to Rad21 peaks, which were split into groups: overlapping with a called Mis4 peak and not overlapping with a called Mis4 peak. Likewise, for called Ssl3 peaks. RPKM values, which were overlapping a called peak in Mis4 or Ssl3 data, were extracted to compute gene expression values for each of these four sets. Note, the peak at 4,494,276 was not taken into consideration for the above analysis as it does not overlap any genes and may be part of the adjacent peak (not considered as such due to a possible data generation artifact, where surrounding probes give 0 signal). Genome-wide ChIP analysis for S. cerevisiae and H. sapiens S. cerevisiae cohesin and cohesin loader ChIP-Seq data was obtained from a previous study (11). Data were mapped to the S. cerevisiae genome SacCer3 using Bowtie with the following parameters: -m 1 -k 1 -v 3 –C (-m: permissible maximal number of alignments found for any read; -k: permissible maximal number of alignments to be reported for any read; -v: permissible number of mismatches end to end of any read; -C: color space alignment). ChIP-Seq data for mammalian cohesin subunit (SMC1) and RNAPII in HB2 cells were extracted from (37). ChIP Seq data was mapped to hg19 with Bowtie using the options -m 1 -k 1 –n 1 -S (-m: permissible maximal number of alignments found for any read; -k: permissible maximal number of alignments to be reported for any read; -n: permissible number of mismatches in seed region of any read; -S: report alignments in SAM output). Gene expression analysis for S. cerevisiae and H. sapiens SCC1 peaks were considered in a RNAPII transcribed region, if they overlapped with any genes from the Saccharomyces Genome Database (SGD). These were obtained from the UCSC table browser. RNAPIII regions are defined as tRNA loci obtained from the SGD. S. cerevisiae gene expression values in RPKM were re-analyzed from previously published study (38) and averaged over two biological repeats. RNA-Seq from HB2 cells was mapped to hg19 using Tophat2 with the following parameters: -G genes_hg19.gtf -g 1 -p 8 –segment-length 15 –no-coverage-search, where genes_hg19.gtf is the gtf file containing all the human genes downloaded from UCSC. The mapped Bam file was further processed to determine RPKM values with Cufflinks2, which were used as gene expression values for human genes. Called SCC2 and SMC1 peaks were extended at each end by 200nt for overlap analysis, as this is the estimated fragment size (39). Meta analysis Peaks for S. cerevisiae data were called using MACS with default parameters. Hyper-ChIPable region in S. cerevisiae had been annotated previously (40). Peaks for human SMC1, NIPBL and RNAPII were called with MACS2 using the –broad parameter (default parameters for SMC1 and NIPBL, –broad-cutoff = 0.01 for RNAPII). Further RNAPII peaks were pooled into one if they were less than 10 kb apart. Genomic distances Genomic distance of Rad21 peaks at Mis4+/Rad21+ and Rad21+ sites to nearest Pds5 peaks were computed as the distance between annotated margins of the respective peaks. Any overlap was taken as distance 0. Distance profiles Distances of called Sfc6 and Pds5 peaks from called cohesive and non-cohesive peaks were plotted as histograms. Data was binned into 5000 bp bins and normalized to the size of dataset under consideration (cohesive or non-cohesive peaks). Boxplots Boxplots were created using R. The median is presented as a line, upper and lower quartiles (q3 and q1 respectively) are presented as a box. The whiskers are given by q3 + 1.5(q3 – q1) (upper) or q1 – 1.5(q3 – q1) (lower). The default of 1.5 corresponds to approximately ±2.7σ and 99.3 coverage if the data are normally distributed. Outliers are plotted if their value is higher than the upper whisker or smaller than the lower. The plotted whiskers extend to the adjacent value, which is the most extreme data value that is not an outlier. Statistical analyses: comparisons of two sets of peaks To identify whether cohesive peaks localize in significantly higher expressed regions than non-cohesive peaks the Wilcoxon Rank Sum test was applied to compute the one-sided P-value. The Wilcoxon Rank Sum test was also applied to the absolute values of peak distances to Pds5 peaks. P-values were further verified by a permutation test. Fluorescent in situ hybridization Wild type (or ts mutants) cells were grown overnight to mid-log phase (OD600 = 0.5). The following day, cells were re-inoculated into fresh medium and incubated for an hour until the culture had an OD600 = 0.30–0.35. Nearly 3 × 107 cells were resuspended in 1.2 M sorbitol and gently swirled for 5–10 min. Cells were fixed in freshly prepared 3% para-formaldehye at 30°C for 30 min and quenched with 104 mM glycine. Subsequently, cells were harvested at 3000 rpm, 5 min and washed twice with PEM buffer (100 mM PIPES, 1 mM EGTA, 1 mM MgSO4). Cells were transferred to 1.5 ml sterile screw cap tubes, washed with PEMS (PEM + 1.2 M sorbitol), spheroblasted with 100 U of Zymolyase 100T in PEMS at 37°C for 1 h and pelleted twice at 2000 rpm for 1 min by turning the caps for each spin to collect maximum cells. Cytoplasm was permeabilized in PEMS + 1% Triton-X-100 for exactly 6 min. Pellets were washed three times with PEM, resuspended in PEMBAL (PEMS + 1% BSA, 0.1% NaN3, 100 mM lysine monohydrochloride) containing 1 mg/ml RNaseA and incubated at 37°C for 2–4 h. Hybridization was performed in solution. RNaseA treated cells were resuspended once in each of 200 μl 2× SSC, 2× SSC + 10% formamide, 2× SSC + 20% formamide and 2× SSC + 40% formamide, 15 min at room temperature. Finally, pellets were resuspended in 100 μl hybridization buffer (10% dextran sulphate, 5× Denhardt's, 50% formamide, 0.5 mg/ml salmon sperm DNA, 2× SSC) and labeled FISH probe was added at a final concentration of 1 ng/μl. The hybridization mix (cells + probe) was denatured at 75°C for 5 min, followed by 2 min on ice and finally, incubated at 40°C overnight in a thermomixer at 700 rpm. Next day, cells were washed three times with 2× SSC for 15 min at 700 rpm, 37°C and DNA stained DAPI. Excess DAPI was removed by a wash in PBS only and cells were resuspended in 30–50 μl PBS. For microscopy, ∼7 μl cells were dropped on poly-l-lysine coated slides and covered with 22 × 22 mm coverslips. Excess liquid was removed by blotting and slides were sealed with nail varnish. When imaging was not performed immediately, slides were stored at 4°C in the dark. Image acquisition was performed using SoftWorx program (Applied Precision) on the DeltaVision microscope (Applied Precision). Cells were visualized with a 100 × 1.35 NA objective lens. For each experiment, z-sections were taken at 200 nm intervals. Background light was corrected by the built-in deconvolution algorithm of the SoftWorx program. Final images represent maximum intensity projections of z-stacks obtained using the ImageJ/Fiji software (NIH). FISH probes were prepared using the FISH-Tag Kit (Molecular Probes, Invitrogen) with either Alexa-488 or Alexa-555 fluorophores. All FISH probes (except the centromeres) spanned ∼10 kb. The centromeric probes targeted the dg region in the outer most centromeric repeats and were ∼2.2 kb. Overlapping PCR fragments (∼2.5 kb) were amplified using Phusion DNA polymerase and pooled in equimolar quantities. All primer pairs used for the probes preparation are listed in Supplementary Table S4. The labeling reaction was performed according to the manufacturers’ instructions. The strategy involved modifying the template DNA by incorporating aminolyated nucleotides, which then acted as a bait to couple the reactive dye. RESULTS Cohesive and non-cohesive loci on chromosomal arms This study builds on previously published ChIP-chip data that provide binding profiles of cohesion proteins across chromosomes 2 and 3 in S. pombe (29). Cohesin co-localizes with its loading complex Mis4-Ssl3 (Figure 1A, inset top panel), but surprisingly, there are also cohesin peaks detected away from its loading sites (Figure 1A, inset bottom panel). It has been shown previously that cohesin can be pushed by ongoing transcription across a gene (26,33). However it remained unclear how cohesin could be re-located from its loading sites across several kilobases and various transcription units. Furthermore, recent studies showed that cohesin could bind to DNA in vitro through spontaneous topological but inefficient interaction. Cohesin loader Mis4 stimulates ATPase activity of cohesin, resulting in its efficient binding to DNA (12). Figure 1. Cohesive and non-cohesive subsets of cohesin on chromosome arms in S. pombe. (A) An illustration of sister chromatids (blue) held together by the ring shaped cohesin complex (yellow) at centromeres and across the arm regions. The inset depicts cohesin enrichment at intergenic regions between convergent genes (grey arrows) on chromosome arms in S. pombe, wherein cohesin localizes with (top panel) or without (bottom panel) its loader Mis4-Ssl3 (black-green). (B) Top: schematic showing cohesed sister chromatids at Mis4+/Rad21+ sites visualized by DNA-FISH (red probes). Bottom: Single locus specific DNA FISH showing single dots at CARs I to VI (Mis4+/Rad21+ sites) in G2 cells (selected by septation index) (n = 200, three biological repeats). Red (Alexa-555) or green (Alexa-488) probes are shown inside DAPI stained nucleus (blue). (C) Top: schematic showing non-cohesed sister chromatids at Rad21+ sites visualized by DNA-FISH (increased separation between red probes). Bottom: Similar amount (±50%) of single-double dots at CAR VI to XII in G2 cells (selected by septation index) (n = 200, three biological repeats). Red (Alexa-555) or green (Alexa-488) probes are shown inside DAPI stained nucleus (blue). (D) Inter-chromatid distance measurement between sister chromatids at Mis4+/Rad21+ and Rad21+ regions. Z-stacks were converted into 2D images and distances between two spots measured using line scan tool of the ImageJ software. (E) DNA-FISH analysis in wild type (WT), mis4-367ts and rad21-K1ts strains after Mis4 and Rad21 inactivation by shift to 37°C. Top: cohesion at the arms, CAR III (green dots) was destabilized but cohesion at centromeres, CAR VI (red dots) was unaffected after 6 h at 37°C. Bottom: defective arm (CAR III) and centromeric cohesion (CAR VI) after 8 h at 37°C. Arm and centromeric cohesion was unaffected in WT throughout (n = 100 cells). (F) DNA-FISH analysis in rad21-K1ts strains after Rad21 inactivation by shift to 37°C. CARs III, VIII and IX were analyzed. Cells showing single or double dots were counted and plotted in a bar graph. * P < 0.05, two-tailed, paired Student's t-test, Error bars represent SD, n = 3. We re-analyzed the distribution of cohesin subunits Rad21 and Psc3; cohesin loader Mis4-Ssl3 and cohesin maintenance protein Pds5 from ChIP-chip data (29). We used our own peak-calling algorithm to match the current S. pombe annotation, to avoid data smoothing and to maintain consistency between datasets. We identified 283 peaks for Rad21 and 150 peaks for Mis4 on chromosome 2 (all peaks shown in Supplementary File 1), which include >90% Rad21 and >82% Mis4 of previously described peaks (see Supplementary File 2 for comparison in log2 and linear scale). In agreement with previous analysis, only 33% of Rad21 peaks overlap with Mis4. We selected six Rad21 positive regions, CAR I to VI that co-localize with Mis4-Ssl3 (Mis4+/Rad21+) and a further six that do not co-localize with Mis4-Ssl3, CAR VII to XII (Rad21+, Supplementary Tables S1 and S2). CAR VI derives from centromeric degenerate repeats (cendg) and is used as a positive control. Detailed profiles of Rad21 (green), Mis4 (red) and a no tag control (blue) across all selected CARs can be seen in Supplementary Figure S1. We employed DNA FISH to test cohesion at selected CARs. A 10 kb long unique probe was fluorescently labeled and hybridized to a single CAR resulting in single (Figure 1B, schematic) or double dots (Figure 1C, schematic). Single dot represents cohesion, whilst double dots are a result of locally separated sister chromatids. We observed 100% of cells exhibiting single dots at Mis4+/Rad21+ loci in G2 cells (selected by septation index = 0) (Figure 1B, CARs I–VI and Supplementary Figure S2A) indicating stable cohesion between chromatids at these sites. In contrast, Rad21+ loci showed ∼50% of cells with single or double dots (Figure 1C, CAR VII–XII and Supplementary Figure S2B) indicating reduced cohesion at these loci. Furthermore, sister chromatids at Mis4+/Rad21+ sites were <0.2 μm apart, while the inter-chromatid distance at Rad21+ loci increased from >0.2 to 0.9 μm (Figure 1D). To confirm that FISH signals do indeed represent functional cohesion, we employed temperature sensitive (ts) Mis4 and Rad21 mutants. First, wild type (WT), mis4-367ts (9) and rad21-K1ts (41) strains were synchronized in G1 by nitrogen starvation, transferred to rich medium and simultaneously shifted to 37°C to inactivate Mis4 and Rad21 proteins. Cell cycle progression for WT and rad21-K1ts was monitored by FACS analysis. Nitrogen starved cells exhibit a ∼2 h lag period before re-entering the cell cycle (42), likewise WT moved from 1C to 2C DNA content between 3–5 h at 37°C (Supplementary Figure S3A), whereas rad21-K1ts completed replication between 4–5 h. Subsequently, DNA content in rad21-K1ts became heterogeneous, predominantly between 6–8 h at 37°C (41). While WT maintained punctate nuclei throughout, mis4-367ts and rad21-K1ts exhibited DNA fragmentation at 6–8 h, indicating loss of cohesion due to the inactivation of Mis4 and Rad21 (Supplementary Figure S3B). Next, we analyzed cohesion on chromosomal arms (CAR III) and centromeres (CAR VI) after 4, 6 and 8 h at 37°C. Cohesion was intact in WT throughout this time course (Figure 1E and Supplementary Figure S3C). In contrast, cohesion on the arms (CAR III) was destabilized in mis4-367ts and rad21-K1ts mutants after 6 h at restrictive temperature (Figure 1E, top panel). Centromeric cohesion was eventually destabilized after 8 h of Rad21 or Mis4 inactivation (Figure 1E, bottom panel). These observations confirm that our FISH assay is biologically functional and specific for physiological cohesion of sister chromatids. FISH analysis of Rad21+ loci resulted in single dots in ∼50% of cells. To test whether single dots at CARs VII-XII are cohesin dependent, we performed FISH experiment in rad21-K1 cells at restrictive temperature. We employed probes specific for one Mis4+/Rad21+ (CARIII) and two Rad21+ (CARs VIII and IX) loci (Figure 1F). We do confirm the loss of cohesion at CARIII in 66% of cells (34% remain with single dot after Rad21 inactivation). The loss of cohesion from 100% to 34% is significant (P < 10−3). Surprisingly, we see no significant loss of cohesion at Rad21+ CARs after Rad21 inactivation. For CARs VIII and IX, the number of cells with a single dot dropped from 47% to 40% and 45% to 39%, respectively. These changes are not statistically significant (P = 0.269 for CAR VIII and P = 0.344 for CAR IX) (Figure 1F). Therefore, we conclude that detected single dots at Rad21+ sites are most likely a result of random chromosome breathing or just general close proximity of sister chromatids. Overall, our DNA FISH analysis demonstrates that Mis4+/Rad21+ loci are associated with stable cohesion, while Rad21+ loci display cohesin's dynamic association with chromatin. Mis4+/Rad21+ sites associate with Pds5 Although cohesin levels across the S. pombe genome seem to be similar, our DNA FISH analysis shows functional differences between Mis4+/Rad21+ and Rad21+ cohesin sub-sets. Therefore, we performed ChIP-qPCR analysis of Rad21-9Pk at all CARs and we detect significant (P < 0.05) enrichment of Rad21 over the background (no tag control) at all tested CARs (Figure 2A). Interestingly, genome wide analysis shows significant cohesin enrichment (P < 0.05) at Mis4+/Rad21+ sites in comparison to Rad21+ only (Figure 2B), suggesting that functional differences between Mis4+/Rad21+ and Rad21+ sites may reflect stoichiometric variation in cohesin levels along the chromosomal arms. Figure 2. Cohesin and Pds5 co-localize at Mis4+/Rad21+ loci. (A) ChIP-qPCR analysis showing Rad21-Pk9 at selected CARs. No tag strain was used to assess the background levels. * P < 0.05, two-tailed, paired student's t-test for Rad21-Pk9 compared to background levels. Error bars represent SD, n = 3. (B) Cohesin enrichment at Mis4+/Rad21+ versus Rad21+ sites. Boxplots comparing sum of log2 Rad21 ChIP signal at Mis4+/Rad21+ versus Rad21+. P < 0.05, one-sided Wilcoxon Rank Sum test. (C) Pds5 enrichment at Mis4+/Rad21+ versus Rad21+ sites. Bar graph shows distances between Pds5 peak margins at Mis4+/Rad21+ and Rad21+ sites in 5000 nt bins. An overlap between intervals spanned by a Mis4+/Rad21+ and Pds5 peak gives a distance of 0. (D) ChIP-qPCR analysis showing Pds5-GFP at selected CARs. No tag strain was used to assess the background levels. *P < 0.05, two-tailed, paired Student's t-test for Pds5-GFP compared to background levels. Error bars represent SD, n = 3. Pds5 is a cohesin associated factor essential for cohesion maintenance and de novo Psm3 acetylation (43–45). To test whether the observed differential cohesion at Mis4+/Rad21+ loci could arise due to Pds5 enrichment, we employed analysis of Pds5 ChIP-chip peaks and show that it significantly overlaps with Mis4+/Rad21+ loci (P < 10−5, Supplementary Figure S4A). Similarly, metagene analysis confirms that Pds5 peaks are closer to Mis4+/Rad21+ loci than to Rad21+ loci (Figure 2C). Next, we employed ChIP-qPCR to test the Pds5 enrichment at selected CARs. We observed significant (P < 0.05) Pds5 enrichment at Mis4+/Rad21+ CARs. Rad21+ loci showed Pds5 close to background levels (Figure 2D). These data suggest that cohesin levels are generally higher at its loading sites, where they overlap with Pds5. Mis4+/Rad21+ sites associate with acetyltransferase Eso1 Psm3 acetylation by Eso1 is a hallmark of cohesion establishment. To test whether Eso1 stabilizes cohesin association with chromatin at CARs I-VI, we performed ChIP-qPCR to detect Psm3 levels at selected CARs in WT and eso1-H17 mutant. Cells were synchronized in early S phase (HU 15mM, 2 h) at 25°C followed by a shift to 37°C for 2 h to inactivate Eso1. Cell cycle progression was monitored by FACS analysis (Supplementary Figure S3D). Cohesin enrichment at all tested CARs was observed in WT and eso1-H17 cells at the permissive temperature (Figure 3A and C). However, Psm3-GFP levels were significantly (P < 0.05) decreased only at Mis4+/Rad21+ loci at the restrictive temperature (Figure 3B and C). Figure 3. Cohesin at Mis4+/Rad21+ loci is stabilized by Eso1. (A) ChIP-qPCR analysis showing Psm3-GFP enrichment at selected CARs in WT and eso1-H17 cells at 25°C. No tag strain was used to assess the background levels. *P < 0.05, two-tailed, paired Student's t-test for Psm3-GFP compared to background levels. Error bars represent SD, n = 3. (B) ChIP-qPCR analysis showing Psm3-GFP enrichment at selected CARs in WT and eso1-H17 cells at 37°C. No tag strain was used to assess the background levels. *P < 0.05, two-tailed, paired Student's t-test for Psm3-GFP compared to background levels. Error bars represent SD, n = 3. (C) ChIP-qPCR analysis showing Psm3-GFP enrichment at centromeric dg repeat (CAR VI) in WT and eso1-H17 cells at 25 and 37°C. Error bars represent SD, n = 3. (D) Left panel: western blot showing Eso1-GFP protein levels in G2/S phase cells. Whole cell extracts were prepared from cycling cells (G2), cells arrested in early S-phase with Hydroxyurea (HU) and subsequently released into G2 at indicated time points. No tag is a control for anti-GFP antibody specificity and Histone H3 is used as a loading control. Right panel: western blot showing Eso1-GFP protein levels in G1 synchronized cells through to G2 at respective time points. H3 is used as a control. (E) ChIP-qPCR analysis showing Eso1-GFP enrichment at selected CARs in G2 synchronized cells (cdc25-22, 37°C, 4 h). No tag strain was used to assess the background levels. * P < 0.05, two-tailed, paired Student's t-test for Eso1-GFP compared to background levels. Error bars represent SD, n = 3. Interestingly, levels of the Eso1 orthologue in S. cerevisiae, Eco1, are regulated by proteasome degradation after S phase (46). To test whether Eso1 is similarly regulated in fission yeast, we compared Eso1 protein levels in G2 and S-phases. FACS analysis indicated S phase arrest after hydroxyurea treatment (Supplementary Figure S3E, 15 mM 2 h at 32°C) followed by progression to G2. Surprisingly, Eso1-GFP protein levels were similar between G2 and S phases (Figure 3D, left panel). We extended this analysis to G1 arrested cells using the cdc10-129 mutant at 37°C, followed by release to S and G2 at 25°C. Cell cycle progression was monitored by FACS analysis (Supplementary Figure S3G). We do detect Eso1 protein levels throughout the cell cycle, although its levels are slightly decreased in G2 (Figure 3D, right panel). Furthermore, we employed ChIP-qPCR and observed significant enrichment of Eso1 in G2 synchronized cells (cdc25-22, Supplementary Figure S3F) at CARs I–VI, but not at CARs VII–XII (Figure 3E). These data together with FISH results suggest that Mis4+/Rad21+ loci are associated with Eso1 and Pds5 and represent cohesive sites on chromosomal arms. Mis4+/Rad21+ sites overlap with highly expressed RNAPII genes Previous analysis demonstrates that Mis4-Ssl3 binding sites on chromosome 2 overlap with highly transcribed RNAPII and RNAPIII genes (29). To delineate this further, we extracted RNAPIII loci from PomBase based on tRNA and 5S rRNA search and assumed all other genes to be RNAPII transcribed. Subsequently, we overlapped Mis4+/Rad21+ and Rad21+ peaks with RNAPIII and RNAPII genes. Interestingly, ∼80% Mis4+/Rad21+ peaks overlap RNAPII genes (Figure 4A), while 4% overlap RNAPIII genes and additional 16% overlap both. Similarly, ∼94% Rad21+ peaks overlap RNAPII genes, with only 1% present at RNAPIII loci (Figure 4B). The slight preference (3%) of Mis4+/Rad21+ versus Rad21+ sites for RNAPIII genes is supported by their proximity to RNAP III transcription factor Sfc6 (Supplementary Figure S4B and C). It should be noted that RNAPI transcribed ribosomal RNA genes were excluded, as all rRNA loci in S. pombe are located on chromosome III. Figure 4. Cohesive sites overlap highly transcribed RNAPII regions. (A) Pie-chart showing overlap between Mis4+/Rad21+ loci and RNAPII genes, with a mere 4% overlap with only RNAPIII genes. (B) Pie-chart showing overlap between Rad21+ loci and RNAPII and RNAPIII genes. (C) Histogram showing the distribution of RNAPII genes versus gene expression in non-cohesin associated genes (blue bars, 1412 genes), cohesive (red bars, 91 genes) and non-cohesive loci (green bars, 187 genes). (D) Boxplots showing expression levels (Reads per Kilobase per Million; RPKM) of genes that overlap with Mis4+/Rad21+ and Rad21+ sites. P < 10−9, Wilcoxon Rank Sum test. (E) Boxplots showing expression level (reads per kilobase per million; RPKM) of genes that overlap with Ssl3+/Rad21+ and Rad21+ sites. P < 10−7, Wilcoxon Rank Sum test. (F) Graph showing Swi6 occupancy at Mis4+/Rad21+ and Rad21+ loci. (G) Boxplots showing Mis4+/Rad21+ sites overlap with highly expressed genes in comparison to Rad21+ sites, including or excluding Swi6 loci. Changes in gene expression of Mis4+/Rad21+/Swi6− loci compared to Mis4+/Rad21+/Swi6+ loci and Rad21+/Swi6− loci compared to Rad21+/Swi6+ loci are not significant. Changes in gene expression of Mis4+/Rad21+/Swi6− loci compared to Rad21+/Swi6− loci and Mis4+/Rad21+/Swi6+ loci compared to Rad21+/Swi6+ loci are significant (P < 10−9, Wilcoxon Rank Sum test). Next, we examined the expression of all RNAPII genes (47) and combined it with Rad21-/RNAPII, Mis4+/Rad21+/RNAPII and Rad21+/RNAPII sites. There is a clear association of Mis4+/Rad21+ sites with highly expressed genes (Reads per kilobase per million, RPKM>600), while the majority of all other RNAPII sites are aggregated at lowly expressed genes (Figure 4C). Despite >90% overlap of Mis4+/Rad21+and Rad21+ peaks with RNAPII, only Mis4+/Rad21+ sites show a significant overlap with highly expressed RNAPII genes (median RPKM > 350, P < 10−7, Figure 4D, Mis4+/Rad21+ and Figure 4E, Ssl3+/Rad21+). Recent reports demonstrate the propensity of misleading ChIP signals observed at highly expressed, open chromatin hyper-ChIPable regions (40). However, it should be noted that in that study, formaldehyde cross-linking was performed for 60 min, while routine ChIP protocols (as used in our study) cross-link for 30 min. Naturally, higher cross-linking times could generate artifacts. However, to eliminate false ChIP signals, we analyzed the distribution of heterochromatin binding protein Swi6 (48) and observed that ∼85% of Mis4+/Rad21+ and ∼90% of Rad21+ sites lack Swi6, in line with overlap with highly expressed genes and not heterochromatin regions (Figure 4F). Furthermore, both Mis4+/Rad21+ and Rad21+ sites exhibit significant differences in gene expression, with or without inclusion of Swi6 peaks (Figure 4G). These data confirm that the association of Mis4+/Rad21+ sites with highly transcribed genes is a biologically valid effect. The correlation between cohesin binding sites and RNAPII regions suggests a possible interplay between RNAPII and cohesin loading/cohesion establishment. Transcription inhibition reduces chromatin bound Mis4 and cohesin on chromosomal arms To test whether association of Mis4+/Rad21+ loci with highly transcribed genes might have a biological relevance; we employed G2 synchronized cells treated with the transcription inhibitor 1,10-phenanthroline. Transcription inhibition was confirmed by a reduction in RNAPII levels at four tested RNAPII promoters after 10 and 30 min of treatment with 1,10-phenanthroline (Figure 5A). Next, we assessed cohesin levels at centromeres. Rad21 signals remained unchanged after transcription inhibition (CAR VI, Figure 5B), presumably because cohesion establishment at centromeres occurs via the RNAi dependent heterochromatin pathway (49). Furthermore, Rad21 levels were significantly (P < 0.05) reduced after 1,10-phenanthroline treatment at CARs I-V but not at CARs VII–XII (Figure 5C). We also performed Mis4 ChIP-qPCR and we detect significant (P < 0.05) Mis4 levels only at Mis4+/Rad21+ loci. Transcription inhibition resulted in reduced levels of Mis4 at CARs I–V, similar to Rad21 (Figure 5D). To test whether decreased levels of Rad21 and Mis4 at Mis4+/Rad21+ loci are not a result of reduced protein levels after transcription inhibition, we performed western blot analysis and show that Rad21 and Mis4 protein levels remained unchanged after treatment with 1,10-phenanthroline (Figure 5E). Figure 5. Transcription inhibition reduces chromatin association of cohesin proteins. (A) ChIP-qPCR analysis showing RNAPII occupancy after 1,10-phenanthroline treatment (300 μg/ml, 10 and 30 min) at selected promoters. *P < 0.05, two-tailed, paired Student's t-test between 0 and 30 min. Error bars represent SD, n = 3. (B) ChIP-qPCR analysis showing Rad21-GFP enrichment at CAR VI after 1,10-phenanthroline treatment (300 μg/ml, 10 and 30 min) in G2 synchronized cells (cdc25-22, 37°C 4 h). No tag strain was used to assess the background levels. *P < 0.05, two-tailed, paired Student's t-test for Rad21-GFP compared to background levels. Error bars represent SD, n = 3. (C) ChIP-qPCR analysis as in (B) at selected CARs. (D) ChIP-qPCR analysis as in (C) showing Mis4-GFP enrichment at selected CARs (E) Western blot showing RNAPII, Rad21-GFP, Mis4-GFP, H3 and tubulin protein levels after transcription inhibition using 1,10-phenanthroline (300 μg/ml, 30 min). Total RNAPII was probed with 8WG16 (ab817), Rad21 and Mis4 with anti-GFP (ab290) and histone H3 with anti-H3 (ab1791) antibody. (F) DNA-FISH analysis of CARs I and II after 1,10-phenanthroline treatment (300 μg/ml, 30 min). Finally, we performed FISH experiment in 1,10-phenantroline treated cells, wherein we do not observe any separation of sister chromatids (Figure 5F). This was an expected result, as cohesion established during S-phase would persist throughout G2 phase. Decreased levels of cohesin after transcription inhibition would most probably have no effect on already established cohesion. Transcription induction facilitates cohesion establishment at heat shock genes. To further validate the role of RNAPII in cohesion establishment, we employed the heat shock response (hsp) gene loci and assessed chromatin occupancy of Mis4 and Psm3 at hsp promoters. Transcription was rapidly induced at hsp70 and hsp9 genes, as we detect increased RNAPII levels at these promoters after heat shock (Figure 6A). To test that RNAPII is not only pausing at hsp promoters but is actively engaged in transcription, we performed ChIP-qPCR and show increased levels of RNAPII phosphorylated at the Serine 5 residue (Ser5), which is a mark of transcriptional activity (Figure 6B). Furthermore, we confirmed transcription of hsp genes by RT-qPCR (Figure 6C). Interestingly, transcription induction resulted in enrichment of Psm3 (Figure 6D), Mis4 (Figure 6E) and Eso1 (Figure 6F) at hsp70 and hsp9 promoters after heat shock. To test, whether the recruitment of cohesin proteins to chromatin is RNA dependent, we performed Psm3 and Mis4 ChIP-qPCR with RNAse A/T treatment and observed the same level of their enrichment at hsp genes after heat shock (Supplementary Figure S5A and B). These data suggest that increased levels of transcription rather than RNA could mediate recruitment of cohesin proteins to chromatin. Figure 6. Transcription induction mediates cohesin proteins recruitment to chromatin. (A) ChIP-qPCR analysis showing RNAPII enrichment at selected promoters after heat shock at 42°C, 30 min. Values are normalized to rpl29. Error bars represent SD, n = 3. *P<0.05, one-tailed, paired Student's t-test comparing 42 to 32°C values. (B) ChIP-qPCR analysis showing enrichment of RNAPII phosphorylated at Ser5 at selected promoters after heat shock at 42°C, 30 min. Values are normalized to act1. Error bars represent SD, n = 3. * P<0.05, two-tailed, paired student's t-test comparing 42°C to 32°C values. (C) RT-qPCR showing act1, hsp70 and hsp9 mRNA levels after heat shock. Error bars represent SD, n = 3. Values are normalized to 32°C signals. *P < 0.05, two-tailed, paired Student's t-test comparing 42 to 32°C values. (D) ChIP-qPCR analysis showing Psm3-GFP enrichment at hsp70 and hsp9 promoters after shifting from 32 to 42°C, 30 min. Values are normalized to rpl29. Error bars represent SD, n = 3. *P < 0.05, one-tailed, paired Student's t-test comparing 42 to 32°C values and GFP signal to no tag control. (E) ChIP-qPCR analysis showing Mis4-GFP enrichment at hsp70 and hsp9 promoters after shifting from 32 to 42°C, 30 min. Values are normalized to rpl29. Error bars represent SD, n = 3. *P < 0.05, one-tailed, paired Student's t-test comparing 42 to 32°C values and GFP signal to no tag control. (F) ChIP-qPCR analysis showing Eso1-GFP enrichment at hsp70 and hsp9 promoters after shifting from 32 to 42°C, 30 min. Values are normalized to rpl29. Error bars represent SD, n = 3. *P < 0.05, one-tailed, paired Student's t-test comparing 42 to 32°C values and GFP signal to no tag control. Next, we employed DNA FISH analysis to test whether de novo recruitment of cohesin, Mis4 and Eso1 in G2 is sufficient to establish stable cohesion. The hsp loci are normally Rad21+ and at 32°C display single dots in ∼50% of cells (Figure 7A). Surprisingly, heat shock induction led to significant (P<0.05) increase in cells with hsp70 (58–88%) and hsp9 (59–83%) single dots, indicating stable cohesion at these loci (Figure 7A). As a control, we analyzed cohesion at Mis4+/Rad21+ and Rad21+ sites (Figure 7B, CARs II and IX), which resulted in cells with 100% and ∼50% cohesion, respectively and were unaffected by heat shock conditions. Figure 7. Transcription induction mediates cohesion establishment. (A) DNA FISH analysis of hsp70 and hsp9 loci. FISH probes (red dots, Alexa-555) are shown inside DAPI stained nucleus (blue). *P < 0.05. two-tailed, paired Student's t-test. n = 100 cells, two biological repeats. (B) DNA FISH analysis of CAR II and CAR IX used as controls for FISH after heat shock induction. 100% cohesion observed at CAR II between 32 and 42°C (left panel) and ±50% cohesion at CAR IX (same as in Figure 1B and C, respectively). n = 100 cells, two biological repeats. (C) ChIP-qPCR analysis showing Eso1-GFP enrichment at hsp70 and hsp9 loci in S synchronized cells (Hydroxyurea) after heat shock. No tag strain was used to assess the background levels. Error bars represent SD, n = 3. *P < 0.05, two-tailed, paired Student's t-test comparing 42 to 32°C values and GFP signal to no tag control. (D) ChIP-qPCR analysis showing Eso1-GFP enrichment at hsp70 and hsp9 loci in G2 synchronized cells (Hydroxyurea block and release) after heat shock. No tag strain was used to assess the background levels. Error bars represent SD, n = 3. *P < 0.05, two-tailed, paired Student's t-test comparing 42 to 32°C values and GFP signal to no tag control. (E) RT-qPCR showing act1, hsp70 and hsp9 mRNA levels at 25, 37 and 42°C. Error bars represent SD, n = 3. *P < 0.05, two-tailed, paired Student's t-test comparing 42 to 25°C values. (F) DNA FISH analysis of hsp70 and hsp9 loci in WT and eso1-H17 cells. Cells were grown o/n at 25°C. Eso1 was inactivated at 37°C for 2 h, followed by heat shock at 42°C. We showed that Eso1 is recruited to hsp genes after heat shock (Figure 6F). In normal conditions cohesion is established during replication in S phase, when Eso1 protein levels are also highest. We employed Eso1 ChIP-qPCR in S or G2 phase synchronized cells prior to heat shock. Eso1 enrichment at hsp genes was moderately higher in S phase than in G2 phase, correlating with its protein levels (Figures 3D and 7C and D). Finally, we tested whether de novo established cohesion at hsp genes is dependent on Eso1. We employed WT and eso1-H17 cells at restrictive temperature followed by a heat shock. RT-qPCR analysis showed only a mild effect of 37°C temperature, which is necessary for Eso1 inactivation, on expression of the heat shock genes (Figure 7E). FISH performed on WT cells confirmed a significant (P < 0.05) increase in the number of cells with single dots at hsp70 (90% of cells) and hsp9 (89% of cells) after heat shock (Figure 7F). In contrast, inactivation of Eso1 resulted in no de novo cohesion at hsp loci after heat shock. These results suggest that increased transcription can lead to functional cohesion establishment, which is Eso1 dependent. RNAPII transcription affects chromatin association of cohesion proteins in human cells. To generalize our results, we cross-compared existing ChIP-Seq data for cohesin subunit SMC1 and RNAPII from human cells (39). ChIP-Seq data was extracted and analyzed as described previously (39), except the reads were mapped against Human_UCSC_GRCh37/hg19 assembly. Interestingly, SMC1 was present within ±1 kb of RNAPII (Supplementary Figure S6A) suggesting that cohesin is in close physical proximity to RNAPII. Furthermore, SMC1 overlapping with cohesin loader NIPBL (human orthologue of Mis4) also associates with highly transcribed genes in human cells (Supplementary Figure S6B). To further substantiate the interplay between cohesin and RNAPII, we assessed chromatin occupancy of cohesin subunits RAD21 and SMC3 in human HEK293T cells after inhibition of RNAPII by α-amanitin (2 ug/ml, 36 h) (34). We observed a >95% drop in RNAPII occupancy at the GAPDH transcription start site (Supplementary Figure S6C, left graph). This coincided with a significant reduction in RAD21 (Supplementary Figure S6C, middle graph, up to 75%) and SMC3 (Supplementary Figure S6C, right graph, up to 82%). Similarly, a decrease in RNAPII at c-MYC TSS (Supplementary Figure S6D, left graph) matched reduced RAD21 (Supplementary Figure S6D, middle graph) and SMC3 signals at the TSS (Supplementary Figure S6D, right graph). While RAD21 dropped at the CTCF+/cohesin+ site upstream (up probe) of c-MYC TSS, SMC3 levels were unperturbed, suggesting that RNAPII mediates chromosomal association of cohesin at highly transcribed genes. Secondary effects due to α-amanitin were ruled out as only RNAPII was degraded (Supplementary Figure S6F), while RAD21, SMC3 and tubulin were unaffected (Supplementary Figure S6F). Next, we assessed cohesin occupancy after transcription induction in human cells, focussing on the well-characterized ERBB2 gene that is actively transcribed in breast cancer derived ZR-75-1 cells, but silent in MCF-7 (50). Coincident with RNAPII enrichment (Supplementary Figure S6E, left graph), RAD21 (Supplementary Figure S6E, middle graph) and SMC3 (Supplementary Figure S6E, right graph) also showed a significant increase at the ERBB2 gene promoter in ZR-75-1 cells compared to MCF7 cells. HPRT was used as a negative control and protein levels were consistent between ZR-75-1 and MCF-7 cells (Supplementary Figure S6G). These results suggest that RNAPII facilitates cohesin association with chromatin in human cells. Cohesin and its loader co-localize with highly expressed RNAPII genes in S. cerevisiae To extend our findings further, we re-analyzed previously published S. cerevisiae ChIP-Seq data (11). First, we observed 92.2% overlap between cohesin and RNAPII loci, in sharp contrast to <0.6% overlap with RNAPIII only (Supplementary Figure S7A), suggesting that most cohesin in budding yeast also associates with RNAPII. Next, cohesin peaks that colocalize with cohesin loader proteins (Scc2+/Scc1+ or Scc4+/Scc1+) showed significant overlap with highly expressed RNAPII genes in contrast to Scc1+ peaks alone (Supplementary Figure S7B, C). Likewise, Scc2+/Smc3+ or Scc4+/Smc3+ overlap highly expressed RNAPII genes in comparison to Smc3+ peaks alone (Supplementary Figure S7D, E). Additionally, ∼80% Scc2+/Scc1+, ∼85% Scc4+/Scc1+ and ∼95% Scc1+ sites (Supplementary Figure S8A) were devoid of the 238 annotated highly expressed, open chromatin hyper-ChIPable regions (40). Similarly, ∼85% Scc2+/Smc3+, ∼85% Scc4+/Smc3+ and ∼95% Smc3+ sites (Supplementary Figure S8B) were also devoid of hyper-ChIPable regions. Furthermore, a comparison of gene expression between Scc2+/Scc1+ versus Scc1+ (Supplementary Figure S8C) and Scc4+/Scc1+ versus Scc1+ (Supplementary Figure S8D) sites, with or without inclusion of hyper-ChIPable regions was similar. Likewise, gene expression differences between Scc2+/Smc3+ versus Smc3+ (Supplementary Figure S8E) and Scc4+/Smc3+ versus Smc3+ (Supplementary Figure S8F) were independent of any hyper-ChIPable artifacts. Overall our results have identified cohesive loci on chromosomal arms in fission yeast. At these sites cohesin co-localizes with its loading complex, Pds5 and Eso1. Furthermore, we have identified RNAPII transcription as a potential functional determinant for cohesin loading and consequent cohesion establishment. DISCUSSION The multi-subunit cohesin complex is highly conserved across all eukaryotes and executes cohesion, transcription and repair/recombination functions (51). Interestingly, cohesin occupancy along chromosomes is highly variable among different organisms (26–30,52). In metazoans cohesin co-localizes at sites of active transcription and insulator elements. In budding yeast, cohesin accumulates at intergenic regions away from loading sites (26,28). In fission yeast, cohesin is divided into sub-sets, (i) cohesin at its loading sites Mis4+/Rad21+, and (ii) cohesin on its own Rad21+ only (29,33). We demonstrate in S. pombe that chromatin-associated cohesin results in functional cohesion of sister chromatids only at the sites of loading (Figure 1). Recent in vitro experiments showed that cohesin exhibits DNA affinity but its loading onto DNA substrates is greatly facilitated by its loader Mis4-Ssl3 (12). Our results show that cohesin association with its loader is indeed required for stable cohesion in vivo. The cohesin loading complex facilitates association of cohesin with DNA during late G1 in S. cerevisiae and telophase in vertebrates (8) but cohesion is established only during S phase (53,54), concomitant with Psm3 acetylation. Additionally, in S. cerevisiae Eco1 (Eso1 in S. pombe) levels are regulated via Cdk1 phosphorylation and proteasome degradation in G2 and M, with maximal expression in S phase (46). We observe significant Eso1 enrichment at cohesive sites (Figure 3E), supported by the presence of Eso1 protein (Figure 3D) in G2 and S phases in S. pombe. This suggests that Eso1 could be recruited to cohesive sites in G2, similar to de novo Eco1 dependent cohesion establishment during DNA damage in budding yeast (55,56). Interestingly, Mis4 and Ssl3 are HEAT- and TPR-repeat containing proteins respectively, inherently capable of several protein-protein interactions (57). Mis4-Ssl3 might play an yet uncharacterized role in chromatin recruitment of Eso1. Genome wide analysis shows that cohesive sites are close to Pds5 sites on chromatin (Figure 2B). In budding yeast and Xenopus, Pds5 is crucial for cohesin release, partly by recruiting Wpl1 (17,58). In contrast, Pds5 is also essential for de novo Psm3 acetylation (43–44,59) and protects cohesin from Hos1 mediated deacetylation during G2 and M phases in yeast (44). We detect significant levels of Pds5 at the cohesin loading sites, exhibiting stable cohesion (Figure 2D). Next, we show that cohesive sites significantly overlap with highly expressed RNAPII transcribed genes (Figure 4) suggesting a possible interplay between cohesion and transcription. We observed a dramatic reduction in chromatin bound cohesin after transcription inhibition (Figure 5) using 1,10-phenanthroline, a commonly used chemical with phenotype similar to an RNAPII ts mutant (60). While much evidence argues for a function of cohesin in the gene expression regulation, here we show that chromosomal association of cohesin and Mis4 depends on RNAPII transcription. Furthermore, a recent study shows direct interaction between RNAPII and cohesin subunit SA1 (homologue of Psc3 in S. pombe) in mammalian cells (61). Once stable cohesion is established (S-phase) it persists until anaphase, wherein only a fraction of cohesin is cleaved. However, we observe a reduction in chromatin bound cohesin upon transcription inhibition in G2, suggesting that RNAPII transcription is likely to facilitate cohesin loading throughout the cell cycle. Next, we show that induction of heat shock transcription leads to recruitment of Psm3, Mis4 and Eso1 at heat shock genes resulting in significant and functional de novo cohesion establishment in G2 (Figures 6 and 7). Recent evidence in budding yeast implicates the RSC chromatin-remodeling complex in establishing nucleosome free regions and facilitating chromatin association of Scc2-Scc4 and cohesin (62). Whether RSC complex performs a similar function in S. pombe is unknown. A direct interaction between cohesin and RNAPII (61) suggests that chromatin association of cohesin could be facilitated by chromatin remodeling but is mediated via RNAPII. To summarize, we propose a model (Supplementary Figure S9) for cohesion establishment on chromosomal arms in fission yeast. De novo synthesis of Rad21 in G1 (Rad21 is cleaved at anaphase onset) coincides with the reassembly of the cohesin ring. RNAPII is likely to mediate recruitment of cohesin through Mis4-Ssl3 to highly transcribed genes throughout the cell cycle. The open chromatin conformation at active genes provides an ambient platform for replication fork assembly and replication initiation. Consequently, the concurrence between newly replicated sister chromatids and the cohesion machinery promotes establishment of stable cohesion at these sites. In contrast, poorly expressed genes fail to recruit Pds5, Mis4 and Eso1. As a result, cohesin at non-cohesive sites maintains transient association and dissociation kinetics and fails to establish stable cohesion. Perhaps these non-cohesive sites could serve as a platform for cohesion establishment in events of DNA damage or stress response. We propose that RNAPII transcription is a key player for cohesin loading and consequently cohesion establishment on chromosome arms in eukaryotes. Supplementary Material SUPPLEMENTARY DATA We are grateful to Mitsuhiro Yanagida and Norihiko Nakazawa for their help with DNA FISH protocol. We also thank Kim Nasmyth, Jean-Paul Javerzat, Frank Uhlmann and Chris Norbury for cell strains and valuable comments. SUPPLEMENTARY DATA Supplementary Data are available at NAR Online. FUNDING Medical Research Council UK Career Development Award [BVRXMJ00 to M.G.]; John Fell Main Award [BVD04090 to M.G.]; EPA research grant (to M.G.); Felix scholarship (to S.B.). Funding for open access charge: Medical Research Council UK. Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Michaelis C. Ciosk R. Nasmyth K. Cohesins: chromosomal proteins that prevent premature separation of sister chromatids Cell 1997 91 35 45 9335333 2. Guacci V. Koshland D. Strunnikov A. A direct link between sister chromatid cohesion and chromosome condensation revealed through the analysis of MCD1 in S. cerevisiae Cell 1997 91 47 57 9335334 3. Haering C.H. Lowe J. Hochwagen A. Nasmyth K. Molecular architecture of SMC proteins and the yeast cohesin complex Mol. Cell 2002 9 773 788 11983169 4. Nasmyth K. Haering C.H. The structure and function of SMC and kleisin complexes Annu. Rev. Biochem. 2005 74 595 648 15952899 5. Nasmyth K. Haering C.H. Cohesin: its roles and mechanisms Annu. Rev. Genet. 2009 43 525 558 19886810 6. Haering C.H. Farcas A.M. Arumugam P. Metson J. Nasmyth K. The cohesin ring concatenates sister DNA molecules Nature 2008 454 297 301 18596691 7. Ivanov D. Nasmyth K. A topological interaction between cohesin rings and a circular minichromosome Cell 2005 122 849 860 16179255 8. Furuya K. Takahashi K. Yanagida M. Faithful anaphase is ensured by Mis4, a sister chromatid cohesion molecule required in S phase and not destroyed in G1 phase Genes Dev. 1998 12 3408 3418 9808627 9. Bernard P. Drogat J. Maure J.F. Dheur S. Vaur S. Genier S. Javerzat J.P. A screen for cohesion mutants uncovers Ssl3, the fission yeast counterpart of the cohesin loading factor Scc4 Curr. Biol.: CB 2006 16 875 881 16682348 10. Arumugam P. Gruber S. Tanaka K. Haering C.H. Mechtler K. Nasmyth K. ATP hydrolysis is required for cohesin's association with chromosomes Curr. Biol.: CB 2003 13 1941 1953 14614819 11. Hu B. Itoh T. Mishra A. Katoh Y. Chan K.L. Upcher W. Godlee C. Roig M.B. Shirahige K. Nasmyth K. ATP hydrolysis is required for relocating cohesin from sites occupied by its Scc2/4 loading complex Curr. Biol.: CB 2011 21 12 24 21185190 12. Murayama Y. Uhlmann F. Biochemical reconstitution of topological DNA binding by the cohesin ring Nature 2014 505 367 371 24291789 13. Weitzer S. Lehane C. Uhlmann F. A model for ATP hydrolysis-dependent binding of cohesin to DNA Curr. Biol. 2003 13 1930 1940 14614818 14. Gruber S. Arumugam P. Katou Y. Kuglitsch D. Helmhart W. Shirahige K. Nasmyth K. Evidence that loading of cohesin onto chromosomes involves opening of its SMC hinge Cell 2006 127 523 537 17081975 15. Kurze A. Michie K.A. Dixon S.E. Mishra A. Itoh T. Khalid S. Strmecki L. Shirahige K. Haering C.H. Lowe J. A positively charged channel within the Smc1/Smc3 hinge required for sister chromatid cohesion EMBO J. 2011 30 364 378 21139566 16. Gligoris T.G. Scheinost J.C. Burmann F. Petela N. Chan K.L. Uluocak P. Beckouet F. Gruber S. Nasmyth K. Lowe J. Closing the cohesin ring: structure and function of its Smc3-kleisin interface Science 2014 346 963 967 25414305 17. Chan K.L. Roig M.B. Hu B. Beckouet F. Metson J. Nasmyth K. Cohesin's DNA exit gate is distinct from its entrance gate and is regulated by acetylation Cell 2012 150 961 974 22901742 18. Chen Z. Cao H. Guo W. Lu Y. Identification of two forms of the Eso1 protein in Schizosaccharomyces pombe Cell Biol. Int. 2014 38 682 688 24375893 19. Feytout A. Vaur S. Genier S. Vazquez S. Javerzat J.P. Psm3 acetylation on conserved lysine residues is dispensable for viability in fission yeast but contributes to Eso1-mediated sister chromatid cohesion by antagonizing Wpl1 Mol. Cell. Biol. 2011 31 1771 1786 21300781 20. Gillespie P.J. Hirano T. Scc2 couples replication licensing to sister chromatid cohesion in Xenopus egg extracts Curr. Biol. 2004 14 1598 1603 15341749 21. Takahashi T.S. Yiu P. Chou M.F. Gygi S. Walter J.C. Recruitment of Xenopus Scc2 and cohesin to chromatin requires the pre-replication complex Nat. Cell Biol. 2004 6 991 996 15448702 22. Bernard P. Maure J.F. Partridge J.F. Genier S. Javerzat J.P. Allshire R.C. Requirement of heterochromatin for cohesion at centromeres Science 2001 294 2539 2542 11598266 23. Nonaka N. Kitajima T. Yokobayashi S. Xiao G. Yamamoto M. Grewal S.I. Watanabe Y. Recruitment of cohesin to heterochromatic regions by Swi6/HP1 in fission yeast Nat. Cell Biol. 2002 4 89 93 11780129 24. Tanaka T. Cosma M.P. Wirth K. Nasmyth K. Identification of cohesin association sites at centromeres and along chromosome arms Cell 1999 98 847 858 10499801 25. Fernius J. Marston A.L. Establishment of cohesion at the pericentromere by the Ctf19 kinetochore subcomplex and the replication fork-associated factor, Csm3 PLoS Genet. 2009 5 e1000629 19730685 26. Lengronne A. Katou Y. Mori S. Yokobayashi S. Kelly G.P. Itoh T. Watanabe Y. Shirahige K. Uhlmann F. Cohesin relocation from sites of chromosomal loading to places of convergent transcription Nature 2004 430 573 578 15229615 27. Misulovin Z. Schwartz Y.B. Li X.Y. Kahn T.G. Gause M. MacArthur S. Fay J.C. Eisen M.B. Pirrotta V. Biggin M.D. Association of cohesin and Nipped-B with transcriptionally active regions of the Drosophila melanogaster genome Chromosoma 2008 117 89 102 17965872 28. Glynn E.F. Megee P.C. Yu H.G. Mistrot C. Unal E. Koshland D.E. DeRisi J.L. Gerton J.L. Genome-wide mapping of the cohesin complex in the yeast Saccharomyces cerevisiae PLoS Biol. 2004 2 E259 15309048 29. Schmidt C.K. Brookes N. Uhlmann F. Conserved features of cohesin binding along fission yeast chromosomes Genome Biol. 2009 10 R52 19454013 30. Pauli A. van Bemmel J.G. Oliveira R.A. Itoh T. Shirahige K. van Steensel B. Nasmyth K. A direct role for cohesin in gene regulation and ecdysone response in Drosophila salivary glands Curr. Biol.: CB 2010 20 1787 1798 20933422 31. Kagey M.H. Newman J.J. Bilodeau S. Zhan Y. Orlando D.A. van Berkum N.L. Ebmeier C.C. Goossens J. Rahl P.B. Levine S.S. Mediator and cohesin connect gene expression and chromatin architecture Nature 2010 467 430 435 20720539 32. Schmidt D. Schwalie P.C. Ross-Innes C.S. Hurtado A. Brown G.D. Carroll J.S. Flicek P. Odom D.T. A CTCF-independent role for cohesin in tissue-specific transcription Genome Res. 2010 20 578 588 20219941 33. Gullerova M. Proudfoot N.J. Cohesin complex promotes transcriptional termination between convergent genes in S. pombe Cell 2008 132 983 995 18358811 34. Nguyen V.T. Giannoni F. Dubois M.F. Seo S.J. Vigneron M. Kedinger C. Bensaude O. In vivo degradation of RNA polymerase II largest subunit triggered by alpha-amanitin Nucleic Acids Res. 1996 24 2924 2929 8760875 35. Cam H.P. Sugiyama T. Chen E.S. Chen X. FitzGerald P.C. Grewal S.I. Comprehensive analysis of heterochromatin- and RNAi-mediated epigenetic control of the fission yeast genome Nat. Genet. 2005 37 809 819 15976807 36. Marguerat S. Bahler J. Coordinating genome expression with cell size Trends Genet.: TIG 2012 28 560 565 22863032 37. Zuin J. Dixon J.R. van der Reijden M.I. Ye Z. Kolovos P. Brouwer R.W. van de Corput M.P. van de Werken H.J. Knoch T.A. van I.W.F. Cohesin and CTCF differentially affect chromatin architecture and gene expression in human cells Proc. Natl. Acad. Sci. U.S.A. 2014 111 996 1001 24335803 38. Freeberg M.A. Han T. Moresco J.J. Kong A. Yang Y.C. Lu Z.J. Yates J.R. Kim J.K. Pervasive and dynamic protein binding sites of the mRNA transcriptome in Saccharomyces cerevisiae Genome Biol. 2013 14 R13 23409723 39. Zuin J. Franke V. van Ijcken W.F. van der Sloot A. Krantz I.D. van der Reijden M.I. Nakato R. Lenhard B. Wendt K.S. A cohesin-independent role for NIPBL at promoters provides insights in CdLS PLoS Genet. 2014 10 e1004153 24550742 40. Teytelman L. Thurtle D.M. Rine J. van Oudenaarden A. Highly expressed loci are vulnerable to misleading ChIP localization of multiple unrelated proteins Proc. Natl. Acad. Sci. U.S.A. 2013 110 18602 18607 24173036 41. Tatebayashi K. Kato J. Ikeda H. Isolation of a Schizosaccharomyces pombe rad21ts mutant that is aberrant in chromosome segregation, microtubule function, DNA repair and sensitive to hydroxyurea: possible involvement of Rad21 in ubiquitin-mediated proteolysis Genetics 1998 148 49 57 9475720 42. Luche D.D. Forsburg S.L. Cell-cycle synchrony for analysis of S. pombe DNA replication Methods Mol. Biol. 2009 521 437 448 19563121 43. Vaur S. Feytout A. Vazquez S. Javerzat J.P. Pds5 promotes cohesin acetylation and stable cohesin-chromosome interaction EMBO Rep. 2012 13 645 652 22640989 44. Chan K.L. Gligoris T. Upcher W. Kato Y. Shirahige K. Nasmyth K. Beckouet F. Pds5 promotes and protects cohesin acetylation Proc. Natl. Acad. Sci. U.S.A. 2013 110 13020 13025 23878248 45. D'Ambrosio L.M. Lavoie B.D. Pds5 prevents the PolySUMO-dependent separation of sister chromatids Curr. Biol. 2014 24 361 371 24485833 46. Lyons N.A. Morgan D.O. Cdk1-dependent destruction of Eco1 prevents cohesion establishment after S phase Mol. Cell 2011 42 378 389 21549314 47. Marguerat S. Schmidt A. Codlin S. Chen W. Aebersold R. Bahler J. Quantitative analysis of fission yeast transcriptomes and proteomes in proliferating and quiescent cells Cell 2012 151 671 683 23101633 48. Cam H. Grewal S.I. RNA interference and epigenetic control of heterochromatin assembly in fission yeast Cold Spring Harbor Symp. Quant. Biol. 2004 69 419 427 16117676 49. Hall I.M. Noma K. Grewal S.I. RNA interference machinery regulates chromosome dynamics during mitosis and meiosis in fission yeast Proc. Natl. Acad. Sci. U.S.A. 2003 100 193 198 12509501 50. Mungamuri S.K. Murk W. Grumolato L. Bernstein E. Aaronson S.A. Chromatin modifications sequentially enhance ErbB2 expression in ErbB2-positive breast cancers Cell Rep. 2013 5 302 313 24120871 51. Nasmyth K. Cohesin: a catenase with separate entry and exit gates? Nat. Cell Biol. 2011 13 1170 1177 21968990 52. Parelho V. Hadjur S. Spivakov M. Leleu M. Sauer S. Gregson H.C. Jarmuz A. Canzonetta C. Webster Z. Nesterova T. Cohesins functionally associate with CTCF on mammalian chromosome arms Cell 2008 132 422 433 18237772 53. Uhlmann F. Nasmyth K. Cohesion between sister chromatids must be established during DNA replication Curr. Biol. 1998 8 1095 1101 9778527 54. Tanaka K. Yonekawa T. Kawasaki Y. Kai M. Furuya K. Iwasaki M. Murakami H. Yanagida M. Okayama H. Fission yeast Eso1p is required for establishing sister chromatid cohesion during S phase Mol. Cell. Biol. 2000 20 3459 3469 10779336 55. Strom L. Sjogren C. Chromosome segregation and double-strand break repair - a complex connection Curr. Opin. Cell Biol. 2007 19 344 349 17466504 56. Unal E. Heidinger-Pauli J.M. Koshland D. DNA double-strand breaks trigger genome-wide sister-chromatid cohesion through Eco1 (Ctf7) Science 2007 317 245 248 17626885 57. Neuwald A.F. Hirano T. HEAT repeats associated with condensins, cohesins, and other complexes involved in chromosome-related functions Genome Res. 2000 10 1445 1452 11042144 58. Shintomi K. Hirano T. Releasing cohesin from chromosome arms in early mitosis: opposing actions of Wapl-Pds5 and Sgo1 Genes Dev. 2009 23 2224 2236 19696148 59. Tanaka K. Hao Z. Kai M. Okayama H. Establishment and maintenance of sister chromatid cohesion in fission yeast by a unique mechanism EMBO J. 2001 20 5779 5790 11598020 60. Grigull J. Mnaimneh S. Pootoolal J. Robinson M.D. Hughes T.R. Genome-wide analysis of mRNA stability using transcription inhibitors and microarrays reveals posttranscriptional control of ribosome biogenesis factors Mol. Cell. Biol. 2004 24 5534 5547 15169913 61. Izumi K. Nakato R. Zhang Z. Edmondson A.C. Noon S. Dulik M.C. Rajagopalan R. Venditti C.P. Gripp K. Samanich J. Germline gain-of-function mutations in AFF4 cause a developmental syndrome functionally linking the super elongation complex and cohesin Nat. Genet. 2015 47 338 344 25730767 62. Lopez-Serra L. Kelly G. Patel H. Stewart A. Uhlmann F. The Scc2-Scc4 complex acts in sister chromatid cohesion and transcriptional regulation by maintaining nucleosome-free regions Nat. Genet. 2014 46 1147 1151 25173104

PMC005xxxxxx/PMC5001583.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2708494510.1093/nar/gkw258Gene regulation, Chromatin and EpigeneticsProtection of CpG islands from DNA methylation is DNA-encoded and evolutionarily conserved Long Hannah K. 12http://orcid.org/0000-0002-5694-0398King Hamish W. 1Patient Roger K. 2Odom Duncan T. 3Klose Robert J. 1*1 Department of Biochemistry, University of Oxford, Oxford, OX1 3QU, UK2 Molecular Haematology Unit, Weatherall Institute of Molecular Medicine, University of Oxford, John Radcliffe Hospital, Oxford, OX3 9DS, UK3 Cancer Research UK Cambridge Institute, University of Cambridge, Cambridge, CB2 0RE, UK* To whom correspondence should be addressed. Tel: +44 1865 613225; Email: rob.klose@bioch.ox.ac.uk19 8 2016 15 4 2016 15 4 2016 44 14 6693 6706 01 4 2016 09 3 2016 13 11 2015 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.DNA methylation is a repressive epigenetic modification that covers vertebrate genomes. Regions known as CpG islands (CGIs), which are refractory to DNA methylation, are often associated with gene promoters and play central roles in gene regulation. Yet how CGIs in their normal genomic context evade the DNA methylation machinery and whether these mechanisms are evolutionarily conserved remains enigmatic. To address these fundamental questions we exploited a transchromosomic animal model and genomic approaches to understand how the hypomethylated state is formed in vivo and to discover whether mechanisms governing CGI formation are evolutionarily conserved. Strikingly, insertion of a human chromosome into mouse revealed that promoter-associated CGIs are refractory to DNA methylation regardless of host species, demonstrating that DNA sequence plays a central role in specifying the hypomethylated state through evolutionarily conserved mechanisms. In contrast, elements distal to gene promoters exhibited more variable methylation between host species, uncovering a widespread dependence on nucleotide frequency and occupancy of DNA-binding transcription factors in shaping the DNA methylation landscape away from gene promoters. This was exemplified by young CpG rich lineage-restricted repeat sequences that evaded DNA methylation in the absence of co-evolved mechanisms targeting methylation to these sequences, and species specific DNA binding events that protected against DNA methylation in CpG poor regions. Finally, transplantation of mouse chromosomal fragments into the evolutionarily distant zebrafish uncovered the existence of a mechanistically conserved and DNA-encoded logic which shapes CGI formation across vertebrate species. cover-date19 August 2016 ==== Body INTRODUCTION DNA methylation on CpG dinucleotides in vertebrate genomes is associated with transcriptional repression (1,2) and is epigenetically inherited during cell division to propagate repressive chromatin states (3–5). Conversely, short contiguous roughly 1–2 kb regions of CpG-rich DNA, known as CpG islands (CGIs), are interspersed throughout the genome and are resistant to DNA methylation (6,7). CGIs are found associated with 60–70% vertebrate gene promoters and are a central and evolutionarily conserved feature at these sites (8,9). CGIs function to recruit a family of ZF-CxxC DNA binding domain-containing proteins that associate with chromatin-modifying activities to remodel chromatin structure at gene promoters and contribute to gene regulation (10–12). Importantly, promoter-associated CGIs are usually free of DNA methylation regardless of the transcriptional state of the associated gene and are hypomethylated in most tissues (13–16), with only a subset of weak promoter-associated CGIs undergoing changes in DNA methylation during development (14,17–19). In addition to promoter-associated CGIs, an additional class of hypomethylated elements has been identified away from gene promoters that appear to function as distal gene regulatory elements, often encompassing enhancers (8,13,15,16,20). These regions tend to be hypomethylated in a subset of tissues, suggesting that the mechanisms underpinning their methylation state may differ from that of classical promoter-associated CGIs. Despite our ever-increasing knowledge of DNA methylation landscapes in diverse vertebrate genomes, the principles and mechanisms that specify these patterns and protect CGIs and distal regulatory elements from DNA methylation is underexplored, particularly with respect to how these mechanisms function in their normal genomic context on chromosomes and during animal development. This is due to the fact that most of our information about these processes has relied on interrogating CGIs and their methylation states by inserting short DNA sequences (natural or synthetic) into engineered acceptor sites in cell culture systems (12,21–24). These studies have suggested that nucleotide features may function as a molecular signal to specify regions that should be refractory to the placement of DNA methylation (12,21–24). Alternatively, in some instances sequence-specific transcription factor binding has been linked to protection of the underlying DNA sequence from DNA methylation (22,23,25–27). Given that genomic context and tissue specific features play central roles in the way DNA elements function, it still remains poorly understood how significant these, or other mechanisms, are in specifying how hypomethylated regions of DNA (HMRs) form at a chromosome scale in animals. To explore the mechanisms that shape CGI methylation state in a genomic context and to directly test whether these properties are evolutionarily conserved in vertebrates we have exploited a transchromosomic mouse animal model (Tc1) (28) in which most of human chromosome 21 has been stably transplanted into the mouse nuclear environment. Using genomic approaches we examined how HMRs are specified within 42 Mb of this human chromosome in developmentally distinct tissues. This revealed that irrespective of host species, CpG-rich promoter-associated CGIs are almost invariantly hypomethylated. In contrast, we discovered that distal elements are prone to alternative DNA methylation states depending on the host species and that this relies on both DNA sequence and transcription factor binding. Strikingly, these observations hold true when mouse chromosomal fragments are transposed into zebrafish, demonstrating for the first time that these general mechanisms are functionally conserved across divergent vertebrate species. Together this reveals that a DNA sequence-encoded logic and evolutionarily conserved mechanisms relying on the interplay between DNA features and transcription factor binding shape the DNA methylation based epigenome in vertebrate chromosomes. MATERIALS AND METHODS Tc1 mouse sample preparation The Tc1 mouse line was maintained as previously described (28). Tc1 mice were bred by crossing female Tc1 mice to male (129S8 × C57BL/6J) F1 mice and were housed in the Biological Resources Unit under Home Office Licence (PPL 80/2197). Genomic DNA was extracted from fresh frozen Tc1 mouse liver and testis tissue using Genomic-tip 100/G kit (QIAGEN). Transposition of BAC DNA sequences into the zebrafish genome Four mouse BACs were engineered to contain a GFP reporter driven by the eF1alpha promoter, and an inverted Tol2 transposition cassette (iTol2) flanking an ampicillin resistance gene (29) by BAC recombineering (Gene Bridges). BAC DNA was prepared using the Nucleobond BAC preparation kit and transposase RNA was generated using the SP6 RNA polymerase mMessage Machine kit (Ambion). BAC DNA and transposase RNA were introduced into zebrafish embryos by microinjection soon after fertilisation. GFP positive embryos were collected at 28–30 hpf and genomic DNA was extracted using the QIAGEN DNeasy Blood and Tissue Kit. BioCAP BioCAP-sequencing was performed as described previously (30) in Tc1 mouse and Tc0 (wildtype) mouse testis and liver tissue in biological duplicate, and for zebrafish embryos containing integrated mouse BAC DNA. Human and mouse BioCAP datasets were detailed previously (GSE43512). HMR peak-calling and differential methylation analysis Tc1 BioCAP-sequencing reads were aligned to a composite genome containing mouse chromosomes and human chromosome 21 using bowtie (31). HMRs were identified in Tc1 mouse tissues from BioCAP-sequencing traces using MACS1.4 (32) with settings –tsize = 50 –bw = 300 –mfold = 10,30 –pvalue = 1e-5 –verbose = 10 -g 4.8e+8 and including the use of an input control. Only HMRs that were identified in both biological replicates were retained. Human HMRs are detailed in (8). The human chromosome 21 in the Tc1 mouse has some rearrangements and duplications (33). To account for this, HMRs overlapping breakpoints or deleted regions were removed from the analysis and read values were scaled appropriately in duplicated regions. Differential methylation between species was identified if an HMR exhibited a greater than 2-fold change in BioCAP reads between the human and Tc1 mouse experiments using BAM files normalised to the same read count across chromosome 21. To compare HMRs on mouse chromosomes, HMR sites were identified in the Tc1 and Tc0 (wildtype) mouse using MACS1.4 (32) and compared as indicated above. Zebrafish Bio-CAP reads were aligned to a composite genome containing zebrafish chromosomes and each individual BAC sequence using bowtie-2 (34). Determining CpG density and GC content CpG density and GC content of individual HMR regions was calculated based on the underlying DNA sequence for HMRs that were shared between human and Tc1 mouse or were species-specific. A matched background control was generated by randomly shifting HMR coordinates to another position on chromosome 21. CpG density and GC content was plotted as frequency distribution plots for liver and testis tissue. Genomic repeat age analysis Repeat elements that overlapped the summit of HMR elements were analysed for repeat age. Repeat age was estimated by determining the number of substitutions from the repeat consensus sequence (Repeat-Masker track from UCSC, milliDiv column) and dividing this number by the estimated mutation rate for mammalian species, which is 2.2 × 10−9 per year (35,36). The distribution of repeat age was displayed as box plots for shared and species-specific HMRs. Transcription factor binding Transcription factor binding sites were identified using the MACS1.4 peak caller with the same settings as were used to identify HMRs above (32). Shared or species-specific transcription factor binding events were determined by intersection of transcription factor binding sites identified in human and Tc1 mouse. Gene expression analysis Gene expression data (reads per kilobase per million mapped reads, RPKM) was used to compare the expression of genes with an HMR overlapping their promoter (TSS ± 500bp) (36). Genes with a promoter-associated HMR were segregated depending on whether the HMR was shared between human and Tc1 mouse, or was species-specific. Graphical representation Heatmaps and metaplots were generated using Homer annotatePeaks and plotted in R. Venn diagrams were plotted in R using the package VennDiagram. Boxplots were plotted in R using the package boxplot. Scatterplots were generated in R. Multiplex bisulfite sequencing Bisulfite conversion of 250 ng human and two replicates of Tc1 mouse liver genomic DNA was performed using the EZ DNA Methylation-Gold Kit (Zymo Research). 10 picogram (pg) unmethylated and 10 pg methylated Arabidopsis thaliana BAC DNA (F24B22 and F19K16 respectively) was spiked into the three conversion reactions to control for bisulfite conversion (Diagenode DNA Methylation control package, EF-100-0040). PCR-amplified DNA was pooled for each replicate sample and Illumina sequencing libraries were prepared using the NEBNext DNA Library Prep Master Mix Set for Illumina (NEB). Indexed libraries were combined and sequenced on a MiSeq machine (Illumina). Sequencing data was analysed using the BisMark software from Babraham Bioinformatics (http://www.bioinformatics.babraham.ac.uk/projects/bismark/) (37). The two Tc1 biological replicates were highly similar, and were averaged for visualisation. Primer sets used for bisulfite sequencing are available on request. RESULTS Hypomethylated regions (HMRs) on human chromosome 21 are largely recapitulated in the transchromosomic mouse model We recently demonstrated that many features of the CGI system are shared across vertebrate species (8).This led us to hypothesize that the mechanisms that specify DNA methylation state at CGIs may be evolutionarily conserved and rely on the underlying DNA sequence. To directly test this hypothesis, we exploited a transchromosomic mouse model (known as Tc1) that has been engineered to contain approximately 42 Mb of human chromosome 21 on a single independently segregating chromosome (28). This interesting trans-species model system provides a unique opportunity to ask how hypomethylated regions of DNA (HMRs) are formed and maintained in their natural genomic context in a developing animal following chromosome-scale introduction of DNA from one vertebrate, human, into the nuclear environment of second vertebrate, mouse. Importantly, it also allowed us to ask whether human chromosomal DNA-encoded information is sufficient to recapitulate the human host DNA methylation patterns in the nuclear environment of mouse, and in doing so test whether these mechanisms are evolutionarily conserved during development. To answer these fundamental questions we isolated DNA from Tc1 mouse liver and testis and used biotinylated CxxC affinity purification (BioCAP) coupled to massively parallel sequencing to isolate regions of the genome containing hypomethylated DNA, and mapped these regions onto the mouse genome plus human chromosome 21 (30,38). We chose liver and testis to use as experimental tissues because their structure is highly similar in mammals in terms of cellular composition (39–41) and gene expression profiles (42–44), enabling a meaningful and direct comparison of BioCAP-seq signal on human chromosome 21. BioCAP experiments were carried out in duplicate with individual experiments correlating extremely well (R2 > 0.95, Supplementary Figure S1), further supporting the robustness of our approach (8,30). To ensure that the presence of the human chromosome did not affect the function of the DNA methylation system in the Tc1 mouse, we compared HMRs on the mouse chromosomes in the Tc1 mouse to HMRs formed in the same tissues in wild-type mice of the same genetic background. Importantly, HMRs on the Tc1 mouse chromosomes were indistinguishable from those of wild-type mice (Supplementary Figure S2A-F). Therefore, the presence of human chromosome 21 in the Tc1 mouse does not affect the function of the host DNA methylation machinery, indicating that the Tc1 mouse is a suitable model to study HMR formation on the newly introduced human chromosome 21. To investigate how HMRs form on human chromosome 21 in the mouse nuclear environment, we identified HMRs in developmentally distinct testis and liver tissue from the Tc1 mouse and directly compared these to HMRs formed on chromosome 21 in the corresponding human tissues (8). This revealed that the majority of human HMRs (85% in testis and 82% in liver) were recapitulated in the Tc1 mouse (Figure 1A and B). This was evident when BioCAP signal was plotted at all human chromosome 21 HMRs in the human and corresponding Tc1 mouse tissues (Figure 1C and D, left, and Figure 1E). To ensure that the HMRs formed on human chromosome 21 in the Tc1 mouse were functioning as they normally do in human tissues, we examined whether they were also modified by H3K4me3, a histone modification placed at HMRs by ZF-CxxC domain containing histone methyltransferase complexes (12,45). In agreement with the observation that the majority of HMRs form normally on human chromosome 21 in the Tc1 mouse tissues, we also observed robust H3K4me3 ChIP-seq signal at these sites in liver and testis tissue (Figure 1C and D, right). Overall, the majority of HMRs on human chromosome 21 and their stereotypical chromatin modifications are faithfully recapitulated when the chromosome is transplanted into mouse. This indicates that, for the most part, DNA sequence is sufficient to shape DNA methylation state at HMRs during vertebrate development, and furthermore that the systems required to achieve HMR formation at these sites are highly conserved between two vertebrate species, human and mouse. Figure 1. Hypomethylated regions (HMRs) on human chromosome 21 are largely recapitulated in the transchromosomic mouse model. (A and B) Profiles of non-methylated DNA (BioCAP-seq) at three regions on human chromosome 21 in human (upper) and in Tc1 mouse (lower, inverted) liver (A) and testis (B) tissues. Genes are depicted above the BioCAP traces. (C and D) Heatmaps depicting BioCAP (left) and H3K4me3 ChIP-seq (right) signal across human chromosome 21 HMRs in the human and Tc1 mouse liver (C) and testis (D) tissues. Signal is ranked according to HMR length and aligned to the centre of the HMR. Scalebar in kb. (E) Scatterplots of BioCAP-seq reads for human and Tc1 mouse at all human HMRs from liver (left) and testis tissue (right). Identification of species-specific HMRs Although HMRs on chromosome 21 were for the most part appropriately specified in the Tc1 mouse, we observed a number of locations where DNA methylation was gained or lost in the Tc1 mouse when compared to the same locus in human (Figures 1E and 2A and B, upper panels). This interesting observation suggests that some species-specific features must also contribute to HMR specification. To explore these changes in more detail, we used targeted deep bisulfite sequencing to examine at single base pair resolution the DNA methylation state at a number of these loci. In all cases, this revealed quantitative differences in DNA methylation across a range of CpG densities in agreement with the altered magnitude of BioCAP-seq signal. This is evident when individual loci were visualized (Figure 2A and B, lower panels) and more broadly at a set of 37 differentially methylated HMRs but not control loci (Supplementary Figure S3). We refer to these differentially methylated loci as species-specific HMRs (ssHMRs) to distinguish them from the majority of HMRs that display invariant methylation patterns when chromosome 21 resides in either Tc1 mouse or human tissues. Figure 2. Identification of species-specific HMRs. (A and B) BioCAP-seq traces across two human-specific (A) and two Tc1-specific (B) HMRs on human chromosome 21. Species-specific HMRs (ssHMRs) are indicated by a horizontal bar below the BioCAP-seq traces (upper). Bisulfite sequencing at species-specific HMRs confirms alterations in methylation at these sites (lower). Bisulfite amplicons (BA) are depicted by a horizontal black bar, CpG dinucleotides by a vertical line and the methylation status of each CpG in human or Tc1 liver is depicted as a vertical line between 0 and 100%. (C) Scatter plot of BioCAP-seq reads for human and Tc1 mouse at all HMRs to illustrate human and Tc1-specific HMRs in liver (upper) and testis (lower). (D and E) Heatmaps of BioCAP-seq signal in human and Tc1 mouse liver (D) and testis (E) tissues illustrate that a subset of HMRs are differentially methylated. Heatmaps are ranked according to HMR length and aligned to the centre of the HMR with shared (upper), human-specific (middle), and Tc1-specific (lower) sites clustered together. Scalebar in kb. (F) Venn diagrams depicting the overlap between human-specific HMRs (upper) or Tc1-specific HMRs (lower) from different tissues. Differential analysis of BioCAP-seq signal over the combined set of Tc1 and human HMRs revealed a surprisingly large number of ssHMRs (244 in liver and 209 in testis) (Figure 2C–E). At these sites there was a tendency for the hypomethylated state to predominate in the Tc1 tissue, accounting for 68–69% of ssHMRs (Figure 2C–E). Interestingly, human-specific HMRs were usually unique to one tissue (96%) (Figure 2F, upper), suggesting that HMR formation at these sites may rely on tissue-specific events. In contrast, fewer Tc1-specific HMRs were unique to one tissue (66%) (Figure 2F, lower) suggesting that species-specific, but tissue-invariant, activities contribute to HMR formation at the remaining (33%) Tc1-specific HMRs. The majority of species-specific HMRs are not directly related to changes in gene expression and are distal to gene promoter regions The identification of ssHMRs on human chromosome 21 in the human and mouse nuclear environments provided a unique opportunity to examine these differences in detail and discover the mechanisms that specify these methylation patterns. Given the historical relationship between CGIs and gene promoters, we first examined whether ssHMRs were associated with gene transcriptional start sites (TSSs). This revealed that ssHMRs are infrequently associated with TSSs (38/244 for liver and 30/209 for testis), and that the majority of these sites are located away from gene promoters (Figure 3A). Furthermore, when we examined the magnitude of change in BioCAP signal at ssHMRs, those found away from gene promoters showed larger alterations suggesting that these sites are more prone to nuclear environment-dependent changes in DNA methylation (Figure 3B). In the small number of instances where ssHMRs overlapped a gene promoter, there was no obvious relationship between the observed differences in DNA methylation and host gene expression patterns, suggesting that these alterations are not related to transcriptional changes (Supplementary Figure S4). Figure 3. The majority of species-specific HMRs are distal to gene promoter regions. (A) Venn diagrams depicting the overlap between HMRs in the human and Tc1 mouse tissue. HMRs are segregated into those located at transcription start site (TSS ± 500bp) (upper) and those found away from TSSs (lower). Human-specific HMRs are coloured in dark grey, shared HMRs in grey and Tc1-specific HMRs in light grey. (B) Boxplots depicting fold-change in BioCAP signal between Tc1 mouse and human at Tc1-specific HMRs located at a gene TSS or elsewhere in the chromosome. Tc1-specific HMRs away from gene TSSs exhibit a greater fold-change in BioCAP signal compared to those at gene TSSs. Significance values calculated using Welch's T-test (*P < 0.05). Species-specific HMRs in the Tc1 mouse are CpG-rich and often associated with young repetitive DNA elements Given the overrepresentation of ssHMRs in regions of the genome distal to TSSs, we examined in more detail the DNA sequence features at these sites to understand if this could explain their species-specific methylation patterns and provide insight into the mechanisms shaping their methylation state. We initially focused on CpG dinucleotide frequency and GC content, based on our previous observations that these nucleotide features are almost universally elevated amongst regions in vertebrate genomes that lack CpG methylation (8). Initially we examined Tc1-specific HMRs as they were the most abundant group of ssHMRs and directly compared their nucleotide features to shared HMRs. In general, shared HMRs on human chromosome 21 in both the human and Tc1 mouse were characterized by classical CpG island nucleotide features, with CpG density and GC content above the genome average, consistent with their invariantly hypomethylated state (Figure 4A and B). Somewhat surprisingly, Tc1-specific HMRs had highly elevated CpG density and GC content, even eclipsing the nucleotide frequencies characteristic of the non-methylated state at most shared HMRs (Figure 4A and B). This indicates that despite their CGI-like nucleotide composition, a subset of CpG and GC-rich regions on human chromosome 21 are actively targeted by the DNA methylation machinery in human tissues, and yet these same sequences evade DNA methylation in the Tc1 mouse. Figure 4. Tc1-specific HMRs are CpG-rich and often associated with young repetitive DNA elements. (A and B) CpG density (A) and GC content (B) plots of liver (upper panel, red) and testis (lower panel, blue) HMRs that are shared between human and Tc1 mouse (dark line) or are Tc1-specific (pale line). A background control is indicated (grey). (C) Boxplots depicting age of repeats (in million years) associated with HMRs in liver (left) and testis tissue (right). HMRs are segregated into those that are shared in human and Tc1 mouse (purple) and are Tc1-specific (pink). The difference in repeat age between shared and Tc1-specific HMRs is significantly different as calculated by a Mann–Whitney U test (*P < 5 × 10−4). (D and E) Snapshots of two Tc1-specific HMRs associated with a repetitive element. (F) A snapshot of a gene promoter on human chromosome 21 with a shared HMR, and which is associated with RNA PolII in both the Tc1 mouse and endogenous human nuclear environment. (G) RNA PolII enrichment at human-Tc1 shared TSS-associated HMRs (upper) and Tc1-specific non-TSS-associated HMRs (lower) in human and Tc1 mouse liver. CpG and GC-rich sequences that are methylated in human tissues are often associated with repetitive or parasitic DNA elements (46). Targeted methylation of these DNA elements is a common evolutionary strategy to suppress the potentially deleterious effects on the host genome (1,47). In the human genome, primate-specific repetitive elements have emerged recently enough that their CpG frequency remains high in the face of evolutionary mechanisms that drive down the overall frequency of CpG dinucleotides in vertebrate genomes (48–52). Therefore some of these elements display an elevated CpG density compared to the surrounding genome, much like CGI elements (51,53). We previously showed that in some cases, primate-specific repeats can possess latent gene regulatory potential and become activated in the mouse (36). When the average age of repeats associated with either shared HMRs, human-specific HMRs, and Tc1 mouse-specific HMRs were compared it was evident that young, newly acquired repetitive elements were overrepresented in Tc1 mouse-specific HMRs (Figure 4C and Supplementary Figure S5). Interestingly, however, loss of methylation at most of these sites was unrelated to gene transcription, as examination of ChIP-seq for RNA polymerase II (Pol II) in liver tissue at Tc1-specific HMRs revealed that only a subset (20%) of these sites acquired RNA Pol II occupancy. This is evident at individual loci (compare Figure 4D and E to F) and through analysis of RNAPII occupancy at Tc1-specific HMRs compared to transcribed shared HMRs (Figure 4G). Together, this indicates that Tc1-specific HMRs do not simply result from unmasking of dormant or cryptic gene promoters with latent regulatory potential (36) and supports the idea that species-specific trans-acting factors actively target DNA methylation to these sites. Therefore, our observations reveal that in the absence of co-evolved mechanisms that specifically methylate young human repeat elements, elevated CpG and GC content is sufficient to protect these sequences from DNA methylation during development. Together these interesting interspecies observations provide new and widespread evidence that mechanisms that sense CpG dinucleotide frequency contribute centrally to evasion of the DNA methylation machinery. Species-specific transcription factor binding is widely associated with species-specific HMR formation Elevated nucleotide features at young human repetitive sequences appear to contribute to their hypomethylated state in the Tc1 mouse, however this cannot explain how a subset of human HMRs on human chromosome 21 acquire DNA methylation in the mouse nuclear environment. To understand the features that drive formation of these human-specific HMRs we again examined their underlying nucleotide features. In contrast to Tc1 mouse-specific HMRs, human-specific HMRs were characterized by CpG density and GC content above the genome average, but significantly below those found at classical CGI elements (Figure 5A–B). This suggests that mechanisms protecting these sites from methylation in human tissues do not rely on classical CGI-like nucleotide frequencies. Figure 5. Species-specific transcription factor binding is widely associated with species-specific HMR formation. (A and B) CpG density (A) and GC content (B) plots of liver (upper panel, red) and testis (lower panel, blue) HMRs that are shared between human and Tc1 mouse (dark line) or are human-specific (dashed line). A background control is indicated (grey). (C) A snapshot illustrating a human-specific HMR corresponding to a human-specific transcription factor binding event as indicated by CEBPA and HNF4A ChIP-seq signal. (D) A 4-way Venn diagram comparing human-specific HMRs to human-specific transcription factor binding sites (TFBSs) for CEBPA, HNF4A and CTCF on human chromosome 21. 31% of human-specific HMRs overlapped a human-specific TFBS. (E) A snapshot illustrating a Tc1 mouse-specific HMR corresponding to a Tc1 mouse-specific transcription factor binding event as indicated by HNF4A and CTCF ChIP-seq signal. (F) A 4-way Venn diagram comparing Tc1-specific HMRs to Tc1-specific transcription factor binding sites for CEBPA, HNF4A and CTCF on human chromosome 21. 28% of Tc1-specific HMRs overlapped a Tc1-specific TFBS. We and others have recently identified a subset of HMRs that are differentially methylated depending on tissue type (8,14–17,54). These tissue-specific HMRs often correspond to distal regulatory elements, including enhancers. Much like the human-specific chromosome 21 HMRs, these sites have low CpG dinucleotide frequency and GC content. Distal regulatory elements usually correspond to sites of transcription factor occupancy, which has been proposed to protect the underlying binding site and surrounding DNA from methylation (20,22,23,26,27,55). Therefore, an interesting possibility is that human-specific HMRs at regions exhibiting an intermediate-to-low CpG frequency and GC content may result from species-specific DNA binding events. To investigate this possibility, we examined the ChIP-seq profiles of three DNA binding factors CEBPA, HNF4A and CTCF in the human and Tc1 mouse liver (36). Despite the conserved nature of these DNA binding factors between human and mouse, we identified a large number of species-specific DNA binding events on chromosome 21, likely due to subtle changes in binding site preference (Supplementary Figure S6A). An examination of these species-specific transcription factor binding events revealed that human-specific HMRs frequently overlap transcription factor binding events that are unique to the human liver. For example, the 3′ end of the WDR4 gene is characterized by HNF4A and CEBPA binding events that only occur in the human liver, and this corresponds to a human-specific HMR (Figure 5C). Strikingly, 31% (23/75) of human-specific HMRs intersected at least one transcription factor binding event that was unique to the human liver tissue (Figure 5D and Supplementary Figure S6B). Therefore, differential transcription factor binding appears to be a central feature of species-specific methylation states at sites with intermediate CpG and GC content on human chromosome 21. Given the clear overlap between transcription factor binding and ssHMRs in human tissue, we wondered whether Tc1 mouse-specific transcription factor binding events may also be implicated in the methylation state of some ssHMRs in mouse. When we examined Tc1-specific HMRs we again observed that nearly a third (47/169) of these sites overlapped with a Tc1-unique transcription factor binding event (Figure 5F). For example, downstream of the human TRPM2 gene an HMR is hypomethylated in the Tc1 mouse and bound by both HNF4A and CTCF. In human these binding events are absent and the region is methylated (Figure 5E and Supplementary Figure S6C). Strikingly, the methylation state of approximately one-third of all species-specific chromosome 21 HMRs appear to be related to the binding profiles of just three transcription factors (Figure 5D and F). This suggests that if more transcription factors were examined, the methylation state of a large proportion of species-specific chromosome 21 HMRs might correspond to these events. Therefore together, our in vivo interspecies experiments demonstrate the breadth with which DNA binding transcription factor occupancy can shape DNA methylation profiles at sites in the genome that have low CpG density and GC content, an observation that is further supported by recent in vitro experiments that also suggest DNA binding events contribute to local methylation profiles (20,22,23). The principles shaping HMR formation are DNA encoded and conserved across vast expanses of divergent vertebrate evolution Although it is clear in the Tc1 mouse that a subset of sites on human chromosome 21 can exist in alternative methylation states depending on the host species, the majority of HMRs throughout the human chromosome form completely normally (Figure 2C–E). This suggests that the general mechanisms specifying DNA methylation patterns and protecting CGIs from methylation are for the most part intact and conserved across 75 million years of evolutionary divergence between human and mouse. We recently demonstrated that features of a CGI-like system exist in most, if not all, branches of vertebrate evolution (8). If DNA sequence is indeed the central driver in HMR formation and the mechanisms driving these methylation states are functionally conserved in divergent vertebrate species then one would predict that the transplantation of mammalian DNA sequences into even more distantly related vertebrates would also result in the accurate specification of HMRs during animal development. To test this hypothesis, we introduced large chromosomal fragments of mouse DNA into the zebrafish genome and used BioCAP-seq to examine how the resulting methylation profiles compared to the profiles observed in mouse tissues. To achieve this, four mouse bacterial artificial chromosomes (mBACs) each containing approximately 200 kb of genomic sequence were introduced into the genome of the fertilized zebrafish zygote by Tol2-mediated transposition (29). When we examined the resulting methylation profiles of the mouse DNA sequences in 28–30 hours post fertilisation (hpf) zebrafish embryos, we observed that promoter-associated mouse HMRs were appropriately specified, including both typical 1–2 kb HMRs (Figure 6A and B) and broad HMRs, which tend to be associated with genes important for development (Figure 6C) and that the BioCAP-seq signal at HMRs on the mouse DNA in zebrafish correlated linearly with the BioCAP-seq signal in mouse tissues (Figure 6F–H). Together this indicates that the principles driving protection of these elements from DNA methylation are DNA encoded and conserved across vertebrate evolution. Figure 6. The DNA encoded principles that underpin HMR formation are conserved across vast expanses of divergent evolution. (A) BioCAP-seq profile of a mouse chromosomal DNA fragment introduced into the zebrafish genome and analysed at 28–30 h post-fertilisation (hpf). The BioCAP signal from three representative mouse cell-types ES cells, liver and testis (green, red and blue traces) and the BioCAP trace observed for this locus in the developing zebrafish (grey) is indicated. CpG density and GC content are depicted in black. All four mouse HMRs form on the mouse BAC DNA in the zebrafish embryo and a cluster of repetitive LTR elements in the centre of the mouse BAC form zebrafish-specific HMRs. (B) A snapshot of a promoter-associated mouse HMR. (C) A snapshot of a broadly hypomethylated region. (D) A snapshot of a zebrafish-specific HMR region that forms at a CpG and GC-rich mouse exonic region that is normally methylated in mouse tissues. (E) A snapshot illustrating a cluster of mouse LTR elements which are CpG dense and become hypomethylated in zebrafish. (F–H) Scatterplots comparing BioCAP-seq read counts at mouse HMRs in zebrafish with mouse liver (F), testis (G) and embryonic stem cells (H). Interestingly, a small number of new HMRs were formed on the mouse BACs integrated into zebrafish, much like the species-specific HMRs that were observed in the Tc1 mouse. When we examined these sites in more detail they again corresponded to CpG and GC rich regions within the mBAC including an exonic region of the Zfp623 gene (Figure 6D). Importantly, several sites where zebrafish specific HMRs formed in the mBACs corresponded to small clusters of CpG rich LTR repeats, including intracisternal A particles (IAPs) and the ERV1 element MMVL30-int, that are restricted to the mouse lineage, and are absent from the zebrafish genome (Figure 6A and E, Supplementary Figure S7A). Many of the LTR retrotransposons present on the mBACs exhibit CpG density comparable to endogenous CpG islands (Supplementary Figure S7B) and, therefore, in a manner similar to young primate specific repeat sequences in the Tc1 mouse, zebrafish presumably lacks co-evolved mechanisms to drive DNA methylation to these mouse repeat regions. Consequently, the elevated CpG and GC content of these repeat sequences appears to render them refractory to methylation further supporting the argument that nucleotide content is a central and evolutionarily conserved DNA encoded feature shaping HMR formation. Therefore, by examining the principles that shape HMR formation in vivo across distinct vertebrate organisms that have undergone extensive divergent evolution, we find that nucleotide content, in particular CpG and GC richness, is a central feature shaping HMR formation. Furthermore, the fact that most mouse HMRs, when situated in their normal genomic context, formed accurately in zebrafish (Figure 6F–H) reveals quite remarkably that the mechanisms driving stereotypical HMR formation, particularly at gene promoters, are highly conserved, despite 450 million years divergent evolution. DISCUSSION The mechanisms that form and maintain epigenomes in vivo are very poorly understood. Recently, cell culture-based approaches inserting small DNA fragments into the genome have suggested some principles that may contribute to formation of DNA methylation states (22–24). However, these approaches do not examine how DNAs behave in their natural genomic context, are limited in the total amount of DNA sequence they interrogate, and do not capture the divergent developmental trajectories that ultimately shape chromosomal DNA methylation patterns in tissues. Therefore, to overcome these limitations we have exploited a transchromosomic mouse model and examined developmentally distinct tissues to discover that the majority of HMRs on human chromosome 21 are appropriately recapitulated when transplanted into the mouse nuclear environment, indicating that the underlying DNA sequence and genomic context is largely sufficient to drive the observed methylation patterns in vertebrate organisms. A subset of regions distal to classical promoter-associated CGIs showed species-specific methylation patterns. Interestingly, CpG and GC-rich regions associated with young repeat sequences became hypomethylated in the mouse nuclear environment, and also in experiments where mouse BACs were integrated into zebrafish, apparently in the absence of mechanisms that would normally drive methylation to these sites. This feature is uniquely revealed through our interspecies experiments and demonstrates that the mechanisms which evolved to recognize repeat elements and methylate them can override the inherent capacity of CpG and GC-rich sequences to evade DNA methylation, presumably as a requirement to protect the genome against the activity of these potentially deleterious elements. These protection mechanisms may rely on the activity of piwi RNA (piRNA) clusters that encode a historical memory of transposition events from invading retroviruses and recruit PIWI proteins to endogenous transposons resulting in their methylation (56,57). Alternatively, a class of KRAB zinc finger DNA binding proteins appears to rapidly evolve the capacity to recognize emerging classes of repetitive DNA sequence and similarly targets them for DNA methylation (58,59). Importantly, in the absence of co-evolved mechanisms that target DNA methylation to these repetitive sequences, it appears that CpG and GC-richness renders these DNA sequences refractory to DNA methylation, much like classical HMRs, supporting the idea that the cell uses CpG and GC richness as a cue to protect against DNA methylation. In contrast to CpG and GC-rich regions, the methylation state of DNA sequences with intermediate CpG and GC content are frequently defined by species-specific DNA binding events, providing new widespread experimental evidence that site-specific DNA binding factor occupancy, in addition to DNA sequence features, can play a central role in shaping DNA methylation patterns at these sites. Together these observations reveal that CpG and GC nucleotide features and protein-based DNA binding events defined by DNA sequence are central determinants in shaping methylation patterns on chromosomal DNA sequences in vivo during development. Furthermore, by transplantation of mouse chromosome fragments into zebrafish we discovered that these general principles are conserved across vast expanses of divergent vertebrate evolution, uncovering the existence of a highly conserved and DNA encoded logic that shapes methylation patterns in the vertebrate epigenome. Our observation that CpG and GC nucleotide content is a central determinant in protecting chromosomal DNA from DNA methylation is in agreement with observations that short synthetic or bacterial CpG and GC-rich DNA sequences can often evade the DNA methylation machinery when inserted into vertebrate genomes (12,21,22). This can in many instances be achieved even if sequences apparently lack motifs necessary for site-specific occupancy of DNA binding transcription factors, suggesting that alternative mechanisms must exist to sense and protect these CpG and GC-rich regions from DNA methylation. This could be achieved by evolutionarily conserved ZF-CxxC domain containing proteins that bind specifically to non-methylated CpG dinucleotides (10,60) and associate with chromatin modifying activities that create chromatin environments that are refractory to DNA methylation. For example, H3K4me3 is targeted to CGIs by several ZF-CxxC proteins (10,12) and can inhibit DNA methyltransferase function on chromatin in vivo (61–63). Furthermore, the KDM2B ZF-CxxC domain-containing protein functions as an H3K36me1/me2 demethylase and is part of the polycomb repressive complex 1 that appears to protect a subset of CGIs from DNA methylation (11,64–68). This occlusion of the DNA methylation machinery is likely reinforced by mechanisms that can actively remove DNA methylation, such as the TET oxygenases that are putative DNA demethylase enzymes and occupy CGIs via their ZF-CxxC domains (69–72). Together these observations suggest that chromatin-modifying activities targeted to CGIs by ZF-CxxC domain containing proteins may play a central and evolutionarily conserved role in protecting CGIs from DNA methylation and shaping DNA methylation landscapes. Furthermore, our demonstration that the mechanisms underpinning specification of HMRs are mechanistically conserved in zebrafish also suggests this tractable developmental model system could be exploited through morpholino or CRISPR based approaches to remove ZF-CxxC proteins or other contributing mechanisms, individually or in combination, to further dissect the mechanisms that shape HMR formation during development. By examining methylation states across vast expanses of the same DNA sequence in two completely distinct vertebrate nuclear environments we provide extensive new evidence that transcription factor binding events are central determinants in shaping DNA methylation profiles at intermediate CpG and GC content regions of the genome. This parallels observations that dynamic and differential methylation is often observed at CpG and GC-poor enhancer elements across tissues of individual organisms (8,14–17,54), a feature that in some cases has been attributed to DNA binding transcription factors occupancy (55,73,74). Importantly, our new trans-species observations provide chromosome-scale experimental evidence that transcription factor occupancy can help to shape DNA methylation patterns. This is supportive of other single gene studies that suggested that SP1 transcription factor occupancy could contribute to the DNA methylation state of the APRT1 gene (26,27), more limited mutational analysis of transcription factor binding sites (20,22,23), and the suggestion that single nucleotide variation associated with human disease affects the capacity of nuclear factors to read the underlying DNA sequence and potentially contributes to alterations in the epigenome (75,76). An important step moving forward will be to determine mechanistically how DNA binding factors influence DNA methylation. It appears unlikely that this will simply result from transcription factors protecting the underlying DNA from the methylation machinery, as HNF4A lacks CpG dinucleotides in its recognition sequence yet influences the methylation of surrounding CpG dinucleotides (73). One interesting possibility is that transcription factors could exploit active mechanisms, perhaps through the function of the TET DNA demethylases, to protect distal gene regulatory sites from the transcriptionally repressive influences of DNA methylation. In support of this possibility a number of recent studies have provided preliminary evidence that this may indeed be the case (55,77). In summary, through exploiting transchromosomic animal experiments we discover that DNA methylation patterns in vivo are primarily informed and shaped by DNA sequence. In doing so we demonstrate that classical CGI like sequences in large genomic regions of DNA can be interpreted by evolutionarily conserved mechanisms, even in distantly related vertebrate organisms, to protect these sequences from DNA methylation and to shape the epigenome during development. ACCESSION NUMBERS Datasets have been deposited in the Gene Expression Omnibus (GEO) with the accession number GSE72208. Supplementary Material SUPPLEMENTARY DATA With thanks to Ricarda Gaentzsch for advice regarding multiplex bisulfite sequencing. We are grateful to Claudia Kutter for providing Tc1 mouse tissue, Michelle Ward for providing analysed data files and to Michelle Ward, Christina Ernst, Sarah Aitken, and Neil Blackledge for fruitful discussion. Present address: Hannah K. Long, Department of Chemical and Systems Biology and Institute for Stem Cell Biology & Regenerative Medicine, Stanford University School of Medicine, Stanford University, Stanford, CA 94305, USA. SUPPLEMENTARY DATA Supplementary Data are available at NAR Online. FUNDING Wellcome Trust [098024/Z/11/Z to R.J.K.]; Lister Institute of Preventive Medicine; Exeter College. Funding for open access charge: Wellcome Trust [098024/Z/11/Z]. Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Jones P.A. Functions of DNA methylation: islands, start sites, gene bodies and beyond Nat. Rev. Genet. 2012 13 484 492 22641018 2. Klose R.J. Bird A.P. Genomic DNA methylation: the mark and its mediators Trends Biochem. Sci. 2006 31 89 97 16403636 3. Sharif J. Muto M. Takebayashi S. Suetake I. Iwamatsu A. Endo T.A. Shinga J. Mizutani-Koseki Y. Toyoda T. Okamura K. The SRA protein Np95 mediates epigenetic inheritance by recruiting Dnmt1 to methylated DNA Nature 2007 450 908 912 17994007 4. Bird A. DNA methylation patterns and epigenetic memory Genes Dev. 2002 16 6 21 11782440 5. Jones P.A. Liang G. Rethinking how DNA methylation patterns are maintained Nat. Rev. Genet. 2009 10 805 811 19789556 6. Cooper D.N. Taggart M.H. Bird A.P. Unmethlated domains in vertebrate DNA Nucleic Acids Res. 1983 11 647 658 6188105 7. Bird A. Taggart M. Frommer M. Miller O.J. Macleod D. A fraction of the mouse genome that is derived from islands of nonmethylated, CpG-rich DNA Cell 1985 40 91 99 2981636 8. Long H.K. Sims D. Heger A. Blackledge N.P. Kutter C. Wright M.L. Grutzner F. Odom D.T. Patient R. Ponting C.P. Epigenetic conservation at gene regulatory elements revealed by non-methylated DNA profiling in seven vertebrates eLife 2013 2 e00348 23467541 9. Antequera F. Structure, function and evolution of CpG island promoters Cell. Mol. Life Sci.: CMLS 2003 60 1647 1658 14504655 10. Long H.K. Blackledge N.P. Klose R.J. ZF-CxxC domain-containing proteins, CpG islands and the chromatin connection Biochem. Soc. Trans. 2013 41 727 740 23697932 11. Blackledge N.P. Zhou J.C. Tolstorukov M.Y. Farcas A.M. Park P.J. Klose R.J. CpG islands recruit a histone H3 lysine 36 demethylase Mol. cell 2010 38 179 190 20417597 12. Thomson J.P. Skene P.J. Selfridge J. Clouaire T. Guy J. Webb S. Kerr A.R. Deaton A. Andrews R. James K.D. CpG islands influence chromatin structure via the CpG-binding protein Cfp1 Nature 2010 464 1082 1086 20393567 13. Illingworth R.S. Gruenewald-Schneider U. Webb S. Kerr A.R. James K.D. Turner D.J. Smith C. Harrison D.J. Andrews R. Bird A.P. Orphan CpG islands identify numerous conserved promoters in the mammalian genome PLoS Genet. 2010 6 e1001134 20885785 14. Weber M. Hellmann I. Stadler M.B. Ramos L. Paabo S. Rebhan M. Schubeler D. Distribution, silencing potential and evolutionary impact of promoter DNA methylation in the human genome Nat. Genet. 2007 39 457 466 17334365 15. Straussman R. Nejman D. Roberts D. Steinfeld I. Blum B. Benvenisty N. Simon I. Yakhini Z. Cedar H. Developmental programming of CpG island methylation profiles in the human genome Nat. Struct. Mol. Biol. 2009 16 564 571 19377480 16. Hodges E. Molaro A. Dos Santos C.O. Thekkat P. Song Q. Uren P.J. Park J. Butler J. Rafii S. McCombie W.R. Directional DNA methylation changes and complex intermediate states accompany lineage specificity in the adult hematopoietic compartment Mol. Cell 2011 44 17 28 21924933 17. Mohn F. Weber M. Rebhan M. Roloff T.C. Richter J. Stadler M.B. Bibel M. Schubeler D. Lineage-specific polycomb targets and de novo DNA methylation define restriction and potential of neuronal progenitors Mol. Cell 2008 30 755 766 18514006 18. Xie W. Schultz M.D. Lister R. Hou Z. Rajagopal N. Ray P. Whitaker J.W. Tian S. Hawkins R.D. Leung D. Epigenomic analysis of multilineage differentiation of human embryonic stem cells Cell 2013 153 1134 1148 23664764 19. Farthing C.R. Ficz G. Ng R.K. Chan C.F. Andrews S. Dean W. Hemberger M. Reik W. Global mapping of DNA methylation in mouse promoters reveals epigenetic reprogramming of pluripotency genes PLoS Genet. 2008 4 e1000116 18584034 20. Stadler M.B. Murr R. Burger L. Ivanek R. Lienert F. Schöler A. Wirbelauer C. Oakeley E.J. Gaidatzis D. Tiwari V.K. DNA-binding factors shape the mouse methylome at distal regulatory regions Nature 2011 480 490 495 22170606 21. Mendenhall E.M. Koche R.P. Truong T. Zhou V.W. Issac B. Chi A.S. Ku M. Bernstein B.E. GC-Rich sequence elements recruit PRC2 in mammalian ES cells PLoS Genet. 2010 6 e1001244 21170310 22. Lienert F. Wirbelauer C. Som I. Dean A. Mohn F. Schubeler D. Identification of genetic elements that autonomously determine DNA methylation states Nat. Genet. 2011 43 1091 1097 21964573 23. Krebs A.R. Dessus-Babus S. Burger L. Schubeler D. High-throughput engineering of a mammalian genome reveals building principles of methylation states at CG rich regions eLife 2014 3 e04094 25259795 24. Wachter E. Quante T. Merusi C. Arczewska A. Stewart F. Webb S. Bird A. Synthetic CpG islands reveal DNA sequence determinants of chromatin structure eLife 2014 3 e03397 25259796 25. Dickson J. Gowher H. Strogantsev R. Gaszner M. Hair A. Felsenfeld G. West A.G. VEZF1 elements mediate protection from DNA methylation PLoS Genet. 2010 6 e1000804 20062523 26. Brandeis M. Frank D. Keshet I. Siegfried Z. Mendelsohn M. Nemes A. Temper V. Razin A. Cedar H. Sp1 elements protect a CpG island from de novo methylation Nature 1994 371 435 438 8090226 27. Macleod D. Charlton J. Mullins J. Bird A.P. Sp1 sites in the mouse aprt gene promoter are required to prevent methylation of the CpG island Genes Dev. 1994 8 2282 2292 7958895 28. O'Doherty A. Ruf S. Mulligan C. Hildreth V. Errington M.L. Cooke S. Sesay A. Modino S. Vanes L. Hernandez D. An aneuploid mouse strain carrying human chromosome 21 with Down syndrome phenotypes Science 2005 309 2033 2037 16179473 29. Suster M.L. Abe G. Schouw A. Kawakami K. Transposon-mediated BAC transgenesis in zebrafish Nat. Protoc. 2011 6 1998 2021 22134125 30. Blackledge N.P. Long H.K. Zhou J.C. Kriaucionis S. Patient R. Klose R.J. Bio-CAP: a versatile and highly sensitive technique to purify and characterise regions of non-methylated DNA Nucleic Acids Res. 2012 40 e32 22156374 31. Langmead B. Trapnell C. Pop M. Salzberg S. Ultrafast and memory-efficient alignment of short DNA sequences to the human genome Genome Biol. 2009 10 R25 19261174 32. Zhang Y. Liu T. Meyer C. Eeckhoute J. Johnson D. Bernstein B. Nusbaum C. Myers R. Brown M. Li W. Model-based Analysis of ChIP-Seq (MACS) Genome Biol. 2008 9 R137 18798982 33. Gribble S.M. Wiseman F.K. Clayton S. Prigmore E. Langley E. Yang F. Maguire S. Fu B. Rajan D. Sheppard O. Massively parallel sequencing reveals the complex structure of an irradiated human chromosome on a mouse background in the Tc1 model of Down syndrome PLoS One 2013 8 e60482 23596509 34. Langmead B. Salzberg S.L. Fast gapped-read alignment with Bowtie 2 Nat. Methods 2012 9 357 359 22388286 35. Kumar S. Subramanian S. Mutation rates in mammalian genomes Proc. Natl. Acad. Sci. U.S.A. 2002 99 803 808 11792858 36. Ward Michelle C. Wilson Michael D. Barbosa-Morais Nuno L. Schmidt D. Stark R. Pan Q. Schwalie Petra C. Menon S. Lukk M. Watt S. Latent regulatory potential of human-specific repetitive elements Mol. Cell 2013 49 262 272 23246434 37. Krueger F. Andrews S.R. Bismark: a flexible aligner and methylation caller for Bisulfite-Seq applications Bioinformatics 2011 27 1571 1572 21493656 38. Illingworth R. Kerr A. Desousa D. Jorgensen H. Ellis P. Stalker J. Jackson D. Clee C. Plumb R. Rogers J. A novel CpG island set identifies tissue-specific methylation at developmental gene loci PLoS Biol. 2008 6 e22 18232738 39. Schmidt D. Wilson M.D. Ballester B. Schwalie P.C. Brown G.D. Marshall A. Kutter C. Watt S. Martinez-Jimenez C.P. Mackay S. Five-vertebrate ChIP-seq reveals the evolutionary dynamics of transcription factor binding Science 2010 328 1036 1040 20378774 40. Ke M. Raju S. Comparative histology of human and cow, goat and sheep liver J. Surg. Acad. 2014 2 10 13 41. Baratta J.L. Ngo A. Lopez B. Kasabwalla N. Longmuir K.J. Robertson R.T. Cellular organization of normal mouse liver: a histological, quantitative immunocytochemical, and fine structural analysis Histochem. Cell Biol. 2009 131 713 726 19255771 42. Wu X. Schmidt J.A. Avarbock M.R. Tobias J.W. Carlson C.A. Kolon T.F. Ginsberg J.P. Brinster R.L. Prepubertal human spermatogonia and mouse gonocytes share conserved gene expression of germline stem cell regulatory molecules Proce. Natl. Acad. Sci. U.S.A. 2009 106 21672 21677 43. Sha J. Zhou Z. Li J. Yin L. Yang H. Hu G. Luo M. Chan H.C. Group S.S. Zhou K. Identification of testis development and spermatogenesis-related genes in human and mouse testes using cDNA arrays Mol. Hum. Reprod. 2002 8 511 517 12029067 44. Anand M. Prasad B.V. The computational analysis of human testis transcriptome reveals closer ties to pluripotency J. Hum. Reprod. Sci. 2012 5 266 273 23531778 45. Milne T.A. Dou Y. Martin M.E. Brock H.W. Roeder R.G. Hess J.L. MLL associates specifically with a subset of transcriptionally active target genes Proc. Natl. Acad. Sci. U.S.A. 2005 102 14765 14770 16199523 46. Lander E.S. Linton L.M. Birren B. Nusbaum C. Zody M.C. Baldwin J. Devon K. Dewar K. Doyle M. FitzHugh W. Initial sequencing and analysis of the human genome Nature 2001 409 860 921 11237011 47. Hancks D.C. Kazazian H.H. Jr Active human retrotransposons: variation and disease Curr. Opin. Genet. Dev. 2012 22 191 203 22406018 48. Deininger P. Alu elements: know the SINEs Genome Biol. 2011 12 236 22204421 49. Coulondre C. Miller J.H. Farabaugh P.J. Gilbert W. Molecular basis of base substitution hotspots in Escherichia coli Nature 1978 274 775 780 355893 50. Shen J.C. Rideout W.M. 3rd Jones P.A. The rate of hydrolytic deamination of 5-methylcytosine in double-stranded DNA Nucleic Acids Res. 1994 22 972 976 8152929 51. Takai D. Jones P.A. Comprehensive analysis of CpG islands in human chromosomes 21 and 22 Proc. Natl. Acad. Sci. U.S.A. 2002 99 3740 3745 11891299 52. Fazzari M.J. Greally J.M. Epigenomics: beyond CpG islands Nat. Rev. Genet. 2004 5 446 455 15153997 53. Illingworth R.S. Bird A.P. CpG islands–‘a rough guide’ FEBS Lett. 2009 583 1713 1720 19376112 54. Lister R. Pelizzola M. Dowen R.H. Hawkins R.D. Hon G. Tonti-Filippini J. Nery J.R. Lee L. Ye Z. Ngo Q.M. Human DNA methylomes at base resolution show widespread epigenomic differences Nature 2009 462 315 322 19829295 55. Feldmann A. Ivanek R. Murr R. Gaidatzis D. Burger L. Schübeler D. Transcription factor occupancy can mediate active turnover of DNA methylation at regulatory regions PLoS Genet. 2013 9 e1003994 24367273 56. Aravin A.A. Bourc'his D. Small RNA guides for de novo DNA methylation in mammalian germ cells Genes Dev. 2008 22 970 975 18413711 57. Molaro A. Falciatori I. Hodges E. Aravin A.A. Marran K. Rafii S. McCombie W.R. Smith A.D. Hannon G.J. Two waves of de novo methylation during mouse germ cell development Genes Dev. 2014 28 1544 1549 25030694 58. Thomas J.H. Schneider S.E. Coevolution of retroelements and tandem zinc finger genes Genome Res. 2011 21 1800 1812 21784874 59. Rowe H.M. Friedli M. Offner S. Verp S. Mesnard D. Marquis J. Aktas T. Trono D. De novo DNA methylation of endogenous retroviruses is shaped by KRAB-ZFPs/KAP1 and ESET Development 2013 140 519 529 23293284 60. Shin Voo K. Carlone D.L. Jacobsen B.M. Flodin A. Skalnik D.G. Cloning of a mammalian transcriptional activator that binds unmethylated CpG motifs and shares a CXXC domain with DNA methyltransferase, human trithorax, and methyl-CpG binding domain protein 1 Mol. Cell. Biol. 2000 20 2108 2121 10688657 61. Ooi S.K. Qiu C. Bernstein E. Li K. Jia D. Yang Z. Erdjument-Bromage H. Tempst P. Lin S.P. Allis C.D. DNMT3L connects unmethylated lysine 4 of histone H3 to de novo methylation of DNA Nature 2007 448 714 717 17687327 62. Morselli M. Pastor W.A. Montanini B. Nee K. Ferrari R. Fu K. Bonora G. Rubbi L. Clark A.T. Ottonello S. In vivo targeting of de novo DNA methylation by histone modifications in yeast and mouse eLife 2015 4 e06205 25848745 63. Hu J.-L. Zhou B.O. Zhang R.-R. Zhang K.-L. Zhou J.-Q. Xu G.-L. The N-terminus of histone H3 is required for de novo DNA methylation in chromatin Proc. Natl. Acad. Sci. U.S.A. 2009 106 22187 22192 20018712 64. Farcas A.M. Blackledge N.P. Sudbery I. Long H.K. McGouran J.F. Rose N.R. Lee S. Sims D. Cerase A. Sheahan T.W. KDM2B links the Polycomb Repressive Complex 1 (PRC1) to recognition of CpG islands eLife 2012 1 e00205 23256043 65. Blackledge N.P. Farcas A.M. Kondo T. King H.W. McGouran J.F. Hanssen L.L. Ito S. Cooper S. Kondo K. Koseki Y. Variant PRC1 complex-dependent H2A ubiquitylation drives PRC2 recruitment and polycomb domain formation Cell 2014 157 1445 1459 24856970 66. Boulard M. Edwards J.R. Bestor T.H. FBXL10 protects Polycomb-bound genes from hypermethylation Nat. Genet. 2015 47 479 485 25848754 67. He J. Shen L. Wan M. Taranova O. Wu H. Zhang Y. Kdm2b maintains murine embryonic stem cell status by recruiting PRC1 complex to CpG islands of developmental genes Nat. Cell Biol. 2013 15 373 384 23502314 68. Wu X. Johansen J.V. Helin K. Fbxl10/Kdm2b recruits polycomb repressive complex 1 to CpG islands and regulates H2A ubiquitylation Mol. Cell 2013 49 1134 1146 23395003 69. Williams K. Christensen J. Pedersen M.T. Johansen J.V. Cloos P.A.C. Rappsilber J. Helin K. TET1 and hydroxymethylcytosine in transcription and DNA methylation fidelity Nature 2011 473 343 348 21490601 70. Wu H. D'Alessio A.C. Ito S. Xia K. Wang Z. Cui K. Zhao K. Eve Sun Y. Zhang Y. Dual functions of Tet1 in transcriptional regulation in mouse embryonic stem cells Nature 2011 473 389 393 21451524 71. Xu Y. Wu F. Tan L. Kong L. Xiong L. Deng J. Barbera A.J. Zheng L. Zhang H. Huang S. Genome-wide regulation of 5hmC, 5mC, and gene expression by Tet1 hydroxylase in mouse embryonic stem cells Mol. Cell 2011 42 451 464 21514197 72. Branco M.R. Ficz G. Reik W. Uncovering the role of 5-hydroxymethylcytosine in the epigenome Nat. Rev. Genet. 2012 13 7 13 22083101 73. Sheaffer K.L. Kim R. Aoki R. Elliott E.N. Schug J. Burger L. Schübeler D. Kaestner K.H. DNA methylation is required for the control of stem cell differentiation in the small intestine Genes Dev. 2014 28 652 664 24637118 74. Taberlay P.C. Statham A.L. Kelly T.K. Clark S.J. Jones P.A. Reconfiguration of nucleosome-depleted regions at distal regulatory elements accompanies DNA methylation of enhancers and insulators in cancer Genome Res. 2014 24 1421 1432 24916973 75. Ziller M.J. Gu H. Muller F. Donaghey J. Tsai L.T.Y. Kohlbacher O. De Jager P.L. Rosen E.D. Bennett D.A. Bernstein B.E. Charting a dynamic DNA methylation landscape of the human genome Nature 2013 500 477 481 23925113 76. Schaub M.A. Boyle A.P. Kundaje A. Batzoglou S. Snyder M. Linking disease associations with regulatory information in the human genome Genome Res. 2012 22 1748 1759 22955986 77. Rasmussen K.D. Jia G. Johansen J.V. Pedersen M.T. Rapin N. Bagger F.O. Porse B.T. Bernard O.A. Christensen J. Helin K. Loss of TET2 in hematopoietic cells leads to DNA hypermethylation of active enhancers and induction of leukemogenesis Genes Dev. 2015 29 910 922 25886910

PMC005xxxxxx/PMC5001594.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2718588610.1093/nar/gkw422Genome Integrity, Repair and ReplicationBromodeoxyuridine does not contribute to sister chromatid exchange events in normal or Bloom syndrome cells van Wietmarschen Niek 1Lansdorp Peter M. 123*1 European Research Institute for the Biology of Ageing, University Medical Center Groningen, University of Groningen, 9713 AV Groningen, The Netherlands2 Terry Fox Laboratory, BC Cancer Agency, Vancouver, BC V5Z 1L3, Canada3 Division of Hematology, Department of Medicine, University of British Columbia, Vancouver, BC V6T 1Z4, Canada* To whom correspondence should be addressed. Tel: +31 50 361 7300; Fax: +31 50 361 7300; Email: p.m.lansdorp@umcg.nl19 8 2016 16 5 2016 16 5 2016 44 14 6787 6793 05 5 2016 30 4 2016 30 3 2016 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Sister chromatid exchanges (SCEs) are considered sensitive indicators of genome instability. Detection of SCEs typically requires cells to incorporate bromodeoxyuridine (BrdU) during two rounds of DNA synthesis. Previous studies have suggested that SCEs are induced by DNA replication over BrdU-substituted DNA and that BrdU incorporation alone could be responsible for the high number of SCE events observed in cells from patients with Bloom syndrome (BS), a rare genetic disorder characterized by marked genome instability and high SCE frequency. Here we show using Strand-seq, a single cell DNA template strand sequencing technique, that the presence of variable BrdU concentrations in the cell culture medium and in DNA template strands has no effect on SCE frequency in either normal or BS cells. We conclude that BrdU does not induce SCEs and that SCEs detected in either normal or BS cells reflect DNA repair events that occur spontaneously. cover-date19 August 2016 ==== Body INTRODUCTION Sister chromatid exchanges (SCEs) are a possible outcome of DNA double-strand breaks (DSBs) that were repaired through homologous recombination. As such, SCEs are considered sensitive indicators for genome instability in cells (1). This is evidenced by increased SCE rates observed in cells treated with mitomycin C (2), exposed to X-rays (3) or ionizing radiation (4). SCEs are classically detected cytogenetically by means of differential staining of sister chromatids in metaphase spreads from cells cultured with BrdU for two cell cycles. Visualization of chromosomes by staining with Hoechst or Giemsa allows for differentiation between sister chromatids for which either one or both DNA strands are labeled with BrdU (2,5,6). However, it has been widely reported that SCEs are induced by culturing cells in the presence of BrdU, raising questions about if and how many SCEs occur naturally during the cell cycle (7–11). This holds especially true for cells derived from Bloom syndrome (BS) patients. BS is a rare recessive genetic disorder caused by mutations in the BLM gene (12). This gene encodes for the BLM protein, which is a member of the RecQ family of helicases and plays an important role in preventing SCE formation during repair of DSBs (13). Cells from BS patients display marked genome instability, evidenced by the high SCE rates (12), as well as delayed speed of DNA replication and cell division (14), elevated mutation rates (15) and disrupted nuclear architecture (16). In the case of BS, it has been reported that the characteristic high SCE rates only occur during the second DNA replication, when BrdU labeled DNA is used as a template for DNA replication (17,18). Conflicting reports state that the high SCE rates in BS cells do occur spontaneously (19,20). We recently described a new technique for SCE detection, called Strand-seq (21). This technique relies on single-cell DNA template strand sequencing to detect chromosomal rearrangements, including SCEs. One of the major strengths of Strand-seq is that it only requires one round of cell division in the presence of BrdU, thus eliminating any effect of DNA replication using BrdU-labeled templates. Furthermore, Strand-seq allows mapping of SCEs at kilobase resolution or higher, which is several orders of magnitude better than detection by cytogenetics. We used Strand-seq to study SCE rates in both normal and BS cells to elucidate the effect of BrdU. We show that the concentration of BrdU used during cell culture has no effect on SCE rates and that SCE rates also do not increase when BrdU is present in DNA template strands. We also show that BS cells do display spontaneously elevated SCE rates that are not affected by the presence BrdU in cell culture medium or in DNA template strands. These results substantiate that SCEs play a biological role in cells and are not artefacts induced by the method used to detect them. MATERIALS AND METHODS Cell culture The following cell lines were obtained from the Corriell Cell Repository: GM07492 (primary fibroblasts, normal), GM03402 (primary fibroblasts, BS), GM12891 (EBV-transformed lymphoblasts, normal) and GM16375 (EBV-transformed lymphoblasts, BS). Fibroblasts were cultured in Dulbecco's modified Eagle's medium (Life Technologies) supplemented with 15% v/v fetal bovine serum (Sigma Aldrich) and 1% v/v penicillin-streptomycin (Life Technologies), lymphoblasts in RPMI1640 (Life Technologies) supplemented with 15% v/v FBS and 1% v/v penicillin-streptomycin. All cells were cultured at 37°C in 5% CO2. BrdU (Invitrogen) was added to cultures at indicated concentrations for indicated periods of time. Flow cytometry Cells were harvested after the BrdU pulse, and nuclei were isolated by suspending cells in nuclei isolation buffer (100 mM Tris–HCl pH7.4, 150 mM NaCl, 1 mM CaCl2, 0.5 mM MgCl2, 0.1% NP-40 and 2% bovine serum albumin). Nuclei were stained with Hoechst 33258 (Life Technologies) and propidium iodide (Sigma Aldrich) at final concentrations of 10 μg/ml. Nuclei were analyzed and sorted based on low PI (G1 phase) and low Hoechst (BrdU-induced quenching, see also Figure 2E) fluorescence on a MoFlo Astrios cell sorter (Beckman Coulter) or a FACSJazz cell sorter (BD Biosciences) directly into 5 μl Pro-Freeze-CDM NAO freeze medium (Lonza) + 7.5% DMSO, in 96-well skirted polymerase chain reaction (PCR) plates (4Titude). Sorted nuclei were stored at −80°C. Library construction Library preparation was performed using the protocols previously described (21) with the following modifications. Enzymatic reactions were performed in smaller volumes but with the same enzyme and buffer concentrations. All clean-ups were performed using AMPure XP magnetic beads (Agencourt AMPure, Beckman Coulter). After adapter ligation and PCR, clean-ups with magnetic beads were performed twice using a 1.2 volume of beads. All pipetting was performed using the Bravo Automated Liquid Handling Platform (Agilent). Illumina sequencing Libraries were pooled for sequencing and 250- to 450-bp size range fragments were purified using a 2% E-Gel Agarose Gel (Invitrogen). DNA quality was assessed and quantified on a High Sensitivity dsDNA kit (Agilent) on the Agilent 2100 Bio-Analyzer and on the Qubit 2.0 Fluorometer (Life Technologies). For sequencing, clusters were generated on the CBot (HiSeq2500) and single-end 50 bp reads were generated were generated using the HiSeq2500 sequencing platform (Illumina). Bioinformatic analysis Indexed bam files were aligned using bowtie2 (22) and further analyzed using the BAIT software package (23). SCEs were detected by BAIT and their presence was confirmed manually. RESULTS Bloom Syndrome cells display high SCE rates during first division with BrdU In order to assess any potential effect of BrdU on SCE rates, we first established baseline SCE rates in both normal and BS cells. We used primary fibroblasts and EBV-transformed lymphoblasts from healthy donors and BS patients for our analysis. Logarithmically growing cells were pulsed with 40 μM for one cell division (18 h for fibroblasts, 24 h for lymphoblasts). Cells were harvested directly after the BrdU pulse, nuclei suspensions were made and single nuclei were sorted by means of flow cytometry. Nuclei from cells in G1 phase (low PI fluorescence) that had incorporated BrdU into their DNA (low Hoechst fluorescence) were sorted and Strand-seq libraries were made. SCEs were detected by the BAIT analysis software (23) and confirmed by visual inspection (see also Supplementary Figure S1). Two representative libraries made from a normal fibroblast (Figure 1A) and a BS fibroblast (Figure 1B) are depicted, displaying 5 and 40 SCEs, respectively. Average SCE rates were calculated across 67–80 Strand-seq libraries made from normal and BS fibroblasts (Figure 1C) and normal and BS lymphoblasts (Figure 1D). These results show that although there are minor differences between the absolute SCE rates detected in the fibroblasts and lymphoblasts, both BS cell lines display roughly a tenfold increase in SCE rates compared to their normal counterparts during the first cell division with BrdU. These results are in agreement with previously published studies, which also showed ∼10-fold increase in the rate of SCE in BS cells. Interestingly, the average SCE rates detected here correspond to roughly half the baseline SCE rates found after two cell divisions in similar lymphoblast (17,18) and fibroblast (20) cells. This suggests that SCE rates are stable over two subsequent cell divisions, as previously suggested (19). Figure 1. Strand-seq confirms high SCE rates in Bloom Syndrome (BS) cells. (A) Schematic diagram explaining the principle of Strand-seq. (I) A single cell with a single chromosome is represented. Homologs are shown in blue and pink shading. For each chromosome, the Watson (negative) strand is indicated in orange and the Crick (positive) strand in green. Cells are pulsed with BrdU for one cell cycle, causing newly synthesized strands during DNA replication to become labeled with BrdU (dashed lines), while the template strands remain unlabeled (solid lines). An SCE on one homolog (red arrow) will cause template strands to be exchanged between sister chromatids, leading each to have partial Watson and Crick template strands after cell division. (II) Single cells are sorted, DNA is isolated and a sequencing library is constructed. During library construction, all BrdU-labeled DNA is removed by treatment with Hoechst + UV radiation. DNA template strand fragments are amplified and sequenced and reads are mapped to the chromosome based on orientation. An SCE is detected as a switch from Watson to Crick, or vice versa, on one of the template strands (red arrow). (B and C) Representative Strand-seq ideograms made from a normal (B) and a BS fibroblast (C). Orange and blue lines correspond to reads aligning to the Watson and Crick strands, respectively. Chromosome numbers are indicated under each ideogram. SCEs are indicated by black arrowheads. (D and E) Average SCE rates for normal and BS fibroblasts (D, P = 5.3*10−47) and lymphoblasts (E, P = 2.8*10−51). P-values were calculated using t-test. BrdU concentration does not affect SCE rates in normal or Bloom Syndrome cells Next, we investigated whether incorporation of BrdU into newly formed DNA strands could induce SCEs. Both normal and BS cells were cultured with increasing concentrations of BrdU. First, the effect of BrdU on proliferation was determined by culturing cells in increasing concentrations of BrdU and tracking proliferation over the course of 7 days. Fibroblasts were seeded at 10% confluency, lymphoblasts at 200 000 cells/ml. The number of live cells was counted using the trypan blue exclusion method every 24 h or until cultures reached confluency. The results (Figure 2A–D) show that BrdU caused a concentration-dependent decrease of cell proliferation in all cell lines tested. Quite strikingly, the effect of BrdU is much stronger in lymphoblasts than in the fibroblast lines; both lymphoblast lines show hardly any proliferation at 200 μM, while both fibroblast lines still continue to divide at this dose. Finally, both BS cell lines displayed decreased cell growth compared to normal cells, independent of BrdU. In addition, BS cells do not appear to be hypersensitive to the presence of BrdU. Figure 2. BrdU does not affect SCE rates in normal or BS cells. (A–D) Relative growth curves for normal fibroblasts (A), BS fibroblasts (B), normal lymphoblasts (C) and BS lymphoblasts (D). Number of live cells for each time point was normalized to t0, each panel represents a single replicate experiment. (E) Example Hoechst and PI fluorescence of asynchronously growing BrdU-labeled, measured by flow cytometry. Cell cycle stages can be distinguished based on PI staining intensity, BrdU-labeled (G1’, S’, G2’) and unlabeled nuclei (G1, S, G2) based on Hoechst intensity. Relative Hoechst quenching is calculated at the decrease in fluorescence between G1 nuclei with (G1’) and without (G1) BrdU labeling. (F) Relative Hoechst quenching of nuclei from normal and BS lymphoblasts pulsed with different concentrations of BrdU. (G) Average SCE rates across Strand-seq libraries made from normal and BS fibroblasts pulsed with different concentrations of BrdU (Normal: n = 15–24, BS: n = 16–28). (H) Average SCE rates across Strand-seq libraries made from normal and BS lymphoblasts pulsed with different concentrations of BrdU (Normal: n = 10–22, BS: n = 18–25). BrdU is thought to affect SCE frequencies when it is present in the DNA, so therefore we attempted to confirm that higher doses of BrdU also lead to higher BrdU incorporation into the DNA during cell division. Because fluorescence of Hoechst bound to DNA decreases when BrdU is present in the DNA (24), we decided to measure Hoechst fluorescence by means of flow cytometry and use relative Hoechst quenching as an indicator of the level of BrdU incorporation into the DNA. Figure 2E shows an example of Hoechst and PI staining of asynchronous nuclei labeled pulsed with BrdU. Relative Hoechst quenching, calculated at the decrease in Hoechst fluorescence between BrdU-labeled and unlabeled cells, is a measure for the amount of BrdU incorporation into the DNA. In order to determine the effect of BrdU concentration, normal and BS lymphoblasts were pulsed with the same range of BrdU concentrations as above for 24 h, after which cells were harvested, nuclei were isolated and stained with PI and Hoechst and fluorescence was measured by flow cytometry. Hoechst quenching was calculated and the results (Figure 2F) show that there is a dose-dependent effect of BrdU concentration on Hoechst quenching in both normal and BS cells. In order to determine if BrdU does indeed induce SCEs, all four cell lines were pulsed with 10–200 μM BrdU for one cell division and Strand-seq was performed to assess SCE rates under the different conditions. The results confirmed the high SCE rates in BS cells at each BrdU concentration, but we did not detect any significant effect of BrdU concentration on SCE rates in any of the cell lines (Figure 2G and H). Presence of BrdU in DNA template strands does not induce SCEs in normal or BS cells It has previously been proposed that DNA replication over BrdU-substituted DNA induces SCEs (10). Although we show that the presence of BrdU in cell culture medium does not affect SCE rates, we cannot exclude any effect of BrdU in DNA template strands based on these results. Fortunately, Strand-seq can also be used to detect SCEs after two cell divisions in BrdU. However, it is not possible to distinguish between SCEs that occurred during the first or the second cell cycle. In order to properly assess SCE rates during the second cell cycle, fibroblasts and lymphoblasts were pulsed with 40 μM BrdU for 36 or 48 h respectively, and single nuclei that underwent either 1 or 2 cell divisions were sorted. Strand-seq was performed on these nuclei after which first and second division libraries could be distinguished based on strand inheritance patterns. SCE rates in first division libraries were doubled to simulate a second cell division without induction of extra SCEs, and these expected SCE rates were compared to observed SCE rates in second division libraries (Figure 3A–D). The observed SCE rates matched the expected SCE rates in each of the four cell lines, supporting the notion that BrdU in DNA template does not induce SCEs in either normal or BS cells at concentrations typically used for SCE detection. Figure 3. No increase in SCE rates during second cell division. (A–D) Average SCE rates after one and two cell divisions in BrdU as expected and observed for (A) normal fibroblasts, (B) BS fibroblasts, (C) normal lymphoblasts (D) and BS lymphoblasts. P-values were calculated using t-test. DISCUSSION The role BrdU plays in SCE formation has been debated for several decades, but has not been resolved convincingly. Here we use Strand-seq to show that BrdU does not induce SCEs in either normal or BS cells and that Strand-seq can be used to detect spontaneously occurring SCEs in cultured cells. SCE detection by means of staining of metaphase spreads requires two rounds of DNA replication in the presence of BrdU. Because of this, the number of SCEs that occurred during a first cell cycle had to be estimated based on complex staining patterns in second or third cell cycle metaphase spreads. Using Strand-seq, it is possible, for the first time, to directly detect SCEs after one cell division. Our results indicate firstly that SCE frequencies are elevated in both BS fibroblasts and lymphoblasts compared to their normal counterparts. Secondly, the concentration of BrdU used during cell culture has no effect on SCE frequencies. Thirdly, SCE frequencies do not increase during a second cell cycle in the presence of BrdU in either normal or BS cells. Based on these results, we conclude that BS cells display spontaneously elevated SCE rates and that BrdU has no effect on SCE rates in either normal or BS cells. These results disagree with the generally accepted notion that BrdU does induce SCEs, either during incorporation into nascent DNA strands or during DNA replication over BrdU-substituted DNA (7–11). Although cytogenetic SCE detection is possible at lower BrdU concentrations than those required for Strand-seq, several studies did report an effect of BrdU at similar concentrations to the ones used here (8,9,19). What factors could explain these different results? It has been shown that the presence of BrdU in DNA increases sensitivity of cells to a wide range of chemicals, including mitomycin C (2), Hoechst (7) and ethylnitrosourea (25), as well as UV radiation (26) and DNase I (27). This suggests that BrdU itself does not induce SCEs in cells, but is capable of sensitizing cells to agents that do. The power of Strand-seq allows us to detect SCE rates after one cell division in BrdU, thus minimizing any outside effect on SCE frequencies. This also highlights the importance of minimizing exposure to exogenous sources of DNA damage while culturing cells in BrdU. It has previously been suggested that BS cells only display elevated SCE rates when replicating DNA over BrdU-labeled template strands, suggesting that the high SCE rates are induced by BrdU and do not occur spontaneously (17,18). In other studies, no such effect was detected and it was concluded that BS cells do show spontaneously elevated SCE rates (19,20). It is unclear exactly why different results were obtained in these studies, but they were obtained based on highly complex staining patterns observed in metaphase spreads after two or three rounds of DNA replication in BrdU. This method is susceptible to misinterpretation of results when SCEs are not properly assigned to the cell cycle during which they occurred (28). Misidentification is even more likely to occur in BS cells due to the high SCE rates and multiple SCEs occurring in close proximity, leading to metaphase spread staining patterns that cannot be reliably analyzed. One major difference between these previously reported studies are the cell types used for experiments: Epstein-Barr virus (EBV) transformed B lymphoblastoid cell lines (17,18), primary lymphocytes (19) and primary fibroblasts (20). It has been suggested that the different results reflect an effect of cell transformation by EBV (19,20). However, we used both EBV transformed lymphoblasts and primary fibroblasts in this study and we found no differences in relative SCE rates and effect of BrdU on SCE rates. The only differences we observed are that both lymphoblast cell lines displayed slightly higher SCE rates than the fibroblast cell lines and that BrdU had a larger effect on cell proliferation in the lymphoblasts. This phenotype was seen in both the normal and BS lymphoblasts, suggesting it was caused by an intrinsic difference, possibly reflecting the ongoing oncogenic stress that occurs as the result of EBV-transformed nature of the cells. Based on these results, we conclude that BS cells display spontaneously elevated SCE rates and that this reflects high levels of genomic instability in patient cells that likely contributes to the wide range of symptoms associated with BS, including the strong cancer predisposition seen in patients (29). Finally, we show that Strand-seq can be used to detect spontaneously occurring SCEs at high resolution, making it a powerful tool for studying genomic instability at the single cell level. ACCESSION NUMBER Sequencing data have been deposited in the European Nucleotide Archive under accession number PRJEB13795. Supplementary Material SUPPLEMENTARY DATA We thank Evert-Jan Uringa and Sarra Merzouk for helpful discussions and critical reading of this manuscript. We also thank Nancy Halsema, Inge Kazemier, Karina Hoekstra-Wakker, Diana Spierings and Marianna Bevova for sequencing assistance and David Porubsky and Victor Guryev for help with data analysis. SUPPLEMENTARY DATA Supplementary Data are available at NAR Online. FUNDING European Research Council (ERC) [ROOTS-Grant Agreement number 294740 to P.M.L.]. Funding for open access charge: ERC [ROOTS-Grant Agreement number 294740]. Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Aguilera A. Gomez-Gonzalez B. Genome instability: a mechanistic view of its causes and consequences Nat. Rev. Genet. 2008 9 204 217 18227811 2. Latt S.A. Sister chromatid exchanges, indices of human chromosome damage and repair: detection by fluorescence and induction by mitomycin C Proc. Natl. Acad. Sci. U.S.A. 1974 71 3162 3166 4137928 3. Abramovsky I. Vorsanger G. Hirschhorn K. Sister-chromatid exchange induced by X-ray of human lymphocytes and the effect of L-cysteine Mutat. Res. 1978 50 93 100 347282 4. Littlefield L.G. Colyer S.P. Joiner E.E. DuFrain R.J. Sister chromatid exchanges in human lymphocytes exposed to ionizing radiation during G0 Radiat. Res. 1979 78 514 521 451170 5. Perry P. Wolff S. New Giemsa method for the differential staining of sister chromatids Nature 1974 251 156 158 4138930 6. Wolff S. Perry P. Differential Giemsa staining of sister chromatids and the study of chromatid exchanges without autoradiography Chromosoma 1974 48 341 353 4141298 7. Stetka D.G. Carrano A.V. The interaction of Hoechst 33258 and BrdU substituted DNA in the formation of sister chromatid exchanges Chromosoma 1977 63 21 31 71978 8. Ishii Y. Bender M.A. Factors influencing the frequency of mitomycin C-induced sister-chromatid exchanges in 5-bromodeoxyuridine-substituted human lymphocytes in culture Mutat. Res. 1978 51 411 418 362180 9. Tice R. Chaillet J. Schneider E.L. Demonstration of spontaneous sister chromatid exchanges in vivo Exp. Cell Res. 1976 102 426 429 976355 10. O'Neill J.P. Heartlein M.W. Preston R.J. Sister-chromatid exchanges and gene mutations are induced by the replication of 5-bromo- and 5-chloro-deoxyuridine substituted DNA Mutat. Res. 1983 109 259 270 6682483 11. Lambert B. Hansson K. Lindsten J. Sten M. Werelius B. Bromodeoxyuridine-induced sister chromatid exchanges in human lymphocytes Hereditas 1976 83 163 174 61954 12. Chaganti R.S. Schonberg S. German J. A manyfold increase in sister chromatid exchanges in Bloom's syndrome lymphocytes Proc. Natl. Acad. Sci. U.S.A. 1974 71 4508 4512 4140506 13. Wu L. Hickson I.D. The Bloom's syndrome helicase suppresses crossing over during homologous recombination Nature 2003 426 870 874 14685245 14. Hand R. German J. A retarded rate of DNA chain growth in Bloom's syndrome Proc. Natl. Acad. Sci. U.S.A. 1975 72 758 762 1054854 15. Warren S.T. Schultz R.A. Chang C.C. Wade M.H. Trosko J.E. Elevated spontaneous mutation rate in Bloom syndrome fibroblasts Proc. Natl. Acad. Sci. U.S.A. 1981 78 3133 3137 6942420 16. Yankiwski V. Marciniak R.A. Guarente L. Neff N.F. Nuclear structure in normal and Bloom syndrome cells Proc. Natl. Acad. Sci. U.S.A. 2000 97 5214 5219 10779560 17. Shiraishi Y. Yosida T.H. Sandberg A.A. Analysis of single and twin sister chromatid exchanges in endoreduplicated normal and Bloom syndrome B-lymphoid cells Chromosoma 1982 87 1 8 6984382 18. Shiraishi Y. Yosida T.H. Sandberg A.A. Analyses of bromodeoxyuridine-associated sister chromatid exchanges (SCEs) in Bloom syndrome based on cell fusion: single and twin SCEs in endoreduplication Proc. Natl. Acad. Sci. U.S.A. 1983 80 4369 4373 6308619 19. Tsuji H. Kojima T. Presence of abnormally high incidences of sister chromatid exchanges in three successive cell cycles in Bloom's syndrome lymphocytes Chromosoma 1985 93 87 93 4064833 20. Tsuji H. Heartlein M.W. Latt S.A. Disparate effects of 5-bromodeoxyuridine on sister-chromatid exchanges and chromosomal aberrations in Bloom syndrome fibroblasts Mutat. Res. 1988 198 241 253 2965297 21. Falconer E. Hills M. Naumann U. Poon S.S. Chavez E.A. Sanders A.D. DNA template strand sequencing of single-cells maps genomic rearrangements at high resolution Nat. Methods 2012 9 1107 1112 23042453 22. Langmead B. Salzberg S.L. Fast gapped-read alignment with Bowtie 2 Nat. Methods 2012 9 357 359 22388286 23. Hills M. O'Neill K. Falconer E. Brinkman R. Lansdorp P.M. BAIT: organizing genomes and mapping rearrangements in single cells Genome Med. 2013 5 82 24028793 24. Latt S.A. Microfluorometric detection of deoxyribonucleic acid replication in human metaphase chromosomes Proc. Natl. Acad. Sci. U.S.A. 1973 70 3395 3399 4128545 25. Rodriguez-Reyes R. Morales-Ramirez P. Sister chromatid exchange induction and the course of DNA duplication, two mechanisms of sister chromatid exchange induction by ENU and the role of BrdU Mutagenesis 2003 18 65 72 12473737 26. Wojcik A. von Sonntag C. Obe G. Application of the biotin-dUTP chromosome labelling technique to study the role of 5-bromo-2′-deoxyuridine in the formation of UV-induced sister chromatid exchanges in CHO cells J. Photochem. Photobiol. B 2003 69 139 144 12633986 27. Stoilov L. Wojcik A. Giri A.K. Obe G. SCE formation after exposure of CHO cells prelabelled with BrdU or biotin-dUTP to various DNA-damaging agents Mutagenesis 2002 17 399 403 12202627 28. Heddle J.A. Influence of false twins on the ratios of twin and single sister chromatid exchanges J. Theor. Biol. 1969 22 151 162 5797565 29. German J. Bloom syndrome: a mendelian prototype of somatic mutational disease Medicine (Baltimore) 1993 72 393 406 8231788

PMC005xxxxxx/PMC5001600.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2722047010.1093/nar/gkw4751Methods OnlineCasHRA (Cas9-facilitated Homologous Recombination Assembly) method of constructing megabase-sized DNA Zhou Jianting 12Wu Ronghai 12Xue Xiaoli 1*Qin Zhongjun 1*1 Key Laboratory of Synthetic Biology, Institute of Plant Physiology and Ecology, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences, Shanghai 200032, China2 University of the Chinese Academy of Sciences, Beijing 100049, China* To whom correspondence should be addressed. Tel: +86 21 54924171; Fax: +86 21 54924176; Email: qin@sibs.ac.cnCorrespondence may also be addressed to Xiaoli Xue. Email: xlxue@sibs.ac.cn19 8 2016 24 5 2016 24 5 2016 44 14 e124 e124 11 5 2016 22 4 2016 19 2 2016 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Current DNA assembly methods for preparing highly purified linear subassemblies require complex and time-consuming in vitro manipulations that hinder their ability to construct megabase-sized DNAs (e.g. synthetic genomes). We have developed a new method designated ‘CasHRA (Cas9-facilitated Homologous Recombination Assembly)’ that directly uses large circular DNAs in a one-step in vivo assembly process. The large circular DNAs are co-introduced into Saccharomyces cerevisiae by protoplast fusion, and they are cleaved by RNA-guided Cas9 nuclease to release the linear DNA segments for subsequent assembly by the endogenous homologous recombination system. The CasHRA method allows efficient assembly of multiple large DNA segments in vivo; thus, this approach should be useful in the last stage of genome construction. As a proof of concept, we combined CasHRA with an upstream assembly method (Gibson procedure of genome assembly) and successfully constructed a 1.03 Mb MGE-syn1.0 (Minimal Genome of Escherichia coli) that contained 449 essential genes and 267 important growth genes. We expect that CasHRA will be widely used in megabase-sized genome constructions. cover-date19 August 2016 ==== Body INTRODUCTION The assembly of small DNA fragments into large constructs, such as the genes involved in biochemical pathways, biological machinery, and even entire genomes, is one of the most fundamental requirements of synthetic biology. State-of-the-art in vitro DNA assembly methods have been developed to fulfil different purposes (1,2). Restriction endonuclease assembly methods (including BioBricks (3), BglBricks (4) and Golden gate (5)) are widely used to assemble standardized biological components, and in vitro overlap assembly methods (including InFusion (6), SLIC (7) and Gibson one-step isothermal assembly (8)) are useful for large DNA constructions. The efficient homologous recombination observed in Saccharomyces cerevisiae has greatly simplified the construction of DNA molecules up to several hundred kilobases (9,10). The techniques required to synthesize a complete genome are important for synthetic biology. The ongoing Synthetic Yeast Genome Project (Sc2.0) has demonstrated the plasticity of the yeast genome by synthesizing chromosome arms (11). As a milestone achievement, the first heavily edited yeast chromosome, synIII (273 kb), was synthesized by 11 successive rounds of DNA assembly in vivo and used to replace the native chromosome III (12). This method avoided the direct assembly of large DNAs and allowed for step-by-step testing of the functionality of the newly synthesized regions. However, a similar strategy would not be suitable for genome synthesis in other organisms without an efficient endogenous homologous recombination system. Because the smallest free-living organisms have genomes larger than 1 Mb (13), the ability to synthesize a megabase-sized genome is of great importance. The first complete 1.08 Mb Mycoplasma mycoides genome was synthesized by three rounds of assembly in yeast following the Gibson procedure, and its functionality was demonstrated in another closely related species, M. capricolum (14). The Gibson procedure of genome assembly allows for the one-step assembly of more than 10 DNA fragments in yeast and has an efficiency of 10–100%. However, the assembly efficiency decreases drastically down to 1/48 at the final stage of genome assembly because this method requires a sufficient quantity of high-quality large subassembly DNA segments from complex in vitro manipulation procedures. All eleven 100 kb assembly intermediates were isolated from yeast, and then two steps of purification and one step of enrichment were performed to remove the host chromosomal DNA contamination (14). These difficult and time-consuming in vitro manipulations resulted in low efficiency. In recent years, the RNA-guided nuclease Cas9 has become widely used in prokaryotic and eukaryotic genome editing (15–20). Guided by designed RNAs, Cas9 can introduce double-stranded DNA breakage at the defined position, which dramatically increases the efficiency of targeted genome deletions, mutations, and insertions (15). Cas9 has also been reported to facilitate direct multiplex integrations of DNA parts assembled in vivo in yeast (21). To avoid the complex and time-consuming in vitro manipulations of large linear DNAs during genome assembly, we have developed a new method in which circular DNAs are used directly in a one-step in vivo assembly. We designated this method CasHRA (Cas9-facilitated Homologous Recombination Assembly). The large circular DNAs were cut by target RNA-guided Cas9 in vivo to release the linear DNAs for direct assembly by overlapping sequences and homologous recombination in yeast. By eliminating the guide RNAs, the assembled circular DNAs could be directly used for the next round of DNA assembly. We tested the efficiency of CasHRA in assemblies with both small and large circular DNAs. To demonstrate the utility of CasHRA for genome assembly, we combined this method with the Gibson procedure of genome assembly as an upstream assembly method to construct a 1.03 Mb MGE-syn1.0 (Minimal Genome of E. coli). MATERIALS AND METHODS Strains, plasmids, media and reagents The yeast strain used in this study was S. cerevisiae VL6-48 (MATα, trp1-Δ1, ura3-Δ1, ade2-101, his3-Δ200, lys2, met14) (22). The yeast cells were cultured in YPAD medium (1% [w/v] Bacto yeast extract, 2% [w/v] Bacto peptone, 2% [w/v] glucose, and 80 mg/L adenine hemisulfate) or synthetic complete medium (SC) (Sigma-Aldrich, St. Louis, MO, USA) lacking the nutrients required by the corresponding auxotrophic strains. The E. coli strain DH10B (F−, λ−, endA1, glnV44, thi-1, recA1, relA1, gyrA96, deoR, nupG, Φ80dlacZΔM15, Δ(lacZYA-argF)U169, hsdR17(rK− mK+), mcrA, mcrBC, mrr, galE) was used for the routine genetic modifications. The plasmids and primers used in this study are listed in Supplementary Tables S1 and S2. The primers were synthesized by Genescript (Nanjing, Jiangsu, China) or JIE LI Biology (Shanghai, China). Phanta Super-Fidelity DNA Polymerase (Vazyme, Nanjing, Jiangsu, China) and KOD-Plus-Neo DNA Polymerase (Toyobo, Osaka, Japan) were used for the standard PCR amplification. KOD-FX DNA Polymerase (Toyobo, Osaka, Japan) was used for the colony PCR confirmation. The Wizard SV Gel and PCR Clean-up System (Promega, Fitchburg, WI, USA) was used for DNA purification. The restriction endonucleases and DNA size markers were purchased from New England BioLabs (Ipswich, MA, USA). All of the chemicals were obtained from Sigma-Aldrich unless otherwise specified. Construction of plasmids Plasmids p426-SNR52p-gRNA.CAN1.Y-SUP4t and p415-GalL-Cas9-CYC1t were gifts from George Church (15) (Addgene plasmids #43803 and #43804). To construct the plasmid pMetcas9-0 for the constitutive expression of Cas9, the two fragments released by HpaI/SpeI digestion of p415-GalL-Cas9-CYC1t were gel purified. The selection marker MET14 was PCR amplified from the S. cerevisiae S288c genome using the primers Met14-F/Met14-R. The strong constitutive promoter TEF1 was amplified by PCR using the template pAG36 (23) and primers Tef1p-F/Tef1p-R. The PCR primers were designed to contain 40 bp overlapping sequences with the adjacent DNA segments. The four DNA segments described were co-introduced into yeast and assembled into pMetcas9-0. To eliminate the guide RNA expression plasmid, the guide RNA targeting the replication origin of pTrp at the S3 site was synthesized by PCR amplification using the template p426-SNR52p-gRNA.CAN1.Y-SUP4t and primers Cas-S3-F/Cas-S3-R. The galactose inducible GAL1 promoter was PCR amplified from p415-GalL-Cas9-CYC1t using GalL-F/GalL-R. The two DNA segments were assembled by fusion PCR and then cloned into pMetcas9-0 at the XbaI site to construct pMet-Cas9 (Supplementary Figure S1). To construct the vector backbone pTrp to facilitate the cloning of constitutively expressed guide RNAs, the auxotrophic marker TRP1 was PCR amplified from the S. cerevisiae S288c genome using the primers Trp1-F/Trp1-R. The 2 μm origin, pBR322 origin and ampicillin resistance gene were PCR amplified from p426-SNR52p-gRNA.CAN1.Y-SUP4t using the primer pairs 2μ ori-F/2μ ori-R, pBR322-F/pBR322-R, and f1ori-F/f1ori-R separately. The four DNA fragments were co-introduced into yeast by transformation and then assembled into pTrp. For the first round of DNA assembly, the Target1 and Target2 DNAs, which constitutively express the guide RNAs targeting the common vector backbone of all of the circular DNAs at the S1 (5′-ATAGTGTCACCTAAATAGCT-3′) and S2 (5′-CGTAGCAACCAGGCGTTTAA-3′) sites, were synthesized (for the detailed sequences, please refer to Supplementary Figure S2). The synthesized DNA segments were released from the cloning vectors by BglII/SpeI and SpeI/NcoI digestion and ligated together with the BglII/NcoI digested vector pTrp to construct pTrp-gRNA (Supplementary Figure S2). For the second round of DNA assembly, the DNA segments that targeted the common vector backbone of all of the circular DNAs at the S4 (5′-CGGCCAACGCGAACCCTTAG-3′) and S5 (5′-TGATGAACCTGAATCGCCAG-3′) sites were synthesized and similarly cloned into pTrp to construct pTrp-gRNA2. The assembly vector pCriv0 (Supplementary Figure S3) was PCR amplified using the primers pCriv0-VF/pCriv0-VR and the DNA template pZQ233 (constructed by our lab, unpublished) to contain the replication origin of chromosome II (24) from Vibrio cholerae N16961 (25), the chloramphenicol resistance marker (Cm), and the yeast origin CEN6 ARS4. The yeast selection marker ADE2 was amplified from the S. cerevisiae S288c genome using the primers ade2-F/ade2-R. The two DNA fragments were introduced into yeast and then assembled into pCriv0 through 40 bp overlaps at the termini. The assembly vector pZJ231 (Supplementary Figure S4) for EMG-syn1.0 was constructed by assembling the following four DNA segments. The replication origin of V. cholerae chromosome II and its partition system as well as the yeast origin CEN6 ARS4 and the selection marker HIS3 were PCR amplified from pZQ233. The kanamycin resistance gene was PCR amplified using pET28 as a template. The 500 bp sequences that overlapped with the large DNA segments Criv7 and TP5 were PCR amplified from the E. coli MDS42 genome (26). Co-introduction of multiple small or large circular DNAs The circular DNAs constructed from our previous work (27) were used for the assembly process in this study (Table 1). Two or three small circular DNAs to be assembled were co-introduced into the yeast cells harbouring pMet-Cas9 by using the lithium acetate (LiAc) transformation procedure (28). Circular DNAs larger than 100 kb were co-introduced into the yeast cells by using the protoplast fusion procedure (29,30). Cells harbouring two and three circular DNAs in addition to pMet-Cas9 were selected on the synthetic drop-out media SC-Met-His-Ura (omitting methionine, histidine, and uracil) and SC-Met-His-Ura-Lys (omitting methionine, histidine, uracil, and lysine), respectively. The cells were allowed to grow for 3 days at 30°C until colonies appeared. Table 1. Efficiency of CasHRA in the assembly of both small and large circular DNAs Circular DNAs joined in assembly Length of assembled DNA (kb)a Name of assembly No. of transformantsb in three independent experiments Assembly efficiencyc Name Length of DNA (kb)a Two small circular DNAs assembly pAEEG5 10 25 pCriv4 27, 232, 173 87 ± 15% pAEEG6 15 Three small circular DNAs assembly pAEEG4 14 pAEEG5 10 39 pCriv5 77, 244, 116 60 ± 31% pAEEG6 15 Two large circular DNAs assembly pSP5 117 302 pCriv6 28, 74, 27 80 ± 17% pTP3-U 185 Three large circular DNAs assembly pTP1 177 pTP2 298 660 pCriv7 7, 10, 6 73 ± 9% pTP3-L 185 Assembly of the 1.03 Mb MGE-syn1.0 (Minimal Genome of E. coli) pCriv7 660 pTP4 185 1028 MGE-syn1.0 4, 7, 11 65 ± 13% pTP5 185 aThe DNA length was calculated without the vector backbone. bTransformants of pCriv4–pCriv7 were plated on triple-drop-out medium SC-Met-Trp-Ade (omitting methionine, tryptophan, and adenine) for the selection of the assembly vector (auxotrophic marker ADE2) in addition to the complete Cas9 system. The auxotrophic marker in the MGE-syn1.0 assembly vector was HIS3; thus, the corresponding transformants were plated on triple-drop-out medium SC-Met-Trp-His (omitting methionine, tryptophan, and histidine). cThe assembly efficiency was evaluated by PCR positive rates from three independent experiments. Assembly booting The unique linear assembly vector was PCR amplified using pCriv0 as the template and primers that included 60 bp sequences that overlapped the adjacent DNA segments. Approximately 1 μg pTrp-gRNA and 1 μg linear assembly vector (except for the first assembly experiment of pCriv4, for which only half the amount of each DNA was used) were co-introduced into the yeast cells containing circular DNAs to be assembled and pMet-Cas9 by using LiAc transformation. Five percent of the transformants were selected on double-drop-out medium SC-Met-Trp (omitting methionine and tryptophan) for selection of the complete Cas9 system (pMet-Cas9 and pTrp-gRNA). The remaining transformants were plated on the triple-drop-out medium SC-Met-Trp-Ade (omitting methionine, tryptophan, and adenine) for selection of the assembly vector as well as the complete Cas9 system. The transformants were allowed to grow for 2 days at 30°C until colonies appeared. MGE-syn1.0 assembly The guide RNA expression plasmid pTrp-gRNA was eliminated using a previously described protocol (15). After the elimination of pTrp-gRNA, the yeast cells containing pMet-Cas9 (auxotrophic marker Met) and pCriv7 (auxotrophic marker Ade) were protoplast fused with a yeast cell derivative that harboured pTP4 (auxotrophic marker Ura) and a yeast cell derivative that harboured pTP5 (auxotrophic marker Lys), respectively. Cells containing all three large circular DNAs and pMet-Cas9 were selected on SC-Met-Ade-Ura-Lys (omitting methionine, adenine, uracil, and lysine) medium. The linear assembly vector was generated by PCR amplification using the template pZJ231 and the primers pCriv8-VF/ pCriv8-VR. Approximately 1 μg pTrp-gRNA2 and 1 μg linear assembly vector were co-introduced into the yeast cells by LiAc transformation. Five percent of the transformants were selected on the double-drop-out medium SC-Met-Trp. The remaining transformants were plated on the triple-drop-out medium SC-Met-Trp-His (omitting methionine, tryptophan, and histidine). The transformants were allowed to grow for 2 days at 30°C until colonies appeared. Confirmation of the correct assemblies Correct assemblies were confirmed by colony PCR using the primers listed in Supplementary Table S2. For each assembly, PCR products from five to eight randomly selected positive colonies were sent for sequencing using the same primers used for the PCR assay. The positive assemblies were further subjected to enzymatic digestion and electrophoresis analysis. Small assemblies (pCriv4 and pCriv5) were isolated from the yeast and introduced into E. coli DH10B by electroporation. The assemblies re-isolated from E. coli were subjected to enzymatic digestion and an electrophoresis analysis. It was difficult to introduce DNAs of several hundred kilobases into E. coli; therefore, the large assemblies pCriv6, pCriv7 and MGE-syn1.0 were isolated directly from the yeast and analysed. Approximately 5 × 108 yeast cells were collected and embedded in agarose plugs using a previously described protocol (14). The linear yeast chromosomal DNAs were separated out of the agarose plugs by pulsed-field gel electrophoresis (PFGE) performed under the following conditions: 6 V/cm, 10–60 s switch time, and 14°C for 20 h. The large circular assembly trapped in the agarose plugs was subjected to restriction enzymatic digestion and then a PFGE analysis performed under the following conditions: 6 V/cm, 1–25 s switch time, and 14°C for 20 h. RESULTS Design of the CasHRA method CasHRA was developed to facilitate the direct use of large circular DNAs for one-step assemblies in vivo and avoid the substantial manipulation of linear DNA segments in vitro. As illustrated in Figure 1, three large circular DNAs were co-introduced into yeast cells harbouring the Cas9 expression plasmid pMet-Cas9 via protoplast fusion (step 1). The assembly booting was achieved by co-introducing pTrp-gRNA and the linear assembly vector into the cells by transformation (step 2). The plasmid pTrp-gRNA constitutively expressed guide RNAs that targeted the common vector backbone of all of the large circular DNAs at the S1 and S2 sites. Once pTrp-gRNA entered the cell, a functional Cas9 system rapidly developed and cleaved the large circular DNAs to release the linear DNA segments with exposed overlaps at the termini. Taking advantage of the highly efficient homologous recombination system of S. cerevisiae, these large DNA segments were then properly assembled together with the linear assembly vector. The plasmid pMet-Cas9 contained a galactose-inducible guide RNA that targets the replication origin of pTrp-gRNA at the S3 site. Transferring the yeast cells into the galactose medium efficiently eliminated pTrp-gRNA (step 3) and allowed the yeast cells containing the assembled circular DNA to be directly used for the next round of DNA assembly. The entire CasHRA process requires 9 days, including 5 days to introduce all of the circular DNAs into the yeast cells, 3 days to obtain the yeast cells with the assembled product, and 1 day to eliminate pTrp-gRNA. Figure 1. Schematic diagram of the CasHRA method. Large circular DNAs were co-introduced into individual yeast cells harbouring the Cas9 expression plasmid pMet-Cas9 by protoplast fusion (step 1). Then, the guide RNA expression plasmid pTrp-gRNA and the linear assembly vector were co-introduced into the cells by transformation. The RNA-guided Cas9 cut the vector backbone of all of the large circular DNAs at the S1 and S2 sites. The released linear DNA segments were joined together with the linear vector through overlaps by taking advantage of the efficient homologous recombination system in yeast. The plasmid pMet-Cas9 contained a galactose-inducible guide RNA targeting the replication origin of pTrp-gRNA at the S3 site. Efficiency of CasHRA for circular DNA assembly We validated the efficiency of CasHRA for the assembly of both small and large circular DNAs as listed in Table 1. These circular DNAs contained 10–298 kb of essential and important growth genes for E. coli and were constructed in our previous work (27). The overlaps in the adjacent DNA segments were designed to be approximately 500 bp. Shorter overlap sequences could also be used; however, longer overlaps tended to increase the assembly efficiency (10). The common vector backbone of these circular DNAs contained CEN6 and ARS4 for replication and single-copy maintenance in yeast and were used to avoid any discrepant assembly in the yeast (10). Additional yeast replication origins were included in every ∼100 kb of E. coli DNA to improve the stability of these large DNA segments in yeast (31). To increase the stability of the large foreign DNAs in the yeast, different auxotrophic selection markers were used to ensure the co-occurrence of the subassemblies in the yeast cells. To properly validate the success rate, all of the assembly reactions were repeated three times. As shown in Table 1, the success rates of pCriv4 and pCriv5, which were assembled from two and three small circular DNAs, were 87 ± 15% and 60 ± 31%, respectively. Similar efficiencies were obtained for the assembly of circular DNAs larger than 100 kb. The success rates of pCriv6 and pCriv7, which were assembled from two and three large circular DNAs, were 80 ± 17% and 73 ± 9%, respectively. Sequencing 82 overlaps from 23 assembly DNA molecules revealed 10 errors, all of which occurred in the 60 bp overlaps between the vector and the leftmost or rightmost subassemblies. The 60 bp overlaps were introduced by the 3′ ends of the primers during PCR amplification of the assembly vector, and errors could have resulted from impure primers. Detailed information on the enzymatic analysis and sequencing of the assemblies of pCriv4∼7 is presented in Supplementary Figures S5–S8 and Supplementary Table S3. Assembly of a 1.03 Mb MGE-syn1.0 genome via CasHRA To gain a deeper understanding of the basic processes required to create synthetic life, we designed a 1.03 Mb MGE-syn1.0 (Minimal Genome of E. coli) that contained 449 essential genes from published experimental and computational studies (27,32–36) and 267 important growth genes, including 151 genes that affect growth (34) and 76 genes that have not been assigned to central metabolism pathways and regulation networks. These genes were PCR amplified from the E. coli MDS42 genome (26), sequenced, and then assembled into five large circular DNAs pTP1–pTP5 (Figure 2) by three rounds of assembly following the Gibson procedure (14). The Gibson procedure of genome assembly method can assemble >10 DNA segments at one time with high efficiency; thus, it is a useful method for the early stages of genome assembly. Our previous attempts to sequentially assemble pTP1–pTP3 (177–298 kb) into pCriv7 (660 kb) using the Gibson procedure of genome assembly failed. However, the application of CasHRA successfully assembled pCriv7, which contained all 449 E. coli genes. Because of the simplicity and efficiency of CasHRA, we applied this method to assemble the 1.03 Mb MGE-syn1.0. To directly use yeast cells containing pCriv7 for the next round of DNA assembly, the guide RNA expression vector pTrp-gRNA was eliminated by simply transferring the cells into galactose medium, and the elimination frequency was usually >90% in one round of induction. The 1.03 Mb MGE-syn1.0 was assembled from pCriv7 (660 kb), pTP4 (185 kb) and pTP5 (185 kb) (Figure 2). The assembly vector for MGE-syn1.0, pZJ231, was constructed to contain approximately 500 bp overlaps with the left terminus of Criv7 and the right terminus of TP5. The success rate of the MGE-syn1.0 assembly was 65 ± 13% (Table 1), and the sequences of all of the overlaps were 100% correct. The circular MGE-syn1.0 DNA was separated from the yeast linear chromosomes and digested with SpeI and XbaI separately. The digestion patterns were consistent with the theoretical calculation, and DNA bands larger than 50 kb were clearly shown in the pulsed-field gel electrophoresis (PFGE) image (Figure 3), although smaller DNA bands were not separated from the short fragments of the yeast genome under the experimental conditions. Figure 2. Schematic diagram of the MGE-syn1.0 (Minimal Genome of E. coli) assembly. Four hundred forty nine essential genes and 267 important growth genes were PCR amplified from the E. coli MDS42 genome, sequenced and assembled into five large, circular DNAs pTP1–pTP5 (177–298 kb via three rounds of assembly following the Gibson procedure. The TP1–TP5 were assembled into the 1.03 Mb MGE-syn1.0 via two rounds of CasHRA. Please note that the DNA length was calculated without the vector backbone. Figure 3. Analysis of the assembled MGE-syn1.0. (A) Map of MGE-syn1.0, assembled from three large DNA segments, Criv7 (660 kb, marked in red), TP4 (298 kb, marked in yellow), and TP5 (184 kb, marked in green). The total size of MGE-syn1.0 (1.03 Mb) and the assembly vector (10 kb) was 1.04 Mb. The SpeI cutting sites are marked by greenish brown lines in the inner circle. The XbaI cutting sites are marked by black lines on the outer circle. To separate out the linear yeast chromosomal DNA, agarose plugs were subjected to PFGE under the following conditions: 6 V/cm, 10–60 s switch time, and 14°C for 20 h. The circular MGE-syn1.0 that was trapped inside the plug was digested with SpeI and XbaI, separately, followed by another round of PFGE under the following conditions: 6 V/cm, 1–25 s switch time, and 14°C for 16 h. For the control, the assembly host VL6-48 was subjected to the same treatment. (B) The SpeI digestion of MGE-syn1.0 released ten DNA fragments. The five larger DNA fragments with sizes of 262, 200, 182, 157 and 98 kb are indicated by red arrows. Another five smaller bands (39, 34, 27, 19 and 18 kb) could not be clearly separated from the short fragments of yeast chromosomes under the experimental conditions. (C) The XbaI digestion of MGE-syn1.0 released nine DNA fragments. The DNA fragments with sizes of 248, 205, 173, 143, 117, 64 and 56 kb are indicated by red arrows. The smaller bands (20 and 11 kb) could not be clearly separated from the yeast chromosomal fragments under the experimental conditions. We have attempted to transplant MGE-syn1.0 from yeast to E. coli by two different methods to test its functionality. First, we attempted to directly isolate and purify MGE-syn1.0 from yeast and introduce it into E. coli DH10B by electroporation. Our pilot study showed that pTP1 (the total size of TP1 and the vector backbone was 187 kb) could be successfully introduced into E. coli and stably maintained. However, the introduction of pTP2 (the total size of TP2 and the vector backbone was 308 kb) into E. coli failed. The size of TP2 (308 kb) exceeded the maximum size reported for E. coli electroporation (37), which might have caused the failure of the TP2 introduction. Second, we attempted to use protoplast fusion to transplant MGE-syn1.0 from yeast to E. coli. We prepared the protoplasts of E. coli (38) and yeast (29) according to reported protocols and mixed them in different ratios. However, successfully fused E. coli cells were not observed in our pilot study. A new method of transplanting giant DNA (>1 Mb) from yeast to E. coli awaits further development. DISCUSSION Conventional DNA assembly methods produce purified linear subassembly DNAs by complex manipulations in vitro, which can be burdensome in the last stage of megabase-sized DNA (e.g., synthetic genome) assembly. CasHRA was developed to directly use large circular DNAs (>100 kb) in a one-step DNA assembly process in vivo that avoids the difficult manipulations in vitro. All of the critical steps of CasHRA have been developed in vivo, including the co-introduction of large circular DNAs into individual cells by protoplast fusion, the release of large linear DNA segments by RNA-guided Cas9, and the subsequent DNA assembly using the yeast homologous recombination system. The time (9 days) and effort required to assemble DNAs with sizes from hundreds of kilobases (pCriv6 and pCriv7) to one megabase (MGE-syn1.0) were similar. The positive rates of the large assemblies (60–80%) were high enough to allow for easy detection of the correct constructs. Moreover, the efficient elimination of the guide RNA expression vector allowed for the direct use of the yeast cells containing the correct assembly in the next round of DNA assembly, which is convenient for large DNA constructions. Because the activities of CasHRA in vivo do not have strict limitations on the size of the subassemblies, we speculate that this method could be used to assemble DNA constructs larger than one megabase. In our experiments, the subassemblies or the fully assembled molecules could be stably maintained in yeast by selection with different auxotrophic markers. Six auxotrophic markers are available for the S. cerevisiae VL6-48 strain. In this study, three auxotrophic markers were replaced by the Cas9 expression plasmid, the gRNA expression plasmid, and the linear assembly vector. The three remaining auxotrophic markers were used for the subassemblies. However, when combined with the commonly used selection markers (e.g. G418, bleomycin and hygromycin), additional subassemblies could be stably maintained in the individual yeast cells. Because of the high performance of Cas9 and the yeast homologous recombination, additional subassemblies could be simultaneously assembled by CasHRA. To ensure the efficiency of CasHRA, the sequence and length of the assembly overlaps must be considered. In this study, the overlap sequences originated from different regions of the E. coli genome (GC content, 51%) to reduce the possible effect of sequence variation in the homology overlaps on the assembly efficiency. Similar efficiencies were obtained for five assembled DNAs with different overlaps, indicating the limited effect of overlap sequence variations on the assembly efficiency in this study. It is important to realize that introducing short overlaps (e.g. in the overlaps between the vector and subassemblies) at the 3′ ends of PCR primers would result in errors in the overlap sequence because of the impurity of the primers. Therefore, pre-cloning the overlaps into the assembly vector will be required if the sequence accuracy of the overlaps are important for the functionality of the assembled DNA. When the assembly efficiencies for both small and large circular DNAs are similar, fewer transformants are obtained from the assembly of large circular DNAs. This result may be caused by a decreased probability of large DNAs finding their counterparts with approximately 500 bp overlaps for homologous recombination. Whether longer overlaps result in more transformants from large DNA assemblies awaits further exploration. In summary, CasHRA is a promising method for performing the in vivo assembly of multiple large DNA segments, and it could be useful in the construction of megabase-sized genomes. Supplementary Material SUPPLEMENTARY DATA We are grateful to Prof. Biao Kan from the Chinese Center for Disease Control and Prevention for kindly providing the genomic DNA of Vibrio cholerae N16961. SUPPLEMENTARY DATA Supplementary Data are available at NAR Online. FUNDING National Key Basic Research Program of China (973 Program) [2011CBA00801, 2012CB721102]; National Natural Science Foundation of China [31421061]. Funding for open access charge: National Key Basic Research Program of China (973 Program) [2011CBA00801, 2012CB721102]; National Natural Science Foundation of China [31421061]. Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Ellis T. Adie T. Baldwin G.S. DNA assembly for synthetic biology: from parts to pathways and beyond Integr. Biol. 2011 3 109 118 2. Liu W. Yuan J.S. Stewart C.N. Jr Advanced genetic tools for plant biotechnology Nat. Rev. Genet. 2013 14 781 793 24105275 3. Shetty R.P. Endy D. Knight T.F. Jr Engineering BioBrick vectors from BioBrick parts J. Biol. Eng. 2008 2 5 18410688 4. Anderson J.C. Dueber J.E. Leguia M. Wu G.C. Goler J.A. Arkin A.P. Keasling J.D. BglBricks: a flexible standard for biological part assembly J. Biol. Eng. 2010 4 1 20205762 5. Engler C. Kandzia R. Marillonnet S. A one pot, one step, precision cloning method with high throughput capability PLoS One 2008 3 e3647 18985154 6. Sleight S.C. Bartley B.A. Lieviant J.A. Sauro H.M. In-Fusion BioBrick assembly and re-engineering Nucleic Acids Res. 2010 38 2624 2636 20385581 7. Li M.Z. Elledge S.J. Harnessing homologous recombination in vitro to generate recombinant DNA via SLIC Nat. Methods 2007 4 251 256 17293868 8. Gibson D.G. Young L. Chuang R.Y. Venter J.C. Hutchison C.A. 3rd Smith H.O. Enzymatic assembly of DNA molecules up to several hundred kilobases Nat. Methods 2009 6 343 345 19363495 9. Gibson D.G. Benders G.A. Axelrod K.C. Zaveri J. Algire M.A. Moodie M. Montague M.G. Venter J.C. Smith H.O. Hutchison C.A. 3rd One-step assembly in yeast of 25 overlapping DNA fragments to form a complete synthetic Mycoplasma genitalium genome Proc. Natl. Acad. Sci. U.S.A. 2008 105 20404 20409 19073939 10. Shao Z. Zhao H. Zhao H. DNA assembler, an in vivo genetic method for rapid construction of biochemical pathways Nucleic Acids Res. 2009 37 e16 19074487 11. Dymond J.S. Richardson S.M. Coombes C.E. Babatz T. Muller H. Annaluru N. Blake W.J. Schwerzmann J.W. Dai J. Lindstrom D.L. Synthetic chromosome arms function in yeast and generate phenotypic diversity by design Nature 2011 477 471 476 21918511 12. Annaluru N. Muller H. Mitchell L.A. Ramalingam S. Stracquadanio G. Richardson S.M. Dymond J.S. Kuang Z. Scheifele L.Z. Cooper E.M. Total synthesis of a functional designer eukaryotic chromosome Science 2014 344 55 58 24674868 13. Moya A. Gil R. Latorre A. Pereto J. Pilar Garcillan-Barcia M. de la Cruz F. Toward minimal bacterial cells: evolution vs. design FEMS microbiology reviews 2009 33 225 235 19067748 14. Gibson D.G. Glass J.I. Lartigue C. Noskov V.N. Chuang R.Y. Algire M.A. Benders G.A. Montague M.G. Ma L. Moodie M.M. Creation of a bacterial cell controlled by a chemically synthesized genome Science 2010 329 52 56 20488990 15. DiCarlo J.E. Norville J.E. Mali P. Rios X. Aach J. Church G.M. Genome engineering in Saccharomyces cerevisiae using CRISPR-Cas systems Nucleic Acids Res. 2013 41 4336 4343 23460208 16. Jiang W. Bikard D. Cox D. Zhang F. Marraffini L.A. RNA-guided editing of bacterial genomes using CRISPR-Cas systems Nat. Biotechnol. 2013 31 233 239 23360965 17. Mali P. Yang L. Esvelt K.M. Aach J. Guell M. DiCarlo J.E. Norville J.E. Church G.M. RNA-guided human genome engineering via Cas9 Science 2013 339 823 826 23287722 18. Qi L.S. Larson M.H. Gilbert L.A. Doudna J.A. Weissman J.S. Arkin A.P. Lim W.A. Repurposing CRISPR as an RNA-guided platform for sequence-specific control of gene expression Cell 2013 152 1173 1183 23452860 19. Ran F.A. Hsu P.D. Lin C.Y. Gootenberg J.S. Konermann S. Trevino A.E. Scott D.A. Inoue A. Matoba S. Zhang Y. Double nicking by RNA-guided CRISPR Cas9 for enhanced genome editing specificity Cell 2013 154 1380 1389 23992846 20. Wang H. Yang H. Shivalila C.S. Dawlaty M.M. Cheng A.W. Zhang F. Jaenisch R. One-step generation of mice carrying mutations in multiple genes by CRISPR/Cas-mediated genome engineering Cell 2013 153 910 918 23643243 21. Jakociunas T. Rajkumar A.S. Zhang J. Arsovska D. Rodriguez A. Jendresen C.B. Skjodt M.L. Nielsen A.T. Borodina I. Jensen M.K. CasEMBLR: Cas9-facilitated multiloci genomic integration of in vivo assembled DNA parts in Saccharomyces cerevisiae ACS Synth. Biol 2015 4 1226 1234 22. Larionov V. Kouprina N. Solomon G. Barrett J.C. Resnick M.A. Direct isolation of human BRCA2 gene by transformation-associated recombination in yeast Proc. Natl. Acad. Sci. U.S.A. 1997 94 7384 7387 9207100 23. Goldstein A.L. McCusker J.H. Three new dominant drug resistance cassettes for gene disruption in Saccharomyces cerevisiae Yeast 1999 15 1541 1553 10514571 24. Egan E.S. Waldor M.K. Distinct replication requirements for the two Vibrio cholerae chromosomes Cell 2003 114 521 530 12941279 25. Heidelberg J.F. Eisen J.A. Nelson W.C. Clayton R.A. Gwinn M.L. Dodson R.J. Haft D.H. Hickey E.K. Peterson J.D. Umayam L. DNA sequence of both chromosomes of the cholera pathogen Vibrio cholerae Nature 2000 406 477 483 10952301 26. Posfai G. Plunkett G. 3rd Feher T. Frisch D. Keil G.M. Umenhoffer K. Kolisnychenko V. Stahl B. Sharma S.S. de Arruda M. Emergent properties of reduced-genome Escherichia coli Science 2006 312 1044 1046 16645050 27. Xue X. Wang T. Jiang P. Shao Y. Zhou M. Zhong L. Wu R. Zhou J. Xia H. Zhao G. MEGA (Multiple Essential Genes Assembling) deletion and replacement method for genome reduction in Escherichia coli ACS Synth. Biol. 2015 4 700 706 25494410 28. Gietz R.D. Schiestl R.H. High-efficiency yeast transformation using the LiAc/SS carrier DNA/PEG method Nat. Protoc. 2007 2 31 34 17401334 29. Curran B.P. Bugeja V.C. Protoplast fusion in Saccharomyces cerevisiae Methods Mol. Biol. 1996 53 45 49 8925003 30. Kouprina N. Larionov V. Selective isolation of genomic loci from complex genomes by transformation-associated recombination cloning in the yeast Saccharomyces cerevisiae Nat. Protoc. 2008 3 371 377 18323808 31. Noskov V.N. Karas B.J. Young L. Chuang R.Y. Gibson D.G. Lin Y.C. Stam J. Yonemoto I.T. Suzuki Y. Andrews-Pfannkoch C. Assembly of large, high G+C bacterial DNA fragments in yeast ACS Synth. Biol. 2012 1 267 273 23651249 32. Gil R. Silva F.J. Pereto J. Moya A. Determination of the core of a minimal bacterial gene set Microbiol. Mol. Biol. Rev.: MMBR 2004 68 518 537 15353568 33. Hashimoto M. Ichimura T. Mizoguchi H. Tanaka K. Fujimitsu K. Keyamura K. Ote T. Yamakawa T. Yamazaki Y. Mori H. Cell size and nucleoid organization of engineered Escherichia coli cells with a reduced genome Mol. Microbiol. 2005 55 137 149 15612923 34. Baba T. Ara T. Hasegawa M. Takai Y. Okumura Y. Baba M. Datsenko K.A. Tomita M. Wanner B.L. Mori H. Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: the Keio collection Mol. Syst. Biol. 2006 2 2006.0008 35. Forster A.C. Church G.M. Towards synthesis of a minimal cell Mol. Syst. Biol. 2006 2 45 16924266 36. Keseler I.M. Mackie A. Peralta-Gil M. Santos-Zavaleta A. Gama-Castro S. Bonavides-Martinez C. Fulcher C. Huerta A.M. Kothari A. Krummenacker M. EcoCyc: fusing model organism databases with systems biology Nucleic Acids Res. 2013 41 D605 D612 23143106 37. Sheng Y. Mancino V. Birren B. Transformation of Escherichia coli with large DNA molecules by electroporation Nucleic Acids Res. 1995 23 1990 1996 7596828 38. Weiss R.L. Protoplast formation in Escherichia coli J. Bacteriol. 1976 128 668 670 789360

PMC005xxxxxx/PMC5001603.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2728844610.1093/nar/gkw531Structural BiologyChanges in conformational dynamics of basic side chains upon protein–DNA association Esadze Alexandre †Chen Chuanying †Zandarashvili Levani Roy Sourav Pettitt B. Montgometry *Iwahara Junji *Department of Biochemistry and Molecular Biology, Sealy Center for Structural Biology and Molecular Biophysics, University of Texas Medical Branch, Galveston, TX 77555-1068, USA* To whom correspondence should be addressed. Tel: +1 409 747 1403; Fax: +1 409 772 6334; Email: j.iwahara@utmb.eduCorrespondence may also be addressed to B. Montgometry Pettitt. Tel: +1 409 772 0723; Fax: +1 409 772 0725; Email: mpettitt@utmb.edu† These authors contributed equally to the paper as first authors.19 8 2016 10 6 2016 10 6 2016 44 14 6961 6970 31 5 2016 26 5 2016 24 3 2016 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Basic side chains play major roles in recognition of nucleic acids by proteins. However, dynamic properties of these positively charged side chains are not well understood. In this work, we studied changes in conformational dynamics of basic side chains upon protein–DNA association for the zinc-finger protein Egr-1. By nuclear magnetic resonance (NMR) spectroscopy, we characterized the dynamics of all side-chain cationic groups in the free protein and in the complex with target DNA. Our NMR order parameters indicate that the arginine guanidino groups interacting with DNA bases are strongly immobilized, forming rigid interfaces. Despite the strong short-range electrostatic interactions, the majority of the basic side chains interacting with the DNA phosphates exhibited high mobility, forming dynamic interfaces. In particular, the lysine side-chain amino groups exhibited only small changes in the order parameters upon DNA-binding. We found a similar trend in the molecular dynamics (MD) simulations for the free Egr-1 and the Egr-1–DNA complex. Using the MD trajectories, we also analyzed side-chain conformational entropy. The interfacial arginine side chains exhibited substantial entropic loss upon binding to DNA, whereas the interfacial lysine side chains showed relatively small changes in conformational entropy. These data illustrate different dynamic characteristics of the interfacial arginine and lysine side chains. cover-date19 August 2016 ==== Body INTRODUCTION DNA recognition by proteins is vital for gene expression, DNA replication and repair. Three-dimensional (3D) structures of protein–DNA complexes show that basic side chains play important roles through electrostatic interactions with DNA phosphates as well as hydrogen-bonding with DNA bases (1–3). Thermodynamic studies also indicate the importance of interfacial basic side chains: they form ion pairs with DNA phosphate groups and cause release of condensed counterions from DNA, which is a driving force for many protein–DNA association processes (4–6). Despite the importance of the basic side chains, their dynamic properties have not been well studied by experimental means. Although some studies by nuclear magnetic resonance (NMR) spectroscopy show significant roles of conformational entropy in macromolecular recognition and association (7–9), such investigations typically probe the dynamics of backbone NH or side-chain CH3 groups only. For side chains that form hydrogen bonds and/or ion pairs, the dynamic properties and their entropic roles remain largely unknown. This represents a bottleneck to thoroughly understand molecular recognition of nucleic acids by proteins, where a large number of intermolecular hydrogen bonds and electrostatic interactions are involved. From this perspective, we conduct a comparative study on the conformational dynamics of arginine (Arg) and lysine (Lys) side chains of the DNA-binding domain of Egr-1 (also known as Zif268) in the free state and in the complex with target DNA. This protein recognizes the target 9-bp DNA sequence via three Cys2His2-class zinc fingers with high affinity (10). For the Egr-1 DNA-binding domain, the dissociation constant of the specific DNA complexes ranges from 10−11 M to 10−8 M, depending on ionic strength (11–13). In the brain, Egr-1 is induced by synaptic signals and activates genes for long-term memory formation and consolidation (14,15). In the cardiovascular system, Egr-1 is a stress-inducible transcription factor that activates genes for initiating defense responses against vascular stress and injury (16,17). The Egr-1–DNA interactions were extensively characterized in previous biophysical and biochemical studies (12,13,18–21) and high-resolution crystal structures are available for the Egr-1–DNA complexes (22–24). The investigations at an atomic level are important particularly because Egr-1 (Zif268) has been used as a major scaffold for zinc-finger (ZF) technology for artificial gene editing and regulation (25–27). In this work, we investigate the internal motions of Lys side-chain NH3+ and Arg guanidino Nϵ-Hϵ moieties in the free and DNA-bound states using NMR spectroscopy and examine changes in mobility of each basic side chain upon Egr-1's binding to the target DNA. The ZF DNA-binding domain of Egr-1 contains 21 basic side chains (15 Arg and 6 Lys residues), 15 of which interact with DNA (Figure 1). The cationic groups exhibit well-isolated NMR signals in 1H-15N heteronuclear correlation spectra for both the free protein and the complex. Thus, this system provides an opportunity for in-depth investigations on dynamic behavior of each basic side chain in the DNA recognition process. Our NMR data provide comprehensive experimental data on changes in conformational dynamics of basic side chains upon protein–nucleic acid association. In conjunction with NMR, we also use molecular dynamics (MD) simulations to gain deeper insight into the side-chain dynamics and conformational entropy in the protein–DNA association process. Characteristic differences between Arg and Lys side chains in DNA recognition dynamics become evident through these experimental and computational investigations. Figure 1. The Egr-1 (Zif268) zinc-finger (ZF)–DNA complex studied in this work. The structure shown is PDB ID: 1AAY (22). The 12-bp DNA duplex contains the target sequence (red) recognized by Egr-1. The ZF domains contain 15 Arg and 6 Lys side chains, which are shown in purple and blue, respectively. The residue numbering schemes are according to Pabo et al. (10,22). MATERIALS AND METHODS Protein and DNA preparation The Egr-1 ZF protein comprised of three zinc fingers (human Egr-1 residues 335–423) was prepared as described previously. 15N or 13C/15N labeled proteins were expressed in Escherichia coli strain BL21 (DE3) cultured in minimal media containing ammonium chloride and glucose as sole nitrogen and carbon sources. The unlabeled 12-bp DNA duplex of dAGCGTGGGCGAT and dATCGCCCACGCT (underline, the Egr-1 recognition site) was chemically synthesized and purified as described (28). NMR samples of the Egr-1–DNA complex were 370-μl solutions containing 0.4 mM protein and 0.6 mM DNA in a buffer of 20 mM potassium succinate (pH 5.8), 20 mM KCl, 0.1 mM ZnCl2. Based on the dissociation constant of this complex (12) and the concentrations of Egr-1 and DNA, more than 99.9% of the protein is expected to be in the DNA-bound state under the buffer conditions used. NMR samples of the free protein (0.4 mM) were prepared with the same buffer compositions. Each sample was sealed in a Norell co-axial tube (diameter, 5 mm) in which D2O for the NMR lock is separately sealed into the inner stem (diameter, 2 mm) to avoid isotope shifts and broadening of 15N resonances due to hydrogen-deuterium exchange (29). NMR experiments NMR experiments were performed with Bruker Avance III NMR spectrometers operated at a 1H frequency of 600, 750 or 800 MHz. The 600 and 800-MHz spectrometers were equipped with a cryogenic probe, whereas the 750-MHz spectrometer was equipped with a room-temperature probe. All NMR data were processed with the NMR-Pipe program (30) and analyzed with the NMR-View program (31). For both the free and DNA-bound proteins, backbone 1H, 13C, and 15N resonances were assigned using 3D HNCO, HN(CA)CO, HN(CO)CA, HNCA, CBCA(CO)NH, HNCACB and HBHA(CO)NH spectra (32). Side-chain 1H and 13C resonances were assigned using 3D HCCH-TOCSY, HCCH-COSY, H(CCO)NH and C(CO)NH spectra (32). These experiments for resonance assignment were performed at 35°C for the complex and at 25°C for the free proteins. Arg side-chain 15Nϵ and 1Hϵ resonances were assigned using broadband HNCACB and 3D 15N-edited NOESY spectra, as described (33). Lys side-chain NH3+ resonances were assigned using Lys-selective 2D HISQC, H2(C)N, (H2C)N(CC)H-TOCSY, 3D H3NCECD and 3D H3NCG spectra, as described (34). The 1H, 13C and 15N resonance assignment data were deposited to Biological Magnetic Resonance Data Bank: the accession numbers are 26808 for the DNA-bound protein and 26807 for the free protein. To determine the rotational diffusion parameters, the backbone 15N R1 and R2 relaxation rates at the 1H frequency of 800 MHz were measured for the free protein at 5°C and 25°C and for the complex at 10°C and 25°C. The 15N relaxation experiments for Arg Nϵ-Hϵ groups were performed with the pulse sequences for NH groups together with selective 15N rSNOB 180° pulses (1.0 ms) (35) in the INEPT schemes. By using 15N carrier position set to 81 ppm together with these selective pulses, the Arg Nϵ-Hϵ resonances were selectively observed in these 15N relaxation experiments. For Arg Nϵ-Hϵ groups, 15N R1 and heteronuclear NOE data were recorded at the 1H frequencies of 750 and 600 MHz and 15N R2 data were recorded at the 1H frequency of 750 MHz. 15N R2 relaxation dispersion experiment for Arg 15Nϵ nuclei was performed at the 1H frequency of 750 MHz using the CW-CPMG scheme (36) with the CPMG frequencies (νCPMG) of 33, 67, 100, 200, 333, 500, 667, 1000, 1333 and 1667 Hz. The Arg relaxation experiments were carried out at 25°C for both the free protein and the complex. The Lys NH3+ relaxation experiments were performed as described in our previous publications (18,37–41). For Lys NH3+ groups, 15N R1 and heteronuclear NOE data were recorded at the 1H frequencies of 800 and 600 MHz and 15N R2 data were recorded at the 1H frequency of 800 MHz. Lys 15N R2 relaxation dispersion experiment (38) was performed at the 1H frequency of 800 MHz with the CPMG frequencies of 33, 67, 100, 200, 333, 500, 667, 1000, 1333 and 1667 Hz. These measurements for the complex were conducted at 10°C under the above-mentioned buffer conditions. The Lys 15N relaxation experiments for the free protein were conducted at 5°C and pH 5.0. The lower temperature and pH were necessary to mitigate broadening of the Lys NH3+ signals due to rapid hydrogen exchange (38,39). On the other hand, the use of 5°C was difficult for the complex because of the poor quality of backbone relaxation data at that temperature. So, we used different temperatures (10 versus 5°C) in the Lys side-chain 15N relaxation experiments for the complex and for the free protein. Judging from our previous temperature-dependence study on the internal motions of Lys NH3+ groups (41), the use of these different temperatures does not significantly interfere with comparative analysis of the NH3+ order parameters. To detect hydrogen-bond scalar coupling between Lys NH3+ and DNA phosphate groups, the two-dimensional H3(N)P experiment was performed for the complex at 10°C using a cryogenic QCI-P (1H, 13C, 15N and 31P) probe at the 1H-frequency of 600 MHz, as described (37). 15N relaxation data analysis Rotational diffusion parameters (D||, D⊥ and two polar angles for the main principal axis) for the axially symmetric diffusion model (42) were determined from the backbone 15N relaxation data using a C program together with GNU Scientific Library, as described (43,44). The effective rotational correlation time τr,eff and the anisotropy of the rotational diffusion r are given by (2D|| + 4D⊥)−1 and D|| / D⊥, respectively (42). For the free protein, this calculation was performed separately for the three ZF domains because they tumble almost independently in the free state. Using MATLAB software, the order parameters for Arg Nϵ-Hϵ groups were calculated from the relaxation data at the two magnetic fields. The 15N chemical shift anisotropy parameter (σ|| - σ⊥) for arginine side-chain 15Nϵ nuclei was set to −114 ppm and the Nϵ-Hϵ distance was set to 1.04 Å according to Trbovic et al. (45). Four spectral density functions were tested for each Arg Nϵ-Hϵ group: two of them were the model-free functions of Lipari and Szabo (Equations. 35 and 43 in Ref. (46)) and the others were the extended model-free functions of Clore et al. (Equations 2 and 4 in Ref. (47) multiplied by 2/5). The best model among the four spectral density functions was selected using Akaike's information criterion calculated for each model (48). Using Mathematica software, the order parameters for Lys side-chain NH3+ groups were calculated from the 15N relaxation data at the two magnetic fields, as previously described in detail by Esadze et al. (38). Molecular dynamics simulations MD simulations of the Egr-1–DNA complex and the free Egr-1 solvated with TIP3P water molecules were performed using NAMD 2.9 software (49) with CHARMM27 all-atom force fields parameters (50–52), as previously described (18). The 1.6 Å resolution crystal structure of the Egr-1–DNA complex (PDB ID: 1AAY) (22) was used for initial structures. For each system, the macromolecule was solvated in a box of TIP3P water molecules of suitable dimensions: 69.0 × 73.0 × 74.0 Å3 (the complex) and 85.8 × 89.3 × 91.3 Å3 (free protein). For the free protein, a larger water box was introduced to ensure all possible conformational states are sufficiently solvated, as the inter-domain displacement became increasingly extended in the first 100 ns from the initial compact structure. For zinc-coordinating cysteine residues, the parameters of the deprotonated thiolate moieties were taken from Foloppe et al. (53). The parameters for the zinc ions were set based on the hydration free energy parameter set of Merz et al. (54). The tautomeric state with protonated Nδ1 and deprotonated Nϵ2 atoms was used for zinc-coordinating histidine residues. The protonation states of other titratable residues were assigned according to their standard protonation states at pH 7.0. The Na+ and Cl− ions were randomly added to neutralize the system at the salt concentration of 0.15 M. Particle Mesh Ewald was used to calculate long-range electrostatic interactions, and van der Waals interactions were truncated at 12 Å. All bonds were constrained using the SETTLE algorithm with a time step of 2 fs. Temperature was controlled with Langevin dynamics with a damping coefficient of 5 ps−1. The Nosé-Hoover method with a Langevin piston was used to maintain a pressure of 1 atm with an oscillation period of 100 fs and a damping time of 50 fs. After energy minimization, the systems were first heated from 25 K to 298 K with restraints on the Cα atom of the protein in the NVT ensemble, and then were switched to the NPT ensemble. The trajectory was saved at an interval of 0.1 ps, and was continued up to 600 ns for the complex and ∼700 ns for the free protein. For the free protein, the last 600 ns were used for analysis. Computation of Lys/Arg order parameters and conformational entropies from MD trajectories Order parameters for Arg Nϵ-Hϵ and Lys Cϵ-Nζ bond vectors were calculated from the MD trajectories using the auto-correlation function for internal motions (46,55) (1) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} }{}\begin{equation*} {C_I}(t)=\langle{P_2}[{\mu({{t_0}+t})\mu(t)}]\rangle, \end{equation*}\end{document} where μ(t0)μ(t0 + t) is the projection of a unit vector pointing along a bond vector at time t0 onto itself at time t0 + t; P2(x) = (3x2 – 1)/2, is the second Legendre polynomial; and the brackets denote a time average over the trajectory. The trajectory frames were first superimposed onto a reference to remove the effects of overall tumbling. For the free Egr-1 protein, the reference frame was individually defined for each ZF domain because the three ZF domains exhibit virtually independent domain motions (56). For the complex, the reference frame for the auto-correlation function was defined with principal axes of the complex. The time dependence of the autocorrelation for the reorientational motion was analyzed with Clore's extended model-free auto-correlation function (47): (2) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} }{}\begin{equation*} {C_I}(t)={S^2}+\left({1-S_f^2}\right){\rm{exp}}({- t/{\tau _f}}) + \left( {S_f^2 - {S^2}} \right){\rm{exp}}({-t/{\tau _i}}) \end{equation*}\end{document} where Sf2 and τf are the amplitude and correlation time due to fast librational motion, S2 and τi are the order parameter and correlation time of the reorientational motion of a bond vector. Lys and Arg side-chain conformational entropies were calculated from the distributions of the dihedral angles sampled during the simulations (45,57): (3) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} }{}\begin{equation*} {S_{conf}} = - R\int P(\vec{\chi }){\rm{ln}}P( {\vec{\chi }} )d\vec{\chi }, \end{equation*}\end{document} where R is the gas constant; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} }{}$P( {\vec{\chi }} )$\end{document} is the probability density as a function of the dihedral angles χ1, χ2, χ3 and χ4 of each Arg or Lys side chain. The dependence of the entropies on the bin size of the integral mesh was tested and 5° was chosen. The bin size of the integral mesh has an effect on the absolute entropy, but not on the change in conformational entropy upon complex formation. RESULTS We compare the dynamics of basic side chains of the Egr-1 (Zif268) ZF protein in the free state and in the complex with a 12-bp DNA duplex containing the target sequence. In our previous study, we analyzed the dynamics of Lys side chains in the Egr-1–DNA complex (18). In the current study, we conduct the dynamics investigations for the Arg side chains in the free protein and in the complex as well as for the Lys side chains in the free protein. These data allow us to investigate changes in mobility of each basic side chain upon Egr-1's binding to the target DNA. In the following description of the Egr-1 ZF and the DNA duplex, we adopt the residue-numbering schemes shown in Figure 1, as previously defined by Pabo et al. (10,22). NMR spectra for side-chain cationic groups of the free protein and the complex As shown in Figure 2A and B, the Lys NH3+ and Arg Nϵ-Hϵ groups of Egr-1 exhibit well-dispersed signals in the 1H-15N heteronuclear in-phase single quantum coherence (HISQC) spectra for both free and DNA-bound states. For both Arg Nϵ-Hϵ and Lys NH3+ groups, the complex exhibited wider distributions in 15N and 1H chemical shifts, presumably due to formation of hydrogen bonds and/or ion pairs with DNA. The observation of well-isolated signals for the free and DNA-bound states under the identical (for Arg) or similar (for Lys) conditions allowed us to study the change in dynamics of basic side chains upon DNA-binding. Figure 2. Arg and Lys side-chain heteronuclear 1H-15N correlation spectra recorded for the free Egr-1 and for the Egr-1 ZF DNA complex. (A) Lys NH3+-selective 1H-15N heteronuclear in-phase single-quantum coherence (HISQC) spectra (29) recorded for the free and DNA-bound Egr-1 proteins. (B) Arg Nϵ-Hϵ selective 1H-15N HISQC spectra. (C) Examples of 15N longitudinal and transverse relaxation data. Data for the K79 NH3+ and R42 Nϵ-Hϵ groups are shown. Data for the free protein and for the complex are shown with open and closed circles, respectively. (D) CPMG R2 relaxation dispersion data for R80 15Nϵ and K61 15Nζ nuclei of the complex. 15N relaxation data of basic side chains We measured 15N relaxation of Arg Nϵ-Hϵ and Lys NH3+ groups of the free and DNA-bound Egr-1 ZFs. The relaxation experiments for Arg side chains were conducted at the 1H frequencies of 750 and 600 MHz; and those for Lys side chains were conducted at 800 and 600 MHz. The relaxation parameters measured for the Arg and Lys side chains are shown in Supplementary Tables S1, 2 and 3 in the Supplementary Data. Figure 2C shows the 15N longitudinal and transverse relaxation data and best-fit curves for R42 and K79 side chains in the free protein and in the complex, as typical examples. Despite the smaller molecular size of the free protein, precision in the Lys NH3+ relaxation measurements for the free protein was slightly worse than that for the complex, presumably due to the faster hydrogen exchange rates. The 15N relaxation data for the free protein and the complex were clearly different due to different molecular rotational correlation times as well as due to changes in internal motions of the side chains upon Egr-1 binding to DNA. Slow side-chain dynamics detected by CPMG relaxation dispersion experiment To detect slow dynamics on a μs–ms timescale and analyze their contribution (Rex) to 15N R2 relaxation rates, we conducted 15N R2 CPMG relaxation dispersion experiments for Arg 15Nϵ and Lys 15Nζ nuclei. As shown in Figure 2D, the CPMG R2 relaxation dispersion data for Arg 15Nϵ and Lys 15Nζ nuclei showed that R80 and K61 undergo slow dynamics in the complex. By applying the two-state fast-exchange model of Loria et al. (i.e. Equation 2 in Ref. (58)) to these data, the exchange rates for R80 and K61 were calculated to be (0.8 ± 0.4) × 103 s−1 and (1.8 ± 0.5) × 103 s−1, respectively. Interestingly, both of these two side chains interact with the Gua2 nucleotide residue. Because terminal base pairs of DNA are know to fray and transiently break the inter-base hydrogen bonds (59), the adjacent Gua2 residue as well as the interacting protein side chains might be influenced by the fraying events. For these residues in the complex, the exchange contribution to R2 relaxation rates were subtracted from the observed R2 in the subsequent analysis of the side-chain order parameters. For the other residues, the Arg and Lys relaxation dispersion data did not show any evidence of slow dynamics. Order parameters for the Arg and Lys cationic groups Using the 15N relaxation data at two magnetic fields, we determined the order parameters for Arg Nϵ-Hϵ and Lys NH3+ groups of the free Egr-1 protein and the Egr-1–DNA complex. For this analysis, the molecular rotational diffusion parameters were determined from backbone 15N relaxation rates R1 and R2 together with the 1.6-Å-resolution crystal structure of the complex (PDB ID: 1AAY). The rotational diffusion parameters determined for the free protein and the complex are summarized in Supplementary Table S4 in the Supplementary Data. For the free protein, these parameters were calculated individually for each ZF domain, because the three ZF domains tumble almost independently in the free state, as reported for a similar protein involving three Cys2His2-class zinc fingers (56). The side-chain 15N relaxation data together with the rotational diffusion data were used to determine the order parameters of the Arg Nϵ-Hϵ and Lys NH3+ groups in the free protein and in the protein–DNA complex. Table 1 lists the order parameters determined for the Arg and Lys side chains. In the following three subsections, we describe changes in mobility of these side chains upon Egr-1's binding to DNA. We categorize the basic side chains into three classes: (i) those that electrostatically interact with DNA backbone; (ii) those that interact with DNA bases; and (iii) those outside the protein–DNA interfaces. Table 1. Order parameters determined for Arg Nϵ-Hϵ groupsa and Lys NH3+ groupsb of Egr-1 in the free and DNA-bound states by NMR Side chains S2 (free protein) S2 (complex) Electrostatically interact with DNA phosphates R3 0.215 ± 0.021 0.393 ± 0.009 R14 0.370 ± 0.018 0.292 ± 0.004 R27 0.220 ± 0.010 0.883 ± 0.016 R42 0.386 ± 0.023 0.666 ± 0.010 R55 0.241 ± 0.013 0.899 ± 0.021 R70 0.339 ± 0.027 0.302 ± 0.006 R78 0.399 ± 0.195 0.630 ± 0.025 K33 0.314 ± 0.008 0.378 ± 0.003c K61 0.332 ± 0.011 0.335 ± 0.003c K79 0.276 ± 0.024 0.258 ± 0.005c Interact with DNA bases R18 0.295 ± 0.010 0.968 ± 0.020 R24 0.245 ± 0.018 0.968 ± 0.033 R46 0.454 ± 0.022 0.962 ± 0.023 R74 0.306 ± 0.012 0.908 ± 0.039 R80 0.238 ± 0.034 0.897 ± 0.015 Outside the interfaces R15 0.462 ± 0.027 0.351 ± 0.004 R38 0.228 ± 0.010 0.108 ± 0.002 R87 0.132 ± 0.024 0.085 ± 0.003 K71 0.279 ± 0.005 0.219 ± 0.003c K83 0.249 ± 0.024 0.322 ± 0.006c K89 0.134 ± 0.003 0.075 ± 0.003c aOrder parameters were determined from 15N relaxation data at 1H-frequencies of 750 and 600 MHz. Arg side-chain 15N relaxation parameters are reported in Supplementary Table S1. bOrder parameters S2 for the NH3+ symmetry axis (i.e. the Cϵ-Nζ bond vector) were determined from 15N relaxation data at 1H-frequencies of 800 and 600 MHz. Lys side-chain 15N relaxation parameters are reported in Supplementary Table S2. cThe Lys NH3+ order parameters for the complex are from Ref. (18). Change in mobility of basic side chains that electrostatically interact with DNA phosphates The crystal structures of the Egr-1–DNA complexes (10,22) show short-range electrostatic interactions with the DNA backbone for seven Arg side chains (R3, R14, R27, R42, R55, R70 and R78) and three Lys side chains (K33, K61 and K79). Changes in the order parameters of these cationic groups upon Egr-1's binding to DNA are shown in red in Figure 3A and B. The majority of Arg side chains (i.e. R3, R27, R42, R55 and R78) showed a large increase (by >0.1) in the Nϵ-Hϵ order parameter upon the complex formation, indicating that these side-chains become significantly less mobile due to the interactions with DNA. Interestingly, all Lys NH3+ groups and 2 Arg Nϵ-Hϵ groups (R14 and R70) showed no or only marginal changes in their order parameters upon Egr-1's binding to DNA, indicating that the side chains retain high mobility even in the complex. For K79, a 1H-31P heteronuclear correlation cross peak arising from the hydrogen-bond scalar coupling between the 31P and 15N nuclei was clearly observed in the H3(N)P spectrum (Figure 3C), indicating the presence of the contact ion pair (CIP) of this side chain and DNA phosphate. Perhaps surprisingly, the K79 NH3+ group exhibits virtually no change in the order parameter upon DNA-binding, despite the presence of the CIP state in the complex. This high mobility is likely due to the dynamic equilibria between the CIP and solvent-separated ion-pair states (4,18). Figure 3. Binding-induced changes in mobility for the Arg Nϵ-Hϵ and Lys NH3+ groups. (A) Changes in the Arg Nϵ-Hϵ order parameters upon Egr-1's binding to DNA. (B) Changes in the Lys NH3+ order parameters upon Egr-1's binding to DNA. Red, side chains that form intermolecular ion pairs with DNA phosphates; magenta, side chains that form contacts with DNA bases; and blue, side chains located outside the protein–DNA interfaces. (C) The H3(N)P spectrum (37) recorded for the Lys NH3+ groups of the complex, which indicates the presence of the contact ion pair formed by K79 and DNA phosphate. The 31P chemical shift is referenced to trimethyl phosphate. Change in mobility of basic side chains that interact with DNA bases In the crystal structures of the Egr-1–DNA complexes, 5 Arg side chains (R18, R24, R46, R74 and R80), but no Lys side chain directly interact with DNA bases. Changes in the Nϵ-Hϵ order parameters for these Arg side chains are shown in magenta in Figure 3A. The guanidino groups of these Arg side chains form two hydrogen bonds with a guanine base for each (Gua10, Gua8, Gua7, Gua4 and Gua2, respectively; see Figure 3), representing the canonical pattern of guanine recognition by Arg side chain (1,60). Upon formation of the complex with DNA, these Arg side chains exhibited a substantial increase (by >0.5) in the Nϵ-Hϵ order parameter S2, indicating that their mobility is substantially restricted by the interactions with DNA. This strong immobilization is likely due to the two hydrogen bonds at distinct N atoms of the guanidino groups as well as the cation–π interaction (61) with the adjacent base aromatic ring onto which the cationic group stacks. In addition to the hydrogen bonds with guanine bases, R18, R46 and R74 side chains also form two more hydrogen bonds with an aspartate side chain (i.e. D20, D48 and D76, respectively) (see the scheme in Figure 3). As rigidification of a ligand often increases binding affinity through a decrease in entropic loss upon complex formation (62,63), one might consider that the role of the auxiliary Asp-Arg ion-pair formation at the interface with DNA bases might be to rigidify the Arg side chains to the active conformation in the free protein. However, our NMR data suggest that this is not the case. In fact, the order parameters indicate that R18, R46 and R74 side chains are mobile in the free state (see Table 1). Changes in mobility of basic side chains outside the protein–DNA interfaces In the crystal structures, R15, R38, K71, K83, R87 and K89 are located outside the protein–DNA interfaces and do not directly interact with DNA. Binding-induced changes in the order parameters of these side chains are shown in blue in Figure 3A and B. Egr-1's binding to DNA did not give a significant impact on mobility for a majority of these side chains (i.e. K71, K83, R87 and K89). This is reasonable because they are far from the binding interfaces. Upon binding, R15 and R38 side chains became slightly (but to a statistically significant degree) more mobile. The reason for this mobility is unclear, but might be related to conformational changes of nearby residues. It should be noted that binding-induced enhancement of mobility were previously reported for backbone amide and side-chain methyl groups of other proteins (e.g. Refs. (9,43,64)). The increase in mobility may partially compensate the entropic loss arising from immobilization of many interfacial side chains. Comparison with MD simulations To gain more insight into the side-chain dynamics of Arg and Lys residues, we analyzed dynamic behavior of each basic side chain from the MD simulations. Because we previously obtained a 600-ns MD trajectory for the Egr-1–DNA complex (18), we carried out a corresponding MD simulation of the same length for the free protein in this study. The MD trajectories provide atomic details of side-chain motions and show the contribution from transient interactions, which are not seen in the crystal structures. The direct contacts of these Arg and Lys side chains are summarized in Supplementary Table S5. The mean lifetimes of the direct contacts between the protein and DNA are on the pico to nano-second timescale (see Supplementary Table S6). Arg is much stronger than Lys in participating in direct contacts with the DNA. In addition, based on electrostatic interactions with the phosphate groups of the DNA, the mean lifetimes of the direct contact of Arg appear longer than Lys, which also indicates that these Arg side chains are less dynamic than Lys upon the complex formation. Using the MD trajectories, we also calculated the order parameters for Arg Nϵ-Hϵ and Lys Cϵ-Nζ bond vectors in the free and DNA-bound states of Egr-1. Values of the MD order parameters for Arg and Lys side chains are shown in Supplementary Table S7. Figure 4A shows the correlation between binding-induced changes in NMR and MD order parameters. As seen in the results of NMR-based order parameters, the results from the MD trajectory showed a similar trend with relatively large changes in Arg Nϵ-Hϵ order parameters and with relatively small changes in Lys Cϵ-Nζ order parameters. R24, R27 and R80 gave outliers in the correlation between the experimental data and computation (Figure 4A). For these side chains, although NMR data show a significant increase in Nϵ-Hϵ order parameter (i.e. immobilization), the MD simulations show a significantly smaller change. Interestingly, however, even for R24 and R80, the conformational entropy of a whole side chain (not just Nϵ-Hϵ) showed substantial decreases upon the complex formation (see the following subsection), and in this sense, the computational data were consistent with the experimental observation of the immobilization. The lower MD-derived order parameter of R80 in the complex could be due to the lack of appropriate representation of cation–π interactions in the classical MD force field (as discussed by Schulten et al. in Ref. (61)) and/or to fraying of the terminal AT base pair. The intermittent breaking of hydrogen bonds and opening of the terminal base pair have been observed and studied by NMR (59), time-resolved Stokes shifts (65) and computer simulations (66). For R24 and R27, the MD-derived order parameters in the free state were significantly larger than those determined by experiment. This might be related to sampling errors, i.e. lack of convergence of time correlation function due to insufficient conformational/configurational sampling. Long time convergence (100's of ns) of simulations are important for statistical agreement with experimental data for processes on a nanosecond timescale (67). Nonetheless, Figure 4A shows a good correlation between the computational and experimental data for the majority of the basic side chains (18 out of 21), for which the root mean squared difference was 0.19. This supports reliability of the model simulations and justifies further analysis. Figure 4. MD trajectory analysis of binding-induced changes in conformational dynamics of the Arg and Lys side chains. (A) Comparison of the MD-derived and NMR-derived changes in Arg and Lys side-chain order parameters upon Egr-1's binding to DNA. Data points for Arg and Lys side chains are shown in open and closed circles, respectively. Uncertainties in the NMR-derived changes are shown in Figure 3A and B. (B) Binding-induced changes in Arg and Lys side-chain conformational entropies. Each entropy value was calculated from the probability distributions of the dihedral angles χ1, χ2, χ3 and χ4 in the MD trajectories for the free protein and the Egr-1–DNA complex. Red, side chains that form intermolecular ion pairs with DNA phosphates; magenta, side chains that contact with DNA bases; blue, side chains located outside the protein–DNA interfaces. Error bars represent standard block errors estimated from calculations for independent 50 ns blocks in the 600 ns trajectories for the free protein and the complex. Loss in side-chain conformational entropy of each basic residue upon binding While the order parameters of Arg Nϵ-Hϵ and Lys NH3+ groups provide information on dynamics of individual side-chain cationic groups, these NMR data do not necessarily reflect mobility in the other parts of Arg and Lys side chains. There are some theoretical models for the relationship between NMR order parameters and conformational entropy (68–70). However, it was proposed that side-chain conformational entropy is not necessarily predictable from NMR order parameters for terminal moieties of long side chains alone because middle parts of the same side chains could remain mobile even if the termini are immobilized (45,71). Since the changes in MD-based order parameters for the cationic moieties were qualitatively consistent with experimental data, we examined Arg and Lys side-chain conformational entropies for the free protein and for the complex using the MD trajectories. To assess the thermodynamic consequences of immobilizing the side-chain cationic groups upon DNA-binding, we calculated side-chain conformational entropy for each basic side chain from the MD trajectories (individual values are reported in Supplementary Table S8 in the Supplementary Data). Figure 4B shows the computed changes in side-chain conformation entropy upon Egr-1's binding to the target DNA. The non-interfacial Arg and Lys side chains showed only marginal changes in conformational entropy. In contrast, many interfacial Arg side chains exhibited significant loss of conformation entropy by ∼8–19 J/mol/K upon protein–DNA association. However, the corresponding entropic loss was smaller for interfacial Lys side chains: K33 and K61 exhibited virtually no loss and K79 exhibited an entropic loss of 8.1 J/mol/K. Thus, these entropic data also illustrate different characteristics of the interfacial Arg and Lys side chains. DISCUSSION Rigid and dynamic interfaces via basic side chains This study demonstrates the diverse dynamic properties of the protein–DNA interfaces via basic side chains. The Arg side chains interacting with DNA bases are strongly immobilized and form rigid interfaces. In contrast, despite the strong short-range electrostatic interactions, the majority of the basic side chains interacting with the DNA phosphates are relatively mobile and form dynamic interfaces. In particular, Lys side-chain NH3+ groups retain high mobility even in the DNA-bound state. Thus, DNA recognition by Egr-1 involves both rigid and dynamic interfaces of the basic side chains. High mobility retained by interfacial Lys side chains It should be entropically favorable that interfacial Lys side chains retain substantial mobility in the DNA-bound state. The retained mobility of Lys side chains could be general in protein–DNA interactions. For example, our previous studies on the HoxD9–DNA and Antp–DNA complexes (18,37,40,41) showed that the interfacial Lys side-chain NH3+ groups at the molecular interfaces are also highly mobile with S2axis < 0.6. The small order parameters suggest that binding-induced change in mobility is relatively small for these interfacial Lys NH3+ groups as well, though the side-chain dynamics of the HoxD9 and Antp proteins in the free state remain to be investigated. Different characteristics of Arg and Lys side-chain interactions with DNA Our data demonstrate the characteristic difference between the interfacial Arg and Lys side chains in the dynamics of DNA recognition. The observed difference can be due to several factors: differences in ability to form hydrogen bonding clusters (1), in charge density, sterics and in desolvation energy (72). For example, pivotal motions through side-chain bond rotations remain possible with a single hydrogen bond, whereas such motions become more difficult with multiple hydrogen bonds. While the multiple hydrogen bonds of an Arg guanidino group with DNA could be favorable in terms of binding enthalpy, the stronger conformational restriction should cause substantial loss in conformational entropy for Arg side chains. In contrast, the Lys side chain possesses only a single charged donor ammonium group, but can adopt various conformations, without substantial loss in side-chain conformational entropy. This might be partially responsible for different spatial distributions of Arg and Lys side chains at interfaces with DNA. Statistical investigations of the 3D structures of protein-DNA complexes showed that the interactions with DNA minor groove prefer Arg side chains over Lys side chains (72). Roh et al. discussed that this preference could be at least partly due to lower desolvation energy for Arg side chains (72). Based on our current data, we speculate that this preference could also relate to the different dynamic properties of Arg and Lys side chains. Insertion in DNA minor groove might diminish the advantage of a Lys side chain in terms of side-chain conformational entropy, because the narrow space in the minor groove would not allow for wide conformational sampling. CONCLUDING REMARKS Our current study delineates the dynamics of the basic side chains in DNA recognition by Egr-1. The Arg side chains interacting with DNA bases are more strongly immobilized and form rigid interfaces. The basic side chains interacting with the DNA phosphates are relatively mobile and form dynamic interfaces. In particular, Lys side-chain NH3+ groups retain high mobility even in the DNA-bound state. Although the Arg side chains can form a larger number of hydrogen bonds, the strong restriction of their mobility renders substantial loss in side-chain conformational entropy. Our data provide atomic-level information of structural dynamics and thermodynamics of the interfacial Arg and Lys side chains in the DNA-binding event. Although Arg or Lys side-chain dynamics were previously studied for some protein–nucleic acid complexes (18,37,40,41,73–75), the binding-induced changes in the dynamics remained to be delineated. Our current work provides the comprehensive experimental data on changes in Arg and Lys side-chain dynamics upon protein–DNA complex formation. To conclude whether or not the characteristic difference between Arg and Lys side chains are general in molecular recognition of nucleic acids by proteins, we require further investigations of other systems. Supplementary Material SUPPLEMENTARY DATA We thank the staff of the Sealy Center for Structural Biology, in particular Dr Tianzhi Wang for the maintenance of the NMR equipment; Dr Ka-yiu Wong for helpful discussions; and Dr Karon Cassidy for editing the manuscript. A portion of the computational research was carried out through the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1053575. SUPPLEMENTARY DATA Supplementary Data are available at NAR Online. FUNDING National Institutes of Health [R01-GM105931 to J.I.; R01-GM107590 to J.I.; R01-GM066813 to B.M.P.]; Welch foundation [H-0037 to B.M.P.]. Funding for open access charge: National Institutes of Health [R01-GM105931 to J.I.]. Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Luscombe N.M. Laskowski R.A. Thornton J.M. Amino acid-base interactions: a three-dimensional analysis of protein-DNA interactions at an atomic level Nucleic Acids Res. 2001 29 2860 2874 11433033 2. Nadassy K. Wodak S.J. Janin J. Structural features of protein-nucleic acid recognition sites Biochemistry 1999 38 1999 2017 10026283 3. Rohs R. Jin X. West S.M. Joshi R. Honig B. Mann R.S. Origins of specificity in protein-DNA recognition Annu. Rev. Biochem. 2010 79 233 269 20334529 4. Iwahara J. Esadze A. Zandarashvili L. Physicochemical properties of ion pairs of biological macromolecules Biomolecules 2015 5 2435 2463 26437440 5. Privalov P.L. Dragan A.I. Crane-Robinson C. Interpreting protein/DNA interactions: distinguishing specific from non-specific and electrostatic from non-electrostatic components Nucleic Acids Res. 2011 39 2483 2491 21071403 6. Record M.T. Jr Ha J.H. Fisher M.A. Analysis of equilibrium and kinetic measurements to determine thermodynamic origins of stability and specificity and mechanism of formation of site-specific complexes between proteins and helical DNA Methods Enzymol. 1991 208 291 343 1779839 7. Frederick K.K. Marlow M.S. Valentine K.G. Wand A.J. Conformational entropy in molecular recognition by proteins Nature 2007 448 325 329 17637663 8. Marlow M.S. Dogan J. Frederick K.K. Valentine K.G. Wand A.J. The role of conformational entropy in molecular recognition by calmodulin Nat. Chem. Biol. 2010 6 352 358 20383153 9. Tzeng S.R. Kalodimos C.G. Protein activity regulation by conformational entropy Nature 2012 488 236 240 22801505 10. Pavletich N.P. Pabo C.O. Zinc finger-DNA recognition: crystal structure of a Zif268-DNA complex at 2.1 A Science 1991 252 809 817 2028256 11. Elrod-Erickson M. Pabo C.O. Binding studies with mutants of Zif268. Contribution of individual side chains to binding affinity and specificity in the Zif268 zinc finger-DNA complex J. Biol. Chem. 1999 274 19281 19285 10383437 12. Esadze A. Iwahara J. Stopped-flow fluorescence kinetic study of protein sliding and intersegment transfer in the target DNA search process J. Mol. Biol. 2014 426 230 244 24076422 13. Hamilton T.B. Borel F. Romaniuk P.J. Comparison of the DNA binding characteristics of the related zinc finger proteins WT1 and EGR1 Biochemistry 1998 37 2051 2058 9485332 14. Bozon B. Davis S. Laroche S. A requirement for the immediate early gene zif268 in reconsolidation of recognition memory after retrieval Neuron 2003 40 695 701 14622575 15. Lee J.L. Everitt B.J. Thomas K.L. Independent cellular processes for hippocampal memory consolidation and reconsolidation Science 2004 304 839 843 15073322 16. Khachigian L.M. Lindner V. Williams A.J. Collins T. Egr-1-induced endothelial gene expression: a common theme in vascular injury Science 1996 271 1427 1431 8596917 17. Yan S.F. Fujita T. Lu J. Okada K. Shan Zou Y. Mackman N. Pinsky D.J. Stern D.M. Egr-1, a master switch coordinating upregulation of divergent gene families underlying ischemic stress Nat. Med. 2000 6 1355 1361 11100120 18. Chen C.Y. Esadze A. Zandarashvili L. Nguyen D. Pettitt B.M. Iwahara J. Dynamic Equilibria of Short-Range Electrostatic Interactions at Molecular Interfaces of Protein-DNA Complexes J. Phys. Chem. Lett. 2015 6 2733 2737 26207171 19. Esadze A. Kemme C.A. Kolomeisky A.B. Iwahara J. Positive and negative impacts of nonspecific sites during target location by a sequence-specific DNA-binding protein: origin of the optimal search at physiological ionic strength Nucleic Acids Res. 2014 42 7039 7046 24838572 20. Zandarashvili L. Esadze A. Vuzman D. Kemme C.A. Levy Y. Iwahara J. Balancing between affinity and speed in target DNA search by zinc-finger proteins via modulation of dynamic conformational ensemble Proc. Natl. Acad. Sci. U.S.A. 2015 112 E5142 E5149 26324943 21. Zandarashvili L. Vuzman D. Esadze A. Takayama Y. Sahu D. Levy Y. Iwahara J. Asymmetrical roles of zinc fingers in dynamic DNA-scanning process by the inducible transcription factor Egr-1 Proc. Natl. Acad. Sci. U.S.A. 2012 109 E1724 E1732 22675124 22. Elrod-Erickson M. Rould M.A. Nekludova L. Pabo C.O. Zif268 protein-DNA complex refined at 1.6 Å: a model system for understanding zinc finger-DNA interactions Structure 1996 4 1171 1180 8939742 23. Hashimoto H. Olanrewaju Y.O. Zheng Y. Wilson G.G. Zhang X. Cheng X. Wilms tumor protein recognizes 5-carboxylcytosine within a specific DNA sequence Genes Dev. 2014 28 2304 2313 25258363 24. Zandarashvili L. White M.A. Esadze A. Iwahara J. Structural impact of complete CpG methylation within target DNA on specific complex formation of the inducible transcription factor Egr-1 FEBS Lett. 2015 589 1748 1753 25999311 25. Durai S. Mani M. Kandavelou K. Wu J. Porteus M.H. Chandrasegaran S. Zinc finger nucleases: custom-designed molecular scissors for genome engineering of plant and mammalian cells Nucleic Acids Res. 2005 33 5978 5990 16251401 26. Pabo C.O. Peisach E. Grant R.A. Design and selection of novel Cys2His2 zinc finger proteins Annu. Rev. Biochem. 2001 70 313 340 11395410 27. Urnov F.D. Rebar E.J. Holmes M.C. Zhang H.S. Gregory P.D. Genome editing with engineered zinc finger nucleases Nat. Rev. Genet. 2010 11 636 646 20717154 28. Takayama Y. Sahu D. Iwahara J. NMR studies of translocation of the Zif268 protein between its target DNA Sites Biochemistry 2010 49 7998 8005 20718505 29. Iwahara J. Jung Y.S. Clore G.M. Heteronuclear NMR spectroscopy for lysine NH3 groups in proteins: unique effect of water exchange on 15 N transverse relaxation J. Am. Chem. Soc. 2007 129 2971 2980 17300195 30. Delaglio F. Grzesiek S. Vuister G.W. Zhu G. Pfeifer J. Bax A. NMRPipe—a Multidimensional Spectral Processing System Based on Unix Pipes J. Biomol. NMR 1995 6 277 293 8520220 31. Johnson B.A. Blevins R.A. Nmr View—a computer-program for the visualization and analysis of Nmr data J. Biomol. NMR 1994 4 603 614 22911360 32. Clore G.M. Gronenborn A.M. Determining the structures of large proteins and protein complexes by NMR Trends Biotechnol. 1998 16 22 34 9470228 33. Iwahara J. Clore G.M. Sensitivity improvement for correlations involving arginine side-chain Nε/Hε resonances in multi-dimensional NMR experiments using broadband 15 N 180° pulses J. Biomol. NMR 2006 36 251 257 17036160 34. Esadze A. Zandarashvili L. Iwahara J. Effective strategy to assign 1 H- 15 N heteronuclear correlation NMR signals from lysine side-chain NH3 + groups of proteins at low temperature J. Biomol. NMR 2014 60 23 27 25129623 35. Kupče E. Boyd J. Campbell I.D. Short selective pulses for biochemical applications J. Magn. Reson. B 1995 106 300 303 7719630 36. Hansen D.F. Vallurupalli P. Kay L.E. An improved 15N relaxation dispersion experiment for the measurement of millisecond time-scale dynamics in proteins J. Phys. Chem. B 2008 112 5898 5904 18001083 37. Anderson K.M. Esadze A. Manoharan M. Brüschweiler R. Gorenstein D.G. Iwahara J. Direct observation of the ion-pair dynamics at a protein-DNA interface by NMR spectroscopy J. Am. Chem. Soc. 2013 135 3613 3619 23406569 38. Esadze A. Li D.W. Wang T. Brüschweiler R. Iwahara J. Dynamics of lysine side-chain amino groups in a protein studied by heteronuclear 1 H-15 N NMR spectroscopy J. Am. Chem. Soc. 2011 133 909 919 21186799 39. Zandarashvili L. Esadze A. Iwahara J. NMR studies on the dynamics of hydrogen bonds and ion pairs involving lysine side chains of proteins Adv. Protein Chem. Struct. Biol. 2013 93 37 80 24018322 40. Zandarashvili L. Nguyen D. Anderson K.M. White M.A. Gorenstein D.G. Iwahara J. Entropic enhancement of protein-DNA affinity by oxygen-to-sulfur substitution in DNA phosphate Biophys. J. 2015 109 1026 1037 26331260 41. Zandarashvili L. Iwahara J. Temperature dependence of internal motions of protein side-chain NH3 + groups: insight into energy barriers for transient breakage of hydrogen bonds Biochemistry 2015 54 538 545 25489884 42. Woessner D.E. Nuclear spin relaxation in ellipsoids undergoing rotational Brownian motion J. Chem. Phys. 1962 37 647 654 43. Iwahara J. Peterson R.D. Clubb R.T. Compensating increases in protein backbone flexibility occur when the Dead ringer AT-rich interaction domain (ARID) binds DNA: a nitrogen-15 relaxation study Protein Sci. 2005 14 1140 1150 15802641 44. Tjandra N. Feller S.E. Pastor R.W. Bax A. Rotational diffusion anisotropy of human ubiquitin from N-15 NMR relaxation J. Am. Chem. Soc. 1995 117 12562 12566 45. Trbovic N. Cho J.H. Abel R. Friesner R.A. Rance M. Palmer A.G. 3rd Protein side-chain dynamics and residual conformational entropy J. Am. Chem. Soc. 2009 131 615 622 19105660 46. Lipari G. Szabo A. Model-free approach to the interpretation of nuclear magnetic-resonance relaxation in macromolecules.1. Theory and range of validity J. Am. Chem. Soc. 1982 104 4546 4559 47. Clore G.M. Szabo A. Bax A. Kay L.E. Driscoll P.C. Gronenborn A.M. Deviations from the simple 2-parameter model-free approach to the interpretation of N-15 nuclear magnetic-relaxation of proteins J. Am. Chem. Soc. 1990 112 4989 4991 48. d'Auvergne E.J. Gooley P.R. The use of model selection in the model-free analysis of protein dynamics J. Biomol. NMR 2003 25 25 39 12566997 49. Phillips J.C. Braun R. Wang W. Gumbart J. Tajkhorshid E. Villa E. Chipot C. Skeel R.D. Kale L. Schulten K. Scalable molecular dynamics with NAMD J. Comput. Chem. 2005 26 1781 1802 16222654 50. Foloppe N. MacKerell A.D. All-atom empirical force field for nucleic acids: I. Parameter optimization based on small molecule and condensed phase macromolecular target data J. Comput. Chem. 2000 21 86 104 51. MacKerell A.D. Bashford D. Bellott M. Dunbrack R.L. Evanseck J.D. Field M.J. Fischer S. Gao J. Guo H. Ha S. All-atom empirical potential for molecular modeling and dynamics studies of proteins J. Phys. Chem. B 1998 102 3586 3616 24889800 52. Mackerell A.D. Jr Feig M. Brooks C.L. 3rd Extending the treatment of backbone energetics in protein force fields: limitations of gas-phase quantum mechanics in reproducing protein conformational distributions in molecular dynamics simulations J. Comput. Chem. 2004 25 1400 1415 15185334 53. Foloppe N. Sagemark J. Nordstrand K. Berndt K.D. Nilsson L. Structure, dynamics and electrostatics of the active site of glutaredoxin 3 from Escherichia coli: comparison with functionally related proteins J. Mol. Biol. 2001 310 449 470 11428900 54. Li P. Roberts B.P. Chakravorty D.K. Merz K.M. Jr Rational Design of Particle Mesh Ewald Compatible Lennard-Jones Parameters for +2 Metal Cations in Explicit Solvent J. Chem. Theory Comput. 2013 9 2733 2748 23914143 55. Case D.A. Molecular dynamics and NMR spin relaxation in proteins Acc. Chem. Res. 2002 35 325 331 12069616 56. Brüschweiler R. Liao X. Wright P.E. Long-range motional restrictions in a multidomain zinc-finger protein from anisotropic tumbling Science 1995 268 886 889 7754375 57. Karplus M. Kushick J.N. Method for estimating the configurational entropy of macromolecules Macromolecules 1981 14 325 332 58. Loria J.P. Rance M. Palmer A.G. A relaxation-compensated Carr-Purcell-Meiboom-Gill sequence for characterizing chemical exchange by NMR spectroscopy J. Am. Chem. Soc. 1999 121 2331 2332 59. Nonin S. Leroy J.L. Gueron M. Terminal base pairs of oligodeoxynucleotides: imino proton exchange and fraying Biochemistry 1995 34 10652 10659 7654719 60. Seeman N.C. Rosenberg J.M. Rich A. Sequence-specific recognition of double helical nucleic acids by proteins Proc. Natl. Acad. Sci. U.S.A. 1976 73 804 808 1062791 61. Zou X. Ma W. Solov'yov I.A. Chipot C. Schulten K. Recognition of methylated DNA through methyl-CpG binding domain proteins Nucleic Acids Res. 2012 40 2747 2758 22110028 62. Chaires J.B. Calorimetry and thermodynamics in drug design Annu. Rev. Biophys. 2008 37 135 151 18573076 63. Velazquez-Campoy A. Todd M.J. Freire E. HIV-1 protease inhibitors: enthalpic versus entropic optimization of the binding affinity Biochemistry 2000 39 2201 2207 10694385 64. Lee A.L. Kinnear S.A. Wand A.J. Redistribution and loss of side chain entropy upon formation of a calmodulin-peptide complex Nat. Struct. Biol. 2000 7 72 77 10625431 65. Andreatta D. Sen S. Perez Lustres J.L. Kovalenko S.A. Ernsting N.P. Murphy C.J. Coleman R.S. Berg M.A. Ultrafast dynamics in DNA: “fraying" at the end of the helix J. Am. Chem. Soc. 2006 128 6885 6892 16719468 66. Zgarbova M. Otyepka M. Sponer J. Lankas F. Jurecka P. Base pair fraying in molecular dynamics simulations of DNA and RNA J. Chem. Theory Comput. 2014 10 3177 3189 26588288 67. Bowman G.R. Accurately modeling nanosecond protein dynamics requires at least microseconds of simulation J. Comput. Chem. 2016 37 558 566 26077712 68. Akke M. Brüschweiler R. Palmer A.G. NMR order parameters and free-energy—an analytical approach and its application to cooperative Ca2+ binding by calbindin-D9k J. Am. Chem. Soc. 1993 115 9832 9833 69. Li Z.G. Raychaudhuri S. Wand A.J. Insights into the local residual entropy of proteins provided by NMR relaxation Protein Sci. 1996 5 2647 2650 8976574 70. Yang D.W. Kay L.E. Contributions to conformational entropy arising from bond vector fluctuations measured from NMR-derived order parameters: application to protein folding J. Mol. Biol. 1996 263 369 382 8913313 71. Li D.W. Brüschweiler R. A dictionary for protein side-chain entropies from NMR order parameters J. Am. Chem. Soc. 2009 131 7226 7227 19422234 72. Rohs R. West S.M. Sosinsky A. Liu P. Mann R.S. Honig B. The role of DNA shape in protein-DNA recognition Nature 2009 461 1248 1253 19865164 73. Berglund H. Baumann H. Knapp S. Ladenstein R. Härd T. Flexibility of an arginine side chain at a DNA-protein interface J. Am. Chem. Soc. 1995 117 12883 12884 74. Wilkinson T.A. Botuyan M.V. Kaplan B.E. Rossi J.J. Chen Y. Arginine side-chain dynamics in the HIV-1 rev-RRE complex J. Mol. Biol. 2000 303 515 529 11054288 75. Wilkinson T.A. Zhu L. Hu W. Chen Y. Retention of conformational flexibility in HIV-1 Rev-RNA complexes Biochemistry 2004 43 16153 16160 15610009

PMC005xxxxxx/PMC5001610.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2731769810.1093/nar/gkw567Nucleic Acid EnzymesBiochemical and genetic analysis of the role of the viral polymerase in enterovirus recombination Woodman Andrew 2†Arnold Jamie J. 2†Cameron Craig E. 2Evans David J. 1*1 Biomedical Sciences Research Complex, North Haugh, University of St. Andrews, St. Andrews KY16 9ST, UK2 Dept. of Biochemistry & Molecular Biology, 201 Althouse Lab, University Park, PA 16802, USA* To whom correspondence should be addressed. Tel: +44 1334 463396; Fax: +44 1334 462595; Email: d.j.evans@st-andrews.ac.uk† These authors contributed equally to this study as First Authors.19 8 2016 17 6 2016 17 6 2016 44 14 6883 6895 14 6 2016 07 6 2016 11 4 2016 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Genetic recombination in single-strand, positive-sense RNA viruses is a poorly understand mechanism responsible for generating extensive genetic change and novel phenotypes. By moving a critical cis-acting replication element (CRE) from the polyprotein coding region to the 3′ non-coding region we have further developed a cell-based assay (the 3′CRE-REP assay) to yield recombinants throughout the non-structural coding region of poliovirus from dually transfected cells. We have additionally developed a defined biochemical assay in which the only protein present is the poliovirus RNA dependent RNA polymerase (RdRp), which recapitulates the strand transfer events of the recombination process. We have used both assays to investigate the role of the polymerase fidelity and nucleotide turnover rates in recombination. Our results, of both poliovirus intertypic and intratypic recombination in the CRE-REP assay and using a range of polymerase variants in the biochemical assay, demonstrate that RdRp fidelity is a fundamental determinant of recombination frequency. High fidelity polymerases exhibit reduced recombination and low fidelity polymerases exhibit increased recombination in both assays. These studies provide the basis for the analysis of poliovirus recombination throughout the non-structural region of the virus genome and provide a defined biochemical assay to further dissect this important evolutionary process. cover-date19 August 2016 ==== Body INTRODUCTION Recombination in single-stranded, positive-sense RNA viruses is a relatively poorly understood driver of extensive genetic change. This group of viruses, typified by poliovirus, exist as viral quasispecies as a consequence of misincorporations by their error-prone RNA dependent RNA polymerases (RdRp) during genome replication coupled with strand transfer events that create hybrids (recombinants) between two viruses replicating in the same cell. As well as being a driver of genetic variation, recombination may have evolved to ‘rescue’ genomes from deleterious mutations that accumulate during error-prone replication (1) though this may only be important for small population sizes such as might occur during inter-host transmission (2). Poliovirus, the prototype picornavirus and aetiological agent of paralytic poliomyelitis, provides a tractable experimental system to study recombination. The 7.5 kb genome encodes a single polyprotein, flanked by non-coding regions (NCR), that is co- and post-translationally processed to generate the structural proteins (the P1 proteins; VP4, VP2, VP3 and VP1) which assemble to form the icosahedral capsid and the non-structural proteins (the P2 and P3 proteins; 2Apro, 2B, 2C, 3A, 3BVPg, 3Cpro and 3Dpol) that subvert the cellular environment and replicate the virus genome. The genome can essentially be considered modular, consisting of structural (P1) and replication (P2 and P3) components, together with the flanking translation and replication determinants occupying the 5′ and 3′ NCRs (3,4). This modularity is emphasized in in vitro studies in which viable recombinants have been engineered or selected (5,6). More compellingly, studies of paralysis induced by vaccine derived recombinant polioviruses (VDRP)—in which the live attenuated Sabin vaccine strain has recombined with a co-circulating species C enterovirus—have demonstrated the importance of recombination in vivo (7). In related enteroviruses, recombinant forms—defined by serotype according to their capsid proteins—have been shown to emerge, prevail and then disappear in temporal epidemiological surveys of globally-distributed serotypes (8–10). In these examples the recombinants are pathogenic (generally associated with poliomyelitis, a range of acute flaccid paralyses or viral encephalitis) and demonstrate maintenance and transmission of the capsid in the population by a range of non-structural ‘modules’ from genetically-related but divergent viruses. An improved mechanistic understanding of recombination is required, both to comprehend the process as a driver of evolutionary change and to develop strategies to prevent or mitigate the consequences of recombination, for example by the design of non-recombinogenic live-attenuated vaccines. We have recently described an in vitro assay (designated the CRE-REP assay) that enabled the identification of early recombination products (11). Briefly, the assay involves two poliovirus genomes each containing a different deleterious (and non-reverting) modification that prevents the production of viable progeny. One genome, a sub-genomic replicon (12), was replication competent but did not encode capsid proteins. The other contained a well-understood mutation (designated SL3) in a critical cis-acting replication element (CRE), a defined stemloop structure essential for positive-strand replication (13,14). Co-transfection of murine cells—permissive for infection, but not-susceptible due to the absence of the poliovirus receptor —resulted in the recovery of viable progeny following a recombination strand transfer event between the non-structural protein (P2 and P3) modules of the sub-genomic replicon (the polymerase donor) and the structural (P1) module of the CRE-defective SL3 genome. By using genomes of different poliovirus serotypes with divergent sequences it was possible to unambiguously identify the recombination junction. To our surprise, the majority of initial viable recombinants were ‘imprecise’ and contained an in-frame partial genome duplication. Subsequent serial passage resulted in the incremental loss of the genome duplication, leaving a range of ‘precise’ junctions with no additional sequences (11). These observations strongly suggest that recombination may be biphasic, involving the formation of an initial imprecise product that, through a process we termed resolution, yields the genome-length recombinants that are typically isolated in vivo. Further studies demonstrated that modifying the polymerase fidelity significantly influenced the yield of recombinants, an observation supported by a recent study in alphaviruses (15). This indicated both a fundamental role for the polymerase in the process of recombination and that the process was replicative and therefore different from previously described non-replicative recombination (16,17). Although the CRE-REP assay enables the role of the polymerase in recombination to be analysed it does not demonstrate that the polymerase is sufficient for strand exchange. Furthermore, although modification of the polymerase fidelity influences recombination yield in the CRE-REP assay, this does not exclude the possibility that this is an indirect consequence of the polymerase interacting with cellular components of the replication machinery. Recombination occurs within membrane-bound replication complexes (RCs), as shown by elegant imaging studies and the demonstration that nocodazole—a drug that prevents RC coalescence—inhibits the process (11,18). Since the components of the RC are not fully understood we reasoned that a defined in vitro recombination assay would enable the role of viral, and possibly cellular, components to be more readily delineated. Using such an assay we demonstrate here that the viral polymerase alone is sufficient for the strand transfer reaction. In addition, by analysis of variants of the polymerase with well-characterised changes to the enzymes’ fidelity or nucleotide turnover, we demonstrate both that this biochemically defined in vitro assay recapitulates aspects of the CRE-REP assay and that the fidelity of the viral polymerase is a key determinant of the recombination process. In support of this conclusion we demonstrate that the majority of in vitro-generated transfer products contain additional nucleotides or mutations at or near the recombination junction. By modifying components of the in vitro assay we show that the strand transfer event is sequence-dependent. Finally, we modified the CRE-REP assay and demonstrate that recombinant yield is influenced by the distance between the deleterious mutations in the parental genomes and can partially compensate for polymerases with reduced nucleotide turnover. These studies will enable the analysis of recombination events throughout the region encoding the non-structural proteins of poliovirus, and—by extrapolation—to related enteroviruses and other positive sense RNA viruses. This study provides the basis for the detailed kinetic and mechanistic analysis of the initial strand-transfer and subsequent resolution events critical for the formation of both viable and competitively fit recombinant viruses. MATERIALS AND METHODS Viruses and cell culture Adherent monolayers of HeLa and L929 fibroblasts were grown in Dulbecco's Modified Eagle Medium (DMEM) or Glasgow Minimum Essential Medium (GMEM supplemented with G418 antibiotic). Media was supplemented with 100 U/ml penicillin, 100 μg/ml streptomycin, 2 mM l-glutamine and 10% heat inactivated (HI)-FBS. All cells were passaged in the presence of trypsin–EDTA. Where stated, guanidine hydrochloride (Sigma) was added to growth media at 4 mM. Poliovirus type 1 (Mahoney) and type 3 (Leon) were recovered following transfection of RNA generated in vitro (see below) from full-length cDNA. Virus was quantified by plaque assay or TCID50 as appropriate and as described previously (19). Virus growth analysis was determined by synchronous infection of HeLa cells at a multiplicity of infection (moi) of 10 pfu/cell, washing with PBS to remove unadsorbed virus and incubation in fresh media at 37°C in an atmosphere containing 5% CO2. Supernatant virus was quantified at various time points post infection by plaque assay. Virus competition assays were conducted by co-infection (at the specified ratios) of HeLa cells with a final moi of 10 pfu/cell. When serially passaging virus, harvested supernatant was diluted 1:4 with fresh media. Plasmids, in vitro transcription, cell transfection and recombinant virus characterisation pRLucWT and pT7Rep3-L are, respectively, cDNAs in pBR-derived plasmids encoding poliovirus type 1 Mahoney and type 3 Leon sub-genomic replicons with a luciferase reporter gene inserted in-frame in place of the P1 capsid coding region (13,20). pRLucWTG64S contains a Gly to Ser substitution at residue 64 of the viral polymerase. Derivatives of these replicons bearing substitution of lysine 359 with arginine (pRLucWTK359R and pT7Rep3-LK359R) were constructed using standard molecular protocols and verified by sequencing. pT7/SL3 has been described previously (13) and consists of a full-length poliovirus type 3 (Leon) cDNA bearing 8 synonymous substitutions in the cis-acting replication element (CRE) in the 2C-coding region. pT7/SL3K359R was constructed by exchange of a relevant restriction fragment from pT7Rep3-LK359R into pT7/FLC, a plasmid carrying a full length cDNA for poliovirus type 3 Leon. A derivative of pRLucWT deleted for the CRE in the native location (pRLucWTΔCRE) was generated by site directed mutagenesis introducing the same 8 substitutions defined in the SL3 mutant of poliovirus previously described (13). A synthetic CRE was introduced into the region encoding the 3′ NCR of pRLucWTΔCRE by first introducing a Bss HII site immediately following the site of polyprotein termination and then adding complementary oligonucleotides for the Synth2 version of the CRE (21) so creating pRLucWTΔCRE_3′CRE. Similar methods were used to introduce a 3′CRE into a full length poliovirus type 3 genome to create pFLCΔCRE_3′CRE. Plasmids encoding type 1 and type 3 poliovirus genomes (full length or sub-genomic) were linearized with Sal I and Apa I respectively, transcribed in vitro using T7 RNA Polymerase (Fermentas), treated with 2 U DNAse Turbo (Ambion) to remove residual DNA template and the RNA transcripts purified using RNeasy Mini Kit (Qiagen) before spectrophotometric quantification. Unless otherwise specified 1μg of RNA was transfected into near-confluent T25 flasks using Lipofectamine 2000 (Invitrogen). Recovered recombinant viruses were isolated by limit dilution and sequenced following RT-PCR amplification across the recombination junction as described previously (11). Viral RNA was extracted from clarified culture supernatant using a Qiagen RNAeasy Mini kit, reverse transcribed using Superscript II reverse transcriptase (Invitrogen) and an oligo-dT primer at 46°C for 50 min with the reaction terminated by incubation for 15 min at 70°C. PCR amplification of recombination junctions used template cDNA and appropriate oligonucleotides as listed in Supplementary Table T1 with KOD XL DNA polymerase (Novagen) used according to the manufacturer's protocol. Growth competition studies between parental type 3 poliovirus and FLCΔCRE_3′CRE were conducted by coinfection at a final moi of 10pfu/cell using 1:1, 10:1 and 1:10 ratios of input, with recovered virus being RT-PCR amplified across the region encoding the 2C-CRE followed by Swa I digestion of the PCR product and agarose gel electrophoresis. Luciferase assays Supernatant was removed from transfected cell monolayers, cells were briefly washed with PBS and lysed using 200 μl 1× Glo Lysis Buffer (Promega®) per well in a 12-well plate. The oxidation reaction was catalysed by the addition of 50 μl cell lysate to 50 μl room temperature Bright-GloTM Luciferase Assay System (Promega®) substrate and shielded from the light for 5 mins. Luciferase activity was measured using a luminometer with values normalised to mock transfection controls. In vitro sym/sub-based template switching assay The sym/sub assay has been described previously (22). Elongation complexes were assembled by incubating 5 μM WT or mutant poliovirus polymerase with 1 μM sym/sub RNA primer-template and 500 μM ATP for 5 min (Mix 1). Unless otherwise specified, template-switching reactions were initiated by addition of 60 μM RNA acceptor template and 500 μM CTP, GTP and UTP (Mix 2) and then quenched at various times by addition of 50 mM EDTA. All reactions were performed at 30°C in 50 mM HEPES, pH 7.5, 10 mM 2-mercaptoethanol, 60 μM ZnCl2, and 5 mM MgCl2. Products were analyzed by denaturing polyacrylamide gel electrophoresis, visualized using a PhosphorImager and the transfer products quantified using ImageQuant software (GE Healthcare). The strand transfer product was excised from the gel and sequenced following the miRCat-33™ protocol (Integrated DNA Technologies). CRE-REP recombination assay The CRE-REP assay for replicative recombination was conducted as described previously (11) and similar conditions were used for the 3′CRE-REP assay developed during this study. Where specified, ribavirin was added to transfected L929 monolayers at 600 μM. Recombinant viruses were quantified by plaque assay. RESULTS Polymerase fidelity and nucleotide turnover influences viral recombination Mutations to the viral polymerase can influence the fidelity and nucleotide turnover rate of the enzyme. Previous studies have demonstrated that substitution of a glycine to serine at position 64 in the poliovirus polymerase (G64S) confers resistance to ribavirin, a guanine-analog (23,24). We have demonstrated that this mutation decreased recombinant yield in an intertypic (poliovirus type 1 versus poliovirus type 3) CRE-REP assay by ∼20-fold whereas the addition of ribavirin, which enhances the polymerase error rate, increased recombinant yield in the same assay (11). To determine whether independently identified polymerase fidelity variants similarly influenced recombination we investigated the consequences of a conserved lysine to arginine substitution at residue 359 (K359R). This residue, located within motif D of the polymerase adjacent to the active site, is proposed to act as the general acid required for the second protonation event during polymerase-catalyzed nucleotidyl transfer (25). Polymerases bearing a K359R substitution retain activity but turn over nucleotides at 5–10% the rate of the unmodified protein and are both genetically-stable and attenuating (26). We additionally studied the influence of a polymerase with an extensively characterized (27) mutator phenotype, bearing a histidine to arginine substitution at residue 273 (H273R). This polymerase substitution has no impact on virus replication (27) but increases the mutation rate 3-fold and markedly attenuates the virus in vivo. Similar low fidelity polymerases have recently been shown to positively influence the yield of recombinant and defective interfering (DI) virus in an alphavirus model (15). We engineered the K359R or H273R substitutions into poliovirus type 1 and type 3 sub-genomic replicons and quantified luciferase production after transfection of HeLa cells with RNA synthesised in vitro. Both K359R-containing replicons exhibited a marked delay in maximal luciferase synthesis and a lower overall yield, reaching ∼30% of the luciferase level generated by the unmodified parental replicons (Figure 1A and B). We additionally engineered the K359R or H273R substitutions into the poliovirus type 3-derived SL3 genome, generating SL3K359R and SL3H273R respectively. We then conducted independent CRE-REP recombination assays using the SL3K359R and SL3H273R acceptor genomes with type 1 or type 3 sub-genomic replicons bearing the matching mutation, monitoring the yield of inter- and intratypic recombinants produced in 48 hours following transfection of murine L929 cells. Figure 1. Replication and recombination analysis of poliovirus bearing a K359R mutation in the viral polymerase. (A and B) Replication kinetics of sub-genomic replicons of poliovirus type 1 (RLucWT – A) or poliovirus type 3 (Rep3-L – B). Unmodified (wild-type) replicons indicated with filled circles, replicons bearing a K359R polymerase mutation with filled squares and wild-type replicons in the presence of 4mM guanidine hydrochloride, an inhibitor of poliovirus RNA replication, with filled triangles. 250ng of RNA generated in vitro was transfected into HeLa cells. Samples were taken at the times indicated and luciferase activity was measured and normalised using a mock transfected control. Error bars indicate standard deviation of two independent samples. (C) Influence of G64S, H273R and K359R polymerase mutants on the yield of recombinants in an intertypic (grey bars) and intratypic (black bars) CRE-REP recombination assay. Yield of recombinants is expressed as pfu/ml with error bars showing standard deviation from at least three independent assays. Note that the intratypic and intertypic assays use different scale Y axes. The X axis indicates the identity of the viral polymerase in both donor (sub-genomic replicon) and acceptor (SL3-derived genome) templates. Plaque assay quantities were independently confirmed by TCID50 (data not shown). Using unmodified (WT) parental SL3 and sub-genomic replicons yielded ∼2200 pfu/ml (±571; all titres determined in three or more independent assays and expressed ± SD) and 114 pfu/ml (±29) in intratypic and intertypic CRE-REP assays respectively, in broad agreement with previous results (11). In parallel assays the previously characterized G64S high fidelity polymerase variant produced 180 pfu/ml (±14) and 7 pfu/ml (±5) in intra- and intertypic recombinants respectively, the latter in agreement with earlier results (11). In the intertypic CRE-REP assay, genomes bearing the H273R substitution yielded 346 pfu/ml (±83; Figure 1C), an increase of ∼3-fold over the unmodified templates. A comparable (∼3.3-fold) increase in recombinant yield was also observed in the intratypic CRE-REP assay with templates bearing the same H273R variant polymerase. In contrast, the presence of the K359R mutation in both donor and recipient genomes prevented the recovery of any viable recombinants in either inter- or intratypic CRE-REP assays (Figure 1C). The results obtained with the low fidelity H273R polymerase, together with the recombination phenotype of the high fidelity K359R and G64S (11) enzymes, strongly support a role for polymerase fidelity as a key determinant of recombination frequency. Engineering the CRE into the 3′ non-coding region of the poliovirus genome The frequency of recombination is presumably related to the opportunity the polymerase has to switch templates. For example, in previous studies we demonstrated that preventing replication complex coalescence (via nocodazole inhibition of microtubule polymerisation) reduced the yield of recombinants. Since the generation of viable recombinant progeny in the CRE-REP assay is dependent upon a strand transfer event occurring within the ∼1kb separating the P1-coding region of the sub-genomic replicon and the defective CRE in the SL3 acceptor template (11) we reasoned that increasing the separation of these selection markers should also increase recombinant yield which would confer advantages in analysis of polymerases with reduced nucleotide turnover, such as K359R. Since the majority of characterised naturally-occurring recombinant enteroviruses exhibit recombination junctions within the P2- or P3-coding regions of the genome we wanted to modify the CRE-REP system to allow analysis of recombinants generated in vitro within a similar region. We therefore altered the original poliovirus type 1 sub-genomic replicon (pRLucWT) to inactivate the native CRE in the 2C coding region by the introduction of 8 synonymous substitutions (identical to those used in the construction of the SL3 variants used here and previously; (13)) to generate pRLucWTΔCRE. We confirmed that, as expected, RNA synthesized in vitro from pRLucWTΔCRE and transfected into murine L929 cells could not replicate (Figure 2A). The luciferase signal generated was similar in the presence or absence of 4 mM guanidine hydrochloride, a potent inhibitor of poliovirus replication (28), indicating it was solely from translation of the transfected RNA. We subsequently engineered a synthetic variant of the CRE sequence into the 3′ NCR of a pRLucWTΔCRE to generate pRLucWTΔCRE_3′CRE. The synthetic CRE sequence used had previously been shown to function when inserted into the 5′ NCR and consisted of the terminal 18 nt. of the wild-type CRE sequence engineered to form the terminal loop on a stem with similar structural stability to the native element (Supplementary Figure S1). This sequence was known to be stable in genomes bearing dual CRE sequences (21). RNA generated in vitro from pRLucWTΔCRE_3′CRE was transfected into murine L929 cells and luciferase synthesis monitored for 8 hours post-transfection. In parallel assays the luciferase signal from pRLucWTΔCRE_3′CRE-derived RNA was marginally greater than that from RNA generated from the parental sub-genomic replicon, pRLucWT, confirming that insertion of the CRE within the 3′ NCR was not incompatible with genome replication (Figure 2A). Figure 2. Replication and recombination of poliovirus with the CRE located in the 3′ non-coding region. (A) Replication of sub-genomic poliovirus type 1 replicons expressing a luciferase reporter gene. Wild-type (RLucWT) replicon in the absence (•) and presence (▴) of 4mM guanidine hydrochloride, together with the replicon with the native CRE inactivated (RLucWTΔCRE; ♦). The luciferase activity expressed by a replicon with a synthetic CRE inserted into the 3′ NCR (RLucWTΔCRE_3′CRE) in the absence (▾) and presence of 4mM guanidine hydrochloride (▪). In each case, 250ng of relevant RNA transcribed in vitro was transfected into L929 cells in duplicate, samples were harvested at the time points indicated and luciferase activity was measured and normalized using a mock transfected control. (B) Representative crystal violet stained Hela cell monolayers inoculated with supernatant obtained following transfection of RNA transcribed in vitro from the relevant full-length cDNAs indicated. Infected HeLa cells were covered with plaque assay overlay medium and stained 3 days post infection. To ensure that the resulting modifications had not introduced unexpected functional defects (e.g. disruption of encapsidation, a process critical for the identification of viable recombinants in the CRE-REP assay) into the virus genome we engineered the same mutations to the native CRE (creating pPV1FLCΔCRE) and 3′ NCR (creating pPV1FLCΔCRE_3′CRE) in a full length infectious poliovirus type 1 cDNA (pPV1FLC). RNA generated in vitro was independently transfected into murine L929 cell monolayers and de novo generated supernatant virus was quantified by subsequent plaque assay on HeLa cells. pPV1FLCΔCRE yielded no measurable virus and subsequent blind serial passage did not lead to the emergence of variants capable of replicating, a result in agreement with comparable studies on poliovirus type 3 (13). In contrast, RNA from pPV1FLCΔCRE_3′CRE generated viral plaques indistinguishable in appearance from the parental unmodified cDNA (Figure 2B) with a yield of 6 × 108 pfu/ml, similar to the 4 × 108 pfu/ml yield from PV1FLC under similar assay conditions. To further confirm that a virus bearing the CRE in the 3′ NCR replicated with similar kinetics to the unmodified parental virus we conducted competition experiments between the two viruses, broadly recapitulating the conditions that would prevail in a mixed infection recombination assay (Supplementary Figure S2). The results demonstrated that the poliovirus CRE can be relocated to the 3′ NCR without apparently compromising virus replication, thereby making it a suitable template for analysis of recombination throughout the region encoding the virus non-structural proteins. Extending the CRE-REP assay The relocated CRE in the 3′ NCR is situated ∼2.9 kb from it's native position, extending the potential region for recombination—the distance separating the VP1 capsid-coding region (absent from the sub-genomic replicon) and a functional CRE (absent from the SL3 parental genome)—from ∼1 kb to almost 4 kb (Figure 3A). We investigated intertypic recombination following transfection of murine L929 cells with RNA generated in vitro from pRLucWTΔCRE_3′CRE and pT7FLC/SL3. In parallel, we conducted a control CRE-REP recombination assay with RNA derived from parental pRLucWT and pT7FLC/SL3 (Figure 3B). The latter generated 157 pfu/ml (±20) whereas the assay with the extended recombination region yielded an increase of ∼5.8-fold to an average of 920 pfu/ml (±317; Figure 3C). Preliminary analysis of recombination junctions in viruses recovered in the extended CRE-REP assay (henceforth designated 3′CRE-REP) indicated they were located throughout the region encoding the non-structural proteins, with both imprecise and precise junctions identified (Supplementary Figure S3). Figure 3. An extended 3′CRE-REP recombination assay. (A) Schematic depiction of the 3′CRE-REP recombination assay. The acceptor genome (dark shading) bearing a defective CRE indicated as a broken line in the 2C-coding region is shown above a representation of the donor genome, a luciferase-encoding sub-genomic replicon (light shading) in which the native CRE has been inactivated and a synthetic CRE (indicated sCRE) inserted into the 3′ NCR. Following co-transfection of permissive cells (indicated by an arrow), a replication competent recombinant genome may be recovered of the generic structure shown, consisting of the 5′ part of the genome derived from full-length, capsid-encoding, acceptor genome and the 3′ part from the luciferase-encoding donor replicon. The crossover may occur within the region indicated. (B) Increased recombinant yields from the 3′CRE-REP assay. 250 ng of the acceptor or each of the indicated donor RNAs were co-transfected into L929 murine cells and supernatant harvested at 48 h post-transfection. Recombinant virus present in similar dilutions of supernatant was compared by plaque assay in HeLa cells and stained 72 h post infection. (C) Comparison of intratypic recombination yields from CRE-REP and 3′CRE-REP assays primed with RLucWT and FLC/SL3 donor and acceptor genomes. The presence of K359R polymerase mutation in both donor and acceptor genomes prevented the generation of recombinants in the 3′CRE-REP assay. The data represents the mean from three independent samples (± standard deviation). All plaque assay results were confirmed by TCID50 (data not shown). We then used the 3′CRE-REP assay to investigate whether the increased yield of recombinants was sufficient to detect recombination in genomes bearing the K359R mutation. Following transfection of RNA derived from pRLucWTΔCRE_3′CRE and pT7FLC/SL3, with both cDNAs modified to include the K359R mutation in the 3Dpol coding region, no recombinants were recovered (Figure 3C). Therefore, although increasing the size of the region in which recombination could occur enhanced the yield of recombinants, it was not sufficient to compensate for the non-recombinogenic K359R mutation in the viral polymerase. A biochemically-defined recombination assay To further investigate the strand-transfer reaction that must occur during recombination we exploited a biochemically-defined assay based upon the sym/sub system (22), a 10 nt. heteropolymeric RNA primer-template symmetrical self-complementary substrate (sym/sub) forming a 6 nt duplex flanked with 4 nt. 5′ unpaired regions (Figure 4A). Previous studies using homopolymeric templates and primers has suggested that purified poliovirus polymerase was sufficient for the strand-transfer event (29). We therefore investigated modification of this well-characterised sym/sub assay by inclusion of an acceptor template containing two potential regions of limited complementarity with the primary extension product. Briefly, the two-stage reaction involves assembling the radiolabeled sym/sub template with purified polymerase and ATP (Mix 1) in a suitable buffer to allow the formation of the elongation complex. Addition of the remaining nucleotides and an acceptor template (Mix 2) enables extension and—potentially—strand transfer, with the reaction products over a time course analysed by polyacrylamide gel electrophoresis. The formation of an n + 1 product indicates functional assembly of the elongation complex (Figure 4A and B) prior to the addition of Mix 2. This results in the rapid formation of a number of >n + 1 products including the fully extended sym/sub template represented by the n + 4 product (‘Strong stop’, Figure 4A and B). The generation of an n + 5 product is due to the non-templated addition to the end of the primer template, as previously described (29). The strand transfer product is represented by a discrete band appearing after the addition of Mix 2, the level of which increased as the reaction was incubated for up to 60 min (Figure 4B). Quantification of the strand transfer product at 30 min allowed the optimization of experimental conditions, including polymerase and acceptor template concentrations (Arnold et al., manuscript in preparation). Figure 4. A biochemically-defined in vitro recombination assay. (A) Scheme for in vitro template-switching assay. A radiolabelled (end labelled, indicated *) primer substrate pair known as sym/sub is the basis for the assembly of an elongation complex by the addition of ATP and purified polymerase protein. Addition of remaining nucleotides allows complete extension of the sym/sub template to form a strong stop product. The latter is considered the donor template to which acceptor template ribo-oligonucleotides are added. The complementary acceptor has six nucleotides (underlined) that have the potential to pair with the 3′ end of the extended sym/sub donor product. An alternative acceptor template with no complementarity was also tested. The products of the assay are analysed by polyacrylamide gel electrophoresis (PAGE). (B) Denaturing PAGE of in vitro reactions quenched at 5, 10, 20, 40 and 60 min after addition of the UTP/CTP/GTP and acceptor template (indicated below) mix. Reactions conducted in the absence (-) or presence (+) of ATP only indicated. The n + 1 product is the assembled elongation complex. The strong stop product is the fully extended sym/sub template. The high molecular weight products result from a template switch to the indicated RNA acceptor. Under the optimized conditions a maximum yield of ∼5% of the strand transfer product was achieved after 30 min when catalyzed by the native virus RdRp at 5 μM (Figures 4B and 5B). We reasoned that the acceptor template needed to be present in molar excess over the sym/sub template (1 μM) to provide maximum opportunity for template switching. As copy-choice recombination occurs during negative-strand synthesis (30) we also reasoned that a 60-fold excess of acceptor RNA over sym/sub template would be biologically relevant. Indeed, previous research has shown that negative strand synthesis is between 40–60-fold in excess of positive strand RNA synthesis during the early stages of RNA replication (31). The generation of a strand transfer product was dependent upon both the presence (Figure 4B) and sequence of the acceptor template. We initially investigated the suitability of an acceptor with a similar base composition (the ‘Random’ acceptor, Figure 4B) and demonstrated it was incompatible with formation of the transfer product. We extended this analysis to investigate the specific involvement of 3′ sequences in the acceptor by using each of three RNA molecules that were truncated by one, two or three nucleotides at the 3′ end, but were otherwise identical to the complementary RNA acceptor (Figure 5C). None could be incorporated into a transfer product. Figure 5. The influence of polymerase fidelity and acceptor template sequence. (A) Representative sym/sub template switching assays using the poliovirus polymerase variants indicated. Denaturing PAGE was used to separate reaction products at 2, 4, 8, 10, 20 and 30 min after addition of the extension mix. (B) Quantification of the transfer product generated by WT (•), H273R (▴) and G64S (▪) polymerases as a percentage of total RNA products per sample over time. Results are the average from three independent experiments and error bars indicate standard deviation. (C) Denaturing PAGE gel of in vitro sym/sub reactions quenched at 30 minutes after the addition of the extension mix containing complementary (WT) or acceptor RNAs truncated at the 3′ end by one, two, or three nucleotides (D31, D32 and D33 respectively) as shown. sym/sub analysis of fidelity polymerase variants The development and optimization of a biochemically defined sym/sub recombination assay confirmed that the viral polymerase was the only protein required for formation of the strand transfer product. This assay therefore provides an experimental environment to investigate the direct influence of polymerase on strand transfer. Using the optimized conditions already established (Arnold et al., manuscript in preparation) we investigated the formation of the sym/sub strand transfer product by G64S, H273R and K359R polymerase fidelity mutants by quantifying the production of the strand transfer product 30 min after addition of acceptor template and nucleotides to the reaction. The high fidelity G64S polymerase variant generated 1.8-fold less of the strand transfer product compared to wild type at 30 min. In contrast, the H273R low fidelity variant increased template transfer product 3-fold. Strikingly, the generation of the strand transfer product by the K359R polymerase was so low that accurate quantification over a 30 minute time course was not possible (Figure 5A and B). Sequence analysis of the strand transfer product The observations in both CRE-REP and sym/sub assays indicated that polymerase fidelity was a key determinant of strand transfer efficiency. The demonstration that the majority of the strand transfer products were of a minimum uniform size suggested there may be some sequence specificity to the in vitro recombination event, presumably involving one or other of the two limited regions of complementarity with the primary extension product (Figure 4A). In order to gain insight into the strand switching event we isolated the transfer product and subsequently cloned and sequenced the RNA species. Over 100 sequences were obtained, analysis of which resulted in two striking observations (Table 1). Firstly, the template switch occurred at the 3′ end of the acceptor template not at the internal region of complementarity. Secondly, a significant number of misincorporations or non-templated nucleotide additions occurred at or near the junction of the sym/sub donor and acceptor templates. Many of these involve the addition of C nucleotides, thereby increasing the potential for C–G base pairing with the 3′ end of the acceptor template. Only one of the sequences obtained showed no evidence of misincorporation or non-templated base additions. These observations support the role of polymerase fidelity in strand transfer and, together with the failure to observe a strand transfer product with random or truncated acceptor templates (Figures 4B and 5C), indicate that minimal regions of sequence complementarity are required for recombination in vitro. DISCUSSION Recombination in mammalian positive sense RNA viruses, such as poliovirus, is a relatively poorly characterized mechanism responsible for the generation of extensive genetic variation in the progeny of co-infections. In a recent study we demonstrated that replicative recombination (i.e. in which both parental viruses are replication-competent, to distinguish it from a distinct non-replicative process that has also been described (16,17)) is a biphasic process that involves distinct strand-transfer and resolution events (11). Both phases are necessary for the generation of genome-length recombinants that circulate in the population (7,10). The CRE-REP assay developed in our previous study enabled both viral and cellular determinants of these phases to be investigated. However, being cell-based, it did not allow the influence of the viral polymerase to be defined and studied in isolation. Furthermore, recombination was only quantifiable within a limited region of the genome. We address both these restrictions here. We describe a modified CRE-REP assay in which recombination throughout the region encoding the viral non-structural proteins can be analysed. In addition, we show that recombination can be recapitulated in a defined biochemical system in which the only protein present is the viral polymerase. We use these two strategies to investigate the role of RNA dependent RNA polymerase fidelity and nucleotide turnover rate in the genetic recombination of poliovirus. We had previously demonstrated that a G64S high fidelity polymerase variant significantly reduced (∼20-fold) the recombination rate observed in the CRE-REP assay. We extended this study by investigating the influence of K359R and H273R polymerase mutants, which are reported to exhibit higher and lower fidelity than the native enzyme respectively (26,27). In all instances, the relevant polymerase mutation was present in both the donor (sub-genomic replicon) and acceptor (SL3-derived) templates. In contrast to the G64S results, the presence of the H273R mutation led to at least a 3-fold increase in recombinant yield (Figure 1C), suggesting that low fidelity increases the rate of template switching by the viral polymerase, a result in agreement with a recent study in alphaviruses (15). Both the G64S and H273R variations have little impact upon the replication of poliovirus or sub-genomic replicons in cell based assays (27,32) indicating that fidelity alone may account for the changes in recombination rates measured in the CRE-REP assay. With the K359R mutation we were unable to detect any intra- or intertypic recombinants using the CRE-REP assay under conditions previously shown to yield viable progeny from G64S-containing templates (Figure 1C). Since the K359R polymerase has a lower fidelity than the G64S enzyme (26), fidelity alone cannot explain the different phenotypes of these enzymes in this assay. However, the K359R mutation also catalyses nucleotidyl transfer at 10% the rate of the unmodified enzyme (25,26), an observation we verified in luciferase assays using poliovirus type 1 or type 3 sub-genomic replicons (Figure 1A and B), which could also influence the recombination process. If there was a directly proportional relationship between replication rate and recombinant yield, we would have expected to detect recombinants in the intratypic CRE-REP assay (where the native enzyme generated ∼103 pfu/ml) even if they were undetectable in the intertypic assay which is ∼20-fold less sensitive. However, when investigated (Figure 1C), the absence of recombinants in the intratypic CRE-REP assay by K359R-containing genomes, indicates that the yield was at least 3log10 lower, implying that replication rate and polymerase fidelity may exert a cumulative influence on recombination. Previous studies have already indicated that there is a direct relationship between elongation rate and polymerase fidelity (26,33). Due to the pleiotropic effect of mutations such as K359R on the activities of the polymerase it is difficult to determine the relative importance of these two polymerase phenotypes on recombination. In an attempt to achieve this we have investigated the consequences of adding the mutagen ribavirin on recombinant recovery in the CRE-REP assay primed with donor and acceptor RNAs bearing the high-fidelity G64S mutation (which exhibits similar processivity to the native enzyme (26)). Ribavirin at 600μM enhanced the yield of recombinants from either the intertypic or intratypic CRE-REP assay 2- to 3-fold (Supplementary Figure S4) providing additional confirmation that fidelity per se is a key determinant of the template switching events in recombination. Ribavirin at a similar concentration could not rescue recovery of recombinants from CRE-REP assays primed with templates bearing the K359R polymerase mutation (data not shown). We reasoned that increasing the region within which recombination could occur in the CRE-REP assay might compensate for the reduced replication of the K359R polymerase mutant. We tested this by further separating the lesions that rendered the donor and recipient templates incapable of alone generating viable progeny, which we achieved by introducing a synthetic variant of the CRE to the 3′ NCR. In doing this, we demonstrated that the poliovirus CRE is functional in the 3′ NCR, 2.9 kb away from the native location and 6.8 kb distant from the 5′ NCR location previously reported (21). The robust replication of RLucΔCRE_3′CRE (Figure 2A and B) convincingly demonstrates the positional independence of the poliovirus CRE, as has also been shown for foot and mouth disease virus (34). Using a donor template (Figure 3A and C) with the CRE in the 3′ NCR increased recombinant yield by ∼5.8-fold in the modified 3′-CRE-REP assay over the original assay. The increased yield of recombinants was approximately proportional to the increase in separation of the genetic lesions used for the selection (5.8-fold versus 4-fold respectively). We interpret this as indicating there were no significant recombination ‘hotspots’ within the non-structural coding region that were absent from the ∼1 kb recombination window in the original CRE-REP assay. This conclusion is supported by preliminary analysis of a limited panel of recombinants from the 3′CRE-REP assay which indicated they were distributed throughout the region encoding the viral non-structural proteins (Supplementary Figure S3), with limited evidence of clustering at the 2C/3A boundary and the region encoding the C-terminus of the polymerase. We have previously suggested that the clustering of imprecise recombinants could be explained by functional constraints on the initial hybrid virus genome (11). The extended opportunity for recombination offered by the 3′CRE-REP assay will allow the influence of RNA structure, sequence identity and these functional constraints to be more fully examined in future studies. Notwithstanding the increased yield of recombinants obtained from the 3′CRE-REP assay with the unmodified polymerase, inclusion of the K359R mutation completely inhibited the formation of detectable recombinant progeny (Figure 1C). This provides further support for this mutation having a non-recombinogenic phenotype, but does not help elucidate the role of reduced nucleotide incorporation and/or increased fidelity in achieving this behaviour. A notable feature of the replication of the K359R polymerase is the apparent ‘stall’ in replication from around 2 to 6 h post-transfection. This was reported in a previous study (26) and also observed in sub-genomic replicons bearing this substitution (Figure 1A and B). During this period replication is occurring, but at a markedly reduced rate when compared to genomes bearing the native polymerase. After 6–8 h, replication gradually picks up, resulting in an overall yield ∼1log10 lower than the wild-type virus (26). The early events in genome replication include the formation and subsequent coalescence of membrane-anchored replication complexes (18,35). If one or more of these events is delayed, either directly by the action or interactions of the K359R polymerase, or indirectly because of reduced genome replication, then the yield of recombinants may be reduced. Nocodazole and cold treatment, which prevents microtubule re-polymerization, inhibits coalescence of RC and yield of recombinants (11,18). More compellingly, reversible stalling of virus replication using guanidine hydrochloride, also resulted in inhibition of the mobility of RCs (36). To try and discriminate between a direct influence of the K359R viral polymerase on the strand transfer event per se, and a defect in one or more of the cell-related events required for recombination, we investigated the process in a system in which the polymerase was the only protein present. We exploited the well-characterised sym/sub assay that has previously been used to analyse the phenotype of the poliovirus polymerase in a biochemically-defined system. Briefly, an elongation complex is assembled on a 10 nt self-complementary template in the presence of ATP alone and the stalled extension relieved by addition of the remaining nucleotides. By inclusion of an acceptor template with limited complementarity to the primary extension product (Figure 4A and B) we were able to demonstrate the formation of a specific strand-transfer product which exhibited the expected reduced mobility, confirming that the polymerase alone is necessary for template switching during recombination. We used this assay to explore the influence of polymerase variants or the sequence of the acceptor template on the production of the strand-transfer product. Our results strongly suggest that interactions of the donor and acceptor templates are important during recombination in vitro. An acceptor oligonucleotide with 6 nt. of complementarity (to the ‘strong stop’ extension product; underlined in Figure 4A) at its 3′ end was incorporated into the strand-transfer product, the length of which (Figure 4B) and subsequent sequencing (Table 1) indicated it is the position at which the polymerase switches template from donor to acceptor. The sequence specificity of this interaction was emphasised by the inability to generate a detectable strand-transfer product in the presence of an acceptor template with no complementarity (Figure 4B) or with acceptor templates truncated by 1–3 nucleotides at the 3′ end (Figure 5C). This analysis clearly demonstrates that the presence of two G nucleotides at the 3′ end of the acceptor—and complementarity between donor and acceptor template 3′ ends—was critical for generation of the transfer product. Table 1. Sequence analysis of strand transfer products # Misincorporation Untemplated Sequence 5′-GCAUGGGCCC … -3′ 24 +C AUGC C AUGC AUGC UUGC N0-3 20 +CC AUGC CC AUGC AUGC UUGC N0-3 2 G→C +C AUCC C AUGC AUGC UUGC N0-3 38 G→C AUCC AUGC AUGC UUGC N0-3 14 (U/G)→C ACC AUGC AUGC UUGC N0-3 4 +A(C/U)C AUGC A(C/U)C AUGC AUGC UUGC 1 AUGC AUGC UUGC 1 ΔG AUC AUGC UUGC The first column indicates the total number of each sequence obtained, all of which were preceded with the core sym/sub sequence shown in the header line. Underlined nucleotides indicate misincorporations or untemplated additions (summarised in the indicated columns on the basis of the differences between the observed and expected [the latter highlighted in bold] sequences) at or near the junction of the sym/sub donor and acceptor sequences. N indicates the untemplated addition of any nucleotide, nucleotides in brackets indicate alternates, ΔG indicates a single nucleotide deletion. Sequence analysis of the strand transfer product (Table 1) provides further insight to the importance of the two G nucleotides at the 3′ end of the acceptor. The majority of the sequences exhibited additional cytidine nucleotides at or near the recombination junction between the sym/sub donor and acceptor templates. Over 50% of the sequences included misincorporations, primarily of G to C, which would increase the sequence complementarity with the G dinucleotide at the 3′ end of the acceptor. In addition, an equivalent proportion (∼48%) of the sequences showed evidence of one or two non-templated C nucleotides at the junction. These presumably account for the reduced mobility apparent in some ‘strong stop’ products (Figures 4B and 5A) and would again increase the sequence complementarity between donor and acceptor molecules. Of the 104 transfer product sequences obtained, only one contained no sequence substitutions, additions or deletions from the expected product (5′-GCAUGGGCCCAUGCAUGCUUGC-3′). This level of variation has not been observed in the sequence of presumed recombination junctions in previous studies (1,37–41) or in our recent analysis using the CRE-REP assay ((11) and Bentley et al., manuscript in preparation), all of which involved analysis of replication-competent genomes. However, these analysed junctions were predominantly located within the polyprotein-coding region which, if disrupted, would be incompatible with virus viability. Over 60% of the strand transfer products sequenced (Table 1) contained additional nucleotides which would have resulted in the premature truncation of the open reading frame (ORF), and those in which the ORF remained intact contained additional amino acids, potentially poorly tolerated in the viral polyprotein. This strongly suggests that stringent functional selection of a diverse population of recombinant molecules may generate a much more limited subset of sequences which are observed in replication-competent viral populations. Although the non-templated addition of nucleotides to the 3′ end of the donor was important in enhancing complementarity with the acceptor it does not alone account for the ability of the polymerase to generate a strand transfer product. The K359R polymerase generates reduced-mobility ‘strong stop’ products, indicative of this untemplated addition, but was unable to generate recombinant products in either the sym/sub or CRE-REP assays (Figures 1C and 5A). Additional polymerase fidelity mutants tested in the sym/sub assay also recapitulated the phenotypes previously characterized in the cell-based CRE-REP assay (Figures 1C, 4A and 5A). The low fidelity H273R polymerase yielded ∼3-fold more transfer product in the sym/sub assay and at least 3-fold more recombinants in the CRE-REP assay when compared with the unmodified polymerase. Similarly, in both the biochemical and cell-based assay the high fidelity G64S mutant polymerase exhibited a reduced yield, by ∼2-fold and ∼20-fold respectively (11). These results emphasise the relevance of the sym/sub recombination assay and allow the detailed dissection of the contributions of polymerase, the donor and acceptor molecules in the strand transfer event. This should provide important insights into the molecular mechanism of recombination in enteroviruses and, by extrapolation, other single-stranded RNA viruses. It may additionally help understand the processes involved in other polymerase/RNA interactions in which discontinuous RNA sequences are juxtaposed in the resulting product, for example during the generation of internally-deleted defective interfering genomes (42) or the involvement of the body and leader transcriptional regulatory sequences implicated in discontinuous subgenomic transcription in coronaviruses (43). It is clear that the phenotype of the K359R mutant polymerase is of wider significance. This mutation lies within motif D, a conserved element present in all RNA dependent RNA polymerases, where the lysine acts as a general acid for nucleotidyl transfer (25,44,45). Not only does K359R attenuate poliovirus neurovirulence (26) but it is now shown that it also inhibits the frequency of template switching. Recombination between live attenuated Sabin vaccines and co-circulating species C enteroviruses is regularly observed and associated with paralytic disease (7,46,47). Once poliovirus eradication is achieved a K359R-containing variant could be considered as a suitable vaccine seed until global poliovirus vaccination programs are terminated. Using the CRE-REP assay we have defined the recombination of enteroviruses as a biphasic process, with distinct recombination and resolution events. Of these, the first generates ‘imprecise’ junctions, effectively consisting of limited duplications within the genome. Sequence analysis of several hundred of these has failed to detect any sequence specificity to the process ((11) and unpublished). In contrast, the junctions resulting from the resolution event always contain ‘ambiguities’ where it is unclear whether the sequence at the junction is derived from the donor or recipient template. This implies that there may be a requirement for complementarity for resolution. Formally, recombination is a trans event involving the joining of separate molecules whereas resolution could occur either in cis or in trans. It is therefore interesting to note the sequence complementarity required for the generation of the strand-transfer product in the sym/sub assay which, by definition, must occur in trans. Further studies will determine whether the biochemically-defined sym/sub recombination assay is actually recapitulating the primary recombination or secondary resolution events described. Supplementary Material SUPPLEMENTARY DATA SUPPLEMENTARY DATA Supplementary Data are available at NAR Online. FUNDING Biotechnology and Biological Sciences Research Council [BB/M009343/1] and PhD. studentship funding for AW (to D.J.E.); NIH [R01 AI045818 from NIAID to C.E.C.). Funding for open access charge: Biotechnology and Biological Sciences Research Council and University of St. Andrews. Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Runckel C. Westesson O. Andino R. DeRisi J.L. Identification and manipulation of the molecular determinants influencing poliovirus recombination PLoS Pathog. 2013 9 e1003164 23408891 2. Simon-Loriere E. Holmes E.C. Why do RNA viruses recombine? Nat. Rev. Microbiol. 2011 9 617 626 21725337 3. Lukashev A.N. Lashkevich V.A. Ivanova O.E. Koroleva G.A. Hinkkanen A.E. Ilonen J. Recombination in circulating Human enterovirus B: independent evolution of structural and non-structural genome regions J. Gen. Virol. 2005 86 3281 3290 16298973 4. Oberste M.S. Maher K. Pallansch M.A. Evidence for frequent recombination within species human enterovirus B based on complete genomic sequences of all thirty-seven serotypes J. Virol. 2004 78 855 867 14694117 5. Rohll J.B. Percy N. Ley R. Evans D.J. Almond J.W. Barclay W.S. The 5′-untranslated regions of picornavirus RNAs contain independent functional domains essential for RNA replication and translation J. Virol. 1994 68 4384 4391 8207812 6. Rohll J. Moon D. Evans D.J. Almond J.W. The 3′-untranslated region of picornavirus RNA - features required for efficient genome replication J. Virol. 1995 69 7835 7844 7494295 7. Kew O. Morris-Glasgow V. Landaverde M. Burns C. Shaw J. Garib Z. André J. Blackman E. Freeman C.J. Jorba J. Outbreak of poliomyelitis in Hispaniola associated with circulating type 1 vaccine-derived poliovirus Science 2002 296 356 359 11896235 8. McWilliam Leitch E.C. Bendig J. Cabrerizo M. Cardosa J. Hyypiä T. Ivanova O.E. Kelly A. Kroes A.C. Lukashev A. MacAdam A. Transmission networks and population turnover of echovirus 30 J. Virol. 2009 83 2109 2118 19091869 9. McWilliam Leitch E.C. Cabrerizo M. Cardosa J. Harvala H. Ivanova O.E. Kroes A.C.M. Lukashev A. Muir P. Odoom J. Roivainen M. Evolutionary dynamics and temporal/geographical correlates of recombination in the human enterovirus echovirus types 9, 11, and 30 J. Virol. 2010 84 9292 9300 20610722 10. McWilliam Leitch E.C. Cabrerizo M. Cardosa J. Harvala H. Ivanova O.E. Koike S. Kroes A.C. Lukashev A. Perera D. Roivainen M. The association of recombination events in the founding and emergence of subgenogroup evolutionary lineages of human enterovirus 71 J. Virol. 2012 86 2676 2685 22205739 11. Lowry K. Woodman A. Cook J. Evans D.J. Recombination in Enteroviruses Is a Biphasic Replicative Process Involving the Generation of Greater-than Genome Length 'Imprecise' Intermediates PLoS Pathogens 2014 10 e1004191 24945141 12. Percy N. Barclay W.S. Sullivan M. Almond J.W. A poliovirus replicon containing the chloramphenicol acetyltransferase gene can be used to study the replication and encapsidation of poliovirus RNA J. Virol. 1992 66 5040 5046 1321286 13. Goodfellow I. Chaudhry Y. Richardson A. Meredith J. Almond J.W. Barclay W. Evans D.J. Identification of a cis -acting replication element within the poliovirus coding region J. Virol. 2000 74 4590 4600 10775595 14. Goodfellow I. Polacek C. Andino R. Evans D.J. The poliovirus 2C cis-acting replication element-mediated uridylylation of VPg is not required for synthesis of negative-sense genomes J. Gen. Virol. 2003 84 2359 2363 12917456 15. Poirier E.Z. Mounce B.C. Rozen-Gagnon K. Hooikaas P.J. Stapleford K.A. Moratorio G. Vignuzzi M. Low-fidelity polymerases of alphaviruses recombine at higher rates to overproduce defective interfering particles J. Virol. 2015 90 2446 2454 26676773 16. Gmyl A.P. Belousov E.V. Maslova S.V. Khitrina E.V. Chetverin A.B. Agol V.I. Nonreplicative RNA recombination in poliovirus J. Virol. 1999 73 8958 8965 10516001 17. Gmyl A.P. Korshenko S.A. Belousov E.V. Khitrina E.V. Agol V.I. Nonreplicative homologous RNA recombination: promiscuous joining of RNA pieces? RNA 2003 9 1221 1231 13130136 18. Egger D. Bienz K. Recombination of poliovirus RNA proceeds in mixed replication complexes originating from distinct replication start sites J. Virol. 2002 76 10960 10971 12368339 19. Minor P.D. Mahy BWJ Virology: A practical approach 1985 Oxford IRL Press 25 41 20. Andino R. Rieckhof G.E. Achacoso P.L. Baltimore D. Poliovirus RNA synthesis utilizes an RNP complex formed around the 5′-end of viral RNA EMBO J. 1993 12 3587 3598 8253083 21. Goodfellow I. Kerrigan D. Evans D.J. Structure and function analysis of the poliovirus cis-acting replication element (CRE) RNA 2003 9 124 137 12554882 22. Arnold J.J. Cameron C.E. Poliovirus RNA-dependent RNA polymerase (3D(pol)). Assembly of stable, elongation-competent complexes by using a symmetrical primer-template substrate (sym/sub) J. Biol. Chem. 2000 275 5329 5336 10681506 23. Pfeiffer J.K. Kirkegaard K. A single mutation in poliovirus RNA-dependent RNA polymerase confers resistance to mutagenic nucleotide analogs via increased fidelity Proc. Natl. Acad. Sci. U.S.A. 2003 100 7289 7294 12754380 24. Pfeiffer J.K. Kirkegaard K. Increased fidelity reduces poliovirus fitness and virulence under selective pressure in mice PLoS Pathog. 2005 1 e11 16220146 25. Castro C. Smidansky E.D. Arnold J.J. Maksimchuk K.R. Moustafa I. Uchida A. Gotte M. Konigsberg W. Cameron C.E. Nucleic acid polymerases use a general acid for nucleotidyl transfer Nat. Struct. Mol. Biol. 2009 16 212 218 19151724 26. Weeks S.A. Lee C.A. Zhao Y. Smidansky E.D. August A. Arnold J.J. Cameron C.E. A polymerase mechanism-based strategy for viral attenuation and vaccine development J. Biol. Chem. 2012 287 31618 31622 22854962 27. Korboukh V.K. Lee C.A. Acevedo A. Vignuzzi M. Xiao Y. Arnold J.J. Hemperly S. Graci J.D. August A. Andino R. RNA virus population diversity, an optimum for maximal fitness and virulence J. Biol. Chem. 2014 289 29531 29544 25213864 28. Pfister T. Wimmer E. Characterization of the nucleoside triphosphatase activity of poliovirus protein 2C reveals a mechanism by which guanidine inhibits poliovirus replication J. Biol. Chem. 1999 274 6992 7001 10066753 29. Arnold J.J. Cameron C.E. Poliovirus RNA-dependent RNA polymerase (3Dpol) is sufficient for template switching in vitro J. Biol. Chem. 1999 274 2706 2716 9915801 30. Kirkegaard K. Baltimore D. The mechanism of RNA recombination in poliovirus Cell 1986 47 433 443 3021340 31. Novak J.E. Kirkegaard K. Improved method for detecting poliovirus negative strands used to demonstrate specificity of positive-strand encapsidation and the ratio of positive to negative strands in infected cells J. Virol. 1991 65 3384 3387 1851886 32. Vignuzzi M. Wendt E. Andino R. Engineering attenuated virus vaccines by controlling replication fidelity Nat. Med. 2008 14 154 161 18246077 33. Campagnola G. McDonald S. Beaucourt S. Vignuzzi M. Peersen O.B. Structure-function relationships underlying the replication fidelity of viral RNA-dependent RNA polymerases J. Virol. 2015 89 275 286 25320316 34. Mason P.W. Bezborodova S.V. Henry T.M. Identification and characterization of a cis-acting replication element (cre) adjacent to the internal ribosome entry site of foot-and-mouth disease virus J. Virol. 2002 76 9686 9694 12208947 35. Bolten R. Egger D. Gosert R. Schaub G. Landmann L. Bienz K. Intracellular localization of poliovirus plus- and minus-strand RNA visualized by strand-specific fluorescent in situ hybridization J. Virol. 1998 72 8578 8585 9765396 36. Egger D. Bienz K. Intracellular location and translocation of silent and active poliovirus replication complexes J. Gen. Virol. 2005 86 707 718 15722531 37. King A.M. Preferred sites of recombination in poliovirus RNA: an analysis of 40 intertypic cross-over sequences Nucleic Acids Res. 1988 16 11705 11723 2463516 38. Pilipenko E.V. Gmyl A.P. Agol V.I. A model for rearrangements in RNA genomes Nucleic Acids Res. 1995 23 1870 1875 7596811 39. Dedepsidis E. Kyriakopoulou Z. Pliaka V. Markoulatos P. Correlation between recombination junctions and RNA secondary structure elements in poliovirus Sabin strains Virus Genes 2010 41 181 191 20640496 40. Romanova L.I. Blinov V.M. Tolskaya E.A. Viktorova E.G. Kolesnikova M.S. Guseva E.A. Agol V.I. The primary structure of crossover regions of intertypic poliovirus recombinants: a model of recombination between RNA genomes Virology 1986 155 202 213 3022471 41. Tolskaya E.A. Romanova L.I. Blinov V.M. Viktorova E.G. Sinyakov A.N. Kolesnikova M.S. Agol V.I. Studies on the recombination between RNA genomes of poliovirus: the primary structure and nonrandom distribution of crossover regions in the genomes of intertypic poliovirus recombinants Virology 1987 161 54 61 2823469 42. Kuge S. Saito I. Nomoto A. Primary structure of poliovirus defective-interfering particle genomes and possible generation mechanisms of the particles J. Mol. Biol. 1986 192 473 487 3031313 43. Sawicki S.G. Sawicki D.L. Siddell S.G. A contemporary view of coronavirus transcription J. Virol. 2007 81 20 29 16928755 44. Castro C. Smidansky E. Maksimchuk K.R. Arnold J.J. Korneeva V.S. Gotte M. Konigsberg W. Cameron C.E. Two proton transfers in the transition state for nucleotidyl transfer catalyzed by RNA- and DNA-dependent RNA and DNA polymerases Proc. Natl. Acad. Sci. U.S.A. 2007 104 4267 4272 17360513 45. Liu X. Yang X. Lee C.A. Moustafa I.M. Smidansky E.D. Lum D. Arnold J.J. Cameron C.E. Boehr D.D. Vaccine-derived mutation in motif D of poliovirus RNA-dependent RNA polymerase lowers nucleotide incorporation fidelity J. Biol. Chem. 2013 288 32753 32765 24085299 46. Yang C.F. Chen H.Y. Jorba J. Sun H.C. Yang S.J. Lee H.C. Huang Y.C. Lin T.Y. Chen P.J. Shimizu H. Intratypic recombination among lineages of type 1 vaccine-derived poliovirus emerging during chronic infection of an immunodeficient patient J. Virol. 2005 79 12623 12634 16188964 47. Zhang Y. Wang H. Zhu S. Li Y. Song L. Liu Y. Liu G. Nishimura Y. Chen L. Yan D. Characterization of a rare natural intertypic type 2/type 3 penta-recombinant vaccine-derived poliovirus isolated from a child with acute flaccid paralysis J. Gen. Virol. 2010 91 421 429 19846676

PMC005xxxxxx/PMC5001619.txt

==== Front Nucleic Acids ResNucleic Acids ResnarnarNucleic Acids Research0305-10481362-4962Oxford University Press 2737878310.1093/nar/gkw601728GenomicsExperimental single-strain mobilomics reveals events that shape pathogen emergence Schoeniger Joseph S. †Hudson Corey M. †Bent Zachary W. †Sinha Anupama Williams Kelly P. *Systems Biology Department, Sandia National Laboratories, Livermore, CA 94551, USA* To whom correspondence should be addressed. Tel: +1 925 294 4730; Fax: +1 925 294 3020; Email: kpwilli@sandia.gov† These authors contributed equally to the paper as first authors.19 8 2016 04 7 2016 04 7 2016 44 14 6830 6839 22 6 2016 15 6 2016 20 4 2016 © The Author(s) 2016. Published by Oxford University Press on behalf of Nucleic Acids Research.2016This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.Virulence genes on mobile DNAs such as genomic islands (GIs) and plasmids promote bacterial pathogen emergence. Excision is an early step in GI mobilization, producing a circular GI and a deletion site in the chromosome; circular forms are also known for some bacterial insertion sequences (ISs). The recombinant sequence at the junctions of such circles and deletions can be detected sensitively in high-throughput sequencing data, using new computational methods that enable empirical discovery of mobile DNAs. For the rich mobilome of a hospital Klebsiella pneumoniae strain, circularization junctions (CJs) were detected for six GIs and seven IS types. Our methods revealed differential biology of multiple mobile DNAs, imprecision of integrases and transposases, and differential activity among identical IS copies for IS26, ISKpn18 and ISKpn21. Using the resistance of circular dsDNA molecules to exonuclease, internally calibrated with the native plasmids, showed that not all molecules bearing GI CJs were circular. Transpositions were also detected, revealing replicon preference (ISKpn18 prefers a conjugative IncA/C2 plasmid), local action (IS26), regional preferences, selection (against capsule synthesis) and IS polarity inversion. Efficient discovery and global characterization of numerous mobile elements per experiment improves accounting for the new gene combinations that arise in emerging pathogens. cover-date19 August 2016 ==== Body INTRODUCTION Bacterial pathogens acquire many of their antibiotic resistance and virulence genes as cargo on mobile elements that transfer between cells, typically through conjugation tubes or phage particles. Once delivered to recipient cells, these DNAs circularize and are then stably maintained either by self-replicating as plasmids or by integrating into the chromosome as genomic islands (GIs). Other mobile elements that move only intracellularly, such as transposons, are nonetheless relevant to pathogen emergence because they can insert into, and thereby augment the cargos of, the inter-bacterial mobile DNAs. While plasmids are readily identified in genome projects by their circular assembly maps, GIs can be more challenging to delineate. A classical GI is the prophage form of bacteriophage lambda. Phage-encoded integrase catalyzes crossover between the phage attP site and the related attB chromosomal target site, leaving the prophage flanked by the recombinant attL and attR sites (1). Upon induction of the lysogen, the reverse excision reaction leaves (i) circular lambda DNA whose circularization junction (CJ) is the regenerated attP and (ii) a deletion junction (DJ) in the chromosome, the regenerated attB (Figure 1). Some bacterial ISs are likewise known to form intermediates with CJs (2–4). Excised phage lambda goes on to replicate intracellularly (5). Figure 1. Detecting excision products by NGS. Island excision is shown, but some ISs circularize also, hence the general terms CJ (circularization junction) and DJ (deletion junction) for excision products. Island excision produces stoichiometric amounts of CJ and DJ, but IS circularization produces a CJ only. The black segment represents the sequence-identical overlap region in which crossover takes place. Several complementary bioinformatic approaches to GI discovery take advantage of common but non-universal island features: (i) their preference for target sites in tRNA and tmRNA genes (6), (ii) their density of phage-associated genes (7), (iii) anomalous nucleotide composition (8) and (iv) sporadic occurrence among closely related strains (9). None of these approaches can find all islands, precisely define their genomic coordinates or guarantee that those found have retained their mobility. Experimental approaches to discovery of active islands often employ the DNA-damaging agent mitomycin C (MMC) as an inducer. The released phages may then be isolated and characterized, but this approach misses non-viral islands like integrative conjugative elements. Predicted elements can be validated experimentally by inducing and detecting with polymerase chain reaction (PCR) the attP of the excised island; however, for some of the bioinformatic approaches the predictions are sufficiently inaccurate that PCR test designs may fail. We present a new experimental approach for the discovery of multiple islands, tranposable elements and (in principle) other mobile DNAs, using high-throughput (next-generation) sequencing (NGS) to map them onto the genome with nucleotide precision. This approach is an unbiased discovery method, that relies on no preconceptions about the nature of the mobile DNA. The data from our method inform questions about the biased distributions of transposition events, mechanisms of DNA mobility and comparative biology of multiple mobile DNAs. MATERIALS AND METHODS Culture A stock of Klebsiella pneumoniae BAA-2146 (Kpn2146) was obtained from American Type Culture Collection (ATCC). Its minimum inhibitory concentration of MMC was measured by broth microdilution at 6 μg/ml. For each experiment, a Kpn2146 colony on an LB agar plate was inoculated into LB broth and shaken at 37°C overnight. Shaking was continued for this overnight culture, and it was also diluted 100-fold into fresh broth and grown to log phase (OD600 = 0.5). At this point, MMC was added to cultures at a low (1 μg/ml) or high (5 μg/ml) level and incubation was continued for indicated periods before harvesting cells. Sequencing ‘Experiment X0’ sequencing data from Kpn2146 genomic DNA purchased from ATCC was described previously (10); read pairs were merged as possible using PEAR, and merged and unmergeable sequences were pooled. For the new experiments (Table 3), genomic DNA was extracted and purified from culture using DNeasy blood and tissue kit (Qiagen) as instructed. Nanodrop 2000 and Qubit DNA high sensitivity quantification kit was used to check quality and quantity. Aliquots (1 μg DNA) of some samples were treated with the exonucleolytic Plasmid-Safe DNase (Epicentre), in 50 μL volumes of the supplied buffer with 10 units DNase, incubated overnight at 37°C. Sequencing libraries were prepared using Illumina Nextera DNA sample preparation kit as instructed. Libraries were sequenced (single end, 150 cycle) on Illumina NextSeq 500 in high output mode or MiSeq. A quality filter (10) was applied to raw read data. Bioinformatics The Juxtaposer software used to find recombination junctions in NGS data is described in SI (‘Juxtaposer software’). Its main download site is bioinformatics.sandia.gov, with source code also available at github.com/sandialabs/Juxtaposer. A first iteration of its use with Kpn2146 identified errors in the genome sequence that were corrected (SI, ‘Genome reassembly’). Once mobile element termini were determined by Juxtaposer, abundance of each element was measured in each sample as the average count for its genome-unique 21-mers among the reads. Additionally, regular expressions (regexes) were designed to distinguish and count attL, attR, attB and attP for each element, with wild card positions that allow any sequence to intervene between IS ends in CJs. AttCt software applies these regexes and their reverse complements to each read, reporting the matching sequences and summary read counts. The above measurements, for each mobile DNA in each sample, were normalized using F, the average counts for the unique 21-mers in 30 kbp at each flank. This normalization factor was more reliable than the average of attL+attB and attR+attB counts; mean max/min was 1.41 for the latter two counts (when both ≥10), while that for F was 1.17, 1.13 and 1.09 using 3, 15 and 30 kbp flanks respectively. attCt/F values, for CJs or for DJs, would have a maximum of one if only excision occurs, however replication of CJ forms could take this value higher. Treatment of exonuclease data is described in SI (‘Exonuclease treatment’). RESULTS NGS monitors DNA mobility that occurs during culture As part of our original genome project for Kpn2146 we had bioinformatically predicted numerous mobile elements (Tables 1 and 2) in addition to its four plasmids (10). Moreover, while characterizing the newly discovered insertion sequence ISKpn21, we had found that some Illumina reads were from the junction region of circularized forms of the IS, indicating that NGS data report on DNA mobility events occurring during culture (Figure 1). Based on this observation new experiments (Table 3) were designed that use MMC to induce excision of GIs, compare log phase to overnight cultures, and use exonuclease treatment to measure dsDNA circular forms. The new experiments are presented along with further analysis of the original sequencing dataset. Details of the results are presented below, but first the NGS methods that demonstrate activity of mobile elements are described. Table 1. Genomic islands of Kpn2146 Juxtaposer AttCt Island Length Type CJ DJ Tp CJ DJ Remap* Kpn40guaA 40 467 Podo. prophage 845 137 0 1856 222 59,79 Kpn16fis 15 995 Sipho. prophage 21 0 1 344 2 −31,−31 Kpn49R 49 134 Sipho. prophage 25 5 0 232 115 Kpn37X 36 567 Myo. prophage 200 3 0 649 6 Kpn38rybB 37 500 Myo. prophage 1 0 0 1 0 Kpn42yraA 42 389 Myo. prophage 311 212 0 2376 324 3235,6 Kpn11L 11 188 Satell. prophage 0 0 0 0 1 Kpn21L 21 481 Satell. prophage 1 0 0 1 0 Kpn23sapBC† 23 301 Satell. prophage 217 841 2 414 1680 Kpn29S 28 544 MobQP ICE 2 2 0 2 4 Kpn55F 54 900 MPF-G ICE 0 0 0 0 2 *Remap: rightward shift of the newly determined terminal coordinates (left, right), relative to original island call. †GI Kpn23sapBC was apparently excised in the founding colony of experiment X3, exaggerating the DJ read count sum. Read counts for mobilizations (CJ, circularization junction; DJ, deletion junction, Tp, transposition) are from two methods, summed over all experiments. Table 2. Transposable elements of Kpn2146 Juxtaposer AttCt Type Length Family Intact copies CJ DJ Tp CJ DJ Ref* IS1F 768 IS1 1 0 0 2 0 0 IS1R 768 IS1 1 0 0 0 0 0 IS1×4 768 IS1 1 0 0 6 0 0 ISKpn14 768 IS1 2 5 0 115 8 0 IS26 820 IS6 11 2 0 170 25 0 (11) IS6100 880 IS6 2 0 1 3 0 0 IS903B 1057 IS5 2 0 0 0 0 0 IS4321 1327 IS110 1 5248 0 5 5241 0 (10) IS5075 1327 IS110 1 403 0 0 401 0 (10) ISKpn18 1303 IS3 1 30 1 17 50 0 ISEcl1 1336 IS3 2 0 0 1 0 0 ISKpn1 1445 IS3 1 20 0 1 31 0 ISEcp1 1656 IS1380 1 0 0 3 0 0 ISKpn21 2278 ISNCY 1 194 0 0 447 0 (8) ISEc22 2454 IS66 1 0 0 1 0 0 IS3000 3236 Tn3 1 0 0 0 0 0 Tn6187 9308 Tn3 1 0 0 1 0 0 *References to circularization of insertion sequence. Layout as in Table 1. Table 3. Experiments and samples, with the indicated harvest time after MMC treatment Experiment Sample Growth phase MMC (μg/ml) Harvest time (min) X0 S1 by ATCC* 0 0 X1 S1 log phase 0 15 X1 S2 log phase 1 15 X2 S1 log phase 0 0 X2 S2 log phase 1 60 X2 S3 log phase 1 120 X2 S4 log phase 5 60 X2 S5 log phase 5 120 X2 S6 overnight 0 0 X2 S7 overnight 1 120 X2 S8 overnight 5 120 X3 S1 log phase 0 0 X3 S2 log phase 5 15 X3 S3 log phase 5 30 X3 S4 log phase 5 60 X3 S5 log phase 5 120 X3 S6 overnight 0 0 *This genomic DNA sample was purchased from American Type Culture Collection (10). Experiment X0 was sequenced in paired end mode, the others in single end mode. Experiments X0 and X1 were sequenced using MiSeq, and X2 and X3 using NextSeq. Each of the X2 samples were sequenced both with and without prior exonuclease treatment. The primary method used here to examine DNA mobility with NGS data is the program Juxtaposer (Supplementary Figure S1). This discovery software finds reads whose left and right ends match non-adjacent genomic locations; such non-standard reads may identify short-range recombination events, such as transpositions, the circularization junction (CJ or attP) of a mobile element, or the chromosomal DJ (attB) left upon deletion of the element. Its main steps are listed in SI; the major filter removes standard reads that fully match a span of the standard genome. The output recombinant reads emphasize mobility events of GIs and ISs (Figure 2). Figure 2. Juxtaposer program reveals DNA mobility events. Each Juxtaposer output read names two juxtaposed genomic coordinates. These coordinates, from all 22 543 output reads for the samples not treated with exonuclease, were counted and mapped to 500-bp genomic windows. The high-count coordinates identified the two ends of several mobile elements (underlined), palindromic PCR artifacts, counts due mainly to high copy number of one plasmid (pMYS, whose copy ratio to the chromosome core was 74.7 in the combined data from these samples), deletion scars for tandem repeats (also potential PCR artifacts) and a plausible target of positive selection (wbaP gene encoding the first enzyme in the capsule polysaccharide synthesis pathway, disrupted by three independent ISKpn14 transposition events). Counts for 23sapBC ends are exaggerated by one experiment in which the island was pre-deleted. Two additional methods were employed: (i) once mobile element ends were discovered and mapped, somewhat more sensitive tests for CJ and DJ reads could be devised, which were counted using our AttCt software. (ii) When levels of free excised copies of single-copy elements rose above the background from the original genomic copy, they could be measured using coverage by the unique k-mers of the whole element in comparison to its flanks. Figure 3 shows examples of results for two mobile elements that were highly active, the GI Kpn40GuaA and the insertion sequence IS4321. Each is shown for a low-activity and high-activity sample, and with an exonuclease treatment that shows resistance of the mobilized DNA, suggesting that it is at least partly in circular form. Figure 3. Methods. NGS analysis methods revealing mobility, for two mobile elements. The top panels are from samples with little mobility activity; values for the element are from Juxtaposer, AttCt and average unique 21-mers (Kmer), all normalized to F (average unique 21-mer count for 30-kbp flanks). The plots show (x-axis) positions along the genome (120 kbp centered at Kpn40guaA and 4 kbp centered at IS4321) and (y-axis) read counts for each genome-unique 21-mer normalized to the average for the flanks. The middle panels are from samples where the elements have mobilized; the excesses over the flanking sequence may be due to replication, either post-excision for the island or as part of the circularization mechanism for the IS. The bottom panels are from the exonuclease-treated sample partners of the middle panel samples, showing resistance of circular molecules; expected values for a similar sized plasmid are given in parentheses. For the genomic island (GI), the observed exonuclease resistance is significantly (∼6-fold) less than expected from plasmids. Genomic island CJs and DJs Table 1 summarizes the Juxtaposer and AttCt results for combined experiments. For 6 of the 11 GIs, Juxtaposer identified numerous CJ reads and smaller numbers of DJ reads, and these counts increased with AttCt. The att site sequences validated by these reads, and potential effects on target gene products and expression, are shown for these islands in Supplementary Figure S2. By providing the recombinant attP and attB sequences, Juxtaposer output precisely maps island termini onto the genome. Coordinates for active islands that the Islander program (6) had mapped bioinformatically to tRNA and tmRNA genes were confirmed, while the new data altered the previously predicted coordinates of three Kpn2146 islands integrated into protein coding genes, in one case by over 3 kbp. These remappings validate the ability of Juxtaposer to discover new mobile DNAs. Figure 4 shows two biological replicates of brief timecourses for the response to a high MMC dose, for the six active islands. Except for a gross difference between the two experiments (Kpn23sapB was apparently pre-deleted in experiment X3), trends were similar, with some differences in extent; Kpn49R and Kpn42yraA were somewhat more active in X3 than X2, with the contrary for Kpn40guaA. These results show differential biology among the islands in terms of extent of excision, onset of excision and CJ/DJ ratio. Excision alone should produce stoichiometric amounts of CJ and DJ. However the typical next step after excision would be theta mode replication of the circular island, at a faster rate than replication of the chromosomal DJ (5). Thus differences in CJ/DJ ratio may reflect different lags or rates of replication. Figure 4. Timecourse. The AttCt timecourse data (normalized to the F measure of chromosomal flanks, see ‘Materials and Methods’ sections) for the high mitomycin dose in log phase (experiment X2:S1,S4,S5 and X3:S1,S4,S5) are shown for six islands, all suspected prophages. Three groups of graphs (underlined) use different y-axis scales, increasing to the right; y-axis is AttCt measures of CJ and DJ reads normalized to F (coverage of flanking sequence). CJs (blue) may be more numerous than DJs (red) due to post-excision island replication. Measurement error was propagated from that of F (standard error in unique 21-mer counts across the flanks). The circular nature of the molecules containing various island att forms was interrogated with an exonuclease that according to the manufacturer is active on linear dsDNA (less so on linear or circular ssDNA), but not on nicked, closed or supercoiled circular dsDNA. Circular DNAs survive extraction protocols dependent on their size, except for chromosome-sized circles which are so large that they virtually never purify intact. The size dependence of exonuclease resistance of DNA circles was internally calibrated using the four endogenous plasmids (which range from 2–141 kbp), separately for each sample (Supplementary Figure S3); even the largest plasmid was more resistant to exonuclease than the chromosome. Islands with CJ-forming activity were resistant to exonuclease, but not to the extent expected from the plasmid calibration curves (Supplementary Table S1). These findings are consistent with replication as modeled by phage lambda, which eventually produces linear genomic DNA production in rolling-circle mode, with the genome length cuts occurring at a site distant from the CJ/attP site (5). IS circles Two of the experiments compared overnight and log phase cultures. For unknown reasons, in one experiment (X2) but not the other (X3), the overnight culture produced a bonanza of CJs for seven types of ISs (Supplementary Figure S4 and Figure 5). Consistent with known transposition mechanisms, no DJs (from simple deletion of the original mobile DNA) were found for any of the ISs. The flanking donor ends that might be left immediately after cut-and-paste transposition would require more specialized (ligation-based) sequence library preparation methods for detection. Figure 5. Left flank size fidelity in IS4321 and IS5075 CJs. Upper panel: genomic sequences at IS ends; red, left flank sequence; blue, right flank sequence. This sequence is the same for both ISs; they target the same conserved end of Tn21-like transposons (11). It had been unclear which underlined T was the source of a particular T present in the usual CJ sequence (11). Lower panel: read counts for CJ sequences, considering only indels (i.e. ignoring base substitutions). Green overline, putative −35 and −10 promoter boxes (11). The predominant sequence is the same as that described previously for CJs (11). The bottom two sequences were the only ones found with indels outside of the flank junction. CJs join the left and right IS ends with a linker sequence from one or the other flank, allowing identification of the active copy among multi-copy ISs (Supplementary Figure S4). This demonstrates differential activity of IS copies: 90% of ISKpn18 CJs came from one of two copies (P = 4.2 × 10−9, binomial distribution), 67% of ISKpn21 CJs came from one of two copies (P = 1.2 × 10−13) and 72% of IS26 CJs came from one of the 11 complete copies (P = 4.6 × 10−14). This strong differential activity is remarkable since the sequences of the IS copies are virtually identical. The ISs show different regularity with respect to the linker length in CJs (Supplementary Figure S4). For ISKpn18, the linker length is always exactly 3 bp, agreeing with the 3-bp direct repeat (DR) found upon its transposition. For ISKpn21, there is an experiment-to-experiment difference in the linker length, suggesting that its mechanism may depend on undetermined physiological variations; in experiment X0 it is 5 bp for all six instances, matching the DR length, while in Experiment X2, it is 1–3 bp for 440 reads. IS26 has widely linker lengths, from either flank, that rarely match its 8-bp DR length. Some IS26 CJ sequence lengths approach the detection limit of read size, suggesting they may reflect only the tail of a larger size distribution, as becomes clearer below. The related ISs IS4321 and IS5075 are special in their site-specificity and their inclusion of canonical short additional sequences at each end, beyond the inverted repeat (11). These are the most rampantly circularizing ISs in these experiments, measured by AttCt at rates of 0.85 CJ/flank for IS4321 and 0.079 for IS5075 in the most productive exonuclease-untreated sample. The numerous CJ sequences allow fine examination of the imprecision of transposases as they produce CJs. Base substitutions that might arise during circularization would be difficult to distinguish from those arising during sequencing; instead indel errors are emphasized, as they are rare artifacts in Illumina sequencing (Figure 5). The CJ sequences provide an internal estimate of 8.6 × 10−6 for the indel sequencing error rate (the rate of indels at positions other than the flank junction). The indel rate is much higher (1.7 × 10−2) at the flank junction itself, and can be ascribed to extra or missing sequence from the left flank; these were assigned to transposase specificity errors. This error rate is still sufficiently low that the proposed −10 promoter box (11) would only rarely be affected. However the flank junction errors resolve mechanistic questions, such as the ambiguity concerning which flank is the source of the underlined T in the CJ sequence AGATAATGAG (11); the progression of error sequences, particularly those where the AGA right flank abuts a non-T from the left flank, suggests that the T in question comes from the left flank. The model IS for circularization, IS911, precedes full circularization with a figure-eight form of the donor, in which only one IS strand is circular and the other is connected to its original donor flanks (3); it is not clear how this form might respond to the exonuclease used here. The CJ molecules of IS4321 and IS5075, unlike those for GIs, respond to exonuclease approximately as predicted from the plasmid calibration curve (O/E values near 1 in Supplementary Table S1). Transpositions The extensive IS circularization suggested that there may have been transposition events during bacterial culture that our methods would also reveal. After Juxtaposer marks reads as CJs, DJs or palindromes, it enables a subsequent search for transposition events among recombinant reads, by identifying those involving the terminus of a potential transposable element, in the appropriate configuration. A reference list of annotated Kpn2146 elements was prepared, comprising 59 transposable elements with at least one end intact, of 27 types and including 26 additional potential mobile DNAs of Kpn2146 (GIs, group II introns and integron cassettes) to allow detection of their promiscuous insertions. This identified 334 reads, which, combining cases of the same juxtaposition in the same experiment (that may have already been present in the culture prior to sample splitting), comprised a minimum of 251 potential transposition junctions. The average read count per transposition junction (1.33) is small, but fluctuations by experiment suggest that the junctions were clonal, arising at different times prior to harvest; only one appeared independently in two experiments. Most junctions (89%) came from three IS types: IS26, ISKpn14 and ISKpn18. Transposition rates were not proportional to circularization rates for IS types (Table 2), despite the opinion that circles are transposition intermediates (2,3,12). IS4321 and IS5075 should be excluded from this question, because their site-specificity and prior occupancy of those sites suffice to explain their non-transposition, as should IS26 whose circles appear as a continuum with transposition events. After these exclusions, ISKpn1 and ISKpn21, which produced little transposition relative to circles, can be contrasted with ISKpn14, which behaved conversely. Transposition events where both ends could be detected Fourteen same-experiment pairs of transposition junctions could be matched as the putative left and right end of the same transposition event (Supplementary Figure S5). The three end-matched cases involving IS26 exhibited the expected 8-bp DR, and the remaining 11 cases involving ISKpn14 showed 9-bp DRs (except for one case of an 8-bp DR); the DR length had not been previously determined for ISKpn14, but is compatible with the 9-bp DR typically found for other members of the IS1 family (13). One two-IS event apparently mobilized a large plasmid segment containing many resistance genes, with ISKpn14 at one end and a partial copy of IS1X4 (also a member of the IS1 family) at the other end. Cases of IS polarity inversion were observed, where the IS gave the appearance of having flipped orientation in situ, both for ISKpn14 and for IS26. At a single new site in cspC both ends of both orientations of ISKpn14 were found in the same sample (Supplementary Figure S5). Plasmid preferences Transposition sites from exonuclease-untreated samples were mapped along the genome, revealing IS specific replicon and regional preferences (Figure 6). Replicon preference was evaluated statistically (Table 4), showing that each of the main active IS types is over-represented in at least one of the four plasmids, in some cases with corresponding under-representation in the chromosome. ISKpn18 shows extreme preference for the conjugative type 1 IncA/C2 plasmid pNDM-US. ISKpn14 transposition is over-represented in pCuAs. IS26 transposition is enriched in the two plasmids where it was originally found (pCuAs and pHg), part of the local action phenomenon described below, but was also enriched in pNDM-US where it was not already present. Figure 6. Hotspots. Transposition sites from exonuclease-untreated samples were ordered by the top three active IS types and then by coordinate on the concatenated replicons. Arrows mark the native copies of each of these IS families (omitting here the numerous IS26 copies on pCuAs and pHg). Black vertical lines mark chromosomal sites showing significant transposition site clustering (SI, ‘Statistics of transposition site bias’), at the capsular polysaccharide synthesis locus cps for ISKpn14 and at the two native chromosomal copies of IS26. Table 4. Enrichment factors for transpositions into replicons Replicon Chromosome pCuAs pHg pMYS pNDM-US IS26 0.10* 35.87† 13.71† 0.38 4.56† ISKpn14 0.91 4.90† 0 0.44 1.01 ISKpn18 0.04* 0 0 0 119.51† *Significant under-enrichment (P < 0.01). †Significant over-enrichment (P < 0.01). Data from the samples not treated with exonuclease were combined. Transposition site clustering Some IS types on some replicons showed little clustering of transposition sites; notably, ISKpn18 sites were randomly distributed on pNDM-US. In other cases there was significant clustering. IS26 showed local action, preferring the region of its only two native chromosomal copies, and pCuAs and pHg, where IS26 is also native. This local action, together with the widely varying IS26 CJs (Supplementary Figure S4), suggests that the IS26 CJs and transpositions represent the same intergraded phenomenon (i.e. the CJs are simply shorter-range local transpositions), thus in the subsequent section IS26 CJ and transposition data are treated together. For ISKpn14 and ISKpn18, whose transpositions do not cluster at their native sites and whose CJs have more discrete flank sizes, we continue to treat transposition as a separate phenomenon from circularization. ISKpn14 on the chromosome was enriched in the cps region, the highly varied multi-gene locus of capsular polysaccharide synthesis (10). The four cps sites were especially significant considering that they occurred only in the two smallest (MiSeq) experiments, which together had yielded a total of only 20 transposition sites for the three main IS types combined. These cps disruptions include the most highly expanded clones for any transposition event in any of the experiments, reaching 25 read counts and sufficient to detect both transposition junctions in two cases (Supplementary Figure S5). Three of the cps transpositions independently disrupted a single gene wbaP whose product catalyzes the first step of O-antigen synthesis, and the fourth disrupted a glycosyl transferase gene. Together these observations suggest that there was selection against capsule production during the two smaller experiments. To more fairly compare replicon preferences, Figure 6 had excluded data from the plasmid-biased exonuclease treatment. However to examine potential IS hotspots within an individual replicon, signals were boosted by pooling all data (Supplementary Figure S6). The extra data did not alter replicon preference for ISKpn18; it still preferred only NDM-US, yet with no particular hotspot. This pattern suggests that ISKpn18 may recognize a unique but dispersed ‘attractant’ feature unique to this 141-kbp replicon. One candidate for such a feature is the pNDM-US-encoded conjugation gene activator AcaCD (14); the P-value threshold 10−6.4 for its predicted binding sites in the Kpn2146 genome leaves 16 sites, all on pNDM-US. For nine of the ten ISKpn18 transpositions into pNDM-US, the average closest distance to one of the 16 AcaCD sites is 1780 bp. For ISKpn14, exonuclease treatment effectively shifted most detected transpositions from the chromosome to the cryptic ColE1-family plasmid pMYS (Supplementary Figure S6). This can be understood primarily as an effect of massive enrichment of tiny pMYS; indeed it is so enriched by exonuclease as to account for 74% of all reads, compared to 2.4% in exonuclease-untreated samples. Overlaid on this shift to pMYS is a preference for a particular region within it. Only part of this regional preference may be explained by known vital plasmid regions; i.e. pMYS bearing an IS in the 555-bp region of the plasmid that encodes the primer RNA of replication should fail to replicate. It is notable that all these transpositions map within or just upstream of the largest pMYS ORF. Although this ORF has no detectable homology with the copy control gene rop found in other ColE1 family members, if it does have a similar function, these transpositions should yield even higher pMYS copy numbers, which might explain the apparent enrichment here. IS26 local activity To unify all relevant IS26 data (CJs, transpositions, and other potential phenomena), all recombinant reads from Juxtaposer were collected in which one hit mapped within 10 bp of any IS26 copy end. Native Kpn2146 IS26 copies are found on the chromosome (copies 1–2), pCuAs (copies 3–5) and pHg (copies 6–12); they have identical terminal sequences, except that copy 8 is partial and copy 5 has a single base substitution 31 bp internally from its right end, marking this copy uniquely. Some transposition sites were difficult to map to the genome because they fell in repetitive DNA (Supplementary Figure S7), including transpositions into internal sites of IS26 itself, although some of these sites could be nonetheless be disambiguated. Fusions of IS26-flanking sequence directly to a new target site, as might occur if IS26 transposase ever behaved like a tyrosine or serine recombinase, could have been detected, but were not. For samples unbiased by exonuclease treatment, transposition sites that mapped unambiguously to the chromosome were spread along it, but not evenly, concentrating at the region of the two native copies of IS26 (Figure 6). A recent study of IS26 emphasizes that replicative transposition within the same replicon yields either an inversion of a flanking segment or deletion of a flanking segment concomitant with its circularization (15). The local concentration of transposition junctions was observed for all four configurations (Supplementary Figure S8A), i.e. left (L) or right (R) IS26 end, aiming upstream (+) or downstream (−). Detailed interpretation of this pattern is somewhat confused by the presence here of two nearby native IS26 copies, in opposite orientations (Supplementary Figure S8C). It is clear however that balance between inversions and circularization/deletions, due to local action of even just one of the two copies, should produce a balance of the four configurations throughout the region, as observed (Supplementary Figures S8C and 9). Circular products of local IS26 action should be resistant to exonuclease treatment, but less so for inversion and deletion products. Its effect was to significantly increase the local concentration, indicating that IS26 circles do indeed arise from this and only this region (Supplementary Figure S8B). Moreover, the configurations were mainly restricted to two types, L+ and R−, those expected for circles of IS26 copy 1; closeups of the mappings confirm that they are consistent with circular forms of copy 1, with several L+ configurations on its right side and R− configurations on its left side (Supplementary Figure S8D). These circles include the CJ reads described above, and circles whose junction would be too long to detect in short reads, i.e. a spectrum of linker lengths. Longer circles may also be produced, but would be less favored by size dependent exonuclease resistance (Supplementary Figure S3). The neighboring, opposite-orientation IS26 copy 2 shows much less activity, although some of its circles are indicated by L− at its right and R+ at its left. The two IS26 copies surround a 4.3-kbp segment containing an integron fragment with an intact aadA2 streptomycin resistance gene. These copies lie within the active island Kpn23sapBC, which itself is amplified relative to the surrounding chromosome. However measuring levels of each island and the remaining chromosome core (using unique 21-mers) shows that amplification of this island target is not sufficient to account for the clustering of IS26 transpositions here, with or without exonuclease treatment. Of the 13 IS26 transpositions into IS26-internal sites (Supplementary Figure S7A), 12 were from exonuclease-treated samples, suggesting they are circular forms. Moreover 11 showed configurations consistent with circles (L− or R+). These may be considered the extreme of the spectrum of local IS26 activity. Neither the end nor the internal site could be mapped to any particular IS26 copy, with one exception, where both could be mapped; this case represented an attack at a non-self copy and is therefore also an exception to local activity. There were also many IS26 transposition events on the IS26-bearing plasmids pCuAs and pHg (Supplementary Figures S10 and 11). Exonuclease resistance does highlight IS26 circles, but not as sharply as for the chromosome, perhaps because the plasmids themselves are also somewhat resistant. On pCuAs the tandem copies IS26.3 and IS26.4 show more local activity than the third nearby oppositely-oriented copy. On pHg the highest concentration of exonuclease-untreated events maps far from any IS26 copies, seemingly contrary to the local activity rule. IS26 copy 6 on pHg has a single base substitution 31 bp internally from its right end, marking this copy uniquely. In the one CJ read whose linker sequence identifies its source as the left end of IS26.6, the right end is the marked one from IS26.6, showing that a single IS26 copy provided both ends of a CJ. Of the 13 events involving the IS26.6 right end, 11 fall in its own replicon pHg; the other two fall in the main sink, the chromosome, but not in its hotspot region. This suggests that copy-local activity reflects self-activity more than an attraction of external copies. Polarity inversion was observed at native IS26 copies with new left ends swapping for the native right ends, specifically detected at native IS26 copies 3, 10, 11 and the partial copy 8; occurrence at the partial copy suggests that the phenomenon is not simple inversion of an existing IS but begins with attack by a remote IS copy. Polarity inversion was also observed with newly transposed IS26 copies, at pCuAs/11932 and at pCuAs/11134. DISCUSSION Classically, microbiologists worked with phage and bacterial stocks and thought of integration as the forward reaction for integrase. Now, ready access to numerous genome sequences inspires testing for excision products, to discover or verify new integrase site specificities. Our approach globally examines such events, providing unbiased experimental avenues to island discovery that complement existing bioinformatic approaches to island identification. It can also be applied to the study of transposition, here suggesting several principles for biased transposition distributions or hotspots, including positive selection, local activity, regional preferences and replicon preferences. Future modifications will allow measurement of the phases for the invertible DNAs that promote phase variation. Although the duration of these experiments was brief, there was sufficient time for an unintended selective pressure, against capsule synthesis, to act detectably. While the characteristic thick capsule of Klebsiella is essential for virulence and important in the organization of biofilms, acapsular mutants accumulate spontaneously during continuous in vitro culture (16). This has also been demonstrated in Streptococcus pneumococcus where 85% of such mutants had point mutations in cpsE (17,18), a homolog of the Klebsiella wbaP gene that accumulated multiple ISKpn14 transpositions here. At another extreme of selective pressure, the transposition events observed need not be stable over extended culture; they may be unstable, even lethal, mutations detected before elimination. As an example, one IS26 event would have truncated by 50% the single-copy glycolysis gene pgk which is essential in Escherichia coli (19). Only short reads were analyzed here, which can detect recombination over short crossover segments, as is catalyzed by site-specific recombinases and transposases. The typical dependence of homologous recombination on longer homologies (dramatic declines in efficiencies for homology < 75 bp (20)) disfavors its detection within short read sequences. Long, PCR-free reads can instead detect recombination across rRNA operons and other long repeats (10). Intramolecular replicative transposition produces rearrangements (circles, inversions and deletions) as it duplicates the transposable element (15); our main method is blind to these events because their detection requires longer-range information, i.e. read lengths longer than the element length (although circularity was addressed here independently through the use of exonuclease). Likewise the TU circles described for tandem IS26 pairs, produced by the transposase through both replicative and non-replicative mechanisms, yield no new DNA sequence juxtapositions (Supplementary Figure S9) and would require other methods to detect (4). Other Kpn2146 plasmids may be mobilizable, but only pNDM-US encodes its own conjugation pilus. ISKpn18 transposed with unexpected frequency into pNDM-US, yet with random distribution around the plasmid. This may indicate a mechanism with evolutionary implications, through which this IS detects and boards a plasmid about to transfer to another bacterial cell. Possible plasmid-marking features could be (i) the complex at the cytoplasmic platform of the conjugation pilus as the plasmid moves, (ii) a factor loaded onto only its replisome or (iii) a unique evenly-spaced plasmid-binding protein. As a possible example of the latter, transpositions were close to the binding sites for the conjugation gene transcription activator AcaCD. Regarding replisome tracking, several transposases have been found to interact with the host beta sliding clamp (21); it will be interesting to learn whether any beta clamp or other factor specifically marks the replisome on this plasmid. IS26 is most commonly found on plasmids, and known for mobilizing flanking antibiotic genes among enterobacteria (4). Although our data show some preference of IS26 for targeting plasmids, local activity can be substantial even in the chromosome. Local transpositions were not inherently biased toward circle-forming configurations, although exonuclease treatment enriched for these. The preference of IS26 for local action should primarily produce rearrangements, deletions and disruptions of local genes (15), which would more likely be tolerated in plasmids than in chromosomes. We focused on the recombination events involving the ends of mobile DNAs, but a larger number of unexplained recombination events are detected by Juxtaposer. They may be partly or mostly due to artifactual PCR-mediated recombination, including palindrome formation, occurring during library preparation; PCR-free library preparation methods should help to resolve this. Alternatively, they may represent a pervasive recombination phenomenon, real non-homologous recombination events that occur with little site specificity. Any such low-level phenomena do not alter our primary results, which stand well above this background (Figure 2). Our approach has some of the same goals as comparative genomics, yet it provides much richer information than can be gleaned from a handful of arbitrarily sampled frozen accidents. It is similar to other recent softwares that look for short reads that support mobility events, breseq and ISMapper [refs], but these treat the sample as having a single uniform genome that may differ from a reference, while Juxtaposer searches for subpopulations within the sample that have undergone different mobility events, and also explicitly searches for CJs. Our approach provides the control aspect of working within a single genetic background, and it can be applied to organisms that may have few related genomes available. It is richly comparative when applied to a rich mobilome, allowing comparison of the behaviors of multiple mobile DNAs at once, or comparison of different physiological settings on mobility mechanisms. In current work, Juxtaposer is providing new insight into the Clostridium difficile mobilome. We recommend routinely including a side MMC-treated culture in bacterial genome sequencing projects, to enable discovery and annotation of GIs. ACCESSION NUMBERS CP006659, CP006660, CP006661, CP006662 and CP006663. Supplementary Material SUPPLEMENTARY DATA SUPPLEMENTARY DATA Supplementary Data are available at NAR Online. FUNDING This research was fully supported by the Laboratory Directed Research and Development program at Sandia National Laboratories. Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy's National Nuclear Security Administration under contract [DE-AC04-94AL85000]. Funding for open access charge: Laboratory Directed Research and Development program at Sandia National Laboratories [DE-AC04-94AL85000]. Conflict of interest statement. None declared. ==== Refs REFERENCES 1. Campbell A.M. Episomes Adv. Genet. 1963 11 101 145 2. Turlan C. Chandler M. IS1 -mediated intramolecular rearrangements: formation of excised transposon circles and replicative deletions EMBO J. 1995 14 5410 5421 7489730 3. Duval-Valentin G. Marty-Cointin B. Chandler M. Requirement of IS911 replication before integration defines a new bacterial transposition pathway EMBO J. 2004 23 3897 3906 15359283 4. Harmer C.J. Moran R.A. Hall R.M. Movement of IS26 -associated antibiotic resistance genes occurs via a translocatable unit that includes a single IS26 and preferentially inserts adjacent to another IS26 Mbio 2014 5 e01801 e01814 25293759 5. Furth M. Wickner S. Hendrix R Roberts J Stahl F Weisberg R Lambda DNA replication Lambda II 1983 NY Cold Spring Harbor Laboratory 145 173 6. Hudson C.M. Lau B.Y. Williams K.P. Islander: a database of precisely mapped genomic islands in tRNA and tmRNA genes Nucleic Acids Res. 2015 43 D48 D53 25378302 7. Zhou Y. Liang Y. Lynch K.H. Dennis J.J. Wishart D.S. PHAST: a fast phage search tool Nucleic Acids Res. 2011 39 W347 W352 21672955 8. Vernikos G.S. Parkhill J. Resolving the structural features of genomic islands: a machine learning approach Genome Res. 2008 18 331 342 18071028 9. Langille M.G. Hsiao W.W. Brinkman F.S. Evaluation of genomic island predictors using a comparative genomics approach BMC Bioinformatics 2008 9 329 18680607 10. Hudson C.M. Bent Z.W. Meagher R.J. Williams K.P. Resistance determinants and mobile genetic elements of an NDM-1-encoding Klebsiella pneumoniae strain PLoS One 2014 9 e99209 24905728 11. Partridge S.R. Hall R.M. The IS1111 family members IS4321 and IS5075 have subterminal inverted repeats and target the terminal inverted repeats of Tn21 family transposons J. Bacteriol. 2003 185 6371 6384 14563872 12. Harmer C.J. Hall R.M. IS26 -mediated precise excision of the IS26-aphA1a translocatable unit Mbio 2015 6 e01866 e01915 26646012 13. Siguier P. Pérochon J. Lestrade L. Mahillon J. Chandler M. ISfinder: the reference centre for bacterial insertion sequences Nucleic Acids Res. 2006 34 D32 D36 16381877 14. Carraro N. Matteau D. Luo P. Rodrigue S. Burrus V. The master activator of IncA/C conjugative plasmids stimulates genomic islands and multidrug resistance dissemination PLoS Genet. 2014 10 e1004714 25340549 15. He S. Hickman A.B. Varani A.M. Siguier P. Chandler M. Dekker J.P. Dyda F. Insertion sequence IS26 reorganizes plasmids in clinically isolated multidrug-resistant bacteria by replicative transposition Mbio 2015 6 e00762 e00815 26060276 16. Mengistu Y. Edwards C. Saunders J. Continuous culture studies on the synthesis of capsular polysaccharide by Klebsiella pneumoniae K1 J. Appl. Microbiol. 1994 76 424 430 17. Schaffner T.O. Hinds J. Gould K.A. Wüthrich D. Bruggmann R. Küffer M. Mühlemann K. Hilty M. Hathaway L.J. A point mutation in cpsE renders Streptococcus pneumoniae nonencapsulated and enhances its growth, adherence and competence BMC Microbiol. 2014 14 210 25163487 18. Shainheit M.G. Valentino M.D. Gilmore M.S. Camilli A. Mutations in pneumococcal cpsE generated via in vitro serial passaging reveal a potential mechanism of reduced encapsulation utilized by a conjunctival isolate J. Bacteriol. 2015 197 1781 1791 25777672 19. Baba T. Ara T. Hasegawa M. Takai Y. Okumura Y. Baba M. Datsenko K.A. Tomita M. Wanner B.L. Mori H. Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: the Keio collection Mol. Syst. Biol. 2006 2 2006.0008 20. Watt V.M. Ingles C.J. Urdea M.S. Rutter W.J. Homology requirements for recombination in Escherichia coli Proc. Natl. Acad. Sci. U.S.A. 1985 82 4768 4772 3161076 21. Gómez M.J. Díaz-Maldonado H. González-Tortuero E. de Saro F.J.L. Chromosomal replication dynamics and interaction with the β sliding clamp determine orientation of bacterial transposable elements Genome Biol. Evol. 2014 6 727 740 24614824

PMC005xxxxxx/PMC5001623.txt

==== Front Sci RepSci RepScientific Reports2045-2322Nature Publishing Group srep3241210.1038/srep32412ArticleAccurate Morphology Preserving Segmentation of Overlapping Cells based on Active Contours Molnar Csaba 1Jermyn Ian H. 2Kato Zoltan 3Rahkama Vesa 4Östling Päivi 4Mikkonen Piia 4Pietiäinen Vilja 4Horvath Peter a141 Synthetic and System Biology Unit, Biological Research Centre of the Hungarian Academy of Sciences, Szeged, Hungary2 Department of Mathematical Sciences, Durham University, Durham, UK3 Department of Mathematics and Informatics, J. Selye University, Komarno, Slovakia4 Institute for Molecular Medicine Finland, University of Helsinki, Helsinki, Finlanda horvath.peter@brc.mta.hu26 08 2016 2016 6 3241211 05 2016 01 08 2016 Copyright © 2016, The Author(s)2016The Author(s)This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/The identification of fluorescently stained cell nuclei is the basis of cell detection, segmentation, and feature extraction in high content microscopy experiments. The nuclear morphology of single cells is also one of the essential indicators of phenotypic variation. However, the cells used in experiments can lose their contact inhibition, and can therefore pile up on top of each other, making the detection of single cells extremely challenging using current segmentation methods. The model we present here can detect cell nuclei and their morphology even in high-confluency cell cultures with many overlapping cell nuclei. We combine the “gas of near circles” active contour model, which favors circular shapes but allows slight variations around them, with a new data model. This captures a common property of many microscopic imaging techniques: the intensities from superposed nuclei are additive, so that two overlapping nuclei, for example, have a total intensity that is approximately double the intensity of a single nucleus. We demonstrate the power of our method on microscopic images of cells, comparing the results with those obtained from a widely used approach, and with manual image segmentations by experts. ==== Body High content analysis of microscopic images is a very active field in computational cell biology12345. While many methods have been developed, the analysis of cell cultures and tissue sections at the single-cell level remains a major challenge. As knowledge of cell-level heterogeneity plays a crucial role in improving the understanding and treatment of human diseases such as cancer, there is an urgent need for methods capable of precisely analyzing images of complex cellular phenotypes at single cell-level. Accurate cell segmentation is the basis of all such analysis, for example the identification of cellular compartments, or feature extraction based on cell morphology, intensity, or texture (Fig. 1). As a result, a great variety of single cell detection algorithms have been proposed. Most simple segmentation methods use local or global thresholding, usually based on the histogram of image intensities, and have therefore the smallest computational requirements6789. Other methods utilize inherent properties of the image intensity values, such as texture, to detect cells with characteristic patterns10. Supervised111213 and unsupervised1415 machine learning methods have proven their practical usefulness in single-cell detection applications: they largely outperform classical segmentation techniques by combining multi-parametric image-derived information and non-trivial decision surfaces. However, these single-cell methods often fail to detect multiple cells in complex spatial arrangements. A possible way to overcome this limitation is to incorporate prior shape information about the objects of interest into the segmentation algorithm. A common approach is to fit rigid predefined shapes (i.e. templates) to the image and identify the best matches1617181920. These methods can, to a certain extent, handle overlapping objects, but they are unable to capture small shape variations such as slightly elongations, which may encode essential phenotypic information. An alternative approach, “active contours”, have proven their popularity and usefulness in medical image analysis21, but the simplest models do not work well on the difficult problems addressed here. However, it is possible to extend simple active contour models, and incorporate different complexities of prior information about the region of interest2223. In particular, the “gas of near circles” model was designed to detect multiple near-circular objects24. In recent years, there has been a growing interest in both academia and industry in developing more complex three dimensional cell culture models. These can better capture the complexity of the tissue, and have the potential to provide more biologically relevant information than two-dimensional models252627. The conventional epifluorescence high-content microscope visualization is often used for shRNA, CRISPR-Cas9 and drug-screening of such 3D cultures, but in these cases it results in images containing many overlapping cells/nuclei. In addition, aggressively growing tumor cells, which have lost contact inhibition; co-cultures of different cell types in 2D; and fluorescently stained tissue samples provide similar challenges. The segmentation methods cited above are not capable of precisely detecting cell nuclei in these cases. Here we present a novel segmentation method, extending the “multi-layer gas of near-circles” (MLGOC) model of Molnar and colleagues28, that can be successfully applied to the counting of overlapping nuclei and to the determination of exact nuclear morphologies from fluorescence microscopy images. Our method uses an important property of most conventional wide-field fluorescence microscopy images: the intensity measured by the microscope at a location is proportional to the density of fluorescent particles, and therefore we assume that using low numerical aperture objectives, the intensity contribution of cells growing on top of each other is approximately the sum of the individual cell contributions. We present a new data term that captures this property and incorporate it into the MLGOC framework. The resulting model is capable of segmenting overlapping nuclei while preserving their morphologies, thus providing a unique platform for high-throughput analysis. We validate the method on synthetic and manually labeled sets of images of a prostate cancer cell line with fluorescently stained nuclei, comparing the results with those obtained from widely-used methods for the segmentation and analysis of single cells. Methods Cell preparation and imaging The PC346C prostate cancer cell line used for the validation experiments was obtained from the Erasmus Medical Centre in Rotterdam2930. 2000 cells in 25 μl of complete medium were delivered to the wells of 384-well plate with the Multidrop Combi Reagent Dispenser (Thermo Fisher Scientific Oy, Finland) using a standard cassette (Thermo Fisher Scientific Oy). After 72 hours, cells were fixed with 4% paraformaldehyde, and stained for 10 min at room temperature with Hoechst33324 (Life Technologies; stock 20 mg/ml, diluted to phosphate-buffered saline, 1:20 000) to detect the cell nuclei. All washing steps were performed with the EL406 Combination Washer Dispenser (Biotek, Germany). The images of fluorescently stained cells were captured with the automated epifluorescence microscope ScanR (Olympus, Germany) with a 150 W Mercury-Xenon mixed gas arc burner, a 20x/0.45 N.A. long working distance objective (Olympus), with 5 ms exposure time (UV channel) without binning, and a 12-bit (1344 × 1024, horizontal × vertical pixels) digital monochrome interline transfer CCD camera C8484-03G01 (Hamamatsu), cooled with the Peltier element. The pixel size at 20x objective is 0.323 μm/pixel. Image formation model for overlapping cell nuclei The image formation model we use is Iobserved = Ibackground + Ioriginal, where it is assumed that illumination problems have already been corrected31. Ibackground is a nearly flat non-zero surface with noise, so called “dark noise”. As described in the Introduction, using low numerical aperture objectives we assume that measured intensity is proportional to the density of fluorescent particles. This means, for example, that the intensity contribution of two cells on top of each other is approximately double that of a single cell. Let μ− and be the mean and variance of the background intensity, and μ+ and be the mean and variance of the measured intensity of a single cell. Let , and . Then, according to the model, the mean of the intensity of multiple cells is given by μ− + nΔμ, and its variance by . The parameters μ−, , Δμ, and Δσ2 are estimated from the corrected images using maximum likelihood estimation. “Multi-layer gas of near circles” model Active contours are popular image segmentation models21 that describe regions by their boundaries. An energy function of this boundary is defined that encodes information about boundary shape (the “geometric term”), to which is then added a “data term” in order to perform segmentation. We describe our geometric term first. In the simplest geometric term, the only interaction is between neighboring points. This allows the model to describe restrictions on boundary length and object area: where γ is the representation of the region boundary, and L and A are boundary length and object area, with weight parameters λC and αC, respectively. This ‘classical’ active contour model has low energy when the object’s area is small and its boundary is smooth. In contrast to this simple model, “higher-order active contour” models use multiple integrals over the boundary in the geometric term. These express long-range interactions between points on the boundary, and therefore can incorporate more specific information about object shape. One of the simplest of these higher-order functionals is32: where t is the tangent vector to the contour; p and p′ are contour parameters; and r(p, p′) is the distance between the points γ(p) and γ(p′). The interaction function Ψ is a monotonically decreasing function that controls the nature of the long-range interaction; it has a parameter d that controls the interaction range. A special parameterization of this higher-order model, the “gas of circles” (GOC) model33, assigns low energy to configurations consisting of several near-circles with approximately a given radius. Rather than represent the object boundary directly, as a parameterized curve, it is convenient to use a level set representation34 known as a “phase field”. The phase field formulation of a model can be used as an equivalent alternative to an active contour formulation35. It possesses many advantages, including easy handling of complex topologies and low computational cost. A phase field represents a subset by a function on the image domain , and a threshold . The phase field geometric term equivalent to Eg can be written in the form36: The above model, whether in its active contour or phase field formulation, has two main limitations. The first one comes from the representation: like most segmentation methods, it cannot represent overlapping object instances. The second arises from the geometric model itself: the non-local energy term, which causes the model to favor near-circular shapes, also results in a repulsive force between neighboring objects separated by a distance comparable to the desired object size. As a result, this method cannot effectively handle objects that are close to each other or overlapping. An extended version of the GOC approach28 uses several independent layers of phase fields to overcome these limitations. The geometric term of the model is simply the sum of the geometric terms of the individual layers, extended by a term that penalizes overlap. However, the overlap penalty was only introduced to solve a problem created by the data term used by Molnar et al.28, which tried to match each layer of the segmentation separately to the image data, sometimes resulting in the same object being segmented multiple times in different layers; the overlap penalty helped to avoid this. The current model tries to match a combination of the individual segmentation layers to the data, and assumes that the measured intensities are a (noisy) additive function of the number of overlapping objects; as a result, these degenerate segmentations do not occur, and we can set the overlap penalty to zero without any negative effects. Overlapping object instances can now be represented by appearing in different layers, while repulsion between objects in different layers is eliminated, even if close or overlapping, since the geometric term contains no inter-layer interactions. New data model for fluorescent microscopy In this section, we introduce a new data model that is adapted to the image formation process in fluorescence imaging. The new model is constructed using the assumption that overlapping cells contribute additively to the image intensity, so that multiple cells on top of each other produce an intensity contribution that is a multiple of that of a single cell. (An early version of this work was presented recently37). Let (Fig. 2), where ϕ(i) is the phase field in the ith layer; this quantity “counts” the number of cells at each point. Let γd be the (positive) weight of the data term; and let I be the intensity of the input image. Using the image formation model described earlier, and a Gaussian model for the image noise, the new data term becomes: Since a phase field takes the values −1 and 1 in its two phases (background and foreground), with a smooth transition between them, the integrand in Eq. (4), which is the energy density, takes a low value when ϕ+ = 0 over regions with background intensity; ϕ+ = 1 over regions of single-cell intensity; and generally, when ϕ+ = n over regions with n cell intensity. We use gradient descent in order to minimize the overall energy and find the optimal phase field configuration, and hence segmentation. The functional derivative of Egeom is given in Molnar et al.28. The functional derivative of Eintensity is: Dependence on initialization Gradient descent methods search for a local optimum of the energy function. The initialization is therefore crucial: a good initialization can greatly increase the accuracy of the resulting segmentation. We use two initialization methods:“Neutral initialization” means that the initial phase field is a realization of Gaussian white noise with mean λf/αf and a small variance. “Seeded initialization” means that we have estimates of the “centers” of the objects, the “seeds”. These could be computed by any available method; they are currently given manually. The initial phase field is defined to be 1 at all points within half the preferred radius of a seed, and zero elsewhere. The seeds are distributed among the layers so as to minimize overlaps. Evaluation method and metrics To evaluate the algorithm and to compare to other methods, three metrics were used. The first two metrics are the precision and recall of object detection. Precision (or positive prediction value) is the ratio of true positives (TP) to the number of detected objects (Precision = TP/(TP + FP)). Recall (or sensitivity) shows what proportion of the objects of interest is found (Recall = TP/(TP + FN)). Values closer to 1 imply better detection. Values are computed as follows. First, a matching is made between the set of ground truth objects and the set of segmented objects. Objects with no overlaps with any other object are deemed not matched. A weight is then assigned to each pair of remaining segmented and ground truth objects, equal to the reciprocal of the area of overlap. The total weight of a matching is defined as the sum of the weights of the matched pairs. This is an instance of the assignment problem; we solve it with the Hungarian algorithm. TP is then the number of segmented objects that have a matching ground truth object, while FP is the number of segmented objects that have no matching ground truth object. FN is the number of ground truth objects that have no matching segmented object. Note that these metrics do not measure morphological accuracy. The third metric measures the morphological accuracy of the segmentations of the individual, correctly-detected objects. For each matched pair of segmented (A) and ground truth (B) objects, we compute the Jaccard index, the ratio of the area of their intersection to the area of their union: . The final measure is then the average of the Jaccard indices of the matched pairs. Annotated images of a prostate cancer cell line To test our model, twenty fluorescence microscopy images of different degrees of complexity, containing ~2000 cells in total, were chosen for quantitative evaluation. For evaluation purposes, ground truth segmentations were generated by manually annotating the images. The annotations were made by two experts using ImageJ ROI Manager with Freehand selection. The image set consists of 20 images: 18 are different images with varying degrees of complexity; the other two images are 180°-rotated copies of two randomly chosen images from the initial 18. Results Synthetic data sets In order to measure the robustness of the proposed method, we generated data sets of synthetic images with different values for the noise variance, the extent of overlap, and varying object ellipticity and size, to create variability similar to that seen in real world observations. Object size selectivity To analyze the ability of our model to select objects of the correct size, a set of images was generated containing circles of two radii: 5 and 15. The expected radius of the GOC model was set to the radius of the larger objects. Figure 2c illustrates the results obtained using the model with the neutral initialization. The proposed model was able to select circles with the desired radius and correctly separate overlapping objects, while at the same time eliminating the circles with the smaller radius. Dependence on initialization To compare “Neutral” and “Seeded” initialization methods, 60 synthetic images (of size 400 × 400 pixels) containing {30, 35, 40} circles of radius 15, and 140 synthetic images (of size 150 × 150) containing 4–10 circles of radius 15 were generated (with 10 dB signal-to-noise ratio level). Background and foreground intensities were chosen from sets {30, 40, 50} and {90, 100, 110} respectively for 8-bit grayscale images. The set of synthetic images thus contains 6300 + 6860 = 13160 perfect circles with different degrees of overlap (0–100%). Figure 3 shows that the manual, seeded initialization makes the model 4–5% more effective at pixel level segmentation accuracy, without loss of object detection accuracy: with an appropriate choice of data weight, the manual initialization performs better than the neutral initialization with respect to all measures. Noise sensitivity It is common in fluorescence microscopy for images to have low contrast and low signal-to-noise ratio (SNR) because of weak fluorescent staining or microscope properties. In order to show the robustness of the proposed method to noise, 50 synthetic images containing 15 circles with up to 20% overlap were generated. Images were distorted with levels of Gaussian white noise, resulting in SNRs of 20 dB, 15 dB, 10 dB, 5 dB, 0 dB, and −5 dB. Note that 0 dB means that signal and noise have equal power, while −5 dB means that the noise power is roughly three times the signal power. The proposed method was able to segment overlapping circles up to 0 dB with minimal error, with the first errors appearing at SNR = −5 dB (Fig. 3). Separation of overlapping objects In order to test the ability of the model to separate overlapping objects accurately, a series of images was created, with each image containing three overlapping circles of the same radius placed at the corners of an equilateral triangle. The distance between the centers of the circles was varied from 0 to 4 times the radius. Additional noise (SNR = 0 dB) was added to the images. The segmentation results show that the accuracy is independent of the extent of the overlap, and that the method is capable of successfully segmenting circles (Fig. 4). Ellipticity test The shapes of cell nuclei vary from circles to more elongated elliptical shapes, depending on various biological aspects, such as the origin of the cell and the phase of the cell cycle38. For example in our annotated image set, the mode (most frequent value) of the minor and major axes ratio of the nuclei was 1.32, and the variance 0.43. The GOC shape model was originally designed to detect near circular shapes with possible slight perturbations24. We tested the ability of the proposed model to capture elongated objects (Fig. 4). Synthetic images were generated containing ellipses with rmin = 10 and rmax = {11, 12, …, 25}. The allowed overlap between two ellipsoids was 10%. As expected, the segmentation becomes worse for larger rmax. The proposed method is suited to most conventional cell nuclei types, but we do not recommend its use when the major/minor axis ratio exceeds 1.75. Semi-real tests: SIMCEP To simulate real cell nuclei we used a framework, SIMCEP, designed to generate and to evaluate algorithms for fluorescent microscopy images39. 50 images were generated by the framework, each containing 30 nuclei with a 50% maximal allowed overlap. The images were segmented by both CellProfiler and the method proposed here (see Fig. 5). The proposed method outperformed CellProfiler by 5–6% at pixel level accuracy and successfully identified cellular morphology on SIMCEP generated data. Real data set Results on annotated images of a prostate cancer cell line We measured the segmentation accuracy on fluorescence microscopy images using precision/recall and the Jaccard index, and compared the results with those obtained from CellProfiler. In CellProfiler, we used a mixture of a Gaussian global threshold method and an intensity-based splitting of clumped objects with object diameters in the range [25 and 35] pixels. In the MLGOC model, the preferred radius was set empirically to 17 (Fig. 6), the mean radius of the manually annotated nuclei. In comparing the MLGOC method to CellProfiler, we treated each image as different; therefore each of the 20 images has separate ground truth. In measuring self-consistency, we designated the original image as ground truth and the rotated version was compared to it, then vice-versa; the single metrics are the averages of the pairs of measures so obtained. A similar method was used to measure inter-expert agreement. The experts’ self-consistency measurements led to mean precision, recall, and Jaccard index values of 0.98, 0.96 and 0.78 respectively. The values for inter-expert agreement were somewhat weaker: 0.93, 0.93 and 0.75 respectively. As shown in Fig. 6, the proposed method outperforms the segmentation accuracy of CellProfiler, and also achieves a better Jaccard index, the value being as good as that arising from the agreement between different field experts. Separation ability is thus improved without losing detection accuracy. Discussion We have presented a novel method for the segmentation of near-circular objects, such as cell nuclei, from fluorescence microscopy images. Unlike previous methods, the model can handle the most difficult cases involving multiple overlapping objects, while still accurately capturing object morphology. This is of major importance for the analysis of the phenotypic behavior of cell populations. In this paper, the method is adapted to the segmentation, from fluorescence microscopy images, of cell nuclei, but could equally be applied to the segmentation of closely interacting or overlapping intracellular organelles, such as endosomes, lysosomes, or lipid droplets. Our image model, although simple, applies to the conventional imaging systems used for high-content screening and tissue scanning with general 10–40x objectives, and in these cases, and for most conventional cell nuclei morphologies, the proposed method represents a state-of-the-art segmentation tool. The image model also applies to other image modalities with related image formation properties, for example electron microscopy. We anticipate that our method would perform well in these cases also. In the future, we will investigate new seed initialization techniques, and an adaptive data term that can automatically handle image-to-image intensity variations and illumination differences within the same image. Additional Information How to cite this article: Molnar, C. et al. Accurate Morphology Preserving Segmentation of Overlapping Cells based on Active Contours. Sci. Rep. 6, 32412; doi: 10.1038/srep32412 (2016). We are most grateful to Dr. Wytske van Weerden and Prof. Guido Jenster from the Erasmus Medical Centre in Rotterdam for the PC346C cell line used for the validation experiments. Cs.M. and P.H. acknowledge the Hungarian National Brain Research Program (MTA-SE-NAP B-BIOMAG). P.M., V.P. and P.H. acknowledge support from the Finnish TEKES FiDiPro Fellow Grant 40294/13. V.P. and P.Ö. have received funding from the European Union’s Seventh Framework Programme (FP7/2007–2013) under grant agreement 258068: EU-FP7-Systems Microscopy. Author Contributions P.H., I.H.J. and Z.K. initiated the project. C.M., P.H. and I.H.J. designed the methodology. C.M. and I.H.J. developed the software and performed experiments. P.M. and V.R. created the ground truth dataset. V.P. and P.Ö. provided microscopy images and the biological background of the project. All authors wrote the manuscript. Figure 1 Comparison of different methods on microscopic images containing overlapping cells. Top row from left to right: (a) Original image; (b) Result (Region of Interest) obtained by adaptive threshold using CellProfiler7; (c) Results of CellProfiler standard segmentation method; (d) Results with the proposed “multi-layer gas of near-circles” method; (e) Precision, recall and Jaccard index of segmented objects (‘o’ and ‘p’ indicate that the metrics are computed at the object and pixel level respectively). Figure 2 Illustration of the proposed data model and behavior of the geometric model. (a) Noisy synthetic image. (b) Phase field representation of the cell configuration in image (a), showing the two layers, and the combined ϕ+ function that “counts” cells. (c) Size selectivity of the “gas of circles” model: using proper settings of the prior and data parameters, it is possible to achieve size-selective segmentation. No initial object seeds were used. Figure 3 Synthetic results with different initialization methods and the robustness of the model to noise. (a) Evaluation of segmentations with different initialization methods. “Neutral” means the initial ϕ was set to a Gaussian white noise with mean λf/αf and small variance (no initial objects are needed). “Seeded” means that small circles inside the objects were used to initialize the phase field. (b) Segmentation results with neutral initialization using 3 layers. (c) Segmentation results with seeded initialization. (d) Synthetic image with overlapping circles and increasing levels of signal-to-noise ratio (+20 dB to −5 dB). (e,f) Evaluation of the results of the proposed model. The segmentation results were evaluated on 50 test images per noise level. Note that different data term weights γd were used for different noise levels. (e) Average precision and recall values of segmentation results for different signal-to-noise ratios. (f) Box and whisker plot of segmentation accuracy for different levels of signal-to-noise ratio. The bottom and top edges of the box indicate the first and third quartiles; the line inside the box indicates the median; the whiskers (lines protruding from the box) indicate the smallest and largest data points whose distance from the box is not greater than 1.5 times the interquartile range. Figure 4 Separation ability of the proposed method and its robustness to elongation of circular objects. (a) Segmentation accuracy on synthetic images containing circles with increasing degrees of overlap. Numbers on the x-axis show the w/r ratio, where w is the distance between circles’ centers and r is radius of the circles. Values of w/r < 2 mean that objects are overlapping. (b) Examples of segmentations of synthetic images containing circles with different degrees of overlap. (c) Average precision and recall values for overlapping ellipses with increasing elongation. The values are computed from 100 test images per value rmax/rmin = {1.1, 1.2, …, 2.5}. (d) Box and whisker plots representing accuracy of segmentation for increasingly elongated ellipses. The bottom and top edges of the box indicate the first and third quartiles; the line inside the box indicates the median; the whiskers (lines protruding from the box) indicate the smallest and largest data points whose distance from the box is not greater than 1.5 times the interquartile range. (e) Examples of segmentations of synthetic images containing ellipses of increasing elongation. Figure 5 Segmentation results on SIMCEP (framework for generating artificial fluorescent microscopy images) data set. (a) Object detection and segmentation accuracy of CellProfiler and the proposed MLGOC method. (b) Left column: images containing overlapping artificial nuclei with possible distortion caused by a microscope applied; middle column: segmentations with CellProfiler; right column: segmentations with the proposed method. Figure 6 Segmentation results on real fluorescence microscopy images. (a) Object detection and segmentation accuracy of CellProfiler and the MLGOC method. (b) Left column: original images of prostate cancer cells containing overlapping nuclei; middle column; segmentation results obtained with the CellProfiler software; right column: results with the MLGOC method. ==== Refs Moffat J. G. . Phenotypic screening in cancer drug discovery - past, present and future . Nat Rev Drug Discov 13 (8), 588 –602 (2014 ).25033736 Celli J. P. . An imaging-based platform for high-content, quantitative evaluation of therapeutic response in 3d tumour models . Sci Rep 4 , 3751 (2014 ).24435043 Singh S. . Increasing the content of high-content screening: An overview . J Biomol Screen 19 (5), 640 –650 (2014 ).24710339 Uhlmann V. . CP-CHARM: segmentation-free image classification made accessible . BMC Bioinformatics 17 (1), 51 (2016 ).26817459 Martin H. L. . High-content, high-throughput screening for the identification of cytotoxic compounds based on cell morphology and cell proliferation markers . PloS One 9 (2), e88338 (2014 ).24505478 Hoggar S. G. Mathematics of digital images: creation, compression, restoration, recognition (Cambridge University Press, 2006 ). Carpenter A. E. . Cellprofiler: image analysis software for identifying and quantifying cell phenotypes . Genome Biol 7 (10), R100 (2006 ).17076895 Collins T. J. Imagej for microscopy . Biotechniques 43 , 25 –30 (2007 ).17936939 Gonzalez R. C. & Woods R. E. Digital Image Processing (3rd Edition) (Prentice-Hall, 2006 ). Wang X. . Cell segmentation and tracking using texture-adaptive snakes. In ISBI 2007: Proceedings of 4th IEEE International Symposium on Biomedical Imaging: from nano to macro, 101–104, Metro Washington DC, USA, IEEE (2007 ). Sommer C. . Ilastik: Interactive Learning and Segmentation Toolkit. In ISBI 2011: Proceedings of 8th IEEE International Symposium on Biomedical Imaging: from nano to macro, 230–233, Chicago, Illinois, USA, IEEE (2011 ). Linder N. . Identification of tumor epithelium and stroma in tissue microarrays using texture analysis . Diagn Pathol 7 (2012 ). Turkki R. . Assessment of tumour viability in human lung cancer xenografts with texture-based image analysis . J Clin Pathol (2015 ). Bamford P. & Lovell B. Unsupervised cell nucleus segmentation with active contours . Signal Processing 71 (2), 203 –213 (1998 ). Linder N. . A malaria diagnostic tool based on computer vision screening and visualization of Plasmodium falciparum candidate areas in digitized blood smears . PloS One 9 (8) (2014 ). Perrin G. . A marked point process model for tree crown extraction in plantations. ICIP 2005: Proceedings of 12th IEEE International Conference on Image Processing, 709–712, Genova, Italy, IEEE (2005 ). Thevenaz P. & Unser M. The snakuscule. In ICIP 2006: Proceedings of 13th IEEE International Conference on Image Processing, 1633–1636, Atlanta, GA, USA, IEEE (2006 ). Thevenaz P. . The ovuscule . IEEE Trans Pattern Anal Mach Intell 33 (2), 382 –93 (2011 ).20513925 Olivo-Marin J.-C. Extraction of spots in biological images using multiscale products . Pattern Recognit 35 (9), 1989 –1996 (2002 ). Kimme C. . Finding circles by an array of accumulators . Commun ACM 18 (2), 120 –122 (1975 ). Kass M. . Snakes - active contour models . Int J Comput Vis 1 (4), 321 –331 (1987 ). Cremers D. . Diffusion snakes: Introducing statistical shape knowledge into the mumford-shah functional .Int J Comput Vis 50 (3), 295 –313 (2002 ). Cremers D. . Diffusion-snakes: Combining statistical shape knowledge and image information in a variational framework. In VLSM'01: Proceedings of 1st IEEE Workshop on Variational and Level Set Methods in Computer Vision, 137–44, Vancouver, BC, USA, IEEE (2001 ). Horvath P. . A higher-order active contour model for tree detection. In ICPR 2006: Proceeding of the 18th International Conference on Pattern Recognition, 130–133 Hong Kong, IEEE (2006 ). Anastasov N. . A 3d-microtissue-based phenotypic screening of radiation resistant tumor cells with synchronized chemotherapeutic treatment . BMC Cancer 15 (1) (2015 ). Nobis M. . Intravital flim-fret imaging reveals dasatinib-induced spatial control of src in pancreatic cancer . Cancer Res 73 (15), 4674 –4686 (2013 ).23749641 Hickman J. A. . Three-dimensional models of cancer for pharmacology and cancer cell biology: Capturing tumor complexity in vitro/ex vivo . Biotechnol J 9 (9), 1115 –1128 (2014 ).25174503 Molnar C. . A Multi-Layer Phase Field Model for Extracting Multiple Near-Circular Objects. In ICPR 2012: Proceedings of 21st International Conference on Pattern Recognition, 1427-1430, Tsukuba, Japan, IEEE (2012 ). Marques R. B. . High Efficacy of Combination Therapy Using PI3K/AKT Inhibitors with Androgen Deprivation in Prostate Cancer Preclinical Models . Eur Urol 67 (6), 1177 –1185 (2015 ).25220373 Marques R. B. . The human PC346 xenograft and cell line panel: a model system for prostate cancer progression. The human PC346 xenograft and cell line panel: a model system for prostate cancer progression . Eur Urol 49 (2), 245 –257 (2006 ).16413664 Smith K. . CIDRE: an illumination-correction method for optical microscopy . Nat Methods 12 (5), 404 –406 (2015 ).25775044 Rochery M. . Higher order active contours . Int J Comput Vis 69 (1), 27 –42 (2006 ). Horvath P. . A higher-order active contour model of a ‘gas of circles’ and its application to tree crown extraction . Pattern Recognit 42 (5), 699 –709 (2009 ). Osher S. & Sethian J. A. Fronts Propagating with Curvature-Dependent Speed - Algorithms Based on Hamilton-Jacobi Formulations . J Comput Phys 79 (1), 12 –49 (1988 ). Rochery M. . Phase field models and higher-order active contours. In ICCV 2005: Proceedings of 10th IEEE International Conference on Computer Vision, 970–976, Beijing, IEEE (2005 ). Horvath P. & Jermyn I. H. A ‘gas of circles’ phase field model and its application to tree crown extraction. In EUSIPCO 2007: Proceedings of 15th European Signal Processing Conference, Poznan, Poland, IEEE (2007 ). Molnar C. . A New Model for the Segmentation of Multiple, Overlapping, Near-Circular Objects. In DICTA 2015: Proceedings of International Conference on Digital Image Computing: Techniques and Applications, 1–5, Adelaide, Australia, IEEE (2015 ). Neumann B. . Phenotypic profiling of the human genome by time-lapse microscopy reveals cell division genes . Nature 464 (7289), 721 –727 (2010 ).20360735 Lehmussola A. . Computational framework for simulating fluorescence microscope images with cell populations . IEEE Trans Med Imaging 26 (7), 1010 –1016 (2007 ).17649914

PMC005xxxxxx/PMC5001625.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756421310.1371/journal.pone.0161846PONE-D-16-05953Research ArticleBiology and Life SciencesImmunologyVaccination and ImmunizationMedicine and Health SciencesImmunologyVaccination and ImmunizationMedicine and Health SciencesPublic and Occupational HealthPreventive MedicineVaccination and ImmunizationBiology and Life SciencesImmunologyVaccination and ImmunizationVaccinesMedicine and Health SciencesImmunologyVaccination and ImmunizationVaccinesMedicine and Health SciencesPublic and Occupational HealthPreventive MedicineVaccination and ImmunizationVaccinesPeople and PlacesGeographical LocationsOceaniaAustraliaBiology and life sciencesOrganismsVirusesDNA virusesPapillomavirusesHuman PapillomavirusBiology and Life SciencesMicrobiologyMedical MicrobiologyMicrobial PathogensViral PathogensPapillomavirusesHuman PapillomavirusMedicine and Health SciencesPathology and Laboratory MedicinePathogensMicrobial PathogensViral PathogensPapillomavirusesHuman PapillomavirusBiology and Life SciencesOrganismsVirusesViral PathogensPapillomavirusesHuman PapillomavirusMedicine and Health SciencesUrologyGenitourinary InfectionsHuman Papillomavirus InfectionMedicine and Health SciencesInfectious DiseasesSexually Transmitted DiseasesHuman Papillomavirus InfectionMedicine and Health SciencesInfectious DiseasesViral DiseasesHuman Papillomavirus InfectionResearch and Analysis MethodsResearch DesignSurvey ResearchSurveysSocial SciencesAnthropologyCultural AnthropologyReligionSocial SciencesSociologyReligionMedicine and Health SciencesOncologyCancers and NeoplasmsGynecological TumorsCervical CancerAttitudes, Knowledge and Factors Associated with Human Papillomavirus (HPV) Vaccine Uptake in Adolescent Girls and Young Women in Victoria, Australia Predictors of HPV Vaccine Uptake in Adolescent Girls in Victoria, AustraliaTung Iris L. Y. 1Machalek Dorothy A. 12Garland Suzanne M. 123*1 Department of Microbiology and Infectious Diseases, The Royal Women’s Hospital, Parkville, Victoria, Australia2 Murdoch Childrens Research Institute, Parkville, Victoria, Australia3 Department of Obstetrics and Gynaecology, The University of Melbourne, Victoria, AustraliaConsolaro Marcia Edilaine Lopes EditorUniversidade Estadual de Maringa, BRAZILCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: ILYT DAM SMG. Data curation: ILYT DAM SMG. Formal analysis: ILYT DAM SMG. Investigation: ILYT DAM SMG. Methodology: ILYT DAM SMG. Project administration: ILYT DAM SMG. Resources: ILYT DAM SMG. Supervision: ILYT DAM SMG. Validation: ILYT DAM SMG. Visualization: ILYT DAM SMG. Writing – original draft: ILYT DAM SMG. Writing – review & editing: ILYT DAM SMG. * E-mail: suzanne.garland@thewomens.org.au26 8 2016 2016 11 8 e016184610 2 2016 12 8 2016 © 2016 Tung et al2016Tung et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Background Human papillomavirus (HPV) vaccination targets high-risk HPV16/18 that cause 70% of all cancers of the cervix. In Australia there is a fully-funded, school-based National HPV Vaccination Program which has achieved vaccine initiation rate of 82% among age-eligible females. Improving HPV vaccination rates is important in the prevention of morbidity and mortality associated with HPV-related disease. This study aimed to identify factors and barriers associated with uptake of the HPV vaccine in the Australian Program. Methods Between 2011 and 2014, females aged 18–25 years, living in Victoria, Australia who were offered HPV vaccination between 2007 and 2009 as part of the National HPV Vaccination Program, living in Victoria, Australia were recruited into a a young women’s study examining effectiveness of the Australian National HPV Vaccination Program. Overall, 668 participants completed the recruitment survey, which collected data of participants’ demographics and HPV knowledge. In 2015 these participants were invited to complete an additional supplementary survey on parental demographics and attitudes towards vaccinations. Results In 2015, 417 participants completed the supplementary survey (62% response rate). Overall, 19% of participants were unvaccinated. In multivariate analyses, HPV vaccination was significantly associated with their being born in Australia (p<0.001), having completed childhood vaccinations (p<0.001) and their parents being main decision-makers for participants’ HPV vaccination (p<0.001). The main reason reported for HPV non-vaccination was parental concern about vaccine safety (43%). Compared with HPV-vaccinated participants, those unvaccinated were significantly more likely to be opposed to all vaccines, including HPV vaccines (p<0.001) and were less likely to consider vaccinating their own children with all vaccines (p = 0.033), including HPV vaccines (p<0.001). Overall, 61% of unvaccinated participants reported that a recommendation from GPs would increase HPV vaccine acceptance. Conclusions Attitudes towards general health, vaccinations in general, as well as HPV vaccines are important in HPV vaccine uptake. Long-term monitoring of the knowledge, attitude and beliefs towards HPV vaccination in the community is critical to ensure a continued high uptake of the vaccine and success of the program. SMG received funding from the Victorian Cancer Agency TS10_04. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityThe minimum dataset has been uploaded as supplementary material with the revised manuscript.Data Availability The minimum dataset has been uploaded as supplementary material with the revised manuscript. ==== Body Introduction Persistent infection with high-risk human papillomavirus (HPV) types HPV16 and HPV18 cause 70% of all cancers of the cervix worldwide, and a portion of cancers of the vagina, vulva, anus, penis and head and neck [1–4]. In 2007, the Australian Government introduced a fully-funded National HPV Vaccination Program using a three dose course of the quadrivalent HPV (4vHPV) vaccine (that protects against infection by HPV types 16, 18, 6 and 11). Between 2007 and 2009, all girls aged 12–18 years were offered vaccination through schools, with a catch-up through community provided for women up to the age of 26 years [5–7]. In 2013, the program was extended to include 12–13 year old boys. Vaccination of girls and boys aged 12–13 is ongoing in schools under the National Immunisation Program [8]. In 2013, data from the National HPV Vaccination Program Register (NHVPR) showed that 86% of Australian adolescent girls, aged 12 to 13 received at least one vaccine dose, with 77% receiving all three doses, whilst 64% of those aged 20 to 26 received at least one dose [9]. High vaccine coverage has also been achieved in some other programmes internationally. For example the United Kingdom (school-based delivery) [10] and Denmark (clinic based) report 91% and 89% of school-aged girls have received at least one dose of the vaccine, with 87% and 74% receiving all three doses, respectively [11, 12]. Contrastingly, in the US, where HPV vaccination delivery is through healthcare providers on a reimbursement system, substantially lower vaccination rates have been achieved. In 2013, 57% of school-aged females in the US had received at least one dose of the vaccine and 38% all three doses [13]. Published research exploring attitudes, knowledge and socio-demographic factors associated with uptake of the HPV vaccine in high-income countries are mostly US-based and yield inconsistent findings [14–18]. In Australia, prior surveys have identified a number of socio-demographic, lifestyle and behavioural factors associated with vaccine uptake in the catch up phase but all have focused on individual factors [19, 20]. Few studies have explored reasons for non-vaccination with respect to parental factors. Given that parental views and attitude are likely to shape those of the child [21] and parental consent for HPV vaccination is required for girls under the age of 18, understanding parental factors towards the HPV vaccine is crucial to maintain high vaccine uptake. This study aimed to investigate individual and parental factors associated with uptake of the HPV vaccine among young women in the 2007 to 2009 catch-up phase of the National HPV Vaccination Program. Methods Recruitment Between September 2011 and December 2014 women aged 18–25 years, living in Victoria, Australia were recruited into the Vaccine Against Cervical Cancer Impact and Effectiveness (VACCINE) Study, through targeted advertisements placed on the social media website Facebook (Fig 1) as previously described [22]. Briefly, Facebook users who clicked on the VACCINE advertisement were taken to a study website, which provided details about the study and offered the opportunity to lodge an expression of interest in participating via a secure online form. Participants were then contacted by telephone by the study investigators and provided with an opportunity for questions and verbal consent. Those who provided verbal consent were sent an email with a link to the online participant information and consent form, which was hosted on the secure website SurveyMonkey.com. As part of the written consent, participants were asked to permit the researchers to verify their self-reported HPV vaccination details with the National HPV Vaccination Program Register (NHVPR). The NHVPR was established to monitor HPV vaccine uptake across Australia [23] and collects data on date of HPV vaccination, number of doses received and location of dose administration. Once informed consent was provided, a unique study number was issued, a further email was sent containing a link to the online questionnaire and a study pack containing materials for vaginal sample self-collection was mailed to the participant’s nominated address. This study was approved by The Royal Women’s Hospital Human Research and Ethics Committees (HREC number 11/15). 10.1371/journal.pone.0161846.g001Fig 1 Flowchart outlining study design and recruitment. For the current study, VACCINE participants who agreed to be contacted for additional research and who were confirmed as being vaccinated or unvaccinated by the NHVPR were invited to complete a supplementary survey (n = 1154). Participants were categorised as being “vaccinated” if they had received three doses of the vaccine, as recorded in the NHVPR. Participants were categorized as being “unvaccinated” if they self-reported not receiving the vaccine and the NHVPR had no record of any HPV vaccine dose being delivered. Those with fewer than three doses were excluded from the supplementary survey. Participants who self-reported that they had received the HPV vaccine overseas and therefore did not have a NHVPR record were excluded from the study. Measures Responses to two web-based surveys, including the original 2011–2014 VACCINE study survey and a supplementary survey administered in this study were collected. The former collected information on participants’ demographics, sexual history, cervical screening history and HPV knowledge [24], while the latter collected information about participants’ health, childhood vaccination history, attitudes towards vaccination, reasons for non-vaccination and parental characteristics including parental country of birth. Attitudes were assessed using seven-point Likert scales. Statistical analyses Univariate and adjusted logistic regression analysis were performed with Odds Ratios (ORs) and 95% confidence intervals (CI) generated to identify factors associated with vaccination. Factors included age at program commencement, country of birth, completion of childhood vaccinations, maternal and paternal countries of birth. All variables that were associated with vaccination at p<0.100 in univariate analysis were included in the initial multivariable model. A final multivariable model was obtained by performing backward elimination of statistically non-significant variables, each time assessing for confounding and co-linearity, until only statistically significant variables remained. Descriptive statistics were used to examine reasons for HPV non-vaccination in unvaccinated participants. Childhood socioeconomic status (lower or upper 50th centile classified as more or less disadvantaged) was derived from the Australian Bureau of Statistics Index of Relative Socioeconomic Disadvantage for each individual’s residential postcode. Childhood residential area (major city or) was based on the Accessibility/Remoteness Index of Australia (ARIA) classification[25, 26]. Chi-squared tests were performed to compare differences between HPV-vaccinated and unvaccinated cohorts. Data analyses were performed using STATA version 13 (Stata corporation, College Station, TX, US). Results Participant recruitment Between 2011 and 2014 1,154 women who were eligible to receive the free HPV vaccine during the 2007–2009 catch-up period were recruited into the VACCINE study. The women in this study aged 18–25 years in 2011 to 2014 would have been 11–21 years of age at the commencement of the program and therefore had access to the vaccine for free through either the school based or catch up program. Overall, 124 unvaccinated and 544 fully vaccinated participants who consented to be contacted for future studies were invited to the 2014 supplementary survey. Of these, 417 completed the survey (Fig 1). Responders and non-responders for the supplementary survey were not significantly different with respect to age at recruitment (p = 0.378), country of birth (p = 0.832), area of remoteness (p = 0.195), socioeconomic status (p = 0.677) and HPV vaccination status (p = 0.594). Cohort characteristics Overall, 337 (81%) participants were fully HPV-vaccinated and 80 (19%) were unvaccinated. Of the vaccinated participants, 68% received their first vaccine dose at school and 32% at a general practice (GP). The median (interquartile range) age at HPV vaccination for the first dose was 16 years (IQR: 15–18), with no significant difference in age between school- and GP-vaccinated participants (p = 0.262). Table 1 shows the distribution of the population characteristics of the 417 women included in this analysis. The median (interquartile range) age of participants at the supplementary survey was 24 (IQR: 22–25), with no differences by vaccination status. Overall, the majority of women (77%) were under 18 years of age at the commencement of the program. The vast majority were born in Australia (87%), with both parents being Australian born (63%), and 80% resided in metropolitan areas. Just over a quarter (28%) of participants reported that neither of their parents practice any religion. 10.1371/journal.pone.0161846.t001Table 1 Demographic, lifestyle and sexual behaviour characteristics among 417 participants who were offered HPV vaccination between 2007 and 2009 as part of the National HPV Vaccination Program, living in Victoria, by NHVPR confirmed vaccination status. Characteristic Total (N = 417) Unvaccinated (n = 80) Fully vaccinated (n = 337) p-value n (%) n (%) n (%) Birth cohort 1 1996–1994 46 (11.0) 8 (10.0) 38 (11.3) 0.318 1993–1992 95 (22.8) 21 (26.3) 74 (22.0) 1991–1990 124 (29.7) 17 (21.3) 107 (31.8) 1989–1988 108 (25.9) 26 (32.5) 82 (24.3) 1987–1986 44 (10.7) 8 (10.0) 36 (10.7) Age at program commencement 11–17 years old 320 (76.7) 59 (73.8) 261 (77.5) 0.482 18–21 years old 97 (23.3) 21 (26.3) 76 (22.6) Country of Birth Australia 361 (87.2) 53 (66.3) 308 (92.2) <0.001 Other4 53 (7.8) 27 (33.8) 26 (7.8) Childhood area of residency2 Major city 239 (66.0) 35 (68.6) 204 (65.6) 0.672 Regional or remote 123 (34.0) 16 (31.4) 107 (34.4) Childhood SES3 More disadvantaged 135 (37.3) 20 (39.2) 115 (37.0) 0.759 Less disadvantaged 227 (62.7) 31 (60.8) 196 (63.0) Parental country of birth Both Australian born 259 (63.0) 37 (48.1) 222 (66.5) <0.001 One parent born overseas5 80 (19.5) 12 (15.6) 68 (20.4) Both parents born overseas5 72 (17.5) 28 (36.4) 44 (13.2) Parental religion Both non-religious 113 (27.6) 24 (31.6) 89 (26.7) 0.517 One parent religious 100 (24.5) 20 (26.3) 80 (24.0) Both parents religious 196 (47.9) 32 (42.1) 164 (49.3) Childhood vaccinations Incomplete 23 (5.5) 13 (16.3) 10 (3.0) <0.001 Complete 394 (94.5) 67 (83.8) 327 (97.0) Main decision maker regarding receiving the HPV vaccine Self 181 (43.4) 54 (67.5) 127 (37.7) <0.001 One or both parents 236 (56.6) 26 (32.5) 210 (62.3) Tobacco use Never smoked 336 (81.8) 57 (72.2) 279 (84.0) 0.014 Past or current smoker 75 (18.3) 22 (27.8) 53 (16.0) Private health insurance No 174 (41.7) 44 (55.0) 130 (38.6) 0.007 Yes 243 (58.3) 36 (45.0) 207 (61.4) Age at first sex 11–17 185 (55.9) 30 (50.0) 155 (57.2) 0.310 18–24 146 (44.1) 30 (50.0) 116 (42.8) Lifetime number of sexual partners 0 81 (20.0) 17 (22.4) 64 (19.5) 0.530 1–3 167 (41.2) 27 (35.5) 140 (42.6) 4+ 157 (38.8) 32 (42.1) 125 (38.0) Current contraceptive use Yes 296 (89.2) 49 (81.7) 247 (90.8) 0.039 No 36 (10.8) 11 (18.3) 25 (9.2) Date of last Pap smear Within last two years 199 (89.6) 31 (81.6) 168 (91.3) 0.073 More than two years ago 23 (10.4) 7 (18.4) 16 (8.7) 1: Birth year corresponds to women between 11–21 years old as of the 1st of January 2007 who were offered HPV vaccination between 2007 and 2009 as part of the National HPV Vaccination Program; 2: based on the Accessibility/Remoteness Index of Australia (ARIA) classification; 3: based on the Australian Bureau of Statistics Index of Relative Socioeconomic Disadvantage for each individuals residential postcode; 4: Other countries of birth include New Zealand, China, Fiji, Finland, Germany, Hong Kong, India, Indonesia, Japan, Kenya, Malaysia, Singapore, Serbia, South Africa, Sri Lanka, Sweden, United Kingdom, US, Vietnam; 5: Overseas countries correspond to New Zealand, Bangladesh, Brunei, Canada, Chile, China, Egypt, Fiji, Finland, France, Germany, Hong Kong, India, Indonesia, Israel, Italy, Kenya, Latvia, Lebanon, Malaysia, Malta, Mauritius, Netherlands, Pakistan, Papua New Guinea, Philippines, Poland, Republic of Malawai, Sweden, Singapore, Spain, South Africa, Sri Lanka, Tanzania, United Kingdom, US, Vietnam, Zimbabwe. Abbreviations: SES: Socioeconomic status; Pap: Papanicolaou. Numbers do not always total 417 because of small amounts of missing data. Compared with unvaccinated women, those HPV-vaccinated were significantly more likely to be born in Australia (<0.001), be non-smokers (p = 0.014) and have private health insurance (p = 0.007). HPV-vaccinated and unvaccinated participants were not different with respect to age at first sex (p = 0.310) and lifetime number of sexual partners (p = 0.530). However, HPV-vaccinated participants were more likely to use contraception (p = 0.039) and have had a Pap smear within the last two years (p = 0.073); however, the latter did not reach significance (Table 1). Factors associated with HPV vaccination The results of univariate and multivariable analyses are presented in Table 2. In univariate analysis, being vaccinated was significantly associated with being born in Australia (p<0.001), having completed childhood vaccinations (p<0.001), having one or both parents being born in Australian (p<0.001) and one or both parents being main decision makers regarding receiving the HPV vaccine (p<0.001). Age, socioeconomic status and area of remoteness during childhood, and parental religion were not significantly associated with being vaccinated. In multivariable analysis, factors that remained significantly associated with HPV vaccination included being born in Australia (p<0.001), having completion of childhood vaccinations (p<0.001) and one or both parents being main decision maker regarding receiving the HPV vaccine (p<0.001). The results were the same when the analysis was limited to only women who were under 18 years of age at commencement of the National HPV Vaccination Program in 2007. 10.1371/journal.pone.0161846.t002Table 2 Factors associated with receipt of the HPV vaccine between 2007 and 2009 as part of the National HPV Vaccination Program among 417 female participants living in Victoria, Australia, overall and stratified by age-group at commencement of the HPV vaccination program. Factors Overall cohort (N = 417) < 18 years old at program commencement (n = 320) 18 years or older at program commencement (n = 97) OR (95% CI) p–value Adjusted a OR (95%CI) p-value Adjusted b OR (95%CI) p-value Adjusted b OR (95%CI) p-value Age at program commencement 11–17 years old 1.00 0.482 1.00 0.949 18–21 years old 0.82 (0.47–1.43) 0.97 (0.42–2.27) Country of birth Australia 1.00 <0.001 1.00 <0.001 1.00 0.002 1.00 0.399 Other 0.17 (0.09–0.31) 0.21 (0.09–0.55) 0.18 (0.06–0.53) 0.38 (0.04–3.89) Childhood SES1 More disadvantaged 1.00 0.759 Less disadvantaged 1.10 (0.56–2.02) Childhood area of residency2 Major city 1.00 0.672 Regional or remote 1.15 (0.61–2.17) Childhood vaccinations Incomplete 1.00 <0.001 1.00 <0.001 1.00 <0.001 1.00 0.130 Complete 6.34 (2.67–15.07) 6.99 (2.71–17.99) 10.81 (3.2–36.32) 4.00 (0.63–20.71) Parental country of birth Both Australia born 1.00 1.00 1.00 1.00 One parent born overseas 0.94 (0.47–1.91) 0.874 1.19 (0.55–2.62) 0.651 1.42 (0.56–3.57) 0.457 0.78 (0.17–3.45) 0.694 Both parents born overseas 0.26 (0.15–0.47) <0.001 0.82 (0.32–2.07) 0.671 0.95 (0.33–2.73) 0.926 0.42 (0.05–3.57) 0.483 Main decision maker regarding receiving the HPV vaccine Self 1.00 <0.001 1.00 <0.001 1.00 0.004 1.00 0.097 One or both parents 3.43 (2.05–5.76) 3.10 (1.66–5.80) 2.72 (1.37–5.38) 6.21 (0.72–53.50) Parental religion Both non-religious 1.00 One parent religious 1.08 (0.55–2.10) 0.824 Both parents religious 1.38 (0.77–2.49) 0.281 1: Socioeconomic status (SES) based on the Accessibility/Remoteness Index of Australia (ARIA) classification; 2: based on the Australian Bureau of Statistics Index of Relative Socioeconomic Disadvantage for each individuals residential postcode; a: Adjusted for age at program commencement, childhood vaccinations, country of birth, parental country of birth, main decision maker regarding HPV vaccination and participant age; b: Adjusted for all the same variables as in (a) except for age at program commencement. HPV knowledge, source of information and attitudes towards HPV vaccination Participants were asked to complete an HPV knowledge test and report their most trusted sources of HPV information. Overall, 97% of participants knew HPV causes cervical cancer and 99% knew regular Pap smears were required after HPV vaccination (Table 3). Two thirds (68%) knew HPV causes genital warts. HPV-vaccinated participants were more likely to correctly identify that HPV vaccination reduces the risk of cervical cancer (p<0.001), that regular Pap smears are required after HPV vaccination (p = 0.015), and that HPV vaccine protects against 70% of cervical cancers (p = 0.027). Overall, 83% of participants reported their GP to be the most trusted source of HPV information. In addition, compared with unvaccinated, vaccinated participants were significantly more likely to report GPs as their most trusted source of HPV information (85.5% versus 71.3%, p = 0.003). 10.1371/journal.pone.0161846.t003Table 3 HPV knowledge among 417 participants who were offered HPV vaccination between 2007 and 2009 as part of the National HPV Vaccination Program, living in Victoria, Australia, stratified by NHVPR confirmed vaccination status. Frequency of correct responses presented. Knowledge Total (N = 417) Vaccinated (N = 337) Unvaccinated (N = 80) p-value n (%) n (%) n (%) HPV infection causes genital warts 261 (67.6) 221 (69.5) 40 (58.8) 0.088 HPV infection causes cervical cancer 374 (96.9) 310 (97.5) 64 (94.1) 0.147 HPV vaccination reduces risk of cervical cancer 398 (95.4) 331 (98.2) 67 (83.8) <0.001 Regular Pap smears are still required after HPV vaccination 393 (98.7) 323 (99.4) 70 (95.9) 0.015 HPV vaccine protects against 70–80% of cervical cancers 287 (72.1) 242 (74.5) 45 (61.6) 0.027 Abnormal Pap smears may still occur after HPV vaccination 370 (93.0) 303 (93.2) 67 (91.8) 0.662 Next, participants were asked to report their agreement with a number of statements about vaccination on a seven-point Likert scale. The results of this are presented in Table 4. Unvaccinated participants were significantly more likely to report neutral to opposing views towards vaccinations in general (p = 0.001) and towards HPV vaccination (p<0.001), than vaccinated participants. In addition, unvaccinated participants were more likely to report neutral or opposing views when asked if they would vaccinate their children with the HPV vaccine now or in the future (p = 0.033). Attitudes around stigma of HPV vaccination and that it implied promiscuity due to HPV being sexually transmitted were not significantly different between the two groups (Table 4). 10.1371/journal.pone.0161846.t004Table 4 Attitude towards vaccination and the HPV vaccine among 417 participants who were offered HPV vaccination between 2007 and 2009 as part of the National HPV Vaccination Program, living in Victoria, Australia, stratified by NHVPR confirmed vaccination status. Total (N = 417) Vaccinated (N = 337) Unvaccinated (N = 80) p-value n (%) n (%) n (%) I am opposed to vaccination in general (any vaccine) Disagree 393 (94.5) 324 (96.4) 69 (86.3) 0.001 Neutral 17 (4.1) 8 (2.4) 9 (11.3) Agree 6 (1.4) 4 (1.2) 2 (2.5) I would not vaccinate my children with any vaccine now or in the future Disagree 393 (94.5) 322 (95.8) 71 (88.8) 0.033 Neutral 17 (4.1) 11 (3.3) 6 (7.5) Agree 6 (1.4) 3 (0.9) 3 (3.8) I am opposed to HPV vaccination Disagree 383 (92.7) 322 (96.7) 61 (76.3) <0.001 Neutral 25 (6.1) 9 (2.7) 16 (20.0) Agree 5 (1.2) 2 (0.6) 3 (3.8) I would not vaccinate my children with the HPV vaccine now or in the future Disagree 378 (91.1) 319 (95.2) 59 (73.8) <0.001 Neutral 31 (7.5) 13 (3.9) 18 (22.5) Agree 6 (1.5) 3 (0.9) 3 (3.8) There is a stigma attached to the HPV vaccine Disagree 296 (71.2) 242 (72.0) 54 (67.5) 0.193 Neutral 107 (25.8) 86 (25.6) 21 (26.3) Agree 13 (3.1) 8 (2.4) 5 (6.3) Having the HPV vaccine implies a person may be or may become sexually promiscuous Disagree 324 (77.9) 268 (79.8) 56 (70.0) 0.083 Neutral 82 (19.7) 62 (18.5) 20 (25.0) Agree 10 (2.4) 6 (1.8) 4 (5.0) Attitudes were measured using a 7-point Likert scales (Strongly disagree and Disagree were grouped as “Disagree”, Neutral, Agree and Strongly Agree were grouped as “Agree”)Numbers do not always total 417 because of small amounts of missing data. Reasons for non-vaccination Unvaccinated participants were asked to report their reasons for why they believed they were not vaccinated. The most commonly reported reasons included parental concern about vaccine safety (42.5%), parental perception of their daughter being at low risk of HPV infection (22.5%) or having a needle phobia (21.3%). Fifteen percent (15.0%) of unvaccinated participants reported practical barriers to HPV vaccination, including being absent at school and forgetting to bring their signed consent forms on the day of vaccination. Only 5.0% reported parental belief that HPV vaccination promoted promiscuity as the reason for not being vaccinated (Fig 2A). When stratified by age at program commencement, parental concern about vaccine safety was reported by 57.6% of women who were <18 years of age, compared to only 4.7% of those who were 18 years and over (p<0.001). 10.1371/journal.pone.0161846.g002Fig 2 Reported reasons for not receiving the HPV vaccine (A) and; measures to improve HPV vaccination rates (B) among unvaccinated female participants who were eligible for the HPV vaccine between 2007 and 2009 as part of the National HPV Vaccination Program. Unvaccinated participants were asked to report factors that would encourage them to get the HPV vaccine (Fig 2B). The most commonly reported factors to encourage HPV vaccination included that vaccination was free (77.5%), and receiving a recommendation by a doctor to get vaccinated (61.3%). Participants reported that they were more likely to accept HPV vaccination, if they believed their family and friends approved of their decision to become vaccinated (71.3%). Some participants believed the HPV vaccine to be still new, but might become more likely to accept the vaccine with time (33.8%). Discussion In this study of 417 young women living in Victoria Australia, we investigated attitudes, knowledge and factors associated with receipt of the HPV vaccine in the National HPV Vaccination Program. Compared with the vaccinated, unvaccinated women were more likely to be smokers and were less likely to use contraception. Unvaccinated women were also more likely to report neutral or negative attitudes towards the HPV vaccine and vaccinations in general. Sexual behaviour and HPV knowledge did not differ between the two groups. Independent factors associated with receipt of the vaccine in the school-based program included being Australia born and having completed childhood vaccinations. Among unvaccinated women, parental concerns about vaccine safety and perceived low risk of HPV infection for their daughters were predominant reasons for non-vaccination. The results of the study emphasize the need for long-term monitoring of the knowledge, attitude and beliefs towards HPV vaccination in the community to ensure high uptake of the vaccine and success of the program. In Australia, those offered HPV vaccination through the school-based program are provided with information on the vaccine at the same time of receiving the consent form [27]. Despite this, nearly half (43%) of unvaccinated participants in this study reported parental concerns about vaccine safety as the main reason for not receiving the vaccine. Published Australian research suggests that when presented with facts about the HPV vaccine, some parents remained hesitant towards HPV vaccination due to an underlying mistrust of the health care system and belief of vaccinations to cause disability [28]. In fact, unvaccinated women in our study were significantly more likely to report incomplete childhood vaccinations in general. These findings are consistent with published Australian and international data suggesting that HPV non-vaccination is partly driven by parental disapproval of vaccination in general [29, 30]. Studies on HPV vaccine acceptability have reported parents’ health behaviours, beliefs and knowledge about HPV are significant predictors of parental intent for vaccination [31, 32]. In a US study of linked electronic health records of girls (aged 9 to 17 years) the authors found that mothers’ attitudes about preventative measures such has the Pap test influenced their adolescent daughters update of the HPV vaccine [32]. In turn, individual attitudes and beliefs around health and vaccination are likely to be influenced by parental views and behaviours [33–35]. Unvaccinated participants in our study were more likely to be smokers, were less likely to use contraception and had incomplete attendance at cervical screening, although the latter did not reach significance. Unvaccinated participants were also more likely to report neutral or negative attitudes towards vaccinations (including the HPV vaccine), and report neutral or negative attitudes towards future intentions of vaccinating their own children, suggesting they themselves were hesitant towards vaccination. Among unvaccinated women who reported neutral or negative attitudes towards vaccinations, the majority took a neutral position. Research into parental opinions on immunisation have identified a continuum of views on vaccine hesitancy ranging from unquestionable acceptor through hesitancy to refuser [36, 37]. Accurate understanding of an individual’s HPV perception is critically important and may help guide effective strategies to promote vaccine acceptance. This is important given the fact that those identified as having a lesser degree of hesitancy are more likely to accept full vaccination uptake [38]. Primary health care provider recommendation remains a key strategy to increase HPV vaccination rates and promote vaccine acceptance [39, 40]. In our study, GPs were the most trusted sources of HPV information, with 61% unvaccinated participants reporting that they would accept HPV vaccination if it were recommended by a GP. Health care providers (GPs or nurses within a practice) play an important role in providing information and services for women, and are well placed to encourage participation in cervical screening. While the HPV vaccine is no longer freely available outside of the target age-range, it is important that heath care providers continue to use opportunistic visits by young women to engage in general discussion around HPV prevention, including vaccination. Providing young women with the opportunity to voice their concerns should form a central part of these discussions. Evidence based frameworks have been developed to assist health professionals when communicating with parents about vaccination [37]. Such frameworks should be expanded to include young adults, regardless of where they are along the decision-making trajectory. In line with previous literature, this study showed that being born overseas was associated with lower HPV vaccine uptake [16]. This is in part driven by families who have arrived in Australia outside of the catch-up program window. However, published research has demonstrated that coverage of various vaccinations are lower in new migrants to Australia. This was attributed to a number of factors including language barriers, lack of specialised migrant health services, lack of awareness in migrants of their rights within the Australian health system, and cultural differences in attitudes towards preventative health care [41]. Australia is ethnically diverse and in 2011, 27% of Australians were first generation migrants [42]. It is therefore essential for public health prevention programs to target new migrants using culturally sensitive approaches, to ensure equitable delivery of vaccination to all population subgroups. This study has several limitations. First, the retrospective nature of the study design means that it may be subjected to recall bias. Furthermore, the views of parents were reported by the participants rather than the parents themselves and may therefore be inaccurate or misconstrued. Second, reasons for non-vaccination are likely to be different in the catch-up program than in the school program. However in our study, the proportion of women who were aged 18 years and over at the time of the catch-up program was small (23%). Therefore, we had limited power to investigate factors associated with HPV vaccination in this group. Third, we did not demonstrate an association between geographic remoteness and uptake of the HPV vaccine. While this is consistent with published Australian data which shows a relatively equal uptake across urban and regional areas, lower vaccine completion rates have been reported among residents of remote areas [43]. It is likely that our study lacked power to demonstrate this association as only 19 women resided in outer regional Australia and none resided in remote Australia at the time of the catch up program. Last, our sample was limited to Victorian women only. While recruitment via social media has been shown to yield a broadly representative sample when compared with age-equivalent census data [44], there are substantial geographical and population variations between states and territories [27, 45] therefore the results may not be generalizable to all Australian women. Furthermore, the results may not be generalizable to other countries due to differences in delivery models of HPV vaccination. Conclusions In summary, attitudes towards health, HPV infection and vaccinations may impact on the success of the HPV vaccination program. It is important for public health campaigns to continue to emphasise the efficacy and safety of the HPV vaccinations as a preventative health measure. Furthermore, long-term monitoring of the knowledge, attitude and beliefs towards HPV vaccination in the community is critical to ensure a continued high uptake of the vaccine and success of the program into the future. Supporting Information S1 File Tung data. (XLS) Click here for additional data file. We would like to acknowledge the Victorian Cancer Agency for their TS10_04 grant– 2011–2013, Monitoring the effectiveness of the Australian cervical cancer vaccine programme: translation to reduction in vaccine-related HPV infection and precancerous cervical lesions. We thank members of the VACCINE study group for project management and coordination, participants of the VACCINE study for their contribution to data of this study and Athena Costa for assisting with the manuscript submission. We thank the NHVPR staff who assisted with the study. The NHVPR is owned by the Australian Government Department of Health and operated by the Victorian Cytology Service. ==== Refs References 1 Bzhalava D , Guan P , Franceschi S , Dillner J , Clifford G . A systematic review of the prevalence of mucosal and cutaneous human papillomavirus types . Virology . 2013 ;445 (1–2 ):224 –31 . 10.1016/j.virol.2013.07.015 23928291 2 Koutsky PL . Epidemiology of Genital Human Papillomavirus Infection . The American Journal of Medicine . 1997 ;102 (5, Supplement 1 ):3 –8 . 3 Lowy DR , Kirnbauer R , Schiller JT . Genital human papillomavirus infection . Proceedings of the National Academy of Sciences of the United States of America . 1994 ;91 (7 ):2436 –40 . 8146136 4 zur Hausen H . Condylomata acuminata and human genital cancer . Cancer research . 1976 ;36 (2 pt 2 ):794 175942 5 Garland S , Steben M , Sings H , James M , Lu S , Railkar R , et al Natural history of genital warts: analysis of the placebo arm of 2 randomized phase III trials of a quadrivalent human papillomavirus (types 6, 11, 16, and 18) vaccine . Journal of Infectious Diseases . 2009 ;199 (6 ):805 –14 . 10.1086/597071 19199546 6 Garland SM , Hernandez-Avila M , Wheeler CM , Perez G , Harper DM , Leodolter S , et al Quadrivalent vaccine against human papillomavirus to prevent anogenital diseases . New England Journal of Medicine . 2007 ;356 (19 ):1928 –43 . 17494926 7 Quadrivalent vaccine against human papillomavirus to prevent high-grade cervical lesions . The New England journal of medicine . 2007 ;356 (19 ):1915 –27 . 17494925 8 Australian Government DoH . Human papillomavirus, Immunise Australia Program Australia : Australian Government ; 2015 [updated 2015 April 20; cited 2016 May 27]; Available from: http://www.immunise.health.gov.au/internet/immunise/publishing.nsf/Content/immunise-hpv 9 National HPV Vaccination Program Register . HPV Vaccination Coverage by Dose 2013 . Australia : National HPV Vaccination Program Register ; [updated 2015 June; cited 2015 July 07]; Available from: http://www.hpvregister.org.au/research/coverage-data/HPV-Vaccination-Coverage-by-Dose-20132 10 Public Health England . Annual HPV Vaccine coverage 2013 to 2014: by PCT, local authority and area team . United Kingdom : Public Health England ; [updated 2014 December 02; cited 2015 August 18]; Available from: https://www.gov.uk/government/statistics/annual-hpv-vaccine-coverage-2013-to-2014-by-pct-local-authority-and-area-team 11 Slåttelid Schreiber SM , Juul KE , Dehlendorff C , Kjær SK . Socioeconomic Predictors of Human Papillomavirus Vaccination Among Girls in the Danish Childhood Immunization Program . Journal of Adolescent Health . 2015 ;56 (4 ):402 –7 . 10.1016/j.jadohealth.2014.12.008 25659994 12 Statens Serum Institut . EPI-NEWS, HPV Vaccination—Coverage 2012 . Copenhagen, Denmark : Statens Serum Institut ; [updated 15 May 2013; cited 2015 August 18]; Available from: http://www.ssi.dk/English/News/EPI-NEWS/2013/No%2020%20-%202013.aspx 13 Centers for Disease Control and Prevention . Human Papillomavirus Vaccination Coverage Among Adolescents, 2007–2013, and Postlicensure Vaccine Safety Monitoring, 2006–2014 . United States : Centers for Disease Control and Prevention ; [updated 2014 July 25; cited 2015 February 15]; Available from: http://www.cdc.gov/mmwr/preview/mmwrhtml/mm6329a3.htm 14 Brewer NT , Fazekas KI . Predictors of HPV vaccine acceptability: A theory-informed, systematic review . Preventive Medicine . 2007 ;45 (2–3 ):107 –14 . 17628649 15 Fisher WA . Understanding Human Papillomavirus Vaccine Uptake . Vaccine . 2012 ;30 , Supplement 5 (0):F149 –F56 .23199958 16 Jeudin P , Liveright E , Del Carmen MG , Perkins RB . Race, ethnicity, and income factors impacting human papillomavirus vaccination rates . Clinical therapeutics . 2014 ;36 (1 ):24 –37 . 10.1016/j.clinthera.2013.11.001 24417783 17 Rambout L , Tashkandi M , Hopkins L , Tricco AC . Self-reported barriers and facilitators to preventive human papillomavirus vaccination among adolescent girls and young women: a systematic review . Prev Med . 2014 ;58 :22 –32 . 10.1016/j.ypmed.2013.10.009 24176938 18 Walhart T . Parents, adolescents, children and the human papillomavirus vaccine: a review . International nursing review . 2012 ;59 (3 ):305 –11 . 10.1111/j.1466-7657.2012.00991.x 22897180 19 Brotherton JML , Piers LS , Vaughan L . Estimating human papillomavirus vaccination coverage among young women in Victoria and reasons for non-vaccination . Sexual health . 2016 ;13 (2 ):190 –2 . 10.1071/SH15131 26615408 20 Canfell K , Egger S , Velentzis LS , Brown JD , O’Connell DL , Banks E , et al Factors related to vaccine uptake by young adult women in the catch-up phase of the National HPV Vaccination Program in Australia: Results from an observational study . Vaccine . 2015 ;33 (20 ):2387 –94 . 10.1016/j.vaccine.2015.01.024 25843203 21 Roberts ME , Gerrard M , Reimer R , Gibbons FX . Mother-Daughter Communication and Human Papillomavirus Vaccine Uptake by College Students . Pediatrics . 2010 ;125 (5 ):982 –989 . Pediatrics. 2010;126(3):617-. 10.1542/peds.2009-2888 20385645 22 Gertig DM , Brotherton JM , Budd AC , Drennan K , Chappell G , Saville AM . Impact of a population-based HPV vaccination program on cervical abnormalities: a data linkage study . (1741–7015 (Electronic)). 23 Gertig DM , Brotherton JML , Saville M . Measuring human papillomavirus (HPV) vaccination coverage and the role of the National HPV Vaccination Program Register, Australia . Sexual health . 2011 ;8 (2 ):171 –8 . 10.1071/SH10001 21592430 24 Osborne SL , Tabrizi SN , Brotherton JML , Cornall AM , Wark JD , Wrede CD , et al Assessing genital human papillomavirus genoprevalence in young Australian women following the introduction of a national vaccination program . Vaccine . 2015 ;33 (1 ):201 –8 . 10.1016/j.vaccine.2014.10.045 25444787 25 Australian Bureau of Statistics . Census of Population and Housing: Socio-Economic Indexes for Areas (SEIFA), Australia, 2011 . Australia : Australian Bureau of Statistics ; 2013 [updated 2014 November 12; cited 2015 March 17]; Available from: http://www.abs.gov.au/ausstats/abs@.nsf/DetailsPage/2033.0.55.0012011 26 Australian Bureau of Statistics . Australian Statistical Geography Standard (ASGS): Correspondences, July 2011 . Australia : Australian Bureau of Statistics ; [updated 2013 February 14; cited 2015 March 13]; Available from: http://www.abs.gov.au/AUSSTATS/abs@.nsf/DetailsPage/1270.0.55.006July%202011?OpenDocument 27 Cooper Robbins SC , Bernard D , McCaffery K , Brotherton J , Garland S , Skinner SR . "Is cancer contagious?": Australian adolescent girls and their parents: making the most of limited information about HPV and HPV vaccination . Vaccine . 2010 ;28 (19 ):3398 –408 . 10.1016/j.vaccine.2010.02.078 20199758 28 Leask J , Chapman S , Hawe P . Concerns about immunisation. Facts are not enough . Bmj . 2000 ;321 (7253 ):109 . 29 Kessels SJM , Marshall HS , Watson M , Braunack-Mayer AJ , Reuzel R , Tooher RL . Factors associated with HPV vaccine uptake in teenage girls: A systematic review . Vaccine . 2012 ;30 (24 ):3546 –56 . 10.1016/j.vaccine.2012.03.063 22480928 30 Walhart T . Parents, adolescents, children and the human papillomavirus vaccine: a review . International nursing review . 2012 ;59 (3 ):305 –11 . 10.1111/j.1466-7657.2012.00991.x 22897180 31 Kahn JA , Rosenthal SL , Hamann T , Bernstein DI . Attitudes about human papillomavirus vaccine in young women . International journal of STD & AIDS . 2003 ;14 (5 ):300 –6 .12803935 32 Chao C , Slezak JM , Coleman KJ , Jacobsen SJ . Papanicolaou screening behavior in mothers and human papillomavirus vaccine uptake in adolescent girls . Am J Public Health . 2009 ;99 (6 ):1137 –42 . 10.2105/AJPH.2008.147876 19372507 33 Mechanic D . THE INFLUENCE OF MOTHERS ON THEIR CHILDREN'S HEALTH ATTITUDES AND BEHAVIOR . Pediatrics . 1964 ;33 :444 –53 . 14131612 34 Rosenthal SL , Rupp R , Zimet GD , Meza HM , Loza ML , Short MB , et al Uptake of HPV Vaccine: Demographics, Sexual History and Values, Parenting Style, and Vaccine Attitudes . Journal of Adolescent Health . 2008 ;43 (3 ):239 –45 . 10.1016/j.jadohealth.2008.06.009 18710678 35 House SH . Transgenerational healing: Educating children in genesis of healthy children, with focus on nutrition, emotion, and epigenetic effects on brain development . Nutrition and Health . 2014 ;22 (1 ):9 –45 . 25005446 36 Benin AL , Wisler-Scher DJ , Colson E , Shapiro ED , Holmboe ES . Qualitative analysis of mothers' decision-making about vaccines for infants: the importance of trust . Pediatrics . 2006 ;117 (5 ):1532 –41 . 16651306 37 Leask J , Kinnersley P , Jackson C , Cheater F , Bedford H , Rowles G . Communicating with parents about vaccination: a framework for health professionals . BMC pediatrics . 2012 ;12 :154 10.1186/1471-2431-12-154 22998654 38 Forbes TA , McMinn A , Crawford N , Leask J , Danchin M . Vaccination uptake by vaccine-hesitant parents attending a specialist immunization clinic in Australia . Human vaccines & immunotherapeutics . 2015 ;11 (12 ):2895 –903 .26366978 39 Balakrishnan VS . Physicians' recommendation affects HPV vaccination uptake . The Lancet Oncology . 2015 ;16 (16 ):e592 . 40 Giles M , Garland S . A study of women's knowledge regarding human papillomavirus infection, cervical cancer and human papillomavirus vaccines . Australian and New Zealand Journal of Obstetrics and Gynaecology . 2006 ;46 (4 ):311 –5 . 16866792 41 Murray SB , Skull SA . Hurdles to health: immigrant and refugee health care in Australia . Australian Health Review . 2005 ;29 (1 ):25 –9 . 15683352 42 Australian Bureau of Statistics . Reflecting a Nation: Stories from the 2011 Census, 2012–2013 Australia : Australian Bureau of Statistics ; 2012 [updated 2012 June 21; cited 2015 2015 July 07]; Available from: http://www.abs.gov.au/ausstats/abs@.nsf/Lookup/2071.0main+features902012-2013 43 Barbaro B , Brotherton JM . Assessing HPV vaccine coverage in Australia by geography and socioeconomic status: are we protecting those most at risk? Aust N Z J Public Health . 2014 ;38 (5 ):419 –23 . 10.1111/1753-6405.12218 24962721 44 Fenner Y , Garland SM , Moore EE , Jayasinghe Y , Fletcher A , Tabrizi SN , et al Web-based recruiting for health research using a social networking site: an exploratory study . Journal of medical Internet research . 2012 ;14 (1 ):e20 10.2196/jmir.1978 22297093 45 Ward K , Quinn H , Bachelor M , Bryant V , Campbell-Lloyd S , Newbound A , et al Adolescent school-based vaccination in Australia . Communicable diseases intelligence quarterly report . 2013 ;37 (2 ):E156 –67 . 24168090

PMC005xxxxxx/PMC5001626.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756438210.1371/journal.pone.0161974PONE-D-16-20343Research ArticleBiology and Life SciencesAnatomyHeadEyesMedicine and Health SciencesAnatomyHeadEyesBiology and Life SciencesAnatomyOcular SystemEyesMedicine and Health SciencesAnatomyOcular SystemEyesMedicine and Health SciencesOphthalmologyEye DiseasesGlaucomaBiology and Life SciencesAnatomyOcular SystemOcular AnatomyLens (Anatomy)Medicine and Health SciencesAnatomyOcular SystemOcular AnatomyLens (Anatomy)Biology and Life SciencesAnatomyOcular SystemOcular AnatomyPupilMedicine and Health SciencesAnatomyOcular SystemOcular AnatomyPupilMedicine and Health SciencesOphthalmologyVisual ImpairmentsScotomaMedicine and Health SciencesOphthalmologyEye DiseasesMedicine and Health SciencesOphthalmologyResearch and Analysis MethodsResearch DesignPilot StudiesVisual Field Testing with Head-Mounted Perimeter ‘imo’ Visual Field Testing with Head-Mounted Perimeter 'imo'Matsumoto Chota 1*Yamao Sayaka 1Nomoto Hiroki 1Takada Sonoko 1Okuyama Sachiko 1Kimura Shinji 2Yamanaka Kenzo 2Aihara Makoto 3Shimomura Yoshikazu 11 Department of Ophthalmology, Kindai University, Faculty of Medicine Osaka-Sayama City, Osaka, Japan2 CREWT Medical Systems, Inc., Tokyo, Japan3 Department of Ophthalmology, Graduate School of Medicine and Faculty of Medicine, The University of Tokyo, Tokyo, JapanAnderson Andrew EditorThe University of Melbourne, AUSTRALIACompeting Interests: The authors have the following interests: this study was funded by CREWT Medical Systems. SK and KY are employees of CREWT Medical Systems. SK and KY are directors and have proprietary interests with CREWT Medical Systems. This is the first study exploring the diagnostic abilities of this new instrument developed by the authors and CREWT Medical Systems. There are no products under development that are relevant to the materials presented in this article, except for the ongoing development of imo. There are patent applications issued by CM and CREWT Medical Systems. This does not alter the authors' adherence to all the PLOS ONE policies on sharing data and materials. Conceptualization: CM MA YS. Data curation: HN SK KY. Formal analysis: HN CM. Funding acquisition: YS CM. Investigation: SY. Methodology: CM. Project administration: CM. Resources: SY ST SO. Software: SK KY. Supervision: CM. Validation: SO. Visualization: CM NH. Writing – original draft: CM. Writing – review & editing: MA HN. * E-mail: chota@med.kindai.ac.jp26 8 2016 2016 11 8 e016197420 5 2016 15 8 2016 © 2016 Matsumoto et al2016Matsumoto et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Purpose We developed a new portable head-mounted perimeter, “imo”, which performs visual field (VF) testing under flexible conditions without a dark room. Besides the monocular eye test, imo can present a test target randomly to either eye without occlusion (a binocular random single eye test). The performance of imo was evaluated. Methods Using full HD transmissive LCD and high intensity LED backlights, imo can display a test target under the same test conditions as the Humphrey Field Analyzer (HFA). The monocular and binocular random single eye tests by imo and the HFA test were performed on 40 eyes of 20 subjects with glaucoma. VF sensitivity results by the monocular and binocular random single eye tests were compared, and these test results were further compared to those by the HFA. The subjects were asked whether they noticed which eye was being tested during the test. Results The mean sensitivity (MS) obtained with the HFA highly correlated with the MS by the imo monocular test (R: r = 0.96, L: r = 0.94, P < 0.001) and the binocular random single eye test (R: r = 0.97, L: r = 0.98, P < 0.001). The MS values by the monocular and binocular random single eye tests also highly correlated (R: r = 0.96, L: r = 0.95, P < 0.001). No subject could detect which eye was being tested during the examination. Conclusions The perimeter imo can obtain VF sensitivity highly compatible to that by the standard automated perimeter. The binocular random single eye test provides a non-occlusion test condition without the examinee being aware of the tested eye. CREWT Medical Systems has funded this work. The funders had a role in providing the device. Co-authors SK and KY are employed by CREWT Medical Systems. Data AvailabilityAll relevant data are within the paper and the Figures.Data Availability All relevant data are within the paper and the Figures. ==== Body Introduction Visual field (VF) testing is essential in diagnosing and monitoring many ophthalmological and neurological diseases. Automated perimeters such as the Humphrey Field analyzer (HFA) (Carl Zeiss Meditec, Dublin, CA) and Octopus perimeter (Haag-Streit, Koeniz, Switzerland) have been widely used in the field of standard automated perimetry (SAP). However, most of these perimeters are stationary type devices that need to be used in a dim testing room with light control. In addition to issues such as portability and space restriction for the standard automated perimeters, patients with special physical conditions may experience difficulty or discomfort trying to physically adapt themselves to the stationary type devices during the test. Therefore, a patient-oriented perimeter with better flexibility in performing the test is a pressing need and a head-mounted perimeter appears to be ideal. Several head-mounted perimeters that do not require a dark room for VF testing have been previously developed for laboratory-based studies[1–3]. However, there is no commercially available head-mounted perimeter devised for clinical setting. We recently developed a new portable head-mounted perimeter named “imo” (CREWT Medical Systems, Tokyo, Japan). imo can perform VF testing without a dark room and under test conditions compatible with those for SAP. It also allows patients to be tested with better comfort at any location. Moreover, as a unique feature of this device, the test target is randomly presented to either eye without occlusion and without the examinee being aware of which eye is being tested (the binocular random single eye test). In this study, we reported the detailed features of our new device. As a pilot study, VF sensitivity values obtained using the imo monocular and binocular random single eye tests were compared with the HFA results in patients with glaucoma. Subjects and Methods The head-mounted perimeter ‘imo’ The head-mounted perimeter imo consists of a main perimeter unit, a user control tablet, and a patient response button. A computer unit and a lithium-ion battery are built in the perimeter unit (W22 cm × D38 cm × H24 cm, 1.8 kg). In a VF test, the examiner operates the control tablet connected to the perimeter unit by Wi-Fi and patient’s responses are obtained using the response button connected by Bluetooth (Fig 1). With these integrated functions, imo realizes a portable high-performance perimeter as compared with the conventional devices. A stationary stand for imo is prepared during the test if the examinee chooses not to wear the device for any special physical reason. 10.1371/journal.pone.0161974.g001Fig 1 The head-mounted perimeter imo. The perimeter imo consists of a main perimeter unit, a user control tablet, and a patient response button. The right and left optical systems in the perimeter unit are completely separated and stimulus presentation and pupil monitoring are independently performed for each eye (Fig 2). The optical system is a wide-angle lens system which can measure the VF within 35°from the fovea. With the function of distortion and field curvature corrections, stimuli can be accurately generated and presented. A telecentric optical system is introduced to equalize the central and peripheral light intensities. 10.1371/journal.pone.0161974.g002Fig 2 The perimeter unit. The perimeter imo has completely isolated optical systems for the right and left eyes. Stimulus presentation is also independently performed for each eye. During the test, a test target was displayed using two sets of full high-definition (HD) transmissive liquid crystal displays and high intensity light emitting diode (LED) backlights separately for the right and left eyes. A test target luminance of 0.032–3183 cd/m2 (0.1–10000 asb) with a background luminance of 10 cd/m2 (31.4 asb) was generated using 10 bit resolutions, and the stimulus duration was 200 msec. The temporal resolution of the transmissive LCD display was 60 Hz and the stimulus intensity reached to a constant luminance level within one frame (1.67 msec). Test targets used the Goldmann size I to V, or any other optimal sizes and shapes available. The spatial resolution of the full HD transmissive LCD used in this device was 1920 x 1080 for each eye. Because the VF beyond the central 35°was masked on the LCD screen, 79800 pixels were actually used for testing. A Goldmann target size III (0.431° visual angle) was displayed using 37 pixels and used as the standard target size in this study. Due to the limitation of the display resolution, target size II and I were respectively displayed using 12 and 2 pixels and it is not currently possible to present a circular size I stimulus. Both right and left pupils were illuminated by near infrared LEDs and these images were obtained using the SXVGA resolution (1280 X 960 pixel) complementary metal–oxide–semiconductor (CMOS) sensor with a maximum frame rate of 60 Hz (Fig 3). 10.1371/journal.pone.0161974.g003Fig 3 A cross-section image showing the structure of the perimeter unit. A test target is displayed on the full HD transmissive liquid crystal displays with high intensity LED backlights for the two eyes. Both pupils are illuminated by near infrared LEDs and these images are monitored by the SXVGA resolution CMOS image sensor. During the examination, both pupil images were continuously monitored at a frame rate of 30 Hz and the images could be used for an eye tracking system. In this study, the subjects were not tested with the eye-tracking mode. Using the spherical lens adjustment dials shown in Fig 2, the lens position inside the perimeter was mechanically moved and a spherical lens correction within the range of -9 to +3 diopters could be performed without using any additional trial lenses. Furthermore, astigmatic correction could be achieved as well using an additional removable magnetic cylindrical lens system. The actual distance between the center of the cornea surface and the lens was 17.5 mm and patient’s viewing distance was set at 1 meter. Using the focus adjustment dial shown in Fig 1, this distance was adjustable within a control range of ± 3 mm according to the shape of the examinee’s face. Even if this distance was adjusted to ± 3 mm, with the imo optical system, the visual angle of the target size would only vary within ± 0.4%. This study used a 30–2 test pattern with 4–2 dB bracketing strategy. The test patterns used for imo are compatible with those for the HFA 30–2, 24–2, 10–2 and 24+ test programs with additional test points in the central 10°VF of the 24–2 program. For thresholding algorisms, the conventional 4–2 dB bracketing strategy and a modified Zipper Estimated Sequential Testing (ZEST) named AIZE (Ambient Interactive ZEST) are available. Furthermore, a binocular random single eye test that is a new testing approach is created for imo as the special feature. Conventional perimetry usually tests the right and left eyes separately. imo not only can test the right and left eyes separately but also presents the test target randomly to either eye under a non-occlusion condition without the examinee being aware of which eye is being tested (Fig 4). 10.1371/journal.pone.0161974.g004Fig 4 Target presentation and the examinee’s view during the binocular random single eye test. The test target was presented randomly to either eye under a non-occlusion condition and the patient was not aware of which eye was being tested. To better demonstrate this test with visible targets, target size V is used in this figure although target size III was the actual size used in this study. Subjects Subjects were 40 eyes of 20 patients with glaucoma (mean age, 60.2 ± 5.8 years; SE, -2.93 ± 2.58 D). The 40 eyes included 16 eyes with primary open angle glaucoma (POAG), 14 eyes with normal tension glaucoma (NTG), 3 eyes with primary angle closure glaucoma (PACG), 3 eyes with secondary glaucoma, and 4 normal contralateral eyes of the patients. Sixteen eyes were in the early stage of glaucoma (MD ≤ 6 dB), 10 eyes were in the moderate stage (6 dB< MD ≤ 12 dB), and 10 eyes were in the advanced stage (12 dB > MD). The exclusion criteria were: best corrected visual acuity < 1.0, refractive error < 6 D sphere and < 2.50 D cylinder, pupil diameter of < 3.0 mm, an ocular surgical history, ocular diseases other than glaucoma that might cause VF loss, and systemic diseases which were likely to affect the patient’s visual functions. Patients with the HFA results with more than 15% positive or negative catch trials and fixation loss, and patients with fusion dysfunction due to strabismus were also excluded. The diagnostic criteria for glaucomatous VF abnormality were as follows: the pattern deviation probability plot that showed a cluster of 3 or more nonedge-contiguous points having sensitivity with a probability of < 5% in the upper or lower hemifield with at least 1 point with a probability of < 1%. The diagnosis of glaucoma was based on the presence of typical glaucomatous optic disc changes, clear nerve fiber bundle defects, and corresponding glaucomatous VF abnormalities by the HFA. All the subjects underwent the HFA SITA standard 30–2 test, and the 30–2 monocular and binocular random single eye tests by imo. The order of the imo monocular and binocular random single eye tests was randomized. VF testing using imo was performed in a regular clinical office setting with an adaptation time of at least 5 minutes including the time for device setting. The examination would be interrupted when the test duration reached to 10 minutes. The results of the imo monocular and binocular random single eye tests were compared with the HFA results. Comparison of the mean sensitivity (MS) values obtained by both methods was performed using the Tukey test and Spearman's rank correlation coefficient. The regression lines were calculated using a Deming regression method. Bland-Altman plots [4] were used to assess the agreement between the results of the imo monocular and binocular random single eye tests. The subjects were asked if they noticed which eye was being tested during the measurement. This study followed the tenets of the Declaration of Helsinki, and all the participants provided written informed consent after the ethics committee of Kindai University Faculty of Medicine (no. 26–239) had approved the study. Results Comparison of test results Fig 5 shows a case of a 61-year-old male with POAG in both eyes. A deep lower nasal sensitivity loss in the right eye, a lower arcade scotoma, and an upper small scotoma near the fixation point in the left eye were detected by the HFA (Fig 5A). The imo binocular random single eye test also detected similar abnormalities in both eyes (Fig 5B). 10.1371/journal.pone.0161974.g005Fig 5 Description of the glaucomatous VF defects by grey scale and actual values for a 61-year-old male with POAG. (A) A deep lower nasal sensitivity loss in the right eye, a lower arcade scotoma and an upper small scotoma near the fixation point in the left eye were detected by the HFA. (B) Similar defects were detected using the imo binocular random single eye test. Comparison of the MS values The MS values for the right eye and the left eye obtained by the HFA, the imo monocular test, and the binocular random single eye test did not significantly differ (P = 0.99, the Tukey test; Table 1). The MS values by the HFA highly correlated with the MS values by the imo monocular test (R: r = 0.96, L: r = 0.94, P < 0.001; Spearman rank-order correlation) and with the values by the binocular random single eye test (R: r = 0.97, L: r = 0.98, P < 0.001; Spearman rank-order correlation) (Fig 6). The MS values by the imo monocular and binocular random single eye tests also highly correlated (R: r = 0.96, L: r = 0.95, P < 0.001; Spearman rank-order correlation) (Fig 7). 10.1371/journal.pone.0161974.t001Table 1 The MS values obtained by the HFA and imo tests. HFA monocular test (imo) binocular random single eye test (imo) p-value the right eye 22.3 ± 7.2 22.4 ± 6.4 22.2 ± 6.7 0.99 the left eye 22.4 ± 5.9 22.2 ± 6.6 22.3 ± 5.7 0.99 10.1371/journal.pone.0161974.g006Fig 6 Correlations between the MS values by the HFA and the imo tests. (A) The MS values for the right eye by the HFA highly correlated with the values by the imo monocular test (the solid regression line: r = 0.96, P < 0.001) and the binocular random single eye test (the dotted regression line: r = 0.97, P < 0.001). The slopes of the Deming regression lines were 1.12 (95% confidence interval [CI], 1.01 to 1.37) for the solid line and 1.08 (95% CI, 0.97 to 1.32) for the dotted line. The intercepts were -2.82 (95% CI, -8.48 to -0.37) for the solid line and -1.64 (95% CI, -7.67 to 1.12) for the dotted line. (B) Similarly, the MS values for the left eye by the HFA highly correlated with the values by the imo monocular test (the solid regression line: r = 0.94, P < 0.001) and the binocular random single eye test (the dotted regression line: r = 0.98, P < 0.001). The slopes of the Deming regression lines were 0.88 (95% CI, 0.70 to 0.98) for the solid line and 1.37 (95% CI, 0.87 to 1.11) for the dotted line. The intercepts were -2.88 (95% CI, 0.92 to 7.24) for the solid line and -0.71 (95% CI, -2.65 to 3.34) for the dotted line. 10.1371/journal.pone.0161974.g007Fig 7 Correlations between the MS values for the right and left eyes by the two imo tests. The MS values by the two imo tests highly correlated for the two eyes (R: r = 0.96, L: r = 0.95, P < 0.001). The slopes of the Deming regression lines were 1.04 (95% CI, 0.89 to 1.26) for the right eye and 0.85 (95% CI, 0.71 to 0.95) for the left eye. The intercepts were -1.09 (95% CI, -6.30 to 2.66) for the right eye and 3.47 (95% CI, 0.92 to 6.35) for the left eye. Bland-Altman plots revealed MS differences of 0.21 dB (95% limits of agreement between -3.6 dB and 4.04 dB) for the right eye and -0.11 dB (95% limits of agreement between -4.39 dB and 4.17dB) for the left eye between the imo monocular and binocular random single eye tests (Fig 8). 10.1371/journal.pone.0161974.g008Fig 8 Bland-Altman plots of the MS differences for the right and left eyes between the two imo tests. Bland-Altman plots revealed MS differences of 0.21 dB (95% limits of agreement between -3.6 dB and 4.04 dB) for the right eye and -0.11 dB (95% limits of agreement between -4.39 dB and 4.17dB) for the left eye between the two imo tests. The average test durations were 17.30 ± 1.25 min for the imo binocular random single eye test and 16.0 ± 1.21 min for the HFA test for the right and left eyes. In addition, no subject could detect which eye was being tested during the imo examination. Discussion Our study clearly showed that the new head-mounted perimeter imo can perform VF testing without a dark room and detect glaucomatous VF abnormalities compatible to the HFA results. Moreover, the equipped binocular random single eye test requires no occlusion for testing and produces results highly correlated with the results by the conventional monocular method. The head-mounted perimeter imo appeared to be a promising new perimetric method. The new perimeter imo has a great advantage: VF testing can be performed under a non-occlusion condition. The conventional monocular test usually requires the untested eye to be occluded. Possible problems caused by the test condition of occlusion have been previously reported. It is suspected that when the dominant eye is occluded, an inhibitory response in the non-occluded eye such as the Ganzfeld blankout or rivalry can be triggered by a binocular interaction [5, 6] although the effect of such phenomena on the SAP result is minimum [6]. In frequency-doubling technology (FDT) perimetry, it is well known that the second eye tested has reduced sensitivity [7–10] and that the delayed light adaptation is the suspected cause for the second eye problem. In addition to these problems, patients sometimes experience visual disturbances in the tested eye with the fellow eye being occluded with an opaque patch during static perimetric testing. With the imo binocular random single eye test, patients are tested under a more comfortable condition with both eyes open and the above-mentioned problems caused by occlusion can be solved. Because the test target is randomly presented to either eye, patients will not be aware of which eye is being tested during the test. Therefore, the binocular random single eye test can also be used for detecting feigned blindness. Indeed, all the subjects in this study could not detect which eye was being tested during the examination. As verified by the Bland-Altman plots in Fig 8, the MS difference between the imo monocular test and binocular random single eye tests was very small. The alternation of the two tests was clinically acceptable because they were within the intersession variability in patients with glaucoma [11, 12]. However, there are some concerns regarding the binocular random single eye test such as the test duration. Theoretically, the test duration for the binocular random single eye test would be twice as long as the test duration for the conventional monocular test. In addition, to reveal the character of the binocular random single eye test results, we used the traditional 4–2 dB bracketing strategy in this pilot study to exclude any possible effect from the strategy used. Therefore, the examination would take a recess when the test duration reached to 10 minutes and this further prolonged the test duration. To solve this problem, imo has a modified ZEST algorism named AIZE that reduces the test duration by about 70% as compared with the duration using the 4–2 dB bracketing strategy. Another issue of the binocular random single eye test is patient’s eye position. In this study, we excluded patients with a fusion problem caused by strabismus. If a patient cannot achieve fusion on the center right and left fixation targets for each eye before the examination, it will be difficult to perform the binocular random single eye test accurately and the testing shall be switched to the traditional monocular test. Although the imo monocular test cannot test both eyes randomly, it still has the advantage of a non-occlusion test condition. Because the right and left optical systems in imo are completely separated, the same background intensity for the tested eye is also available for the fellow eye without occlusion in an imo monocular test. The perimeter imo is an easy-to-use device for VF testing and all the subjects in this study could easily wear this device. However, it may not be suitable for everyone. For example, patients with a neck injury will not be recommended to use this device. As an alternative solution, a stationary stand for imo is prepared during the test for those who choose not to wear the device for testing. With this stand, patients can be tested without wearing the device. In recent years, population ageing has become a serious healthcare problem in Japan. With the feature of portability, imo has the potential to be the ideal device for patients who are receiving medical care at home. Even in hospital, imo can also be useful for patients who cannot undergo the regular perimetric testing due to their physical conditions. With imo, perimetric testing can be performed at the bed side and obtain test results compatible to those by the traditional SAP. Since imo does not require a dark room for testing, mass screening also becomes possible with this device. This pilot study has several limitations. To understand the characteristics of this new device, a reliability comparison between the new and conventional devices will be necessary and informative. Besides, repeatability performance that requires further investigation is another important issue in perimetric methods. However, these could not be achieved in this study because of the limitation of the total test durations and the small number of subjects. As a start, we used VF sensitivity as a measure to evaluate the basic characteristics of imo and the method behind although the systematic bias in MS between the two instruments may not be clinically significant. In the future, we are planning additional normative and repeatability studies using the modified zest algorism with shorter test durations. In conclusion, we have demonstrated that the newly developed head-mounted automated perimeter imo is an ideal device that performs perimetric testing regardless of the location and patient’s physical condition. More importantly, it obtains VF results highly compatible to those by the HFA in patients with glaucoma. With the equipped binocular random single eye test in imo, occlusion is no longer required. All these unique features of imo indicated that imo has the potential to be the ideal automated perimeter of the future. ==== Refs References 1 Hollander DA , Volpe NJ , Moster ML , Liu GT , Balcer LJ , Judy KD , et al Use of a portable head mounted perimetry system to assess bedside visual fields . Br J Ophthalmol . 2000 ;84 (10 ):1185 –90 . 11004108 2 Nagata S , Kani K . A new perimetry based on eye movement In: A. H , editor. Perimetry Update 1988/1989 . Amsterdam Berkeley, Milano : Kugler & ghedini ; 1997 p. 337 –40 . 3 Wroblewski D , Francis BA , Sadun A , Vakili G , Chopra V . Testing of visual field with virtual reality goggles in manual and visual grasp modes . Biomed Res Int . 2014 ;2014 :206082 10.1155/2014/206082 25050326 4 Bland JM , Altman DG . Statistical methods for assessing agreement between two methods of clinical measurement . Lancet . 1986 ;1 (8476 ):307 –10 . .2868172 5 Fuhr PS , Hershner TA , Daum KM . Ganzfeld blankout occurs in bowl perimetry and is eliminated by translucent occlusion . Arch Ophthalmol . 1990 ;108 (7 ):983 –8 . .2196038 6 Spry PG , Furber JE , Harrad RA . The effect of ocular dominance on visual field testing . Optom Vis Sci . 2002 ;79 (2 ):93 –7 . .11871400 7 Adams CW , Bullimore MA , Wall M , Fingeret M , Johnson CA . Normal aging effects for frequency doubling technology perimetry . Optom Vis Sci . 1999 ;76 (8 ):582 –7 . .10472964 8 Anderson AJ , Johnson CA . Effect of dichoptic adaptation on frequency-doubling perimetry . Optom Vis Sci . 2002 ;79 (2 ):88 –92 . .11868852 9 Anderson AJ , McKendrick AM . Quantifying adaptation and fatigue effects in frequency doubling perimetry . Invest Ophthalmol Vis Sci . 2007 ;48 (2 ):943 –8 . 10.1167/iovs.06-0685 .17251498 10 Kogure S , Membrey WL , Fitzke FW , Tsukahara S . Effect of decreased retinal illumination on frequency doubling technology . Jpn J Ophthalmol . 2000 ;44 (5 ):489 –93 . .11033126 11 Artes PH , Iwase A , Ohno Y , Kitazawa Y , Chauhan BC . Properties of perimetric threshold estimates from Full Threshold, SITA Standard, and SITA Fast strategies . Invest Ophthalmol Vis Sci . 2002 ;43 (8 ):2654 –9 . .12147599 12 Luithardt AF , Meisner C , Monhart M , Krapp E , Mast A , Schiefer U . Validation of a new static perimetric thresholding strategy (GATE) . Br J Ophthalmol . 2015 ;99 (1 ):11 –5 . 10.1136/bjophthalmol-2013-304535 .25053761

PMC005xxxxxx/PMC5001627.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756423410.1371/journal.pone.0161970PONE-D-16-21758Research ArticleBiology and Life SciencesBiochemistryProteinsChaperone ProteinsResearch and Analysis MethodsExtraction TechniquesProtein ExtractionResearch and Analysis MethodsBioassays and Physiological AnalysisBiochemical AnalysisEnzyme AssaysBiology and Life SciencesBiochemistryEnzymologyEnzymesBiology and Life SciencesBiochemistryProteinsEnzymesPhysical SciencesPhysicsClassical MechanicsMechanical StressThermal StressesBiology and Life SciencesBiochemistryEnzymologyEnzymesOxidoreductasesLuciferaseBiology and Life SciencesBiochemistryProteinsEnzymesOxidoreductasesLuciferaseBiology and Life SciencesBiochemistryEnzymologyEnzymesOxidoreductasesDehydrogenasesBiology and Life SciencesBiochemistryProteinsEnzymesOxidoreductasesDehydrogenasesPhysical SciencesChemistryChemical CompoundsOrganic CompoundsAlcoholsPhysical SciencesChemistryOrganic ChemistryOrganic CompoundsAlcoholsA Novel Method for Assessing the Chaperone Activity of Proteins Chaperone DiscoveryHristozova Nevena 12Tompa Peter 123Kovacs Denes 12*1 Structural Biology Department, Flemish Institute of Biotechnology, Brussels, Belgium2 Structural Biology Department, Free University Brussels, Brussels, Belgium3 Institute of Enzymology, Hungarian Academy of Sciences, Budapest, HungaryBrodsky Jeffrey L EditorUniversity of Pittsburgh, UNITED STATESCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: NH PT DK. Data curation: NH DK. Formal analysis: NH DK. Funding acquisition: PT. Investigation: NH DK. Methodology: NH DK. Project administration: DK. Resources: PT. Supervision: DK. Validation: NH PT DK. Visualization: NH DK. Writing – original draft: NH. Writing – review & editing: NH PT DK. * E-mail: dkovacs@vub.ac.be26 8 2016 2016 11 8 e016197030 5 2016 15 8 2016 © 2016 Hristozova et al2016Hristozova et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Protein chaperones are molecular machines which function both during homeostasis and stress conditions in all living organisms. Depending on their specific function, molecular chaperones are involved in a plethora of cellular processes by playing key roles in nascent protein chain folding, transport and quality control. Among stress protein families–molecules expressed during adverse conditions, infection, and diseases–chaperones are highly abundant. Their molecular functions range from stabilizing stress-susceptible molecules and membranes to assisting the refolding of stress-damaged proteins, thereby acting as protective barriers against cellular damage. Here we propose a novel technique to test and measure the capability for protective activity of known and putative chaperones in a semi-high throughput manner on a plate reader. The current state of the art does not allow the in vitro measurements of chaperone activity in a highly parallel manner with high accuracy or high reproducibility, thus we believe that the method we report will be of significant benefit in this direction. The use of this method may lead to a considerable increase in the number of experimentally verified proteins with such functions, and may also allow the dissection of their molecular mechanism for a better understanding of their function. http://dx.doi.org/10.13039/501100003130Fonds Wetenschappelijk Onderzoek1.2.734.13Kovacs Denes http://dx.doi.org/10.13039/501100003130Fonds Wetenschappelijk OnderzoekG.0029.12Tompa Peter This work was supported by the Odysseus grant G.0029.12 from the Fonds Wetenschappelijk Onderzoek (FWO, http://www.fwo.be) to PT, and DK is supported by a FWO postdoctoral fellowship 1.2.734.13. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction Molecular chaperones are a diverse group of proteins that play critical roles in assisting the folding and assembly of nascent protein chains [1], refolding proteins, assisting protein translocation through membranes [2], and often facilitating protein degradation [3]. Besides physiological processes, these helper proteins also fulfil crucial functions under unfavorable environmental conditions, such as temperature or osmotic stresses, but also in disease conditions and in response to a pathogen attack [3]. Chaperones are involved in the molecular response to stress in plants [4], bacteria [5], and animals [6], and were identified as key players in host-pathogen interaction and in the formation of innate and adaptive immunity [7,8]. During unfavorable environmental conditions, infection or disease, animals rely on a wide variety of chaperones to prevent and reduce the damage caused to cellular compartments and molecules [9]. Overexpression of chaperones in flies and mice so far proved effective in suppressing neurodegeneration [10,11]. In some cardiovascular diseases, cataract, alcoholic hepatitis, cystic fibrosis, phenylketonuria, and in a range of human cancers the protein homeostasis machinery was also shown to be involved as the first line of defense against the disease, thus rendering it a molecular marker for diagnostics [12,13,14]. This makes proteins with chaperone function potential drug targets, due to their function in various pathological and physiological processes inside the cell. Therefore functional and mechanistic characterization is crucial. The current state of the art in chaperone identification and discovery combines in silico homology screens [14], analysis of transcriptomics data [15], and co-purification with other proteins [16]. Such methods are effective in pooling proteins with stress-related function, but they have limited reproducibility and inherent high error rates of false positives, and require detailed in vitro studies for verification [17]. Methods for chaperone activity testing are under patenting (e.g. application number PCT/US2012/037131), nonetheless, the lack of fast and reliable high-throughput methods for assessing the protective capacity of proteins considerably hinders the discovery and characterization of novel proteins with such functionality. Thus, we propose a semi-high throughput in vitro method for the fast and reliable assessment of chaperone capacity, based on measuring enzyme activity in a 96-well plate format, where the protective activity of a putative chaperone is defined by the retention of enzyme activity during adverse conditions. The proposed method will be of use when a high number of proteins or conditions need to be screened, efficient concentration screens (efficacy) are to be done, or extensive mutation analysis needs to be carried out, for example. In order to show the use of such a method (i.e. to provide proof of principle), we have selected four widely used substrates from the chaperone field: citrate synthase [18], luciferase [19], lactate dehydrogenase [20] and alcohol dehydrogenase [21], optimized the enzymatic assays for the plate reader format and set up the chaperone assay within the plate. We then designed a configuration of the measurements that provides adequate statistical power, and enables us to compare the activity of various enzyme-chaperone mixtures. In addition to testing the technique on well-established chaperones, we demonstrate the use of the proposed method by characterizing the putative chaperone ERD10. The Early Response to Dehydration protein 10 is a Late Embryogenesis Protein (LEA), described as a dehydrin [22]. Dehydrins are plant proteins primarily involved in cellular and systemic responses to drought stress via various, not fully understood, mechanisms [23]. Kovacs et al. recently explored the putative proteostatic activity of ERD10 and provided formal proof that ERD10 can prevent the aggregation of a range of proteins [24]. Thus, we aim to further explore the protective effect of dehydrins, also demonstrating the advantages of the new technique. We also saw the potential of this technique in measuring the protective activity of complex systems, such as whole-cell protein extracts. In vitro experiments traditionally dominated the field of measuring the chaperone activity of purified proteins. Recently, the proteostatic activity of proteins is also studied in vivo, via overexpression in cell lines, and follow its effect via phenotype screens (i.e. stress resistance), addressing their subcellular localization, interaction with other proteins or organelles, and protection of cellular proteins [25–27]. We propose to measure the protein activity protection of a total protein extract from cells grown under physiological and stress conditions when endogenous chaperones are being overexpressed. One of the most prominent ways to discover stress-related proteins is by microarrays. In eukaryotes subjected to stress, the first and highest scoring genes to be expressed and respectively identified, are stress-related proteins and signaling kinases [28]. Bacteria, like other organisms, also express a number of chaperones even when they are not subjected to environmental stress [29]. Our assay is expected to quantify the chaperone activity of the complete protein machinery in a comparative manner. Materials and Methods Expression and purification of ERD10 ERD10 was cloned, expressed and purified as described in Kovacs et al [24]. After obtaining pure protein (as analyzed via SDS-PAGE electrophoresis and mass-spectrometry, ERD10 sample was lyophilized and subjected to elemental analysis for quantification in the Department of Analytical chemistry in the Vrije Universiteit Brussel. Activity measurements In order to study the chaperone activity of a protein, we developed a set of fast, reliable and inexpensive protocols. They are all based on standard enzymatic reactions and the ability of the putative chaperone to protect enzyme activity under stress conditions. We optimized these protocols for four standard enzymes, widely distributed among different species–Alcohol dehydrogenase (ADH, EC 1.1.1.1), Citrate synthase (CS, EC 2.3.3.1), Lactate dehydrogenase (LDH, 1.1.1.27), and Luciferase (Luc, EC 1.13.12.7). We altered and optimized the protocols of enzyme activity measurement to fit the format of a microplate reader [30–34]. Comparative definition of enzyme activity All the protocols follow spectrophotometric reading after the start of the enzymatic reaction. The measurements were performed in a Biotek Synergy Mx microplate reader at wavelengths characteristic for the reactions, until a stable plateau was reached. In all cases, the activity of the corresponding enzyme was defined by the slope of the initial, linear phase of the curve. 100% activity was set to be the activity of the enzyme prior to stress and without the addition of chaperones. Alcohol dehydrogenase activity ADH catalyzes the oxidation of alcohols to aldehyde or ketone using NAD+ as proton acceptor. The reaction leads to production of NADH, which can be followed at 340nm. The reaction was performed using 0.1 μM yeast ADH (A7011, Sigma), 6mM NAD+ (N7004, Sigma), and 3.75mM EtOH in 100mM TrisHCl, supplemented with 250mM NaCl, at pH 8.8 [34]. Data collection started immediately after the addition of NAD+ and EtOH to the enzyme solution and was continued for 3 minutes at RT to obtain the initial part of the activity curves. Citrate synthase activity The activity of citrate synthase was measured based on the method described by Srere [33]. Porcine citrate synthase is the first enzyme of the citric acid cycle and exists in almost all cell and tissue types. It converts Acetyl-CoA and oxaloacetate into citrate and releases reduced coenzyme-A. The latter reacts with the Ellman’s reagent (DTNB) enabling the reaction to be followed at 412 nm where reacted DTNB absorbs. The reaction was performed with 6 nM citrate synthase (C3260, Sigma), 0.45 mM Acetyl-coA (A2056, Sigma), 0.5 mM oxaloacetate (O4126, Sigma), and 0.1 mM DTNB (D8130, Sigma) in 50 mM Tris-HCl buffer supplemented with 250mM NaCl, at pH 7.5, and followed for 3 min at 30°C. Lactate dehydrogenase activity LDH is present in almost all cell types and organisms; it catalyzes the reversible conversion of pyruvate to lactate. The enzyme uses NADH cofactor and oxidizes it to NAD+. The reaction solution contained 10μM bacterial LDH (L3888, Sigma), 6.6 mM NADH (AE12.1, Carl Roth), and 30 mM Sodium pyruvate (8793.1, Carl Roth), in 200mM Tris-HCl, at pH 7.3 [35]. The reaction was followed by monitoring the decrease in absorbance at 340 nm for 3 min at RT. Luciferase activity The light-producing enzyme luciferase is present in various bioluminescent organisms, where it can have a true respiratory function when the major protein components of the respiratory pathway are impaired [36]. The enzymatic activity was followed in 1mM PBS, at pH 6.8 and RT at following the concentrations: 0.5μM firefly luciferase (L9506, Sigma), 1.3 mM NADH (AE12.1, Carl Roth), 0.042 mM FMN (F2253, Sigma), and 0.0025% decanal (D7384, Sigma) [37]. The enzyme activity was defined by the amount of the emitted light, recorded in a luminometric experiment in a plate reader with 1s integration time and at identical gain settings between the different samples. Heat-induced inactivation Once the activities of the enzymes were recorded, another aliquot of the same batch was subjected to high-temperature stress in a thermo block (PST60HL, Boeco). We have tested various ways for enzyme deactivation, and found that without a direct contact with the heating unit, the level of inactivation is not even due to a temperature gradient across the plate. The heating block used in these experiments transfers the heat via a direct contact with the plate, which is further enhanced by continuous mild shaking (250 rpm). The temperatures and times optimized for deactivation in the given setup were first selected from literature and were further optimized as to limit evaporation and increase reproducibility. Therefore, enzymes were deactivated at following temperatures: ADH at 46°C for 60min; CS at 44°C for 40min; LDH at 56°C for 60min, Luc at 44° for 30min. As the reaction volumes are rather small, even limited evaporation would lead to a change in the reagent concentrations in the solution and compromise the measurements. Thus, we sealed the plates at the heating step with a VIEWseal thermo-resistant microtiter plate sealing foil (676070, Greiner Bio-One). Co-factors and reporter agents (where applicable) were added just before the measurement to avoid thermal decomposition. Enzyme activities have always been compared between different samples e.g. before and after stress; in the presence or absence of a putative chaperone; at various concentrations; different strength and/or duration of the stress condition. As a proof of concept, we used well-known molecular chaperones, reported to be effective against temperature deactivation, among other types of stress. Namely, we used GroEL, Hsp70, and Hsp90 in equimolar ratios or in excess to the enzyme (client) [14,38–45]. Chaperone assay with CS in the presence of ERD10 We used the CS activity assay described above to study the protective effect of ERD10 against thermal deactivation of the enzyme. Lyophilized ERD10 was dissolved in 50 mM aqueous solution of 50mM Tris-HCl, 250 mM NaCl at pH 7.5, and added to the CS sample in different ratios–from equimolar to a great excess of ERD10 as compared to CS. Activity of CS was measured as described above, before and after the protein mix was subjected to stress at 44°C for 40 min. The enzymatic activity was compared between samples with and without the addition of ERD10, and between samples with different concentrations of ERD10. Total protein extract preparation To prepare total protein extracts, we used non-transformed E. coli BL21 (DE3) cells strains. The cells were grown on full LB media at 37°C with shaking, until OD600 = 1. After reaching the required optical density, the cells were transferred to 42°C and aliquots were removed before stress and after 3 hours of incubation at high temperature. The samples were then centrifuged at 13000 rpm at 4°C in a bench-top centrifuge, for 1h to pellet the cells. The pelleted cells were resuspended in a lysis buffer containing 50mM Tris-HCl, 50mM NaCl, and 5mM DTT at pH 7.5, supplemented with protease inhibitors, DNAse (50μg/ml) and were lysed by sonication (10 pulses/10s each, followed by 30s on ice). The lysate was loaded into a 3kDa cut-off dialysis membrane, and dialyzed against a buffer of 50mM Tris-HCl, 50 mM NaCl at pH 7.5 buffer, supplemented with protease inhibitors overnight. The total protein extract was quantified using the Bradford assay method (B6916, Sigma). Chaperone assay in the presence of total protein extract from E. coli We used the CS activity assay described above to study the protective effect of total protein extracts from E.coli against thermal deactivation of the enzyme. Final amount of 0.02 mg of protein extract per reaction was added to the CS assay instead of a purified chaperone. In order to anticipate the increased background activity of cell extracts, we also added the total protein extract to a blank reaction mix (containing substrates and reporter agent, but no CS). The assay was further performed as already described above, before and after subjecting the protein mix to stress at 44°C for 40 min. The enzymatic activity of CS was compared between samples with and without the addition of total E.coli protein extract. Data analysis and representation All assays were performed with replicates, ranging from 5 to 12 replicate per experimental condition to allow robust statistical analysis. The slopes of each reaction were calculated using GraphPad Prism ver.6 and were subsequently normalized to the activity of the corresponding enzyme before it was subjected to stress. For each sample, mean and 95% confidence interval was calculated before normalization. Normalized values were plotted as curves to illustrate the gradual deactivation of each enzyme. The statistical significance of the differences between samples was calculated via one-way ANOVA and t-test. Results Chaperone assays are widely used to assess the putative function of proteins being expressed or overexpressed in the cell under stress conditions, such as high temperature, oxidative stress, high salinity, UV, etc. The chaperone activity of a protein is determined by the protection of a client protein either against aggregation or the loss of activity evoked by stress. Although chaperone assays are widely used, they usually require tedious experiments due to the high standard deviation of each measurement mostly attributed to experimental, and not biological, error. Classical measurements usually involve the withdrawal of an aliquot from an enzyme assay, followed by a subsequent evaluation of specific properties of the sample and comparison of results between assays with and without the chaperone. Such assays are very sensitive to changes in concentration, to the uniformity of stress applied on the sample and also to other experimental parameters such as time passed between mixing and start of photometric detection. Accordingly, to reduce experimental error of such an assay, both the number of technical replicates and the accuracy of the assay has to be increased. To this end, we aimed to transfer activity assays to a plate reader, thereby increasing the number of parallel samples and also the reliability of measurements by keeping hands-on time low. We aimed to utilize enzyme activity assays commonly used in the field. As a result, we report the activity assays of four standard enzymes on a plate reader, following their heat-induced deactivation. To demonstrate the efficacy of the plate reader format, we have tested the protective activity of HSP90, HSP70 and GroEL/ES on CS as substrate. Activity assays on a plate reader Activity of an enzyme solution is defined by the initial, maximum velocity of a reaction, measured at 10-fold excess of the substrate over the Km. Under such conditions, the measured reaction velocity is above 90% of the Vmax, and remains linear over a longer period of time, due to the large excess of substrate over the product. In practice, the velocity is measured as the slope of the initial linear phase of the reaction (Fig 1). For comparison of different samples, it is important to define the velocity exactly at the same time point after mixing the components, in order to have an accurate assessment of the effect of the chaperone. Nonetheless, due to technical limitations, the first few seconds of the reaction cannot be observed, which is why the slopes of the curves converge at the negative side of the timeline. The point where the fitted slopes cross shows that a linear fit has been done on the linear phase of the reaction curve. The slopes can then be compared. It is also apparent from Fig 1 that the experimental setup is suitable to measure changes in the velocity (see next section). Confidence intervals of technical repeats illustrate the accuracy of the measurements. 10.1371/journal.pone.0161970.g001Fig 1 Activity of ADH (A), CS (B), LDH (C), and Luc (in black on panel D). The activity of each enzyme was recorded at the specific wavelength for each reaction. The different data points on panels (A), (B) and (C) represent the raw data readout from each time-point during the deactivation by high temperature. The lines represent the slopes calculated for the initial linear phase of each reaction. Due to the necessary mixing step in our protocol we miss the first few seconds of the reaction, thus we extrapolated the slopes to time zero of each reaction. The observed activity of the enzymes drop as shown by the decreasing slopes as a function of time at high temperature. (D) The deactivation of each enzyme was shown as % relative remaining activity: ADH (red), CS (blue), LDH (green), and Luc (black). The data for each enzyme was normalized to the activity of the non-stressed enzyme. The data was plotted as mean and 95% CI. Heat-induced inactivation of enzymes Once we optimized the enzymatic assays for a plate reader, we checked whether these assays would be suitable for a chaperone activity assay. To this end, we applied high temperature stress as described in Materials and Methods and compared activities before and after stress. To ensure sufficient dynamic range for the assays, we have optimized the temperature and the time of incubation so that a drop in activity to about 20% of the original activity occurs (Fig 1D and Table 1). The kinetics of deactivation of the different enzymes are different, thus optimal temperatures and times differ as follows: ADH, 46°C for 60 min; CS, 44°C for 40 min; LDH, 56°C for 60 min, and Luc, 44° for 30 min (Fig 1). For example, the activity of CS was observed to drop from 100% to under 20%, in 40 minutes at 44°C. After repeating the deactivation of each enzyme several times and measuring several points along the deactivation, we concluded that measuring two data points is sufficient to make a conclusive comparison of initial activity and activity after deactivation (for details, view deactivation kinetics of the respective enzyme, Figs 1 and 2). Fig 1D shows a decrease of activity as a function of time upon high-temperature stress. While the activity of LDH drops below 10% in one hour at 56°C, the activity of ADH drops to about 25% in the same time at 43°C. Luciferase activity, unlike that of ADH, CS and LDH, is measured by recording emitted light, which is then converted to relative activity. Luciferase is well known for being temperature sensitive, thus it loses 75% of its initial activity in 10 minutes at 44°C. The corresponding temperatures and incubation times for each enzyme were selected so that the deactivation can be observed in a relatively short time, without inducing an immediate aggregation of the proteins (Table 1). Nonetheless, we aimed to achieve a significant deactivation of the enzymes without prolonged exposure to high temperatures to avoid evaporation, which would lead to an increase in error due to concentration changes of the reagents in each reaction, rendering comparisons impossible. Each measurement was performed multiple times (at least 6, at most 12 technical repeats), so a statistical evaluation of the data can be performed by calculating the 95% confidence interval, and running an one-way ANOVA tests. 95% CI in combination with hypothesis testing is a powerful statistical test for showing differences between two sample populations. 10.1371/journal.pone.0161970.t001Table 1 Loss of activity of each enzyme as a function of time at high temperature. ADH CS LDH Luc Time (min) Average SD Time (min) Average SD Time (min) Average SD Time (min) Average SD 0 100 7.72 0 100 3.67 0 100 5.27 0 100 19.69 10 78.88 6.29 25 60.71 14.22 10 66.44 4.95 10 26.45 5.42 20 58.25 3.57 35 18.88 9.17 20 32.17 2.37 15 16.99 2.36 30 42.71 1.51 45 7.14 4.38 30 9.94 1.74 20 14.07 0.17 40 33.12 0.72 40 2.08 3.66 25 13.94 1.14 50 26.50 0.95 50 2.23 4.44 30 13.86 1.94 60 21.68 0.60 10.1371/journal.pone.0161970.g002Fig 2 Hsp70 (A), Hsp90 (B), and GroEL (C) protect citrate synthase against temperature deactivation. The activity of the enzyme after temperature stress alone (CS AS) or in the presence of different ratios of the respective chaperone (and excess ATP where noted) was measured as described in the Materials and Methods section, and plotted on each panel. The bars represent the mean of the parallel measurements; the error bars indicate the 95% confidence intervals of each measurement. The data was normalized to the activity of the enzyme alone before it was subjected to temperature stress (CS BS). One-way ANOVA analysis was performed in each case to assess the significance of the differences of the protection effect of each chaperone compared to the CS AS. Panel D shows examples of the concentrations of chaperones used in previous experiments to assess protective activity. Chaperones protect enzymes against activity loss Once the enzyme assays were optimized, we tested whether such assays are suitable to measure the protective activity of known chaperones. To this end, we have selected three well-known chaperones from the literature, HSP70 [46], HSP90 [39], and the GroEL/GroES chaperonin complex [47]. According to previous publications, the selected chaperones usually show protective effect between 1:1 and 1:5 concentration ratios (Fig 2). We have measured the activity of CS at the same concentration ratios before and after temperature stress, and found that at 1:1 ratio it shows a rather limited protective effect in all the three cases: CS in the presence of GroEL retains about 30% of its original activity; while at equimolar concentrations of Hsp70 and Hsp90, it is not protected at all (Fig 2). The protective effect increases significantly at 1:2 molar ratio, where two-fold molar excess of GroEL retains the initial CS activity completely; whereas 47% and 30% of the enzymatic activity is retained in the presence of Hsp90 and Hsp70 respectively. At 1:5 ratio, Hsp90 can protect CS against temperature inactivation completely in our assay (Fig 2). We also compared the efficacy of Hsp70 and Hsp90 in the absence and presence of ATP (Fig 2A and 2B). We found that the addition of ATP increased the protection of the enzyme at 1:1 ratio in both the cases of Hsp70 and Hsp90, but did not at a 5 times excess of the chaperone. Chaperone functions of ERD10 ERD10 was already shown to have a protective effect against protein aggregation by Kovacs et al [24]. In this work, we also formally demonstrated the protective effect of ERD10 on enzyme activity, apart from the protection against aggregation. The protective effect of ERD10 on CS activity was measured at various molar ratios which were then compared to the control reaction where CS was alone (Fig 3). After 40min at 44°C, CS alone loses around 80% of its activity, while this is not the case when excess of ERD10 is added to the sample: in the presence of 20x molar excess of ERD10, CS retains its activity completely. This protective effect of the dehydrin increases with increasing concentrations, as shown in Fig 3. The apparent increase of activity of CS above 100% in the presence of ERD10 is believed to be due to reactivation of inactive species in the commercial enzyme product, which appeared to be homogeneous even when subjected to size-exclusion chromatography, apparently devoid of unfolded, misfolded or aggregated species. This phenomenon deserves further investigation as it has been observed in several other experiments with chaperones [48,49]. 10.1371/journal.pone.0161970.g003Fig 3 Enzymatic activity of CS is protected by ERD10 against temperature deactivation. The activity of CS was recorded before (A) and after stress (B) of 44°C for 40min in the presence (■) and absence (•) of different concentration of ERD10 as described in the Material and Methods section. Activity is presented as percentage of the activity of the non-stressed CS without the addition of ERD10. Concentration of CS is constant at 4.5nM. The symbols represent the mean, and the error flags—the 95%CI. Chaperone capacity of total protein extract from E. coli We also examined if our assay is useful to measure the protein protection capacity of cellular extracts. This can prove useful for comparing how the proteome of a cell or tissue changes upon the effect of a stress factor, facilitating the discovery of novel chaperones. The protective effect of a bacterial protein extract on CS activity was measured and compared to the control reaction where CS is alone (Fig 4). After 40min at 44°C, CS alone loses its activity, while this is not the case when excess of total protein extract is added to the sample. We observe a complete protection of the CS activity by the extracts from both stressed and non-stressed cells. We believe that this is due to constitutively expressed chaperones in E. coli. Additionally we also tested for non-specific reactions attributed to the enzymes present in the extract, without the addition of purified CS. We found such background reactivity to be comparable to the reactions where all reagents but the enzyme was added (S1 Fig and S1 Dataset). Thus, comparison between the different samples is reasonable and conclusive. 10.1371/journal.pone.0161970.g004Fig 4 Effect of E. coli total protein extract on the deactivation of CS by high temperature. The activity of citrate synthase was compared before and after temperature stress, between samples with and without the addition of total protein extract from cells grown at 37°C or 42°C as indicated in the grid. The addition of either type of extract significantly protects the activity of CS (shown by t-test). A slight, but not statistically significant, difference was observed between the protective effects of various cell extracts. The height of the bars represents the mean of at least 6 technical repeats, the error flags– 95% CI. The significance of the differences was assessed via an one-way ANOVA. Discussion The prevailing pipeline of the identification and characterization of (novel) chaperones relies on homology-based identification of conserved regions and domains from known proteins with such function, and/or experimental testing of proteins upregulated during stress conditions [4,50–52]. While these standard approaches to predicting, identifying and characterizing protein chaperones have brought much success, we address the lack of a quick and reliable method for the (semi) high-throughput in vitro screening. Here we report on assays optimized for a plate reader, operating in a semi high-throughput design. We believe that the reported protocols would allow both screening for novel chaperones and gaining a better understanding of the activity of known ones. In the optimization, we have taken into account the possible factors that might compromise the accuracy of the measurement, and took necessary precautions to tackle them. We show that such an assay is capable of pinpointing protective activity of a variety of proteins. Whereas the current protocol applies heat stress, it can be easily adapted to other stress conditions, provided the stress condition can be applied uniformly across all wells, i.e. it does not form a gradient within the plate. Protective effects measured for the classical chaperones used in this assay are comparable with those reported previously (Fig 2D) [53,40–45]. Chaperone activity is an important cellular function of proteins, involved in practically all physiological, and many pathological, processes [12]. Here we established a convenient and time effective tool for the reliable screen of protein activity, which may lead to the identification of novel cellular chaperones. It should not be missed, though, that several cases are known where proteostatic mediators play crucial roles outside the cell, in body fluids, to maintain proteostasis in the intracellular space [54,55]. Known examples are the role of HSP70 in the immune response, and the contribution of such proteins in the prevention of neurodegenerative amyloid diseases. Currently, ELISA-based tests are used to measure the level of plasma chaperones, but they are usually based on lengthy protocols with, at best, semi-quantitative results, only giving information on the approximate amount, but not their activity [56]. While testing of our technique on human samples has not been performed yet, we anticipate it will be of use in clinical diagnostics as an indicative tool. Supporting Information S1 Dataset Raw data of chaperone activity measurements. Dataset includes all raw measurement data obtained throughout the study, separated into sheets according to the type of the measurement. (XLSX) Click here for additional data file. S1 Fig Activity of E. coli whole cell extracts. The background activity of whole cell extracts containing endogenous CS was measured to estimate the cross-reactivity of the extract to the recombinant porcine CS. These background activity measurements were performed for each experiment where whole cell extracts were used (see Materials and Methods section). For each bar, the temperature indicated signifies the temperature at which the cells were grown at before extraction. For every sample we also tested the deactivation of the endogenous CS at high temperature (marked as AS) similarly as it has been performed with the recombinant CS. In every case, the data was normalized to the activity of the recombinant porcine CS, before stress. The data was plotted as bars indicating mean and error flags–the 95%CI. (TIF) Click here for additional data file. We thank Karolien Van Belle for the technical support, Guy Vandenbussche (Structure and Function of Biological Membranes, Free University Brussels (ULB)) for the MS analyses of ERD10, Natacha Brion (Analytical and Environmental Geochemistry (AMGC), Free University Brussels) for the elemental analysis of ERD10. ==== Refs References 1 Frydman J , Nimmesgern E , Ohtsuka K , Hartl FU . Folding of nascent polypeptide chains in a high molecular mass assembly with molecular chaperones . Nature . 1994 7 14 ; 370 (6485 ):111 –7 . 8022479 2 Agarraberes FA , Dice JF . A molecular chaperone complex at the lysosomal membrane is required for protein translocation . J Cell Sci . 2001 7 1 ; 114 (13 ):2491 –9 .11559757 3 Goldberg AL . Protein degradation and protection against misfolded or damaged proteins . Nature. Nature Publishing Group ; 2003 12 18 ; 426 (6968 ):895 –9 . 4 Wang W , Vinocur B , Shoseyov O , Altman A . Role of plant heat-shock proteins and molecular chaperones in the abiotic stress response . Trends Plant Sci . 2004 5 ;9 (5 ):244 –52 . 15130550 5 Macario AJ . L. The molecular chaperone system and other anti-stress mechanisms in Archaea . Front Biosci . 2001 2 1 ; 6 (1 ):d262 .11171552 6 Ni M , Lee AS . ER chaperones in mammalian development and human diseases . FEBS Lett . 2007 7 31 ; 581 (19 ):3641 –51 . 17481612 7 Henderson B , Allan E , Coates ARM . Stress wars: the direct role of host and bacterial molecular chaperones in bacterial infection . Infect Immun . 2006 7 1 ; 74 (7 ):3693 –706 . 16790742 8 Srivastava P . INTERACTION OF HEAT SHOCK PROTEINS WITH PEPTIDES AND ANTIGEN PRESENTING CELLS: Chaperoning of the Innate and Adaptive Immune Responses . Annual Reviews 4139 El Camino Way, P.O. Box 10139, Palo Alto, CA 94303–0139, USA ; 2003 11 28 9 Macario AJ . L. Molecular chaperones: Multiple functions, pathologies, and potential applications . Front Biosci . 2007 1 1 ; 12 (1 ):2588 .17127265 10 Muchowski PJ . Protein Misfolding, Amyloid Formation, and NeurodegenerationA Critical Role for Molecular Chaperones? Neuron . 2002 7 3 ; 35 (1 ):9 –12 . 12123602 11 Barral JM , Broadley SA , Schaffar G , Hartl FU . Roles of molecular chaperones in protein misfolding diseases . Semin Cell Dev Biol . 2004 ; 15 (1 ):17 –29 . 15036203 12 Macario AJL , Conway de Macario E . Sick chaperones, cellular stress, and disease . N Engl J Med . 2005 10 6 ; 353 (14 ):1489 –501 . 16207851 13 Slavotinek AM , Biesecker LG . Unfolding the role of chaperones and chaperonins in human disease . Trends Genet . 2001 9 ; 17 (9 ):528 –35 . 11525836 14 Ahsen Pfanner . Molecular chaperones: towards a characterization of the heat-shock protein 70 family . Trends Cell Biol . 1997 3 ; 7 (3 ):129 –33 . 17708923 15 Jakob U , Muse W , Eser M , Bardwell JC . Chaperone Activity with a Redox Switch . Cell . 1999 2 ; 96 (3 ):341 –52 . 10025400 16 Valent QA , Kendall DA , High S , Kusters R , Oudega B , Luirink J . Early events in preprotein recognition in E. coli: interaction of SRP and trigger factor with nascent polypeptides . EMBO J . 1995 11 15 ; 14 (22 ):5494 –505 . 8521806 17 Marini I , Moschini R , Del Corso A , Mura U . Chaperone-like features of bovine serum albumin: a comparison with alpha-crystallin . Cell Mol Life Sci . 2005 12 ; 62 (24 ):3092 –9 . 16314918 18 Buchner J , Grallert H , Jakob U . Analysis of chaperone function using citrate synthase as nonnative substrate protein . Methods Enzymol . 1998 1 ; 290 :323 –38 . 9534173 19 Pinto M , Morange M , Bensaude O . Denaturation of proteins during heat shock. In vivo recovery of solubility and activity of reporter enzymes . J Biol Chem . 1991 7 25 ; 266 (21 ):13941 –6 . 1906889 20 Dong A , Prestrelski SJ , Allison SD , Carpenter JF . Infrared spectroscopic studies of lyophilization- and temperature-induced protein aggregation . J Pharm Sci . 1995 4 ; 84 (4 ):415 –24 . 7629730 21 Markossian KA , Golub N V , Khanova HA , Levitsky DI , Poliansky NB , Muranov KO , et al Mechanism of thermal aggregation of yeast alcohol dehydrogenase I: role of intramolecular chaperone . Biochim Biophys Acta . 2008 9 ; 1784 (9 ):1286 –93 . 10.1016/j.bbapap.2008.04.030 18515108 22 Kim SY , Nam KH . Physiological roles of ERD10 in abiotic stresses and seed germination of Arabidopsis . Plant Cell Rep . 2010 2 ; 29 (2 ):203 –9 . 10.1007/s00299-009-0813-0 20054552 23 Hanin M , Brini F , Ebel C , Toda Y , Takeda S , Masmoudi K . Plant dehydrins and stress tolerance: versatile proteins for complex mechanisms . Plant Signal Behav . 2011 10 ;6 (10 ):1503 –9 . 10.4161/psb.6.10.17088 21897131 24 Kovacs D , Kalmar E , Torok Z , Tompa P . Chaperone Activity of ERD10 and ERD14, Two Disordered Stress-Related Plant Proteins1[OA] . Plant Physiol. American Society of Plant Biologists ; 2008 ;147 (1 ):381 –90 . 25 Alder NN , Shen Y , Brodsky JL , Hendershot LM , Johnson AE . The molecular mechanisms underlying BiP-mediated gating of the Sec61 translocon of the endoplasmic reticulum . J Cell Biol . 2005 1 31 ; 168 (3 ):389 –99 . 15684029 26 Bloch M , Johnson K . Identification of a molecular chaperone in the eukaryotic flagellum and its localization to the site of microtubule assembly . J Cell Sci . 1995 11 1 ; 108 (11 ):3541 –5 .8586665 27 Soltys BJ , Gupta RS . Cell surface localization of the 60 kDa heat shock chaperonin protein (hsp60) in mammalian cells . Cell Biol Int . 1997 5 ; 21 (5 ):315 –20 . 9243807 28 Sørensen JG , Nielsen MM , Kruhøffer M , Justesen J , Loeschcke V . Full genome gene expression analysis of the heat stress response in Drosophila melanogaster . Cell Stress Chaperones . 2005 1 ; 10 (4 ):312 –28 . 16333985 29 Ellis RJ , Laskey RA , Lorimer GH , editors. Molecular Chaperones . Dordrecht : Springer Netherlands ; 1993 ; Available from: http://link.springer.com/10.1007/978-94-011-2108-8 30 Deluca M , McElroy WD . Bioluminescence and Chemiluminescence . Methods in Enzymology. Elsevier ; 1978 3 –15 p. 31 HAUGAARD N . D- and L-lactic acid oxidases of Escherichia coli . Biochim Biophys Acta . 1959 1 ; 31 (1 ):66 –72 . 13628604 32 Rasouli S , Hosseinkhani S , Yaghmaei P , Ebrahim-Habibi A . Effects of sucrose and trehalose on stability, kinetic properties, and thermal aggregation of firefly luciferase . Appl Biochem Biotechnol . 2011 9 ; 165 (2 ):572 –82 . 10.1007/s12010-011-9276-1 21617898 33 Faloona GR , Srere PA . Escherichia coli citrate synthase. Purification and the effect of potassium on some properties . Biochemistry. American Chemical Society ; 1969 11 ; 8 (11 ):4497 –503 . 34 Wagner FW , Burger AR , Vallee BL . Kinetic properties of human liver alcohol dehydrogenase: oxidation of alcohols by class I isoenzymes . Biochemistry. American Chemical Society ; 1983 4 ; 22 (8 ):1857 –63 . 35 Borgmann U , Moon TW , Laidler KJ . Molecular kinetics of beef heart lactate dehydrogenase . American Chemical Society ; 2002 . 36 Makemson JC . Luciferase-dependent oxygen consumption by bioluminescent vibrios . J Bacteriol . 1986 3 ; 165 (2 ):461 –6 . 3944057 37 Nealson KH , Platt T , Hastings JW . Cellular control of the synthesis and activity of the bacterial luminescent system . J Bacteriol. American Society for Microbiology ; 1970 10 ; 104 (1 ):313 –22 . 38 Schröder H , Langer T , Hartl FU , Bukau B . DnaK, DnaJ and GrpE form a cellular chaperone machinery capable of repairing heat-induced protein damage . EMBO J . 1993 11 ; 12 (11 ):4137 –44 . 7900997 39 Mayer MP , Bukau B . Molecular chaperones: The busy life of Hsp90 . Curr Biol . 1999 ; 9 (9 ):R322 –5 . 10322107 40 De Los Rios P , Ben-Zvi A , Slutsky O , Azem A , Goloubinoff P . Hsp70 chaperones accelerate protein translocation and the unfolding of stable protein aggregates by entropic pulling . Proc Natl Acad Sci U S A . 2006 4 18 ;103 (16 ):6166 –71 . 16606842 41 Skowyra D , Georgopoulos C , Zylicz M . The E. coli dnaK gene product, the hsp70 homolog, can reactivate heat-inactivated RNA polymerase in an ATP hydrolysis-dependent manner . Cell . 1990 ;62 (5 ):939 –44 . 2203539 42 Schumacher Robert J. ‡,§, Hansen William J. , Freeman Brian C. ⊥, Alnemri Emad #, Litwack Gerald # and, Toft David O. * ‡. Cooperative Action of Hsp70, Hsp90, and DnaJ Proteins in Protein Renaturation† . American Chemical Society ; 1996 . 43 Freeman BC , Morimoto RI . The human cytosolic molecular chaperones hsp90, hsp70 (hsc70) and hdj-1 have distinct roles in recognition of a non-native protein and protein refolding . EMBO J . 1996 6 17 ; 15 (12 ):2969 –79 . 8670798 44 Li S , Bai J-H , Park Y-D , Zhou H-M . Aggregation of creatine kinase during refolding and chaperonin-mediated folding of creatine kinase . Int J Biochem Cell Biol . 2001 ;33 (3 ):279 –86 . 11311859 45 Teshima T , Kondo A , Fukuda H . Reactivation of thermally inactivated enzymes by free and immobilized chaperonin GroEL/ES . Appl Microbiol Biotechnol. Springer-Verlag ; 1997 7 28 ; 48 (1 ):41 –6 . 46 Mayer MP , Bukau B . Hsp70 chaperones: cellular functions and molecular mechanism . Cell Mol Life Sci . 2005 3 ; 62 (6 ):670 –84 . 15770419 47 Hayer-Hartl MK , Weber F , Hartl FU . Mechanism of chaperonin action: GroES binding and release can drive GroEL-mediated protein folding in the absence of ATP hydrolysis . EMBO J . 1996 11 15 ; 15 (22 ):6111 –21 . 8947033 48 Yang W , Zhang L , Lv H , Li H , Zhang Y , Xu Y , et al The K-segments of wheat dehydrin WZY2 are essential for its protective functions under temperature stress . Front Plant Sci. Frontiers ; 2015 1 11 ; 6 :406 . 49 Kovacs D , Agoston B , Tompa P . Disordered plant LEA proteins as molecular chaperones . 2008 ;3 (9 ):710 –3 . 50 Caspers G-J , Leunissen JAM , de Jong WW . The expanding small heat-shock protein family, and structure predictions of the conserved “α-crystallin domain.” J Mol Evol . 1995 3 ; 40 (3 ):238 –48 . 7723051 51 Benjamin IJ , McMillan DR . Stress (Heat Shock) Proteins: Molecular Chaperones in Cardiovascular Biology and Disease . Circ Res . 1998 7 27 ; 83 (2 ):117 –32 . 9686751 52 Quan S , Bardwell JCA . Chaperone discovery . Bioessays . 2012 11 ; 34 (11 ):973 –81 . 10.1002/bies.201200059 22968800 53 Freeman BC , Toft DO , Morimoto RI . Molecular chaperone machines: chaperone activities of the cyclophilin Cyp-40 and the steroid aporeceptor-associated protein p23 . Science . 1996 12 6 ;274 (5293 ):1718 –20 . 8939864 54 Calderwood SK , Mambula SS , Gray PJ . Extracellular heat shock proteins in cell signaling and immunity . Ann N Y Acad Sci . 2007 10 ; 1113 :28 –39 . 17978280 55 Wyatt AR , Yerbury JJ , Ecroyd H , Wilson MR . Extracellular chaperones and proteostasis . Annu Rev Biochem. Annual Reviews ; 2013 1 7 ; 82 :295 –322 . 56 Li W , Tsen F , Sahu D , Bhatia A , Chen M , Multhoff G , et al Extracellular Hsp90 (eHsp90) as the actual target in clinical trials: intentionally or unintentionally . Int Rev Cell Mol Biol . 2013 1 ; 303 :203 –35 . 10.1016/B978-0-12-407697-6.00005-2 23445811

PMC005xxxxxx/PMC5001628.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756372210.1371/journal.pone.0161823PONE-D-16-01718Research ArticleBiology and Life SciencesAnatomyMusculoskeletal SystemSkeletonSkullMedicine and Health SciencesAnatomyMusculoskeletal SystemSkeletonSkullBiology and Life SciencesPhysiologyPhysiological ProcessesBone RemodelingOssificationMedicine and Health SciencesPhysiologyPhysiological ProcessesBone RemodelingOssificationBiology and Life SciencesOrganismsAnimalsVertebratesAmniotesBiology and Life SciencesAnatomyHeadEarsSemicircular CanalsMedicine and Health SciencesAnatomyHeadEarsSemicircular CanalsBiology and Life SciencesPaleontologyFossilsEarth SciencesPaleontologyFossilsBiology and Life SciencesAnatomyHeadJawMedicine and Health SciencesAnatomyHeadJawBiology and Life SciencesBiomechanicsBiological LocomotionBurrowingBiology and Life SciencesPhysiologyBiological LocomotionBurrowingMedicine and Health SciencesPhysiologyBiological LocomotionBurrowingBiology and Life SciencesAnatomyDigestive SystemMouthPalateMedicine and Health SciencesAnatomyDigestive SystemMouthPalateCranial Morphology of the Carboniferous-Permian Tetrapod Brachydectes newberryi (Lepospondyli, Lysorophia): New Data from µCT Cranial Morphology of the Carboniferous-Permian Tetrapod Brachydectes newberryiPardo Jason D. *Anderson Jason S. Department of Comparative Biology and Experimental Medicine, University of Calgary, Calgary, Alberta, CanadaNi Xijun EditorInstitute of Vertebrate Paleontology and Paleoanthropology Chinese Academy of Sciences, CHINACompeting Interests: The authors have declared that no competing interests exist. Conceptualization: JDP JSA. Data curation: JDP JSA. Funding acquisition: JSA. Investigation: JDP JSA. Methodology: JDP JSA. Project administration: JDP JSA. Resources: JSA. Supervision: JSA. Visualization: JDP. Writing – original draft: JDP JSA. Writing – review & editing: JDP JSA. * E-mail: jdpardo@ucalgary.ca26 8 2016 2016 11 8 e016182313 1 2016 14 8 2016 © 2016 Pardo, Anderson2016Pardo, AndersonThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Lysorophians are a group of early tetrapods with extremely elongate trunks, reduced limbs, and highly reduced skulls. Since the first discovery of this group, general similarities in outward appearance between lysorophians and some modern lissamphibian orders (specifically Urodela and Gymnophiona) have been recognized, and sometimes been the basis for hypotheses of lissamphibian origins. We studied the morphology of the skull, with particular emphasis on the neurocranium, of a partial growth series of the lysorophian Brachydectes newberryi using x-ray micro-computed tomography (μCT). Our study reveals similarities between the braincase of Brachydectes and brachystelechid recumbirostrans, corroborating prior work suggesting a close relationship between these taxa. We also describe the morphology of the epipterygoid, stapes, and quadrate in this taxon for the first time. Contra the proposals of some workers, we find no evidence of expected lissamphibian synapomorphies in the skull morphology in Brachydectes newberryi, and instead recognize a number of derived amniote characteristics within the braincase and suspensorium. Morphology previously considered indicative of taxonomic diversity within Lysorophia may reflect ontogenetic rather than taxonomic variation. The highly divergent morphology of lysorophians represents a refinement of morphological and functional trends within recumbirostrans, and is analogous to morphology observed in many modern fossorial reptiles. http://dx.doi.org/10.13039/501100000038Natural Sciences and Engineering Research Council of CanadaDiscovery Grant 327756-2011Anderson Jason S This research was supported by National Science and Engineering Council of Canada Discovery Grant 327756-2011 awarded to JSA. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction The Paleozoic origins of modern lissamphibians (Caudata, Anura, and Gymnophiona) have been a matter of substantial debate for over a hundred years. Recent effort has resulted in an emerging consensus that modern lissamphibians are amphibamid temnospondyls [1–6] on the basis of cranial anatomy, dental morphology [7], inner ear morphology [1], and life history [6,8–10]. However, the substantial morphological gap between amphibamids and the earliest representatives of the Caudata [11], Anura [12,13], and Gymnophiona [2,14] has left room for skepticism of this prevailing view. A second clade of Paleozoic tetrapods, the Lepospondyli, has been identified as a possible stem-group of all lissamphibians [15–17] or of the Gymnophiona specifically [3,18–21], on the basis of general similarity of the palate and postcranial skeleton, and extent of cranial ossification. Within the Lepospondyli, specific attention has been given to the Lysorophia, a group of lepospondyls with elongate bodies and reduced limbs. Early reports of lysorophian fossils identified these animals as early aquatic salamanders [22–23] or caecilians [24] but the first major attempts at describing lysorophian anatomy made it clear that lysorophian anatomy is inconsistent with direct ancestry to any specific lissamphibian group, and that lysorophians are either relatives of Lissamphibia more generally [25] or are morphologically-specialized early tetrapods without clear affinities to any modern tetrapod group [26–28]. A close relationship between lysorophians and modern lissamphibians has reemerged in some phylogenetic analyses [15–16, 29] and individual characteristics of lysorophians, such as the zygokrotaphic skull, have reemerged as possible support for a close relationship between lysorophians and caecilians [19] (but see [30] for further discussion), but a majority of analyses continue to find lysorophians to have no direct relevance to questions of lissamphibian origins [1–5, 21, 31]. In addition to an uncertain relationship between lysorophians and lissamphibians, the relationships of lysorophians within Lepospondyli are also unclear. A number of workers have suggested a relationship between lysorophians and recumbirostran ‘microsaurs’ (a clade of lepospondyls that includes pantylids, gymnarthrids, ostodolepids, and brachystelechids [3]) on the basis of occipital, vertebral, and pectoral morphology [28, 32], but a close phylogenetic relationship between these taxa has only been recovered by a minority of analyses [15–17]. Instead, a majority of analyses place lysorophians within a larger assemblage of ‘lepospondyls’ with elongate bodies and reduced limbs, alongside the Nectridea and Aïstopoda [3–4, 21, 31, 33–34]. Neither of these results are particularly satisfying. Whereas a relationship between lysorophians and recumbirostrans makes intuitive sense on the basis of vertebral, occipital, and appendicular morphology, this relationship lacks wholesale support within recent lepospondyl phylogenies. Alternately, a relationship with aïstopods and nectrideans is generally supported by recent phylogenetic analyses, but these taxa show few consistent similarities beyond general body shape and vertebral morphology, and much of this topology appears to be driven by characters describing skeletal reduction and axial elongation. The difficulties of resolving the relationships of lysorophians to Lissamphibia and to other lepospondyl taxa both stem from general reduction of the skeleton and loss of morphology characteristic of other major groups. Anderson [21] has noted that this problem may be intractable due to methodological biases in datasets overwhelmed by convergence in “loss” characters. Similar intractable phylogenetic problems have been solved in lungfishes [35–36], caecilians [2, 37–38], and crocodilians [39] by studying variation in the often poorly-described but character-rich braincase. Recent studies have focused attention on the braincase of some lepospondyls [33, 40–43] and have identified variation that may be indicative of relationships. Despite the potential importance of lysorophians in early tetrapod phylogeny, intensive study of the braincase has been lacking. This is, in large part, due to the preservational quality of the majority of lysorophian fossils available for study. Most well-preserved lysorophian fossils come from either Upper Carboniferous cannel coals, such as the Upper Freeport Coal (Allegheny Group, Upper Carboniferous) of the Dunkard Basin, or from the Lower Permian redbeds of Oklahoma and Texas. The two-dimensional preservation of the cannel coal specimens, and the uniform radio-opacity of the redbeds specimens, makes further study of this material using new imaging techniques (such as μCT) difficult. Abundant new lysorophian material has recently been reported from terrestrial paleosol horizons within the earliest Permian Eskridge Shale of Nebraska [44] and Speiser Shale of Kansas [45–46] within the Council Grove Group (CGG), representing lowstand sequences within a larger sequence of alternating terrigenous and nearshore marine sediments. Vertebrate bone from these deposits typically shows little diagenetic alteration, and the surrounding matrix generally exhibits little to no diagenetic precipitation of iron (unlike redbeds fossils), a fact which has made possible detailed study of CGG fossils using micro-computed x-ray tomography (μCT) [33, 36]. Lysorophians are the most common tetrapods in vertebrate-bearing horizons in CGG paleosols and are generally found partially or completely articulated, typically within flask-shaped burrow structures [44–46]. Many well-preserved, three-dimensional skulls have been collected from these localities, permitting detailed study of braincase morphology in B. newberryi at a range of sizes. Here we describe in detail the morphology of the skull and braincase of a partial ontogenetic sequence of these skulls from the Council Grove Group using μCT. Materials and Methods Geological Context The CGG spans the latest Carboniferous (Gzhelian) through the earliest Permian, representing the entirety of the Asselian in Kansas and Nebraska, and possibly extending into the Sakmarian [47–48]. These sediments record a series of fifth-order transgression-regression sequences bounded by well-developed fossil soils representing low-stand deposits. These fossil soils, which are typically gray-green, but sometimes also reddish-brown or mottled, are classifiable as vertisols or aridisols [49] and represent periodically waterlogged muds with limited organic content and high groundwater influence. Traces of roots [49] and vertebrate burrows [36, 44–46, 50] have been reported from these fossil soil horizons. Fossil vertebrates are abundant in these horizons, including gnathorhizid lungfishes [36, 44, 50], dvinosaurian temnospondyls [44, 51–54], amphibamid temnospondyls [44, 55], recumbirostran ‘microsaurs’ [33, 44, 56], diplocaulid nectrideans [57, 58], and rare diadectids and synapsids [44, 57], as well as numerous specimens of the lysorophian Brachydectes newberryi. These vertebrate-bearing localities have been interpreted as seasonal wetland systems [44, 46], similar to either vernal pools or playa lakes. The vertebrate assemblage at these localities appears to be largely autochthonous, although some taxa (such as the synapsids, diadectids, and recumbirostran ‘microsaurs’) may have been only occasional users of the wetland system and are represented primarily by rare isolated fragments (but not always, see [33]), in comparison with abundant obligately aquatic species (such as lungfishes), which are typically found articulated in burrow structures likely associated with aestivation [36, 44, 50, 59]. Lysorophian fossils from the CGG are typically found articulated in burrow structures as well [44–46] but generally occur at the margins of inferred ephemeral pond deposits [46] along with other tetrapods [59], whereas burrows of gnathorhizid lungfishes are typically concentrated at the center of these ponds [59]. Fossils of Brachydectes sampled in this study were collected from two vertebrate-bearing horizons in Lower Permian deposits of the upper CGG, one within the Eskridge Shale, and one within the Speiser Shale (Fig 1). Fossils from the Eskridge Shale come from a single horizon (Paleosol 2 of [49]), which has been sampled at three fossil-bearing localities in Richardson County, Nebraska, collectively the Humboldt Localities [44]. The Eskridge Shale is well-constrained to the early Asselian on the basis of marine microfossil biostratigraphy [47–48]. Fossils from the Speiser Shale come from a single horizon at a single locality west of Eskridge, Kansas. The Speiser Shale is constrained to the early Sakmarian [48] on the basis of conodont biostratigraphy. No permits were required for the described study, which complied with all relevant regulations. 10.1371/journal.pone.0161823.g001Fig 1 Stratigraphic and geographic context of specimens studied. A, geographic location of specimen localities; B, stratigraphic position of localities, after [48]. 1, Humboldt, NE, localities (undivided); 2, Eskridge, KS locality. Light gray indicates Upper Carboniferous surface deposits; dark gray indicates Lower Permian surface deposits. Specimens studied To fully image internal morphology of the skull of Brachydectes newberryi, we scanned nine skulls attributable to this species (S1 Table), representing the full range of sizes observed in the Council Grove Group (Fig 2). These specimens represent the bulk of Council Grove Group lysorophians from the Denver Museum of Nature and Science, Denver, Colorado (DMNH), the University of Kansas Natural History Museum and Biodiversity Institute, Vertebrate Paleontology Division, Lawrence, Kansas (KUVP), and the University of Nebraska State Museum, Lincoln, Nebraska (UNSM). All specimens were scanned in a Skyscan1173 conical beam desktop μCT (Bruker Corporation) and reconstructed as a stack of transverse slices in NRecon v. 1.6.6.0 (Skyscan, 2011). The resulting image stacks were downsampled using an interval of two and cropped in ImageJ using the virtual stack function to reduce computation burden. We then imported the resulting image stacks into Amira 5.3.3 (Visage Imaging, Inc.) for visualization in 3D. Volume renders of complete specimens were produced using the Volren module. Individual elements were then segmented using the LabelField module. Segmented elements were then isolated using the Arithmetic module and rendered using the Volren module, or were rendered as a three-dimensional surface using the SurfaceGen module. 10.1371/journal.pone.0161823.g002Fig 2 Volume renders of selected skulls studied, showing representative size variation in specimens, dorsal view. A, DMNH 47854; B, DMNH 521121; C, KUVP 49541; D, UNSM 32149; E, comparison of skull sizes. Scale bar equals 1 cm. Braincase anatomy In our study of the braincase of Brachydectes, we deviate from the traditional practice of describing individual bones and instead describe braincase morphology in terms of embryonic cartilaginous precursors of neurocranial ossifications. This approach has been previously employed in the study of the recumbirostran braincase [43] as it offers several advantages over traditional descriptive approaches. First, general braincase morphology is broadly conserved among tetrapods [60] permitting broad comparability between taxa on the basis of topological relationships between conserved structures (pathways of cranial nerves, presence of fenestrae, etc.). Second, this approach avoids a priori assumptions of homology between ossifications of the braincase in similar locations. Third, this facilitates comparison of conserved structure that may be preserved within nonhomologous elements, allowing us to focus on morphology of the structure rather than nomenclature of the bone. Finally, some lepospondyl taxa (e.g. the brachystelechids Carrolla craddocki [41] and Quasicaecilia texana [43], and phlegethontiids, [40]) as well as some lissamphibians (e.g. Gymnophiona, [37, 38]) exhibit massively co-ossified otoccipital regions (an os basale), and many early tetrapods massively co-ossified their braincase more generally [61–63], making description of braincase structure rather than individual bones necessary for any comparative efforts which include these taxa. We do, however, recognize individual bones where possible, following the revised terminology of recent studies of recumbirostran morphology [41–43]. This allows us to recognize common ossification centres as well as identify which ossification center has expanded into a particular cartilage. By reporting both the structure of the cartilaginous precursors and the ossifications they are incorporated into, we maximize the available anatomical data in order to best understand the anatomy and relationships of Brachydectes. Systematic Paleontology Lepospondyli Zittel 1888 Recumbirostra Anderson 2007 Molgophidae Cope 1875 Cocytinidae Cope 1875 Paterosauridae Broili 1908 Lysorophidae Williston 1908 Revised diagnosis. Elongate recumbirostrans with the following characteristics: large temporal emargination present contacting the parietals, temporal emargination confluent with orbit due to loss of posterior circumorbital bones (postfrontal, postorbital, and jugal), squamosal narrow and anteriorly-canted, orbitosphenoids and pila antotica robust and oriented vertically, ectopterygoids absent, pineal foramen absent, quadratojugal absent, intercentra absent, cultriform process of parasphenoid broad and rectangular. Phylogenetic definition. Tetrapods more closely related to Brachydectes newberryi than to Carrolla craddocki, Rhynchonkos stovalli, Dvellecanus carrolli, Cardiocephalus sternbergi, Euryodus primus, Huskerpeton englehorni, or Pantylus cordatus. This is a stem-based definition. Discussion. There has been a great deal of uncertainty in the literature concerning the correct name for the family containing Brachydectes. A number of familial designations have been employed in the literature, including Molgophidae [64], Lysorophidae [23], Paterosauridae [65], and Cocytinidae [64]. Of these names, Molgophidae appears first in the literature [64]. In a revision of the taxon, Wellstead [28] elected to maintain the name Cocytinidae, but it is unclear why, because there is no greater history of use of the name Cocytinidae and no lapse in usage of the name Molgophidae would necessitate suppression of the family-group Molgophidae in favor of Cocytinidae [ICZN Article 23.9.1]. Although Molgophis is now considered a junior synonym of Brachydectes, Article 40.1 of the ICZN states that synonymy of the type genus is not alone reason to change the family-group name. We therefore recommend that subsequent works use the name Molgophidae and treat Cocytinidae, Lysorophidae, and Paterosauridae as junior synonyms. We deviate from previous workers by not recognizing the Lysorophia as a distinct order. Although phylogenetic position of Brachydectes is under some debate, the possibility exists that Brachydectes falls well within ‘Microsauria’ [17] and possibly within Recumbirostra [43], and its divergent morphology may represent extreme specialization within Recumbirostra, rather than a distinct lineage separate from Recumbirostra. Brachydectes Cope 1868 Molgophis Cope 1868 Cocytinus Cope 1871 Pleuroptyx Cope 1875 Lysorophus Cope 1877 Type Species. Brachydectes newberryi Cope 1868 Diagnosis. As for Molgophidae. Brachydectes newberryi Cope 1868 Molgophis macrurus Cope 1868 Cocytinus gyrinoides Cope 1871 Pleuroptyx clavatus Cope 1875 Molgophis brevicostatus Cope 1875 Lysorophus tricarinatus Cope 1877 in part Brachydectes elongatus Wellstead 1991 in part Holotype. American Museum of Natural History (AMNH) 6941, partial skull. Holotype locality. Cannel coal above the Upper Freeport Coal (Allegheny Group, Moscovian, Pennsylvanian) of Linton Diamond Mine, Jefferson County, Ohio [66, 67]. Attributed specimens. Mayer Quarry (Humbolt, Richardson County, Nebraska, USA), Eskridge Shale (Council Grove Group, Asselian): DMNH 36599, right dentary; DMNH 43125, right dentary; DMNH 43224, skull with complete axial skeleton; DMNH 47854, skull with partial vertebral column; DMNH 49902, partial skull and partial skeleton; DMNH 50142, partial skull; DMNH 51121, skull with partial skeleton; DMNH 51122, partial skull with disarticulated skeleton; DMNH 52081, partial skull and skeleton. Shot in the Dark Quarry (Humbolt, Richardson County, Nebraska, USA), Eskridge Shale (Council Grove Group, Asselian): UNSM 32100, partial skull; UNSM 32115, partial skull; UNSM 32147, partial skull with partial skeleton; UNSM 32149, partial skull and associated postcranial elements. Eskridge Locality (Eskridge, Waubunsee County, Kansas, USA), Speiser Shale (Council Grove Group, Sakmarian): KUVP 49537, skull and partial skeleton; KUVP 49538, skull and partial skeleton; KUVP 49539, skull and partial skeleton; KUVP 49540, skull and partial skeleton; KUVP 49541, skull and partial skeleton. Revised diagnosis. As for family. We provisionally synonymize all molgophids here as Brachydectes newberryi, as existing diagnoses are insufficient to distinguish between identified species. The presence of populations with vastly different numbers of trunk vertebrae [28] strongly suggests the presence of multiple species of molgophid in the Carboniferous-Permian if North America, but how this character is distributed amongst name-bearing specimens is difficult to ascertain. Description Skull Roof and Cheek The skull of Brachydectes newberryi is extremely derived in comparison with that of other early tetrapods (including other lepospondyls) and exhibits significant reduction of dermal skull elements, especially along the orbit and cheek (Figs 3C and 4). Most strikingly, the lateral region of the cheek lacks dermal ossification between the orbit and the squamosal, resulting in a deep lateral emargination of the ventral cheek that is continuous with the orbit and extends dorsally to the parietals (Fig 3C). Previous workers have disagreed on the extent of the orbit within this region, and most recently Marjanovic and Laurin [16] suggested that the orbit had expanded to encompasse the entire lateral unossified regionto accommodate adductor musculature (although this was subsequently amended [17]). In the Council Grove skulls, a low postorbital process of the parietal is present near the prefrontal-parietal suture, confirming that the open cheek is formed through the loss of the postorbital bar, and that the orbit is very small and restricted to the very anterior extent of this unossified region (Fig 3). The remainder appears to be a greatly enlarged lateral cheek emargination, similar to that seen in the hapsidopareiontids Hapsidopareion and Llistrofus [32, 68, 69], the ostodolepids Tambaroter and Pelodosotis [32, 68, 70], and juveniles of the gymnarthrid Cardiocephalus (JSA, pers obs), although significantly deeper than seen in these taxa. This emargination may have accommodated enlarged adductor mandibulae externus musculature, in which case the morphology seen in B. newberryi may indicate further increase in the size of this muscle in Brachydectes in comparison with other lepospondyls. 10.1371/journal.pone.0161823.g003Fig 3 External skull morphology of Brachydectes newberryi KUVP 49541, rendered from μCT. A, dorsal view; B, anterior view; C, right lateral view; D, occipital view; E, palatal view, with lower jaws removed. Scale bar equals 5 mm. Abbreviations: ang, angular; ap, median ascending process of supraoccipital; bo, basioccipital; cbic, foramen serving the cerebral branch of the internal carotid artery; ch, choana; d, dentary; en, external naris; eo, exoccipital; eoc, exoccipital condyle; fm, foramen magnum; lac, lacrimal; mf, mandibular fenestra; mx, maxilla; op, opisthotic; pal, palatine; pmx, premaxilla; porp, postorbital process; ppl, postparietal lappet; preart, prearticular; prf, prefrontal; psph, parasphenoid; pt, pterygoid; q, quadrate; qsp, stapedial process of quadrate; qtr, trochlea of quadrate; rap, retroarticular process; sa, surangular; so, supraoccipital; sq, squamosal; sqpf, posterior flange of squamosal; t, tabular; v, vomer. Scale bar equals 1 cm. 10.1371/journal.pone.0161823.g004Fig 4 Reconstruction of the skull of Brachydectes newberryi. A, lateral view; B, dorsal view; C, palatal view. Image not to scale. The skull roof of B. newberryi was well described by previous authors [26, 28]. The median skull roof consists of paired nasals, frontals, parietals, and postparietals, as in the majority of early tetrapods. The nasals, frontals, parietals, and postparietals are all subequal in length. The prefrontal is well integrated into the anterior skull roof, forming a bridge between the frontals and maxillae, whereas posteriorly the tabular is reduced and integrated into the dermatocranial support for the suspensorium. Anteriorly the nasals contact the premaxillae along a broad suture. Lateral to this is a broad, laterally-flaring emargination associated with the dorsal margin of the external naris. Posteriorly, the nasals overlap the frontals in a deep interdigitating suture. The median suture between the nasal pair is a straight butt joint without significant interdigitation. The frontal is roughly rectangular, and is flanked laterally by the prefrontal, which excludes it from both the orbit, and the temporal emargination. The median suture between the frontal pair is straight but shows tongue-in-groove interdigitation in transverse section in larger specimens. Posteriorly, the frontal is sutured to the parietal via a series of deep lappets that incise posteriorly into the parietal. Some variation exists in the general shape of the frontal pair. In some specimens, the frontal pair is roughly rectangular, whereas in others the frontal pair is trapezoidal and tapers anteriorly. In a few specimens, the frontal pair expands anteriorly. The parietal pair is also broadly rectangular, and together the two parietals are as wide as the prefrontals and frontals combined. No pineal foramen is present and no pineal fossa is present on the ventral surface of the parietals. The median suture between parietals is essentially straight, with minimal interdigitation. Posteriorly, the parietal interdigitates with deep incisures in the postparietal pair. The postparietals lie directly posterior, and are equal in width, to the parietals. The postparietals contribute to the sloping dorsal portion of the occiput, with a deep fossa on the occipital surface of each postparietal accommodating the epaxial musculature. No posttemporal fenestra is present in the occiput between the postparietals, the occipital arch, and the otic bones. The ventral surface of the postparietals contains impressions of the semicircular canals and may represent investment of the dorsal otic capsule by the postparietals. The temporal region consists of two bones: the squamosal and the tabular (Fig 3). The squamosal makes up the majority of this region. It is a slender bone that descends from the occiput anteriorly towards the jaw articulation, intersecting the coronal plane at the level of the palate at an angle of approximately 40 degrees. The ventral end of the squamosal wraps around the lateral surface of the quadrate. A strong ridge is present on the occipital surface of the squamosal, and likely served as the point of origin of the depressor mandibulae. The tabular is a small triangular bone that overlaps the dorsal portion of the squamosal and the lateral surface of the postparietal (Figs 2 and 3). The lateral surface of the bone is marked by a large fossa that appears to be an extension of the lateral emargination of the cheek. Posteriorly, the tabular projects beyond the occipital surface. This structure is not equivalent to the “tabular horns” seen in some nectrideans or the tabular prongs seen in adelospondyls. The posterior projection of the tabular forms the margin of a fossa for the epaxial musculature rather than constituting an actual projection from the posterior skull roof. Upper Dental Arcade The premaxilla is small with a reduced pars dorsalis (Fig 3). No vomerine shelf of the premaxilla is present. The premaxillary dentition is reduced, with only three to four large, conical teeth. The pars dorsalis is aligned along the midline without an internarial fontanelle. The external nares are large, face anteriorly, and restrict the contact between the premaxilla and nasal to a small suture along the midline. The maxilla is also reduced (Fig 3). No palatal shelf of the maxilla is present, but a tight suture does exist between the maxilla and palatine. The maxillary dentition is relatively reduced, with five to eight large conical teeth. The maxilla makes up the ventral margin of the orbit, but does not extend posterior to the postorbital process of the skull roof. The pars dorsalis is relatively limited, but is tightly sutured to the lacrimal in the antorbital region. Anteriorly, the maxilla participates in the margin of the external naris and choana. Braincase The braincase of B. newberryi is well-ossified and robust (Figs 5–7). The orbitosphenoids, prootics, opisthotics, supraoccipital, exoccipitals, and basioccipital are all present as distinct elements, whereas regions of the braincase associated with the presphenoid, basisphenoid, and pila antotica are co-ossified with the parasphenoid. 10.1371/journal.pone.0161823.g005Fig 5 Braincase morphology of Brachydectes newberryi, DMNH 521121, 3D renders segmented from μCT. A, left lateral view; B, right medial view; C, ventral view; D, dorsal view; E, posterior view. Scale bar equals 5 mm. Abbreviations: bo, basioccipital; eo, exoccipital; eoc, exoccipital condyle; fme, foramen metoticum; n.V, foramen serving the trigeminal nerve (undivided); n.XII, foramen serving the hypoglossal nerve; op, opisthotic; os, orbitosphenoid; pa, pila antotica; po, prootic; psph, parasphenoid; so, supraoccipital. 10.1371/journal.pone.0161823.g006Fig 6 Braincase of Brachydectes newberryi, KUVP 49541, 3D renders segmented from μCT. A, dorsal view; B, anterior view; C, left lateral view; D, left anterolateral view; E, ventral view; F, occipital view. Scale bar equals 5 mm. Abbreviations: ap, median ascending process of the supraoccipital; bo, basioccipital; boc, basioccipital cotyle; bptp, basipterygoid process of parasphenoid; cbic, path of the cerebral branch of the internal carotid artery; cp, cultriform process of the parasphenoid; crbo, conical recess of the basioccipital; eo, exoccipital; eoc, exoccipital condyle; fenv, fenestra vestibularis; fm, foramen magnum; fme, foramen metoticum; hf, hypophyseal fossa; mcv, foramen serving middle cerebral vein; n.I, foramen serving olfactory nerve; n.II, foramen serving optic nerve; n.V, foramen serving trigeminal nerve (undivided); n.VII, foramen serving facial nerve; n.XII, foramen serving hypoglossal nerve; op, opisthotic; os, orbitosphenoid; pa, pila antotica; pbic, sulci serving the palatal branch of the internal carotid artery; po, prootic; psph, parasphenoid; psphe, presphenoid; so, supraoccipital; vs, vidian sulcus. 10.1371/journal.pone.0161823.g007Fig 7 Braincase morphology of Bracydectes newberryi UNSM 32149, 3D renders segmented from μCT. A, dorsal view; B, left anterolateral view; C, left lateral view; D, occipital view; E, right lateral view; F, left posterolateral view; G, ventral view. Scale bar equals 5 mm. Abbreviations: ap, median ascending process of the supraoccipital; bo, basioccipital; crbo, conical recess of the basioccipital; eo, exoccipital; eoc, exoccipital condyle; fm, foramen magnum; fme, foramen metoticum; hf, hypophyseal fossa; lap, lateral ascending process of the supraoccipital; mcv, foramen serving the middle cerebral vein; n.I, foramen serving the olfactory nerve; n.V, foramen serving the trigeminal nerve (undivided); n.VII, foramen serving facial nerve; n.XIII, foramen serving hypoglossal nerve; op, opisthotic; os, orbitosphenoid; pa, pila antotica; po, prootic; pop, paroccipital process; psph, parasphenoid; psphe, presphenoid; so, supraoccipital. The presphenoid (Figs 6 and 7) is a lozenge-shaped element at the base of the anterior region of the braincase. It is closely associated with the cultriform process of the parasphenoid, and these two elements may, in fact, be co-ossified, as no separate presphenoid ossification has been identified by previous workers [25, 26]. Dorsally, the presphenoid invades the columella ethmoidalis (Figs 6 and 7). Ossification of the columella ethmoidalis fills the entire space between the olfactory nerve foramina in the largest specimen studied (Fig 7) but is restricted to the areas surrounding the olfactory nerves in smaller specimens (Figs 5 and 6) and absent entirely in the smallest specimens. Otherwise, the presphenoid is dorsoventrally flattened along its length and does not form an interorbital septum, in contrast with the condition seen in brachystelechids [41, 43] but consistent with the morphology seen in a number of other recumbirostrans [33, 42, 71]. A presphenoid has previously been described in Carrolla craddocki as the sphenethmoid [41] and appears to be present in a variety of other microsaurs as well [42, 43]. Paired orbitosphenoids make up the lateral wall of the anterior sphenoid region, extending from the presphenoid posteriorly to the optic foramen (Fig 6). The orbitosphenoid in B. newberryi is a long, low, robust vertical wall similar in morphology to that of Carrolla craddocki [41] and Quasicaecilia texana [43] rather than the thin, bulging structure seen in more conservative ‘microsaurs’ [33, 42]. A deep notch is present at the posteroventral corner of the bone for accommodation of the optic nerve (Figs 6 and 7). Unlike the condition seen in many recumbirostrans [33, 42], no descending flange of the frontal articulates with the orbitosphenoid. Anteriorly, the surface of the orbitosphenoid is smooth and lacking in facets for articulation with the trabecular plate dorsally and the antorbital cartilage ventrally, in contrast with the condition seen in lissamphibians [60]. It appears likely that no lateral connection between the orbitosphenoid and the nasal capsule was present in Brachydectes. Similar morphology has been observed in a number of other recumbirostrans [33, 41–43, 71], and this may represent a more general ‘microsaur’ condition. The pila metoptica is unossified, consistent with the morphology found in Carrolla craddocki [41], Quasicaecilia texana [43], Huskerpeton englehorni [33], and Rhynchonkos stovalli [42]. The pila antotica is fully ossified, tall, robust, teardrop-shaped in cross section, and firmly sutured to the ventral surface of the parietal (Figs 5–7). Between the pilae antoticae, the hypophyseal fossa is broad and weakly posteriorly-directed (Fig 6). The dorsum sellae delineates the posterior margin of the shallow hypophyseal fossa at the midline, but ascend laterally into broad ridges along the medial surface of the pilae antoticae, and reach the anterior contact between the pilae antoticae and the parietals. Posterior to the pila antotica, there is a large antotic fenestra (Figs 5–7), which in larger specimens is partially to completely bisected by a bony bridge into a dorsal and ventral foramen. Anteriorly, the dorsal foramen opens into a sulcus on the ventral surface of the parietals, and posteriorly passes dorsally and medially to the semicircular canals of the inner ear (Fig 7, sov). The ventral foramen has been identified by previous authors as serving the postganglionic trunk of the trigeminal nerve [41–43], whereas the dorsal foramen has generally been interpreted as serving a vein [41, 43], likely the middle cerebral vein. Both fenestrae are very large in comparison with other recumbirostrans, whereas in other recumbirostrans these foramina are of moderate size [41–43]. The otic capsules consist of distinct prootic and opisthotic elements (Figs 5–7). In smaller specimens, the prootic and opisthotic are small and relatively distinct (Fig 5), but in the largest specimen (UNSM 32149) these elements are more completely ossified, and approach each other ventral to the horizontal semicircular canal (Fig 7). The prootic is roughly triangular, with the base sutured to the parasphenoid at the level of the basipterygoid processes. The opisthotic is tall and pillar-like, with a strong vertical crista along the medial surface, here interpreted as a crista interfenestralis (Fig 8). The prootic and opisthotic co-ossified with the squamosal and parietal dorsally and with the parasphenoid ventrally (Fig 7). The fenestra vestibularis is roughly oval in shape and fits the stapes closely both anteriorly and posteriorly, seemingly without space for an opercular cartilage. Laterally, the crista parotica exists as a strong horizontal ridge on the prootic, bracing against the squamosal. This ridge may be homologous to the paroccipital process of amniotes [72], although the latter involves more significant contributions from the opisthotic. 10.1371/journal.pone.0161823.g008Fig 8 Selected otic elements of the left otic capsule of Brachydectes newberryi, UNSM 32149, in medial view. Scale bar equals 1 mm. Abbreviations: amascc, fossa enclosing ampulla of the anterior semicircular canal; ascc, sulcus for anterior semicircular canal; crint, crista interfenestralis; fme, foramen metoticum; hscc, sulcus for horizontal semicircular canal; lr, lagenar recess; op, opisthotic; po, prootic. The stapes has a round to oval footplate and a well-developed columella, and is well-ossified in larger specimens (Fig 9). A shallow groove is present along the perimeter of the footplate. The columella flares distally, forming a broad articulation with the dorsal portion of the quadrate. A small foramen halfway along the length of the columella likely transmitted the stapedial artery (Fig 9). A robust dorsal process is present in the largest specimen (Fig 9K–9O), but not in smaller specimens (Fig 9A–9E), which extends proximally from the midpoint of the columella towards the crista parotica, although it does not directly articulate with the otic capsule. The deep notch between the dorsal process of the stapes and the base of the columella would likely have encased the lateral head vein. In some specimens, the columella is weakly mineralized between the footplate and the distal tip, and in some cases the distal tip of the columella appears to be a separate distinct ossification. It is possible that this element may be a distinct extracolumella (as is seen in some reptiles and amphibians), or that it represents a second ossification center in an element that co-ossifies in fully adult specimens. 10.1371/journal.pone.0161823.g009Fig 9 Morphology of the stapes in small, subadult, and skeletally mature Brachydectes newberryi. A-E, DMNH 521121, left stapes in A, dorsal; B, left lateral; C, posterior; D, anterior; and E, ventral view; F-J, KUVP 49541, right stapes in F, dorsal; G, right lateral; H, posterior; I, anterior; and J, ventral view; K-O, UNSM 32149, left stapes in K, dorsal; L, left lateral; M, posterior; N, anterior; and O, ventral view. Scale bar equals 1 mm. Abbreviations: col, columellar process of stapes; dp, dorsal process of columella of stapes; stf, stapedial foramen. The occipital arch consists of a well-ossified basioccipital, paired exoccipitals, and a single median supraoccipital (Figs 5–7). Surfaces on the basioccipital and exoccipitals contribute to the occipital cotyle, which is unpaired and crescentic. The occipital cotyle is weakly concave, and accepts the odontoid process of the atlas. The bases of the exoccipitals are swollen anteriorly along the suture with the parasphenoid. The metotic foramen passes through the suture between the exoccipital and opisthotic, and served as the passage for the vagus nerve and the jugular vein. A second foramen, which pierces the exoccipital just lateral to the occipital cotyle, served the hypoglossal nerve (Figs 6 and 7). The synotic tectum is occupied by a single broad supraoccipital bone, as in other ‘microsaurs’ [33, 42]. A pointed ascending process of this ossification extends anteriorly beneath the postparietals and is exposed medially up to the posterior margin of the parietals. Posttemporal fossae are fully closed. No supraoccipital sinus is present. A posterior shelf of the supraoccipital roofs the foramen magnum; the dorsolateral surfaces of this shelf serve as an articular surface for the proatlas. The supraoccipital is present as a single ossification across all ontogenetic stages studied here. Palate The parasphenoid of Brachydectes newberryi is a broadly triangular element in ventral view, and represents the majority of the ossified palate. The cultriform process is laterally expanded to the extent that it is nearly as broad as the distance between the basipterygoid processes, and extends to the anterior end of the orbitosphenoid ossification, where it tapers rapidly to a point. The basal plate of the parasphenoid is roughly rectangular, without the broad triangular posterior expansion seen in some ‘microsaurs’ [33, 41–43]. The basipterygoid processes lie directly ventral to the pilae antoticae, and are together narrower than the widest extent of the basal plate of the parasphenoid. Posteriorly, the parasphenoid is deeply notched, exposing the basioccipital. The parasphenoid is firmly sutured to the ventral surface of both the anterior and posterior regions of the braincase. The branches of the internal carotid artery leave clear grooves and canals on and through the ventral and lateral surface of the parasphenoid (Fig 6), recording the pattern of some of the cranial circulation in Brachydectes. The route of the internal carotid artery coursed along a sulcus on the ventral surface of the parasphenoid, just medial to the basipterygoid processes. At the level of the latter, this sulcus gives rise to another deep groove medially that terminates at a foramen that pierces the basal plate of the parasphenoid and emerges from the medial surface of the pleurosphenoid within the hypophyseal fossa (Fig 6E). This canal would have enclosed the cerebral branch of the internal carotid artery. The main sulcus continues laterally, anterior to the basipterygoid process, and follows a shallow groove on the lateral surface of the cultriform process. This groove would have housed the palatal branch of the internal carotid artery. In contrast to the condition seen in temnospondyls [73] and lissamphibians [74] where the internal carotid divides into cerebral and palatal branches either within the parasphenoid itself or within the hypophyseal fossa, the internal carotid artery of Brachydectes appears to have been divided into cerebral and palatal branches prior to entering the braincase, consistent with the morphology seen in some stem and crown amniotes [75]. The vomers are narrow, elongate bones that contact each other anteriorly at the midline of the palate, anterior to the cultriform process of the parasphenoid (Fig 10). A short premaxillary process extends along the midline of the anterior palate to meet the premaxilla, but neither a premaxillary shelf nor a maxillary process or shelf is present. Posteriorly, the palatine process of the vomer is narrow and extends just lateral to the cultriform process. A series of six to eight large, robust teeth extends along the midline of each vomer. 10.1371/journal.pone.0161823.g010Fig 10 Selected skulls of Brachydectes newberryi in palatal view. A, KUVP 49539; B, UNSM 32149. Scale equals 5 mm. Abbreviations: ch, choana; mx, maxilla; pal, palatine; psph, parasphenoid; pt, palatine; q, quadrate; v, vomer. The palatine is a short, strutlike element that connects the maxilla to the palatine ramus of the pterygoid (Fig 10). The palatine and maxilla are typically found together in specimens even when disarticulated from the remainder of the skull, suggesting that these elements may be more tightly sutured than the remainder of the palate. Neither teeth nor denticles are present on the palatines. The pterygoids are greatly reduced in comparison with those of other early tetrapods (Figs 3E, 10 and 11). The palatine ramus is foreshortened, but still retains a lateral contact with the palatine and an anterior contact with the vomers. The body of the pterygoid extends adjacent to the posterior portion of the cultriform process of the parasphenoid, eliminating any interpterygoid vacuity. The quadrate process is anteriorly displaced and descends obliquely along the medial surface of the quadrate. The medial surface of the pterygoid is reflected dorsally to form a low dorsal process lateral to the pila antotica. The basal process is reduced to a shallow facet at the anterior extremity of the dorsal process. No transverse process extends into the subtemporal vacuity. Teeth and denticles are absent from the body of the pterygoid. 10.1371/journal.pone.0161823.g011Fig 11 Right palatal and suspensorial elements of Brachydectes newberryi, UNSM 32149. A, lateral view; B, anterior view; C, posterior view. Scale equals 5 mm. Abbreviations: ape, ascending process of epipterygoid; dlpt, dorsal lamina of pterygoid; epi, epipterygoid; prpt, palatine ramus of pterygoid; pt, pterygoid; stp, stapedial process of quadrate. An epipterygoid is present and ossified only in the largest specimen scanned, UNSM 32149 (Fig 11). The epipterygoid is a simple element, with a broad, roughly-horizontal ventral facet for articulation with the pterygoid, and a simple, slender dorsal stalk, comparable to the condition seen in amniotes. The epipterygoid does not articulate with the skull roof and appears to serve a reduced role in support of the palate in comparison with the recumbirostrans Pantylus cordatus [76], Huskerpeton englehorni [33], Rhynchonkos stovalli [42], Aletrimyti gaskillae [42], and Dvellecanus carrolli [42]. The epipterygoid contributes substantially to the conus recessus. The quadrate is relatively simple in morphology (Fig 11). The condyle is trochlear and is somewhat mediolaterally compressed. The anterior surface of the quadrate is weakly concave, with a weak transverse ridge running from the lateral trochlea of the condylar surface to the medial surface. Posteriorly the quadrate tapers to a sharp ridge, which extends to a tubercle-like stapedial process in the largest specimen. A shallow sulcus present on the medial surface may be equivalent to the sulcus assigned to the so-called “chorda tympani” nerve. Lower Jaw The lower jaw of Brachydectes newberryi has been adequately described by previous workers [25, 28] and little difference exists between the described morphology and that observed in the Council Grove skulls (Fig 12). The number of ossifications is greatly reduced from the basal tetrapod complement; only a dentary, angular, surangular, prearticular, and coronoid are present. The dentary is relatively short with a deep coronoid expansion that is continued posteriorly by the surangular. As previously described, Brachydectes newberryi has a relatively low tooth count, with approximately five to six teeth. The dentary is the primary bone of the lower jaw, and forms, with the coronoid, a large well-developed coronoid process. It appears to be the only bone contributing to the symphysis. A single, small Meckelian foramen pierces the dentary around the posterior extent of the tooth row. The coronoid is simple, with a small anterior process, but not precluding extensive contact between the dentary and prearticular. Teeth and denticles are both absent on the coronoid. The prearticular makes up the majority of the medial surface of the jaw. The articular is co-ossified with the prearticular and is narrow but deep, fitting closely into the trochlea of the quadrate. Posterior to the articular fossa, the angular forms an elongate, straight retroarticular process with a median contribution from the articular. A large mandibular fenestra is present laterally, between the dentary, angular and surangular. 10.1371/journal.pone.0161823.g012Fig 12 Right lower jaw of Brachydectes newberryi, KUVP 49541, rendered from μCT. A, lateral view; B, dorsal view; C, medial view; D, ventral view. Scale equals 5 mm. Abbreviations: af, articular fossa; ang, angular; cor, coronoid; d, dentary; mf, mandibular fenestra; mnf, mental foramen; ms, medial shelf of prearticular; rap, retroarticular process; sa, surangular; sym, mandibular symphysis. Cranial Endocast In order to better characterize the morphology of the cavum cranii and otic capsule, we produced a virtual endocast of KUVP 49541 (Fig 13). The cavum cranii is roughly rectangular in dorsoventral view, and wedge-shaped in lateral view, without significant curvature of the AP axis. The cavum cranii is expanded anteriorly, constricted in the region of the hypophysis, then expands again posteriorly. 10.1371/journal.pone.0161823.g013Fig 13 Cranial endocast of Brachydectes newberryi, KUVP 49541, rendered from μCT. A-B, position of cranial endocast within complete skull, A, dorsal view; B, lateral view; C, endocast in dorsal view; D, endocast in right lateral view; E, endocast in left lateral view; F, endocast in ventral view; G, endocast in endocast in left posteroventral oblique view; H, left anteroventral oblique view. Scale equals 5 mm. Abbreviations: cbic, passage of the cerebral branch of the internal carotid artery; fme, passage of the foramen metopticum; hsc, horizontal semicircular canal; hyp, hypophyseal fossa; lag, fossa enclosing lagena; mcv, passage serving middle cerebral vein; n.I, passage of the olfactory nerve; n.II, passage of the optic nerve; n.V, passage of the trigeminal nerve; n.VI, passage of the facial nerve; n.XII, passage of the hypoglossal nerve; psc, posterior semicircular canal;; stym, fossa possibly accommodating scala tympani; ut, utricular fossa. The telencephalon region of the endocast, enclosed laterally by the orbitosphenoids, contains the olfactory bulbs and cerebrum. The canals enclosing the olfactory nerves are widely-spaced and diverge anteriorly. Unlike the condition seen in some recumbirostrans [33, 42], there are not distinct fossae accommodating the olfactory lobes and cerebral hemispheres, but the passage for the optic nerve serves as an absolute posteriormost limit on the size of the olfactory bulbs. The passage for the optic nerve is situated at the bottom of the ventral surface of the forebrain region, approximately halfway between the hypophyseal fossa and the passages for the olfactory nerves. The diencephalon, enclosed laterally by the ossified pila antotica, appears as a marked medial constriction between the cerebrum and otic capsules. Ventrally there is a shallow expansion of the endocast accommodating the hypophysis. The passages for the cerebral branches of the internal carotid artery converge anteriorly to meet the hypophysis. Dorsally, neither a canal nor a fossa is present to accommodate the epiphysis, suggesting that this structure may have been highly reduced in comparison with other early tetrapods. The medial wall of the otic capsule is largely unossified, preventing clear delineation of structures in the midbrain. However, the endocast does expand dorsally and somewhat laterally posterior to the pila antotica. This expansion of the endocast may accommodate the cerebellum and optic tectum, or cerebellum alone, depending on the organization of the diencephalon and midbrain. A canal, possibly serving the middle cerebral vein, meets the endocast lateral to this expansion, passing between it and the ampulla of the anterior semicircular canal of the otic capsule. The canal serving the trigeminal emerges ventral to this canal, just posterodorsal of the hypophyseal fossa. A small canal lateral to this and entering the endocast ventromedial to the space accommodating the lagena may represent the course of the facial nerve. Posterior to the otic capsules, the dorsal surface of the endocast is arched, producing an endocast that is roughly triangular in transverse section. Two canals are present in the hindbrain region: a large canal representing the path of the metotic foramen and a smaller canal representing the path of the hypoglossal nerve. The metotic canal, which would have enclosed the glossopharyngeal, vagus, and accessory nerves as well as the internal jugular vein, meets the endocast ventral to the posterior semicircular canal. The canal serving the hypoglossal nerve meets the hindbrain region of the endocast posterior to the metotic canal. The medial surface of the otic capsule of UNSM 32149 preserves osteological correlates of the structure of the inner ear, permitting creation of a virtual endocast of the inner ear and description of the gross morphology of this structure in greater detail (Fig 14). Traces of the courses of the semicircular canals and saccular region are preserved as sulci along the medial surface of the otic bones. The semicircular canals are restricted to the upper portion of the otic capsule and directly abut the utriculus, with no intervening bone. Ampullae are preserved as weak swellings in the fossae for the semicircular canals, but lack distinct morphology. The horizontal semicircular canal is laterally exposed just ventral to the crista parotica. A large saccular region is present ventral to the utriculus, and is expanded greatly below the semicircular canals. A medial projection of the saccular region is preserved as a pit in the basioccipital bone, and may have housed the scala tympani (Fig 14). The well-developed crista interfenestralis forms a solid ossified barrier between the inner ear and metotic fossa and precludes a posterior path of the re-entrant fluid circuit. The medial wall of the otic capsule is incompletely ossified, preventing detailed characterization of the paths of the facial, vestibulocochlear, and abducens nerves. 10.1371/journal.pone.0161823.g014Fig 14 Left otic endocast of Brachydectes newberryi, KUVP 49541, rendered from μCT. A, left lateral view; B, medial view; C, dorsal view. Scale equals 5 mm. Abbreviations: aasc, ampulla of the anterior semicircular canal; amhsc, ampulla of horizontal semicircular canal; asc, anterior semicircular canal; ds, dorsal sulcus; fo, surface of fenestra vestibularis; hsc, horizontal semicircular canal; lag, fossa accommodating lagena; mcv, passage of middle cerebral vein; psc, posterior semicircular canal; stym, fossa possibly accommodating the scala tympani, Discussion Affinities of Brachydectes The phylogenetic relationships of lysorophians have generally been considered in terms of three main issues: the relationship between lysorophians and modern lissamphibians, the relationship between lysorophians and other elongate-bodied lepospondyls, and the relationship between lysorophians and ‘microsaurs.’ The endocranial data presented here suggests a resolution to these problems may be attainable, although an expansive phylogenetic analysis is outside the scope of this study and will appear elsewhere. Brachydectes and lissamphibian origins Differences between the braincases of Brachydectes and of lissamphibians are conspicuous and substantive. Major differences include: the composition of the occipital arch, structure of the occipital condyles, site of insertion of the hypaxial musculature on the ventral braincase, the course of the internal carotid and its branches, the course of the perilymphatic circulation, the course of the lateral head vein, and the structure of the cartilages of the ethmoid trabeculae. The occipital arch of Brachydectes newberryi consists of paired exoccipitals, a median ventral basioccipital, and a definitive amniote supraoccipital that compares well with the morphology described for the eureptiles Captorhinus [72], Petrolacosaurus [77], and Youngina [78]. In contrast, the occipital arch of lissamphibians consists of only paired exoccipitals, which in some species invade the synotic tectum and basioccipital region partially or completely [79]. No basioccipital or supraoccipital element is present in any lissamphibian, fossil or modern. The presence of a basioccipital in Brachydectes is plesiomorphic for tetrapods and thus phylogenetically uninformative, but the presence of a well-developed median supraoccipital is restricted to the amniote crown [80] and recumbirostran ‘microsaurs’. Although the supraoccipital of Brachydectes and ‘microsaurs’ has traditionally been considered convergent with the amniote supraoccipital, new data from μCT have demonstrated that the ‘microsaur’ supraoccipital shares a number of morphological details with early amniotes, and early eureptiles in particular, and is likely homologous with the amniote element [42]. This homology does not extend far down the amniote stem, as seymouriamorphs lack a supraoccipital and ‘anthracosaurs’ generally exhibit paired elements within the synotic tectum [80]. The atlantoccipital articulation in Brachydectes has been identified as exhibiting intermediate morphology between the atlantoccipital articulation of early tetrapods and the atlantoccipital articulation of modern lissamphibians [23, 25]. However, again, the detailed morphology presented here suggests strongly that this is not the case. The atlantoccipital articulation of lissamphibians consists of paired occipital condyles that form the primary articulation between the occiput and atlas. There is no basioccipital articulation with the atlas, nor is there a proatlantal arch at any stage of development. In Brachydectes, in contrast, the basioccipital cotyle forms the primary articulation with the atlas, with shelflike processes of the exoccipital forming accessory articulations with the atlas as well as the proatlantal arch (Fig 15). Small processes of the exoccipital are found in some early amniotes [81] and serve as a site of articulation with the proatlas. Additionally, although the basal amniote atlantoccipital articulation consists of a well-developed basioccipital condyle, a pit receiving the notochord is present within the basioccipital condyle in some modern reptiles (lost in birds and lepidosaurs) as well as many basal amniotes [82]. Although the presence of a proatlas itself is plesiomorphic for tetrapods, the presence of distinct processes of the occiput that articulate with it is not found within the amniote stem (e.g. seymouriamorphs and diadectomorphs) and in no taxon with exoccipital condyles do they articulate with the proatlas in any manner. 10.1371/journal.pone.0161823.g015Fig 15 Atlantoccipital articulation of Brachydectes newberryi, and comparison with selected early tetrapods. A-B, atlantoccipital articulation of Brachydectes newberryi, KUVP 49541, in right lateral view; A, rendered from μCT; B, interpretive drawing; C, Araeoscelis gracilis, after [81], occipital view; D, Quasicaecilia texana, after [43], occipital view; E, Brachydectes newberryi, occipital view; F, Acheloma dunni, after [79], occipital view; G, Hynobius amjiensis, occipital view. A-B, scale bar equals 5 mm; C-G, not to scale. Abbreviations: atc, centrum of atlas; atn, neural arch of alas; bo, basioccipital; boc, basioccipital condyle or cotyle; eo, exoccipital; eoc, exoccipital condyle or shelf; op, opisthotic; pat, proatlas; soc, supraoccipital. The evolution of the hypaxial cervical musculature in early tetrapods has been generally overlooked, although it has been discussed in some detail by Olson [83]. The hypaxial musculature of lissamphibians inserts along the ventral surface of the otic capsules, a feature shared with many early tetrapods [83]. In contrast, the cervical musculature of amniotes inserts on the basioccipital [29]. An occipital insertion of the hypaxial musculature has been recently identified in the microsaur Quasicaecilia texana [43], and paired fossae lateral to the basioccipital cotyle of Brachydectes suggest a similar amniote-like condition of the lysorophian occiput. Although branching patterns of the internal carotids are highly variable in modern tetrapod taxa [84], the course of the common internal carotid and the location of the first cerebral branch of the internal carotid is generally conserved among early tetrapods [75, 85]. In lissamphibians, the common internal carotid enters the braincase through the basal plate of the parasphenoid or through the optic wall and then branches within the cavum cranii [86–87]. Branching of the common internal carotid into cerebral and palatal arteries outside of the skull, as in Brachydectes, is an apomorphy of the amniote stem and crown [75]. The presence of a well-developed crista interfenestralis in Brachydectes divides the otic capsule from the metotic fossa. Similar morphology of the otic region has recently been described in the microsaur Dvellecanus carrolli [42], as well as various amniotes [72, 75]. Importantly, the crista interfenestralis precludes a posterior course of the perilymph circulation, seen in all lissamphibians [1] and some temnospondyls [88], and thus suggests an anterior path of the perilymph circulation, as in amniotes. The morphology of the anterior braincase, particularly the trabecular cartilages, has been a subject of some debate in the past [60] but has not been implemented in phylogenetic analyses to date. Lissamphibians, in contrast to amniotes, have broadly-spaced ethmoid trabeculae which never meet in the midline, forming a broad, flat, fenestrate base to the anterior braincase, a condition called platytraby [60]. In amniotes, the ethmoid trabeculae are fused posterior to the nasal capsules to form a trabecula communis for part or all of their length (tropitraby), producing a narrow, keeled floor of the braincase, often well-developed into an interorbital septum. Although at first glance, the broad, flat cultriform process of the parasphenoid of Brachydectes appears platytrabic, the presence of a well-developed midline presphenoid element suggests the presence of a trabecula communis, which could indicate that Brachydectes is tropitrabic. Similar morphology has been observed in a number of recumbirostrans as well [41–43]. It is unclear whether the basal condition in tetrapods is tropitraby or platytraby [61], but we consider it important to note that, while the broad cultriform process of the parasphenoid of Brachydectes is superficially similar to that of salamanders and caecilians, this does not reflect a lissamphibian-like organization of the neurocranium itself. In summary, neurocranial morphology does not support a close relationship between Brachydectes and lissamphibians. The braincase of Brachydectes shows a number of synapomorphies associated with crown amniotes as well as recumbirostran ‘microsaurs,’ and lacks a number of neurocranial synapomorphies shared between lissamphibians and some temnospondyls. This may reflect broader phylogenetically-informative patterns in uncharacterized neurocranial variation in early tetrapods, in which case the fact that Brachydectes exhibits a braincase that is less amphibian-like and more amniote-like than the basal tetrapod condition will have to be addressed if the lepospondyl hypothesis of lissamphibian origins, as currently understood, is to remain viable. Brachydectes and other lepospondyls Morphology of the braincase of Brachydectes suggests a close relationship with the brachystelechid ‘microsaurs’ Carrolla craddocki [41] and Quasicaecilia texana [43], within the Recumbirostra. Brachydectes shares with Carrolla and Quasicaecilia a robust ossification within the columella ethmoidalis, a pillar-like pila antotica that braces against the skull roof, no participation of a descending flange of the frontal in the suture between the orbitosphenoid and the frontal, robust orbitosphenoid walls without fossae to accommodate the cerebral hemispheres or olfactory tracts, an anterior-facing fenestra antotica, passage of the trigeminal nerve through a foramen restricted to the ventral portion of the fenestra antotica, and relatively ventral position of the optic foramen. Additional cranial characteristics supporting this relationship include a reduced number of enlarged teeth in the maxilla and dentary, and a mandibular fenestra between the dentary, surangular, and angular. A close relationship between Brachydectes and brachystelechids has been suggested in the past [15–17] on the basis of similar reduction of the posterior skull and palate, as well as similarities in the axial and appendicular skeleton. In contrast, the braincase of Brachydectes shows numerous dissimilarities to the braincases of aïstopods, with which it has been allied in the past [3, 21, 31, 33]. Anatomy of the braincase has been described in the aïstopods Phlegethontia [40] and Sillerpeton permianum [89]. In aïstopods, the antotic fissure faces laterally, and the foramen serving the trigeminal nerve is laterally directed as well. Furthermore, in aïstopods, the olfactory nerve exits the braincase far ventrally [40], whereas in Brachydectes and in microsaurs, the foramina serving the olfactory nerve are floored by the median anterior braincase bone and the cultriform process of the parasphenoid [33, 41–43]. In Brachydectes and in microsaurs, the basal plate of the parasphenoid covers most of the ventral surface of the otic region and participates in the ventral margin of the fenestra vestibuli [32, 33, 42], whereas in aïstopods, the parasphenoid is restricted to the sphenoid region only, leaving the otic capsules exposed ventrally [40]. A close relationship has also been found between Brachydectes and the Nectridea [3, 21, 31, 33], but is not supported by the morphology described here. The braincase is poorly known in most nectrideans, but has been partially described in diplocaulids [58, 90, 91]. Brachydectes differs from nectrideans in having ossified orbitosphenoid, basioccipital, and supraoccipital elements (all present in other microsaurs but absent in nectrideans). Nectrideans also demonstrate participation of the exoccipital in the fenestra vestibuli and full enclosure of the metotic foramen by the exoccipital (possibly associated with invasion of the ventral otic capsule by the exoccipital) [58, 90, 91], conditions not seen in Brachydectes nor any other recumbirostran. Although no phylogenetic analysis is presented here, the overwhelming similarity between Brachydectes and brachystelechid microsaurs indicates a number of characteristics which will be incorporated into a future analysis currently in preparation. Moreover, the overwhelming dissimilarity between Brachydectes and both aïstopods and nectrideans suggests that prior skepticism towards a lysorophian-aïstopod clade (as seen in some phylogenetic analyses, such as [21]) may be well-founded, and that less-common phylogenetic results that unite lysorophians and brachystelechids (e.g. [15]) may in fact be correct. This would suggest that axial elongation was as common a trend among early tetrapods as among modern ones, possibly reflecting widespread ecological specialization. Is there evidence of miniaturization or neoteny in Brachydectes? Reconstruction of a partial growth sequence is made possible by the substantial range of sizes in the sample of specimens examined here, with the smallest skull studied exhibiting a total skull length of 10.5 mm and the largest skull representing an animal with a skull length over 30 mm in length. Most skulls are approximately 10 mm or 20 mm in length; intervening sizes are not widely represented in this sample. Ossification of all dermal bones has already been completed by the smallest studied specimen. In another lepospondyl (the aïstopod Phlegethontia) [40], a number of dermal bones ossify relatively late in development. Either this is not the case in Brachydectes, or else all specimens sampled here represent a relatively advanced stage of development, despite small size. Ossification of some endochondral elements progresses through the ontogenetic sequence sampled here. In smaller specimens (e.g. DMNH 52081) the columella ethmoidalis is completely unossified. Ossification of this cartilage is partial in larger specimens (e.g. KUVP 49541), where it forms the medial surface of the foramen housing the olfactory nerve. Completion of ossification of the columella ethmoidalis is seen in the largest specimen (UNSM 32149), in which the entire space between the olfactory foramina is completely invaded by the presphenoid. Complete ossification of the otic capsule is also ontogenetically delayed. The prootic and opisthotic are small floating elements in the smallest specimens, but ossify more completely at larger sizes, suturing to the parasphenoid, squamosal, and parietal in KUVP 49541, with many of the sutures completely obliterated in the largest specimen, UNSM 32149. In UNSM 32149, there is also a distinct crista parotica extending from the lateral wall of the prootic and opisthotic to brace against the squamosal; this structure is only weakly developed in KUVP 49541 and completely absent from smaller specimens. Two specific structures initiate ossification late in ontogeny, and are observed only in UNSM 32149. The epipterygoid is fully ossified in UNSM 32149, but is unossified in all other specimens of Brachydectes studied here. Similarly, the distal portion of the columella, including its dorsal process, is only seen in UNSM 32149. In smaller specimens, the columella is a short, weakly-developed process extending from the footplate, and in the smallest specimens, the columella is essentially absent. These observations present several implications. The first is that all but one of the Brachydectes specimens surveyed here likely represent immature specimens, with smaller specimens likely representing extremely immature animals. It is important to note that the majority of specimens of Brachydectes studied by previous authors [25–26, 28] are much smaller than UNSM 32149, and in some cases even smaller than the smallest specimens studied here. For characteristics that develop relatively late in ontogeny (the dorsal process of the columella, epipterygoid, crista parotica), descriptions focusing on juvenile material (and phylogenetic analyses relying on those descriptions) may miss important anatomical structures that were present in adult Brachydectes. A second implication is that some of the taxonomic diversity previously identified among lysorophians may instead represent ontogenetic variation. Wellstead [28] reviewed lysorophian diversity and concluded that only three taxa could be consistently identified: Brachydectes newberryi, which primarily included small specimens with a skull length of less than one centimeter, Brachydectes elongatus (= Lysorophus tricarinatus), which primarily included specimens with a skull length between one and two centimeters, and Pleuroptyx clavatus, which consisted of a few very large specimens with a skull length likely exceeding two centimeters. Wellstead [28] diagnosed Brachydectes as distinct from Pleuroptyx on the basis of a broader pair of parietals, a more robust pectoral girdle, and well-developed alae on the ribs in the latter taxon. Parietal pair width appears to be variable, even within Wellstead’s sample, and does not reliably correspond with these other characteristics. Rib alae are observed in UNSM 32149 (Fig 16C and 16D) but not smaller specimens (Fig 16A and 16B), and may be ontogenetic as well. As such, Pleuroptyx clavatus appears to represent the large adult morphology of Brachydectes newberryi and may ultimately represent a junior synonym. 10.1371/journal.pone.0161823.g016Fig 16 Rendered μCT volumes of selected elements of the axial skeleton of Brachydectes newberryi. A-B, dorsal vertebrae and ribs of a small specimen of Brachydectes newberryi, DMNH 51121, left lateral view; A, volume render; B, interpretive drawing; C, external and D, internal, view of four dorsal ribs of the largest specimen of Brachydectes newberryi, UNSM 32149. Abbreviations: al, rib alae; na, neural arch; pl, pleurocentra; r, rib. Scale bar equals 1 cm. Within Brachydectes, Wellstead [28] identified two species: B. newberryi and B. elongatus. B. newberryi is distinguished from B. elongatus on the basis of the size of the lateral mandibular fenestra, the relative width of the parietal pair, the number of presacral vertebrae, and the presence of an unforked second ceratobranchial (“epibranchial” of [28]) in the latter taxon. The lateral mandibular fenestra does appear to decrease in size during ontogeny (Fig 12), and relative width of the parietal pair is highly variable, with substantial overlap within the material surveyed by Wellstead [28] as well as the material studied here. Variation in presacral vertebral counts reported by Wellstead [28] is likely indicative of interspecific variation, but it is difficult to align this with variation in cranial morphology. Few skulls described by Wellstead [28] and no skulls studied here are articulated with a complete set of presacral vertebrae, limiting the use of vertebral counts in species identification. The morphology of the second ceratobranchial does appear to differ between B. newberryi and B. elongatus, but it is unclear whether this represents taphonomic, ontogenetic, or intraspecific variation. If Pleuroptyx clavatus and Brachydectes elongatus are to be retained as valid taxa, rather than relegated as junior synonyms of Brachydectes newberryi, then these taxa will need to be revised and consistent morphological differences between them will have to be identified. The limited ontogenetic timing data provides for some comparison with other early tetrapods. Anderson [40] has suggested that miniaturization in Phlegethontia may have been accomplished via early ossification of endochondral elements, which would have restricted later ossification of the dermal skeleton. The ossification sequence data reported here suggests that ossification in Brachydectes followed a different trajectory. Ossification of all dermal bones occurs very early in development, whereas ossification of some endochondral elements (specifically the presphenoid, otic capsule, stapedial columella, and epipterygoid) is delayed until late in skeletal maturity. Moreover, adult specimens of Brachydectes are large compared to many recumbirostrans, and are substantially larger than typical “miniaturized” taxa. We consider it unlikely that the relatively unique morphology exhibited by Brachydectes is the result of miniaturization alone. The possibility that the unique morphology of Brachydectes reflects cessation of ontogeny at an early stage of development [6, 16–17] is also not supported by the morphology we present here. Several bones found in many early tetrapods are missing from the dermal skull of Brachydectes (the postfrontal, postorbital, jugal, and quadratojugal), and the maxilla is dramatically shortened, resulting in an open postorbital and temporal region that resembles that of neotenic salamanders. However, even small specimens of Brachydectes preserve the clavicle and scapula, elements which are considered indicative of skeletal maturity in other lepospondyls [10]. Additionally, Brachydectes demonstrates extensive ossification of endochondral bone (both in the chondrocranium and postcranium) in larger individuals, in contrast to the condition observed in neotenic salamanders. We thus consider it unlikely that the gross morphology of the skull of Brachydectes indicates a heterochronic origin of lysorophian skull morphology. Fossorial adaptations of Brachydectes Previous workers [26, 28] have commented on possible fossorial adaptations of Brachydectes. The presence of distinct recumbirostran characteristics within the braincase permits a more precise discussion of how and why the morphology of Brachydectes differs from closely-related early tetrapods, and thus a better understanding of its ecology. That Brachydectes inhabited burrows at least part of the year is supported by direct evidence; skeletons of Brachydectes are readily found within burrow structures [44–46, 92]. These burrows have generally been interpreted as estivation burrows [44–46] excavated in soft subaqueous sediment, partly due to the sedimentology of these localities [44–46] and partly due to the common interpretation of Brachydectes as an aquatic gill-breathing early tetrapod. However, the skeletal evidence for fossoriality in Brachydectes has remained somewhat equivocal. Bolt and Wassersug [26] argued for a greater role of fossoriality in the ecology of Brachydectes, based on comparisons with modern amphisbaenian squamates, and specifying two general lines of anatomical evidence: the skull and palate were heavily reinforced to withstand compression and torsion stresses, and the jaw and suspensorium were modified to allow the mouth to open within enclosed spaces. Morphology identified in support of the former includes a dramatically thickened skull roof, deep sinuous sutures between skull roofing elements, and robust connections between the braincase, skull roof, and palate, whereas the latter is supported by the ventrally-recessed articular fossa and anteriorly-canted suspensorium. Bolt and Wassersug [26] also make note of the roughly wedge-shaped skull of Brachydectes, and a close relationship between the stapes and jaw articulation. They subsequently concluded that Brachydectes likely burrowed within soft sediment in freshwater environments [26]. Wellstead [28] rejected much of this argument, and argued that the morphology identified by Bolt & Wassersug [26] could be better explained by buccopharyngeal pumping (incorrectly identified by Wellstead as aquatic inertial feeding) rather than facultative or obligate fossoriality. Foremost among these characteristics is the anteriorly-canted suspensorium and ventrally recessed articular fossa, which Wellstead regarded as adaptations for increasing speed of jaw depression and gape size, respectively [28]. Structural adaptations involved in the reinforcement of the braincase were reinterpreted by Wellstead as adaptations associated with small size rather than fossoriality. If, as suggested here, Brachydectes is ultimately found within the Recumbirostra, this permits more specific and constrained comparison of lysorophian morphology with the skulls of more conservative relatives, primarily the brachystelechids Carrolla and Quasicaecilia, for which skulls have recently been described in detail [41, 43], as well as the more generalized recumbirostrans Nannaroter mckinzei [71], Huskerpeton englehorni [33], Rhynchonkos stovalli [42], Aletrimyti gaskillae [42], and Dvellecanus carrolli [42]. Recent redescriptions of recumbirostran morphology have found strong morphological support for a fossorial lifestyle throughout this group [33, 41–43, 71]. In addition to gross morphology (shovel-like snout, expanded surfaces for insertion of epaxial musculature, shortened limbs, elongate body), a number of specific characteristics have been recognized. The anterior sphenoid and posterior ethmoid region are heavily ossified into presphenoid and mesethmoid bones, bracing the nasal region against dorsoventral compressive forces in several recumbirostrans [41–43]. The orbitosphenoids are also heavily ossified, and form an overlapping articulation with the skull roof (frontals and sometimes parietals) and enlarged cultriform process of the parasphenoid, in a ‘cranial box’ arrangement reminiscent of modern fossorial squamates [33, 41–43, 71]. The endochondral bones of the otoccipital region are also co-ossified in some forms [42–43], even incorporating the basisphenoid and parasphenoid in some taxa [41, 43]. Anterior displacement of the lower jaw permits jaw opening within the confines of a burrow [42], and is accomplished despite inferior mechanical advantage by an enlarged m. depressor mandibulae as indicated by the greatly enlarged retroarticular process seen in many recumbirostrans [42]. Similarly, jaw closure under inferior mechanical advantage would have been accomplished by expanded internal and external mm. adductores mandibulae, which would have been accommodated by the enlarged temporal emarginations seen in some recumbirostran taxa [33, 42, 70–71]. Conspicuous interdigitation and scarf joints within the skull roof would have served to resist deformation of the skull roof under compressive or torsional stress [41–42, 71]. Although the morphology of Brachydectes is consistent with these more general patterns, it differs from other recumbirostrans in a few key regards. In dorsal aspect, the skull is much longer and narrower than any other recumbirostran, especially in the temporal/otic region. This is accomplished by completely adpressing the squamosal and quadrate against the lateral wall of the otic capsule. The cheek is completely open from the orbit to the suspensorium, accommodating a greatly-enlarged m. adductor mandibulae externus (mame). An emarginated cheek is observed in various microsaurs, including hapsidopareiontids [32, 68–69], ostodolepids [32, 68, 71], Tambaroter carrolli [70], and Huskerpeton englehorni [33], but the condition seen in Brachydectes newberryi is exaggerated in comparison with other microsaurs, and extends dorsally to the parietals. This is partially accomplished via the loss of the quadratojugal, jugal, postorbital, and postfrontal, and via the modification of the squamosal into a thin strap restricted to the lateral surface of the quadrate. The cultriform process of Brachydectes is also laterally expanded in comparison with other microsaurs, including brachystelechids, making the orbitosphenoids and pleurosphenoids vertical rather than angled dorsolaterally. The bones of the skull roof (particularly the frontals, parietals, and postparietals) are thickened and strongly interdigitated, and articulate tightly with much of the dorsal portion of the braincase, suggesting further refinement of the ‘cranial box’ arrangement of more generalized recumbirostrans. Morphology of the braincase in Brachydectes suggests greater resistance to dorsoventral forces on the skull centered on the frontals and parietals and a reduction of the role of a rostral scoop in excavation, suggesting a different mode of burrowing from shovel-snouted recumbirostrans such as Pelodosotis [32] or Nannaroter [71] as well as the four-step excavation cycle recently proposed for Quasicaecilia texana [43]. Skull morphology in modern fossorial squamates corresponds closely with burrowing mode [93, 94] and substrate type [95], permitting inference of burrowing mode and substrate type in extinct fossorial taxa such as Brachydectes. In gymnophthalmids and amphisbaenids, a steeply-inclined shovel-like skull is typically associated with burrowing in sandy substrate, whereas a lower, bullet-shaped skull is associated with burrowing in denser soil [95]. Bracing of the skull for dorsoventral compression is associated specifically with shovel-type burrowing modes [93, 96]. Axial elongation is also typical in amphibaenians employing shovel-type burrowing, permitting a greater total mass of the epaxial musculature associated with head lifting without also increasing the cross-sectional area of the body [97]. We consider the morphology of Brachydectes to be consistent with shovel-type headfirst burrowing in consolidated sediment. This both clarifies and complicates interpretations of Brachydectes burrowing traces. Burrows containing lysorophian skeletons have been known for some time [27, 44–46, 92, 98]), and have been regularly cited as evidence of aestivation in this taxon. These burrows preserve irregularly-spaced nodes consistent with headfirst burrowing (specifically dorsad adpression), corroborating the osteology-based inference of burrowing mode. Although the burrows are traditionally interpreted as simple subaqueous excavations [45–45, 92], several features of these burrows are more consistent with excavations above the water table. Distinct node structures in the walls of lysorophian burrows are consistent with their excavation in consolidated sediment [45–46] and the presence of multiple individuals in some burrows [92] may support alternate interpretations of these burrow structures. Lysorophian burrows are also sedimentologically distinct from lungfish burrows, which preserve a mudstone-and-organic ‘shell’ surrounding the estivation chamber [45] and are found primarily on the edge of fossil ponds rather than within the center of them [46]. We do not argue here for a specific reinterpretation of lysorophian burrow traces, but rather caution against strong inference of lysorophian ecology and physiology based on prior interpretations of lysorophian burrowing behaviors. Supporting Information S1 Table μCT scan parameters for Brachydectes newberryi specimens used in this study. Abbreviations: DMNH, Denver Museum of Nature and Science (Denver, Colorado, USA); KUVP, University of Kansas Natural History Museum and Biodiversity Institute, Vertebrate Paleontology Division (Lawrence, Kansas, USA); UNSM, University of Nebraska State Museum (Lincoln, Nebraska, USA). (XLSX) Click here for additional data file. ==== Refs References 1 Maddin HC , Anderson JS (2012 ) Evolution of the amphibian ear with implications for lissamphibian phylogeny: insight gained from the caecilian inner ear . Fieldiana Life and Earth Sciences 5 : 56 –76 . 2 Maddin HC , Jenkins FA Jr, Anderson JS (2012 ) The braincase of Eocaecilia micropodia (Lissamphibia, Gymnophiona) and the origin of caecilians . PLOS ONE 7 (12 ) e50743 10.1371/journal.pone.0050743 23227204 3 Anderson JS , Reisz RR , Scott D , Fröbisch NB , Sumida SS (2008 ) A stem batrachian from the Early Permian of Texas and the origin of frogs and salamanders . Nature 453 : 515 –518 . 10.1038/nature06865 18497824 4 Ruta M , Coates MI (2007 ) Dates, nodes and character conflict: addressing the lissamphibian origin problem . J Syst Paleontol 5 : 69 –122 . 5 Sigurdsen T , Green DM (2011 ) The origin of modern amphibians: a re-evaluation . Zool J Linn Soc 162 : 457 –469 . 6 Schoch RR , Milner AR (2004 ) Structure and implications of theories of the origin of lissamphibians In: Arratia G , Wilson MVH , Cloutier R , editors. Recent Advances in the Origin and Early Radiation of Vertebrates . München : Verlag Dr. Friedrich Pfeil pp. 345 –377 . 7 Bolt JR (1969 ) Lissamphibian origins: possible protolisssamphibian from the Lower Permian of Oklahoma . Science 166 : 888 –891 . 17815754 8 Schoch RR , Carroll RL (2003 ) Ontogenetic evidence for the Paleozoic ancestry of salamanders . Evol & Devel 5 : 314 –324 . 9 Fröbisch NB , Carroll RL , Schoch RR (2007 ) Limb ossification in the Paleozoic branchiosaurid Apateon (Temnospondyli) and the early evolution of preaxial dominance in tetrapod limb development . Evol & Devel 9 : 69 –75 . 10 Olori JC (2013 ) Ontogenetic sequence reconstruction and sequence polymorphism in extinct taxa: an example using early tetrapods (Tetrapoda: Lepospondyli) . Paleobiol 39 : 400 –428 . 11 Skutchas P , Martin T (2011 ) Cranial anatomy of the stem salamander Kokartus honorarius (Amphibia: Caudata) from the Middle Jurassic of Kyrgyzstan . Zool J Linn Soc 161 : 816 –838 . 12 Rage JC , Roček Z (1989 ) Redescription of Triadobatrachus massinoti (Piveteau, 1936) an anuran amphibian from the Early Triassic . Palaeontographica Abt. A 206 : 1 –16 . 13 Shubin NH , Jenkins FA Jr (1995 ) An Early Jurassic jumping frog . Nature 377 : 49 –52 . 14 Jenkins FA Jr, Walsh DM , Carroll RL (2007 ) Anatomy of Eocaecilia micropodia, a limbed caecilian from the Early Jurassic . Bull Mus Comp Zool 158 : 285 –365 . 15 Vallin G , Laurin M (2004 ) Cranial morphology and affinities of Microbrachis, and a reappraisal of the phylogeny and lifestyle of the first amphibians . J Vertebr Paleontol 24 : 56 –72 . 16 Marjanović D , Laurin M (2008 ) A reevaluation of the evidence supporting an unorthodox hypothesis on the origin of amphibians . Contr Zool 77 : 149 –199 . 17 Marjanović D , Laurin M (2013 ) The origin(s) of extant amphibians: a review with emphasis on the “lepospondyl hypothesis” . Geodiversitas 35 : 207 –272 . 18 Carroll RL , Currie PJ (1975 ) Microsaurs as possible apodan ancestors . Zool J Linn Soc 57 : 229 –247 . 19 Nussbaum RA (1983 ) The evolution of a unique dual jaw-closing mechanism in caecilians: (Amphibia: Gymnophiona) and its bearing on caecilian ancestry . J Zool 199 : 545 –554 . 20 Carroll RL (2007 ) The Palaeozoic ancestry of salamanders, frogs and caecilians . Zool J Linn Soc 150 (S1 ):1 –140 . 21 Anderson JS (2007 ) Incorporating ontogeny into the matrix: a phylogenetic evaluation of developmental evidence for the origin of modern amphibians In: Anderson JS , Sues H-D , editors. Major Transitions in Vertebrate Evolution . Indiana : Indiana University Press pp. 182 –227 . 22 Case EC (1908 ) Notes on the skull of Lysorophus tricarinatus Cope, 1908 . Bull. Am Mus Nat Hist 24 : 531 –533 . 23 Williston SW (1908 ) Lysorophus, a Permian urodele . Biol Bull 15 : 229 –240 . 24 Moodie RL (1909 ) Vertebrate paleontology . Am Nat 43 : 116 –124 . 25 Sollas WJ (1920 ) XI.—On the structure of Lysorophus, as exposed by serial sections . Phil Trans R Soc London Series B 209 : 481 –527 . 26 Bolt JR , Wassersug RJ (1975 ) Functional morphology of the skull in Lysorophus: a snake-like Paleozoic amphibian (Lepospondyli) . Paleobiol 1 : 320 –332 . 27 Olson EC (1971 ) A skeleton of Lysorophus tricarinatus (Amphibia: Lepospondyli) from the Hennessey Formation (Permian) of Oklahoma . J Paleontol 45 : 443 –449 . 28 Wellstead CF (1991 ) Taxonomic revision of the Lysorophia, Permo-Carboniferous lepospondyl amphibians . Bull Amer Mus Nat Hist 209 : 1 –90 . 29 Laurin M , Reisz RR (1995 ) A reevaluation of early amniote phylogeny . Zool J Linn Soc 113 : 165 –223 . 30 Kleinteich T , Maddin HC , Herzen J , Beckman F , Summers AP (2012 ) Is solid always best? Cranial performance in solid and fenestrated caecilian skulls . J Exper Biol 215 : 833 –844 .22323206 31 Anderson JS (2001 ) The phylogenetic trunk: maximal inclusion of taxa with missing data in an analysis of the Lepospondyli (Vertebrata, Tetrapoda) . Syst Biol 50 : 170 –193 . 12116927 32 Carroll RL , Gaskill P (1978 ) The order Microsauria . Memoirs of the American Philosophical Society 126 : 1 –211 . 33 Huttenlocker AK , Pardo JD , Small BJ , Anderson JS (2013 ) Cranial morphology of recumbirostrans (Lepospondyli) from the Permian of Kansas and Nebraska, and early morphological evolution inferred by micro-computed tomography . J Vertebr Paleontol 33 : 540 –552 . 34 Ruta M , Jeffery JE , Coates MI (2003 ) A supertree of early tetrapods . Proc R Soc B 270 : 2507 –2516 . 14667343 35 Friedman M (2007 ) The interrelationships of Devonian lungfishes (Sarcopterygii: Dipnoi) as inferred from neurocranial evidence and new data from the genus Soederberghia Lehman, 1959 . Zool J Linn Soc 151 : 115 –171 . 36 Pardo JD , Huttenlocker AK , Small BJ (2014 ) An exceptionally preserved transitional lungfish from the Lower Permian of Nebraska, USA, and the origin of modern lungfishes . PLOS ONE 9 (9 ): e108542 10.1371/journal.pone.0108542 25265394 37 Maddin HC , Russell AP , Anderson JS (2012 ) Phylogenetic implications of the morphology of the braincase of caecilian amphibians (Gymnophiona) . Zool J Linn Soc 166 : 160 –201 . 38 Maddin HC (2011 ) Deciphering morphological variation in the braincase of caecilian amphibians (Gymnophiona) . J Morph 272 : 850 –871 . 10.1002/jmor.10953 21538474 39 Gold MEL , Brochu CA , Norell MA (2014 ) An expanded combined evidence approach to the Gavialis problem using geometric morphometric data from crocodilian braincases and eustachian systems . PLOS ONE 9 (9 ): e106793 .25269021 40 Anderson JS (2002 ) Revision of the aïstopod genus Phlegethontia (Tetrapoda: Lepospondyli) . J Paleontol 76 : 1029 –1046 . 41 Maddin HC , Olori JC , Anderson JS (2011 ) A redescription of Carrolla craddocki (Lepospondyli: Brachystelechidae) based on high-resolution CT, and the impacts of miniaturization and fossoriality on morphology . J Morph 272 : 722 –743 . 10.1002/jmor.10946 21484854 42 Szostakiwskyj M , Pardo JD , Anderson JS (2015 ) Micro-CT study of Rhynchonkos stovalli (Lepospondyli, Recumbirostra), with description of two new genera . PLOS ONE 10 (6 ): e0127307 10.1371/journal.pone.0127307 26061187 43 Pardo JD , Szostakiwskyj M , Anderson JS (2015 ) Cranial morphology of the brachystelechid ‘microsaur’ Quasicaecilia texana Carroll provides new insights into the diversity and evolution of braincase morphology in recumbirostran ‘microsaurs.’ PLOS ONE 10 (6 ): e0130359 10.1371/journal.pone.0130359 26107260 44 Huttenlocker AK , Pardo JD , Small BJ (2005 ) An earliest Permian nonmarine vertebrate assemblage from the Eskridge Formation, Nebraska . New Mexico Museum of Natural History and Science Bulletin 30 : 133 –143 . 45 Hembree DI , Hasiotis ST , Martin LD (2005 ) Torridorefugium eskridgensis (new ichnogenus and ichnospecies): amphibian aestivation burrows from the Lower Permian Speiser Shale of Kansas . J Paleontol 79 : 583 –593 . 46 Hembree DI , Martin LD , Hasiotis ST (2004 ) Amphibian burrows and ephemeral ponds of the Lower Permian Speiser Shale, Kansas: evidence for seasonality in the midcontinent . Palaeogeog Palaeoclim Palaeoeco 203 : 127 –152 . 47 Sawin RS , West RR , Franseen EK (2006 ) Carboniferous-Permian boundary in Kansas, Midcontinent, U.S.A . Curr Res Earth Sci, Bull 252 :1 –13 . 48 Boardman DR II, Wardlaw BR , Nestell MK (2009 ) Stratigraphy and conodont biostratigraphy of the Uppermost Carboniferous and Lower Permian from the North American Midcontinent . Kansas Geol Surv Bull 255 : 1 –145 . 49 Joeckel RM (1991 ) Paleosol stratigraphy of the Eskridge Formation: Early Permian pedogenesis and climate in southeastern Nebraska . J Sediment Petr 59 : 469 –481 . 50 McAllister JA (1992) Gnathorhiza (Dipnoi): life aspects, and lungfish burrows, pp.91-105. In Mark-Kurik E, Fossil Fishes as Living Animals. Akademia I, Tallinn. 51 Hotton N (1959 ) Acroplous vorax, a new and unusual labyrinthodont amphibian from the Kansas Permian . J Paleontol 33 : 161 –178 . 52 Coldiron RW (1978 ) Acroplous vorax Hotton (Amphibia, Saurerpetontidae) restudied in light of new material . Amer Mus Nov 2662 : 1 –27 . 53 Foreman BC (1990 ) A revision of the cranial morphology of the Lower Permian temnospondyl amphibian Acroplous vorax Hotton . J Vertebr Paleontol 10 : 390 –397 . 54 Englehorn J , Small BJ , Huttenlocker AK (2008 ) A redescription of Acroplous vorax (Temnospondyli: Dvinosauria) based on new specimens from the Early Permian of Nebraska and Kansas, U.S.A . J Vertebr Paleontol 28 : 291 –305 . 55 Huttenlocker AK , Pardo JD , Small BJ (2007 ) Plemmyradytes shintoni, get. et sp. nov., an Early Permian amphibamid (Temnospondyli: Dissorophoidea) from the Eskridge Formation, Nebraska . J Vertebr Paleontol 27 : 316 –328 . 56 Foreman BC , Schultze H-P (1981 ) A new gymnarthrid microsaur from the Lower Permian of Kansas with a review of the tuditanomorph microsaurs (Amphibia) . Occ Papers Mus Nat Hist Univ Kansas 91 : 1 –25 . 57 Foreman BC , Martin LD (1988 ) A review of Paleozoic tetrapod localities of Kansas and Nebraska . Kans Geol Surv Guide Ser 6 : 133 –145 . 58 Gubin YM (1999 ) A new diplocaulid (Nectridea, Amphibia) from the Early Permian of Kansas (USA) . Paleontol Journ 33 : 630 –637 . 59 McCahon TJ , Miller KB (2015 ) Environmental significance of lungfish burrows (Gnathorhiza) within Lower Permian (Wolfcampian) paleosols of the US midcontinent . Palaeogeog Palaeoclim Palaeoecol 435 : 1 –12 . 60 de Beer SG (1937 ) The development of the vertebrate skull Oxford : Clarendon Press , 522 p. 61 Clack JA (1994 ) Earliest known tetrapod braincase and the evolution of the stapes and fenestra ovalis . Nature 369 : 392 –394 . 62 Clack JA , Holmes R (1988 ) The braincase of Archeria crassidisca with comments on the interrelationships of primitive tetrapods . Palaeontol 31 : 85 –107 . 63 Dempster WT (1935 ) The brain case and endocranial cast of Eryops megacephalus (Cope) . J Comp Neur 62 : 171 –196 . 64 Cope ED (1875 ) Synopsis of the extinct Batrachia from the Coal Measures . Ohio Geol Surv 2 : 349 –411 . 65 Broili F (1908 ) Systematische und biologische Bemerkungen zu der permischen Gattung Lysorophus . Anat Anzeig 33 : 290 –298 . 66 Cope ED (1868 ) Synopsis of the extinct Batrachia of North America . Proc Acad Nat Sci 20 : 208 –221 . 67 Hook RW , Baird D (1986 ) The Diamond Coal Mine of Linton, Ohio, and its Pennsylvanian-age vertebrates . J Vertebr Paleontol 6 : 174 –190 . 68 Daly E (1973 ) A Lower Permian vertebrate fauna from southern Oklahoma . J Paleontol 47 : 562 –589 . 69 Bolt JR , Rieppel O (2009 ) The holotype skull of Llistrofus pricei Carroll and Gaskill, 1978 (Microsauria: Hapsidopareiontidae) . J Paleontol 83 : 471 –483 . 70 Henrici AC , Martens T , Berman DS , Sumida SS (2011 ) An ostodolepid ‘microsaur’ (Lepospondyli) from the Lower Permian Tambach Formation of central Germany . J Vert Paleontol 31 : 997 –1004 . 71 Anderson JS , Scott D , Reisz RR (2009 ) Nannaroter mckinziei, a new ostodolepid ‘microsaur’ (Tetrapoda, Lepospondyli, Recumbirostra) from the Early Permian of Richards Spur (Ft. Sill), Oklahoma . J Vert Paleontol 29 : 379 –388 . 72 Heaton MJ (1979 ) Cranial anatomy of primitive captorhinid reptiles from the Late Pennsylvanian and Early Permian of Oklahoma and Texas . Ok Geol Sur Bul 127 : 1 –84 . 73 Yates AM , Warren AA (2000 ) The phylogeny of the ‘higher’ temnospondyls (Vertebrata: Choanata) and its implications for the monophyly and origins of the Stereospondyli . Zool J Linn Soc 128 : 77 –121 . 74 Francis ETB (1934 ) The Anatomy of the Salamander . Ithaca : Society for the Study of Amphibians and Reptiles , 465 pp. 75 Müller J , Sterli J , Anquetin J (2011 ) Carotid circulation in amniotes and its implications for turtle relationships . N J Geol Palӓont Abh 261 : 289 –297 . 76 Romer AS (1969 ) The cranial anatomy of the Permian amphibian Pantylus . Breviora 314 : 1 –37 . 77 Reisz RR (1981 ) A diapsid reptile from the Pennsylvanian of Kansas . Univ Kans Mus Nat Hist Spec Publ 7 : 1 –74 . 78 Gardner NM , Holliday CM , O’Keefe FR (2010 ) The braincase of Youngina capensis (Reptilia, Diapsida): new insights from high-resolution CT scanning of the holotype . Palaeontol Electr 13 ; 19A : 1 –16 . 79 Maddin HC , Reisz RR , Anderson JS (2010 ) Evolutionary development of the neurocranium in Dissorophoidea (Tetrapoda: Temnospondyli) an integrative approach . Evol & Devel 12 : 393 –403 . 80 Berman DS (2000 ) Origin and early evolution of the amniote occiput . J Paleontol 74 : 938 –956 . 81 Vaughn PP (1955 ) The Permian reptile Araeoscelis restudied . Bull Mus Comp Zool 114 : 36 –89 . 82 Rieppel O (2000 ) Turtles as diapsid reptiles . Zool Scr 29 : 199 –213 . 83 Olson EC (1936 ) The dorsal axial musculature of certain primitive Permian tetrapods . J Morph 59 : 265 –311 . 84 Jamniczky HA , Hallgrímsson B (2011 ) Modularity in the skull and cranial vasculature of laboratory mice: implications for the evolution of complex phenotypes . Evol & Devel 13 : 28 –37 . 85 Shishkin MA (1968 ) On the cranial arterial system of the labyrinthodonts . Acta Zool 49 : 1 –22 . 86 Paterson NF (1939 ) The head of Xenopus laevis . J Cell Sci s2 -81 : 161 –232 . 87 Fox H (1954 ) Development of the skull and associated structures in the Amphibia with special reference to the urodeles . Trans Zool Soc London 28 : 241 –304 . 88 Sigurdsen T (2008 ) The otic region of Doleserpeton (Temnospondyli) and its implications for the evolutionary origins of frogs . Zool J Linn Soc 154 : 738 –751 . 89 McGinnis HJ (1967 ) The osteology of Plegethontia, a Carboniferous and Permian aistopod amphibian . Univ Calif Publ Geol Sci 71 : 1 –46 . 90 Beerbower JB (1963 ) Morphology, paleoecology, and phylogeny of the Permo-Pennsylvanian amphibian Diploceraspis . Bull Mus Comp Zool Harvard Univ 130 : 31 –108 . 91 Germain D (2010 ) The Morroccan diplocaulid: the last lepospondyl, the single one on Gondwana . Hist Biol 22 : 4 –39 . 92 Olson EC , Bolles K (1975 ) Permo-Carboniferous fresh water burrows . Fieldiana Geology 33 : 271 –290 . 93 Kearney M (2003 ) Systematics of the Amphisbaenia (Lepidosauria: Squamata) based on morphological evidence from recent and fossil forms . Herpetol Monogr 17 : 1 –74 . 94 Kearney M , Stuart BL (2004 ) Repeated evolution of limblessness and digging heads in worm lizards revealed by DNA from old bones . Proc R Soc Lond B 271 : 1677 –1683 . 95 Barros FC , Herrel A , Kohlsdorf T (2011 ) Head shape in Gymnophthalmidae: does habitat use constrain evolution of cranial design in fossorial lizards? J Evol Biol 24 : 2423 –2433 . 10.1111/j.1420-9101.2011.02372.x 21883615 96 Gans C (1968 ) Relative success of divergent pathways in amphisbaenian specialization . Am Nat 102 : 345 –362 . 97 Navás CA , Antoniazzi MM , Carvalho JE , Chaui-Berlink JG , James RS , Jared C , et al (2004 ) Morphological and physiological specialization for digging in amphisbaenians, an ancient lineage of fossorial vertebrates . J Exp Biol 207 : 2433 –2441 . 15184515 98 Olson EC (1956 ) Fauna of the Vale and Choza: 11, Lysorophus: Vale and Choza, Diplocaulus, Cacops and Eryopidae: Choza . Fieldiana Geol 10 : 313 –322 .

PMC005xxxxxx/PMC5001629.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756408410.1371/journal.pone.0161884PONE-D-16-11001Research ArticleMedicine and Health SciencesTropical DiseasesNeglected Tropical DiseasesDengue FeverMedicine and Health SciencesInfectious DiseasesViral DiseasesDengue FeverMedicine and Health SciencesDiagnostic MedicineSigns and SymptomsHemorrhageMedicine and Health SciencesPathology and Laboratory MedicineSigns and SymptomsHemorrhageMedicine and Health SciencesVascular MedicineHemorrhagePeople and PlacesDemographyDeath RatesBiology and Life SciencesPopulation BiologyPopulation MetricsDeath RatesPeople and placesGeographical locationsSouth AmericaBrazilBiology and Life SciencesAnatomyBody FluidsBloodPlateletsMedicine and Health SciencesAnatomyBody FluidsBloodPlateletsBiology and Life SciencesPhysiologyBody FluidsBloodPlateletsMedicine and Health SciencesPhysiologyBody FluidsBloodPlateletsMedicine and Health SciencesHematologyBloodPlateletsBiology and Life SciencesCell BiologyCellular TypesAnimal CellsBlood CellsPlateletsMedicine and Health SciencesHematologyThrombocytopeniaMedicine and Health SciencesEpidemiologyDisease SurveillanceMedicine and Health SciencesDiagnostic MedicineClinical Laboratory SciencesClinical LaboratoriesMortality Predictors in Patients with Severe Dengue in the State of Amazonas, Brazil Mortality Predictors of Severe Dengue in Amazonas, BrazilPinto Rosemary Costa 1de Castro Daniel Barros 12de Albuquerque Bernardino Cláudio 1Sampaio Vanderson de Souza 1dos Passos Ricardo Augusto 13da Costa Cristiano Fernandes 1Sadahiro Megumi 1Braga José Ueleres 245*1 Health Surveillance Foundation of Amazonas State (Fundação de Vigilância em Saúde do Amazonas, FVS), Manaus, Brazil2 Sérgio Arouca National School of Public Health (Escola Nacional de Saúde Pública Sérgio Arouca), FIOCRUZ, Rio de Janeiro, Brazil3 Laboratory of Physiology and Control of Arthropod Vectors (Laboratório de Fisiologia e Controle de Artrópodes Vetores), Instituto Oswaldo Cruz, FIOCRUZ, Rio de Janeiro, Brazil4 Institute of Social Medicine (Instituto de Medicina Social), Rio de Janeiro State University (Universidade do Estado do Rio de Janeiro, UERJ), Rio de Janeiro, Brazil5 PECTI-SAÚDE/Research Foundation of the State of Amazonas (Fundação de Amparo à Pesquisa do Estado do Amazonas, FAPEAM), Manaus, BrazilCox Dermot EditorRoyal College of Surgeons in Ireland, IRELANDCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: RCP BCA MS JUB. Data curation: DBC VSS RAP CFC MS JUB. Formal analysis: DBC VSS MS JUB. Funding acquisition: RCP BCA MS JUB. Investigation: DBC VSS MS JUB. Methodology: RCP DBC BCA VSS RAP CFC MS JUB. Project administration: RCP BCA MS JUB. Resources: RCP BCA MS JUB. Software: DBC VSS RAP CFC MS JUB. Supervision: RCP BCA MS JUB. Validation: DBC VSS MS JUB. Visualization: DBC VSS MS JUB. Writing – original draft: RCP DBC BCA VSS RAP CFC MS JUB. Writing – review & editing: RCP DBC BCA VSS RAP CFC MS JUB. * E-mail: ueleres@gmail.com26 8 2016 2016 11 8 e016188416 3 2016 12 8 2016 © 2016 Pinto et al2016Pinto et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Dengue is a major public health problem in tropical and subtropical areas worldwide. There is a lack of information on the risk factors for death due to severe dengue fever in developing countries, including Brazil where the state of Amazonas is located. This knowledge is important for decision making and the implementation of effective measures for patient care. This study aimed to identify factors associated with death among patients with severe dengue, in Amazonas from 2001 to 2013. We conducted a retrospective cohort study based on secondary data from the epidemiological surveillance of dengue provided by the Fundação de Vigilância em Saúde do Amazonas, FVS (Health Surveillance Foundation) of the Secretaria de Saúde do Amazonas, SUSAM (Health Secretariat of the State of Amazonas). Data on dengue cases were obtained from the SINAN (Notifiable Diseases Information System) and SIM (Mortality Information System) databases. We selected cases of severe dengue with laboratory confirmation, including dengue-related deaths of residents in the state of Amazonas from January 1, 2001, to December 31, 2013. The explanatory variables analyzed were sex, age, level of education, spontaneous hemorrhagic manifestations, plasma extravasation and platelet count. Patients who died due to severe dengue had more hematuria, gastrointestinal bleeding, and thrombocytopenia than the survivors. Considering the simultaneous effects of demographic and clinical characteristics with a multiple logistic regression model, it was observed that the factors associated with death were age >55 years (odds ratio [OR] 4.98), gastrointestinal bleeding (OR 10.26), hematuria (OR 5.07), and thrombocytopenia (OR 2.55). Gastrointestinal bleeding was the clinical sign most strongly associated with death, followed by hematuria and age >55 years. The study results showed that the best predictor of death from severe dengue is based on the characteristic of age >55 years, together with the clinical signs of gastrointestinal bleeding, hematuria, and low platelet count. Fundação de Amparo à Pesquisa do Estado do Amazonas (BR)Braga Jose Ueleres http://dx.doi.org/10.13039/501100004916Fundação de Amparo à Pesquisa do Estado do Amazonasde Castro Daniel Barros This work was supported by the government of the state of Amazonas, through the Fundação de Amparo à Pesquisa do Estado do Amazonas, FAPEAM (Research Foundation of the State of Amazonas), for granting scholarships by the RH-Doctoral Program (DBC has a fellowship) and publishing support by Journal Articles Publishing Support Program (PAPAC number 015/2014). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityThe data used in this study are not readily available because the researcher agreed to only use this data to produce the results of the study. The authors signed a data use of terms commitment pledging to maintain confidentiality and anonymity of participants as recommended by Resolution 466/2012 of the Council National Health of the Brazilian Government. More details contact at e-mail: ueleres@gmail.com.Data Availability The data used in this study are not readily available because the researcher agreed to only use this data to produce the results of the study. The authors signed a data use of terms commitment pledging to maintain confidentiality and anonymity of participants as recommended by Resolution 466/2012 of the Council National Health of the Brazilian Government. More details contact at e-mail: ueleres@gmail.com. ==== Body Introduction Dengue is an arbovirus recognized as a growing public health problem owing to its wide geographic dispersion and high incidence in various countries [1]. It is estimated that 390 million cases occur annually, of which 96 million display signs of severity [2], causing nearly 20,000 deaths in developing countries [3]. In Brazil, in 2014, there were 591,000 probable cases of dengue and 410 deaths from the disease [4]. In that year, 3,803 cases were confirmed and 9 deaths were reported to be caused by dengue in the state of Amazonas, where 1,658 cases were confirmed and 6 deaths occurred in its capital, Manaus [5]. In the last decade, the mortality and hospitalization rates due to dengue increased in several regions of the country [6,7]. It is noteworthy to highlight that the first dengue epidemic in the state of Amazonas occurred between the years 1998 and 1999 [8]. Since the introduction of the serotype DEN-2 in 2001, there has been an increase in the recording of severe cases and deaths from the disease [9,10]. In 2011, when the worst epidemic in the state occurred, the circulation of four serotypes was observed, with 57,805 confirmed cases and 23 deaths [11]. To reduce the mortality rate of dengue, the World Health Organization and the Ministry of Health proposed a classification of dengue cases based on a set of indicative clinical and laboratory characteristics of severity, thus allowing the recommendation of the best treatment for each situation [12,13]. While this classification might be useful for the clinical management of patients, it is known that such predictors of severity may vary according to different epidemiological settings [14,15]. In addition, the evolution to death of severe cases may be related to the presence of clinical and laboratory characteristics that are still not fully understood [16]. Considering the lack of information on the predictors of death due to dengue in the state of Amazonas and the importance of this information for decision making and the implementation of effective measures for patient care, this study aimed to identify factors associated with the occurrence of death among patients who presented with severe dengue in the Amazonas. Materials and Methods The state of Amazonas is located in the northern region of Brazil and has a total area of 1,559,161 km2. It is formed by 62 municipalities and 9 health regions. In 2010, the state had 3,480,937 inhabitants, of whom 1,802,525 resided in the capital city of Manaus [17]. Data on dengue cases were obtained from the SINAN (Notifiable Diseases Information System) database. Information about mortality was obtained from the SINAN and SIM (Mortality Information System) databases, provided by the FVS (Health Surveillance Foundation) of the SUSAM (Health Secretariat of the State of Amazonas). A retrospective cohort study was conducted by using secondary data from the epidemiologic surveillance of dengue in the state of Amazonas. This study was approved by the Institutional Review Board of the Adriano Jorge Foundation Hospital, under protocol number 1162956 (CAAE no. 47148715.4.0000.0007), on July 7, 2015. This institutional review board waived the need for written informed consent from participants as the study involved only secondary data and the confidentiality of the patients’ identities was protected. Cases of severe dengue with laboratory confirmation were selected (positive test in IgM serology, NS1 rapid test or enzyme-linked immunosorbent assay, virus isolation, polymerase chain reaction, or immunohistochemistry) and dengue-related deaths of residents in the state of Amazonas, regardless of sex or age, were recorded during the period from January 1, 2001, to December 31, 2013. Any duplicate records or incomplete data were excluded from the analysis. We considered as severe dengue those confirmed cases recorded by SINAN presenting with the “final classification” variable filled out as “dengue with complications—DC”, “dengue hemorrhagic fever—DHF”, or “dengue shock syndrome—DSS”. In order to deal with dengue cases that do not fill WHO criteria for DHF and are not classical dengue as well, Ministry of Health (MoH) adopted the “dengue with complications” classification that is characterized by presenting at least one of the following clinical and laboratory changes: cardio-respiratory dysfunction, liver failure, gastrointestinal bleeding, neurological abnormalities, a leukocyte count equal to or less than 1,000 cells/ml, a platelet count less than 20,000 cells/mm3, pleural effusion, pericardial effusion or ascites. DHF is characterized by the following manifestations: high fever, hemorrhagic phenomena, thrombocytopenia, and plasma leakage. The DSS cases are characterized by all of the four criteria for DHF, plus evidence of circulatory failure manifestation. The classification of severe cases followed the criteria established by the Ministry of Health during the study period [13]. Comparing the WHO 2009 guidelines and MoH classification of “dengue with complications”, we can see that the clinical changes pointed out could be accepted as severe organ involvement. So we decided to exclude the patients classified as “dengue with complications” that present exclusively laboratorial findings like “leukocyte count equal to or less than 1,000 cells/ml” or “platelet count less than 20,000 cells/mm3”. This way, severe dengue definition used here met WHO 2009 criteria. Cases considered to be deaths due to dengue were those reported to SINAN database with the “progress” variable filled out as “death due to dengue,” or from SIM database records where the underlying cause of death was filled out with the code “A90” or “A91,” according to the 10th International Classification of Diseases (ICD-10). So, dengue deaths were considered in two situations: (i) cases reported in SINAN as deaths due to dengue (n = 60) and (ii) cases registered in SIM with dengue as the underlying cause of death, but no death reported in SINAN (n = 1). It was not considered as dengue deaths: (a) SIM records that had dengue as the underlying cause of death, but not notified as dengue case in SINAN (n = 7) and (b) dengue cases reported in SINAN who died, but not due to dengue (n = 3). The explanatory variables analyzed were sex, age, level of education, occurrence of spontaneous hemorrhagic manifestations (epistaxis, gingivorrhagia, petechiae, hematuria, gastrointestinal bleeding), platelet count (lower platelet count during hospitalization), and plasma extravasation, evidenced by the occurrence of at least one of the following signals: hemoconcentration (increase by 20% from baseline hematocrit at the time of patient admission), body cavity effusion (detected by physical examination or radiography), or hypoproteinemia (serum albumin lower than 3.0 mg/dL). These signs, symptoms, laboratory findings and therapeutic procedures were written down in the reporting and investigation forms of dengue cases according to clinical management guidelines of the Brazilian MoH [13]. Prior exploratory analysis examined the relationship between predictors and possible dengue death outcome by univariate approach. A multiple logistic regression model was used to identify independent relationships between demographic and clinical characteristics of the patients and the outcome among severe cases. Using a stepwise approach, explanatory variables were selected if they were associated with the outcome at a level of significance of 0.2. In order to reach the final model prediction, we elected the explanatory variables that remained associated with the outcome with a 5% significance level when considered simultaneously [18]. In this manner, adjusted odds ratios were calculated for each studied variable. Area under the ROC curve (AuROC) was calculated in order to estimate clinical performance of the predictive model. To perform the analysis, we used the STATA statistical package version 13 (StataCorp, College Station, Texas, USA). Results From January 1, 2001, to December 31, 2013, 105,459 cases of dengue with clinical and laboratory confirmation were reported. Of these, 1,605 (1.5%) were categorized as severe dengue. During this period, 62 deaths from dengue were recorded, of which 61 cases (98%) were diagnosed as severe cases with laboratory confirmation, and 1 case (2%) did not meet the classification criteria for severe cases (Fig 1). 10.1371/journal.pone.0161884.g001Fig 1 Number of dengue cases and deaths due to dengue in the years 2001 to 2013 in the state of Amazonas). Severe dengue cases occurred predominantly in young women (15–55 years old) with a level of education of >4 years schooling (Table 1). Concerning day of illness, patients present to the hospital on average on ninetieth day of illness, mode equal to four days and the range from 0 to 34. The most common clinical signs were epistaxis, petechiae, and plasma extravasation. Patients who died had more plasma extravasation, gastrointestinal bleeding, petechiae, and thrombocytopenia (Table 1). 10.1371/journal.pone.0161884.t001Table 1 Descriptive characteristics of severe dengue cases reported in the state of Amazonas, according to the occurrence of deaths in the period from 2001 to 2013. Characteristics Death cases Non-death cases Total Total number (%) Total number (%) Total number (%) Demographic Sex Female 33 (3.9) 817 (96.1) 850 (100) Male 28 (3.7) 727 (96.3) 755 (100) Age <15 years 20 (2.7) 729 (97.3) 749 (100)   15–55 years 32 (4.1) 752 (95.9) 784 (100) >55 years 9 (12.5) 63 (87.5) 72 (100) Years of schooling ≤4 years 7 (2.5) 277 (97.5) 284 (100) >4 years 20 (4.2) 460 (95.8) 480 (100) Not applicablea 9 (3.6) 244 (96.4) 253 (100) Ignored/missing 25 (4.3) 563 (95.7) 588 (100) Clinical Epistaxis No 37 (10.2) 327 (89.8) 364 (100) Yes 8 (0.9) 917 (99.1) 925 (100) Ignored/missing 16 (5.1) 300 (94.9) 316 (100) Gingivorrhagia No 38 (3.7) 991 (96.3) 1.029 (100) Yes 7 (2.7) 250 (97.3) 257 (100) Ignored/missing 16 (5.0) 303 (95) 319 (100) Petechiae No 22 (3.3) 640 (96.7) 662 (100) Yes 22 (3.4) 616 (96.6) 638 (100) Ignored/missing 17 (5.6) 288 (94.4) 305 (100) Hematuria No 28 (2.5) 1.098 (97.5) 1.126 (100) Yes 18 (12.2) 130 (87.8) 148 (100) Ignored/missing 15 (4.5) 316 (95.5) 331 (100) Gastrointestinal bleeding No 17 (1.6) 1.047 (98.4) 1.064 (100) Yes 28 (13.0) 188 (87.0) 216 (100) Ignored/missing 16 (4.9) 309 (95.1) 325 (100) Plasma extravasation No 14 (2.9) 477 (97.1) 491 (100) Yes 35 (4.0) 832 (96.0) 867 (100) Ignored/missing 12 (4.9) 235 (95.1) 247 (100) Platelet count ≥20,000 cells/mm3 39 (3.2) 1.175 (96.8) 1.214 (100) <20,000 cells/mm3 22 (5.6) 369 (94.4) 391 (100) a Patients younger than 7 years. In the univariate analysis, among the demographic characteristics studied, the only characteristic associated with the occurrence of death is age >55 years. Gastrointestinal bleeding, hematuria, and thrombocytopenia were the clinical manifestations related to death (Table 2). 10.1371/journal.pone.0161884.t002Table 2 Relationship between deaths due to severe dengue and demographic and clinical factors in the state of Amazonas, in the period from 2001 to 2013. Factors Univariate analysis Multivariate analysis (final model) Crude OR 95% CI Adjusted OR 95% CI Demographic             Sex Female 1 Reference value - - Male 0.95 0.57–1.59 - - Age ≤55 years 1 Reference value 1 Reference value >55 years 4.06 1.91–8.60 4.98 1.78–13.87 Years of schooling ≤4 years 1 Reference value - - >4 years 1.72 0.72–4.13 - - Clinical             Epistaxis No 1 Reference value - - Yes 0.99 0.49–2.00 - - Gingivorrhagia No 1 Reference value - - Yes 0.73 0.32–1.65 - - Petechiae No 1 Reference value - - Yes 0.9 0.54–1.51 - - Hematuria No 1 Reference value 1 Reference value Yes 5.41 2.91–10.09 5.07 2.54–10.09 Gastrointestinal bleeding No 1 Reference value 1 Reference value Yes 9.21 4.9–17.16 10.26 5.32–19.76 Plasma extravasation No 1 Reference value - - Yes 1.42 0.76–2.68 - - Platelet count ≥20,000 cells/mm3 1 Reference value 1 Reference value <20,000 cells/mm3 1.79 1.05–3.02 2.55 1.33–4.89 Considering the simultaneous effects of demographic and clinical characteristics by using a multiple logistic regression model, it was observed that the factors associated with death were also age >55 years, hematuria, gastrointestinal bleeding, and thrombocytopenia. Gastrointestinal bleeding was the clinical sign most strongly associated with death, followed by hematuria and age >55 years (Table 2). The model algorithm is as follows: -4.945+1.605*[age>55] + 1.623*[hematuria] + 2.328*[GI- bleeding] + 0.938*[thrombocytopenia]. In this model, an adjusted R2 equal to 0.22 was obtained. The final model seems to have a good performance since AuROC (c statistic) was 0.843. Models with reasonable prediction capacity have values greater than 0.7 and strong prediction capacity when more than 0.8 [19]. Discussion The clinical signs that were able to predict death from severe dengue among patients diagnosed with the disease in the state of Amazonas between 2001 and 2013 were gastrointestinal bleeding, hematuria, and platelet count. Age was also a factor that predicted an increased risk of death, as patients with severe dengue who were older than 55 years had a higher possibility to die of the disease. Unfortunately, dengue remains one of the leading causes of death in developing countries, although both the infection of the virus and the evolution of the disease to death could be prevented. In the study population, the mortality rate of severe dengue was 3.8%. This rate is higher than the level desired by the Ministry of Health, whose goal is to reduce the mortality of severe cases to <1% [20]. It is noteworthy to highlight that the main strategy to reduce dengue mortality is the early identification of warning signs and appropriate clinical management of patients with a severe form of the disease. Paixão et al. indicated that the risk of death caused by dengue in Brazil increased significantly between 2000 and 2011 in all regions of the country, and that both mortality and fatality rates were higher in the last year, during one of the greatest epidemics recorded in Brazil [7]. These authors questioned whether reducing the fatality rate of severe dengue in Brazil to <1% is indeed possible, given the current level of therapeutic knowledge and the capacity to accurately estimate the fatality rate. However, we believe that deaths due to dengue might be prevented by adopting appropriate clinical management. On the other hand, the occurrence of deaths due to dengue is an indicator of the weakness in health-care networks, which needs immediate attention [6]. The most common characteristics in severe dengue cases were not associated with death due to the disease. Whereas patients with severe dengue were mostly women younger than 55 years, and the most prevalent clinical signs were the appearance of petechiae, epistaxis, and plasma extravasation, the characteristics associated with death due to dengue were age >55 years, gastrointestinal bleeding, hematuria, and platelet count <20,000 cells/mm3. Figueiró et al. analyzed the degree of actions and health services implementation, as well as the technical and scientific quality of care given to patients who died of dengue in the public health-care network of two municipalities in northeastern Brazil. The authors concluded that the occurrence of death was directly related to the clinical management of the cases, that patient care has not reached the level expected in any of the evaluated services, and that the recommendations from the Ministry of Health for the management of dengue cases are not being followed. According to the authors, the warning signs of dengue and those of shock due to the disease are not routinely investigated, and professionals have not used the clinical classification as recommended by the Ministry of Health [21]. In the same direction, the investigation of 94 deaths, conducted by the Ministry of Health in 2010, showed that aspects related to the organization of the services appear to be determining factors for the occurrence of deaths. These include a low participation of primary care as the preferred entry door of the system, the need to seek care in more than two health centers, and the lack of recognition of warning signs [22]. The importance of screening approaches and clinical management that can identify severe cases in patients seeking health services for a diagnosis, particularly those at a higher risk of complications and worse prognosis, has been already recognized. Passos et al. reported their experience of real-time monitoring of severe cases of dengue in Manaus during the 2011 epidemic, in which they found that clinical support led to the effective management of cases. The authors concluded that the rapid dissemination of strategic information for the prevention and control of dengue, as well as the real-time monitoring of severe dengue patients in health facilities in Manaus, allowed carrying out the appropriate management of the patients in a timely manner. Consequently, there were fewer deaths than what was expected in other previous outbreaks [23]. In our study, patients aged >55 years were more likely to die than younger patients. This finding corroborates those of other studies conducted in Brazil and other countries. Moraes et al. studied 12,321 cases of severe dengue reported between 2000 and 2005 in Brazil, and concluded that patients >50 years old are two times more likely to die. However, it is noteworthy to highlight that the largest epidemics occurred after that study period [24]. Garcia-Rivera and Rigau-Pérez pointed out that the elderly are at a greater risk of death during dengue epidemics, owing to the association with other morbidities that are more prevalent in this age group [25]. Our findings did not indicate any association between level of education and death due to dengue. This result can be attributed to the statistical power of the study, as well as to the occurrence of two major epidemics in the state of Amazonas. It should be noted that in a case-control study conducted in Brazil, it was found that patients with <4 years of education were 1.83 times more likely to die than those with higher levels of education [24]. In this study, patients with severe dengue had no increased risk of death in relation to sex, whereas another Brazilian study reported that women were less likely to die. It is important to note that, in our study, we considered as severe cases of dengue only those that were reported with laboratory confirmation, whereas Moraes et al. also included severe cases that did not meet the criteria of the Ministry of Health, i.e., “dengue with complications,” “hemorrhagic dengue Fever,” or “dengue shock syndrome” [24]. In this study, gastrointestinal bleeding, hematuria, and thrombocytopenia were identified as the clinical markers of risk of death due to dengue. Gastrointestinal bleeding was the clinical sign most strongly associated with an increased risk of death. This association was not found in the case-control national data from 2000 to 2005, which reinforces the importance of developing prediction models in specific epidemiological settings. Another possible explanation may be the greater attention given to prediction signs in the epidemics of recent years, the period during which our study took place. Severe dengue cases presenting with hematuria had a five times greater risk of death than the other cases. The association between clinical signs related to bleeding has been found to be plausible and already accepted by many researchers [26,27]. Another important finding of our study was that thrombocytopenia is a risk factor for death due to dengue. This result confirms the importance of using a cutoff point for platelet count of between 50,000 and 100,000 cells/mm3 as a criterion for the hospitalization of patients with dengue, as also discussed by other authors [24,28]. Our study indicates that, among deaths due to dengue that were reported in the Amazonas, a high proportion (88%) had been previously identified by health services as cases of severe disease. Thus, more careful attention to the signs of dengue could contribute to reducing the mortality of the disease. It is known that the timely identification of dengue cases is fundamental for decision making and the implementation of prevention and control measures, with the aim to mainly prevent deaths. The effective organization of health services, in an area of epidemiological surveillance and assistance, is essential for reducing the mortality caused by dengue, as well as to allow knowing the status of the disease in each area. This study has some limitations, mainly related to the analysis of secondary data. Underreporting of dengue cases is a reality in Brazil and, with regard to severe cases, there is a need for improvement in the proper filling out of report forms. There is also a need for improvement in the closing of cases by SINAN, as well as the proper recording of deaths by SINAN and SIM. It is likely that incorrect classifications of severe cases occurred, and the lack of laboratory confirmation might have influenced the formation of the study sample. We chose to include only cases that were confirmed by laboratory tests, as the reporting of cases during the epidemics might increase the overall number of incorrectly classified dengue cases. Regarding the missing data problem, it is difficult to assess whether there was selective withdraw or losses were by random. We assessed whether individuals with missing and complete data for the main predictors were different in relation to sex and age. The results indicate that there is no statistically significant difference between groups (data not shown). These findings weaken the plausibility of selection bias due to missing data. This study sheds light on the individual and clinical characteristics that can predict death among patients with severe dengue, in the historical and epidemiological contexts of the disease in which health services need to prepare the best surveillance strategies in order to reduce mortality to the lowest possible levels. In conclusion, considering our findings, the best prediction of death among severe dengue cases could be done based on individual characteristic of older age (>55 years), simultaneously with clinical signs of gastrointestinal bleeding, hematuria, and thrombocytopenia. This prediction model can be used both in improving the clinical management of severe dengue and to adopt specific surveillance strategies to recognize individuals with higher risk of death during episodes of dengue to be diagnostic in forthcoming epidemics. So the external validity of this prediction model could be performed in this moment. Confronting new epidemics could be benefited by training clinicians and other professionals who handle severe dengue using locally recognized predictors rather than those identified in clinical and epidemiological settings distinct from that observed in the Amazon. We would like to thank the Núcleo de Sistemas de Informação em Saúde da Fundação de Vigilância em Sáuse do Amazonas (Center for Health Information Systems of the Health Surveillance Foundation of Amazonas) for allowing us to access the SINAN and SIM database, and to the government of the state of Amazonas, through the Fundação de Amparo à Pesquisa do Estado do Amazonas, FAPEAM (Research Foundation of the State of Amazonas), for granting scholarships. DBC is a fellow of the RH-Doctoral Program at FAPEAM. The authors declare that there is no conflict of interest. ==== Refs References 1 Brady OJ , Gething PW , Bhatt S , Messina JP , Brownstein JS , Hoen AG , et al Refining the global spatial limits of dengue virus transmission by evidence-based consensus . PLOS Negl Trop Dis . 2012 ;6 : e1760 10.1371/journal.pntd.0001760 22880140 2 Bhatt S , Gething PW , Brady OJ , Messina JP , Farlow AW , Moyes CL , et al The global distribution and burden of dengue . Nature . 2013 ;496 : 504 –7 . 10.1038/nature12060 23563266 3 Gubler DJ , Meltzer M . Impact of dengue/dengue hemorrhagic fever on the developing world . Adv Virus Res . 1999 ;53 : 35 –70 . 10582094 4 Brazilian Ministry of Health . Monitoramento dos casos de dengue e febre de chikungunya até a Semana Epidemiológica 53, 2014 (in Portuguese) Boletim Epidemiológico Brasília ; 2015 . 5 Fundação de Vigilância em Saúde do Amazonas. Boletim de Vigilância em Saúde (in Portuguese)—Bol. Number 15. Manaus—AM; 2014. 6 Costa MDCN , Barreto FR , Coelho GE , Ba ML , Teixeira MG . Change in the epidemiologic pattern of occurrence of dengue and dengue hemorrhagic fever in Brazil . Journal of Epidemiology & Community Health . 2011 pp. A361 –A361. 10.1136/jech.2011.142976m.35 7 Paixão ES , Costa M da CN , Rodrigues LC , Rasella D , Cardim LL , Brasileiro AC , et al Trends and factors associated with dengue mortality and fatality in Brazil . Rev Soc Bras Med Trop . 48 : 399 –405 . 10.1590/0037-8682-0145-2015 8 De Figueiredo RMP , Thatcher BD , de Lima ML , Almeida TC , Alecrim WD , Guerra MV de F . Exanthematous diseases and the first epidemic of dengue to occur in Manaus, Amazonas State, Brazil, during 1998–1999 . Rev Soc Bras Med Trop . 37 : 476 –9 . 15765597 9 Rocha L , Tauil P . Dengue em criança: aspectos clínicos e epidemiológicos, Manaus, Estado do Amazonas, no período de 2006 e 2007 . Rev Soc Bras Med Trop . 2009 ; 10 Martins V do CA , Bastos M de S , Ramasawmy R , de Figueiredo RP , Gimaque JBL , Braga WSM , et al Clinical and virological descriptive study in the 2011 outbreak of dengue in the Amazonas, Brazil . PLOS ONE . United States; 2014 ;9 : e100535 10.1371/journal.pone.0100535 24978469 11 Fundação de Vigilância em Saúde do Amazonas. Boletim de Vigilância em Saúde (in Portuguese)—Bol. Number 12. Manaus—AM; 2012. 12 WHO T . Dengue: guidelines for diagnosis, treatment, prevention and control Geneva WHO Libr. 2009 ; 13 Fundação Nacional de Saúde . Dengue: diagnóstico e manejo clínico 1st ed. Brasília —DF ; 2002 . 14 Pongpan S , Wisitwong A , Tawichasri C , Patumanond J . Prognostic indicators for dengue infection severity . Int J … . 2013 ; 15 Vicente CR , Lauar JC , Santos BS , Cobe VM , Cerutti C Jr. Factors related to severe dengue during an epidemic in Vitoria, State of Espirito Santo, Brazil, 2011 . Rev Soc Bras Med Trop . Brazil; 2013 ;46 : 629 –632 . 10.1590/0037-8682-1579-2013 24270254 16 Díaz-Quijano FA , Waldman EA . Factors associated with dengue mortality in Latin America and the Caribbean, 1995–2009: an ecological study . Am J Trop Med Hyg . 2012 ;86 : 328 –34 . 10.4269/ajtmh.2012.11-0074 22302870 17 Instituto Brasileiro de Geografia e Estatística—IBGE . Censo Demográfico 2010: Características da População e dos Domicílios Rio de Janeiro , RJ ; 2011 . 18 Hosmer D, Lemeshow S. Applied logistic regression. 2004; 19 Jr DH. Model building strategies and methods for logistic regression. … Logist Regression, …. 2000; Available: http://onlinelibrary.wiley.com/doi/10.1002/9781118548387.ch4/summary 20 Brazilian Ministry of Health . Diretrizes Nacionais para Prevenção e controle de Epidemias de Dengue Brasília —DF ; 2009 . 21 Figueiró AC , Hartz ZM de A , Brito CAA de , Samico I , Siqueira Filha NT de , Cazarin G , et al Death from dengue fever as a sentinel event for evaluation of quality of healthcare: a case study in two municipalities in Northeast Brazil, 2008 . Cad saúde pública . 2011 ;27 : 2373 –85 . 22218580 22 Brazilian Ministry of Health . Dengue: diagnóstico e manejo clínico: adulto e criança 4th ed. Brasília —DF ; 2011 . 23 dos Passos RA , Pinto RC , Albuquerque BC . Monitoramento dos casos graves de dengue em Manaus–AM em tempo real e redução da mortalidade na epidemia de 2011 usando a tecnologia Nokia Data Gathering (NDG) convencionsalud2012.sld.cu. Brasília —DF ; 2012 . 24 Moraes GH , de Fatima Duarte E , Duarte EC . Determinants of mortality from severe dengue in Brazil: a population-based case-control study . Am J Trop Med Hyg . United States; 2013 ;88 : 670 –676 . 10.4269/ajtmh.11-0774 23400577 25 García-Rivera EJ , Rigau-Pérez JG . Dengue severity in the elderly in Puerto Rico . Rev Panam salud pública = Pan Am J public Heal . 2003 ;13 : 362 –8 . 26 Hung N , Lan N . Improvement of Case-Management: A Key Factor to Reduce Case-Fatality Rate of Dengue Haemorrhagic Fever in Southern Viet Nam . Dengue Bull . 2003 ; 27 Bethell DB , Gamble J , Pham PL , Nguyen MD , Tran TH , Ha TH , et al Noninvasive measurement of microvascular leakage in patients with dengue hemorrhagic fever . Clin Infect Dis . 2001 ;32 : 243 –53 . 10.1086/318453 11170914 28 Lye DC , Chan M , Lee VJ , Leo YS . Do young adults with uncomplicated dengue fever need hospitalisation? A retrospective analysis of clinical and laboratory features . Singapore Med J . 2008 ;49 : 476 –9 . 18581021

PMC005xxxxxx/PMC5001630.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756441710.1371/journal.pone.0161990PONE-D-16-13233Research ArticleBiology and Life SciencesAgricultureLivestockSwineBiology and Life SciencesOrganismsAnimalsVertebratesAmniotesMammalsSwineResearch and Analysis MethodsResearch DesignField TrialsMedicine and Health SciencesPulmonologyPneumoniaBiology and Life SciencesVeterinary ScienceVeterinary DiseasesPeople and placesGeographical locationsEuropeUnited KingdomBiology and Life SciencesAgricultureFarmsPeople and placesGeographical locationsEuropeUnited KingdomEnglandMedicine and Health SciencesDiagnostic MedicineSigns and SymptomsPeritonitisMedicine and Health SciencesPathology and Laboratory MedicineSigns and SymptomsPeritonitisPig Abattoir Inspection Data: Can It Be Used for Surveillance Purposes? Pig Abattoir Inspection Data Usage for Surveillance PurposesCorreia-Gomes Carla 1*Smith Richard P. 2Eze Jude I. 1Henry Madeleine K. 1Gunn George J. 1Williamson Susanna 3Tongue Sue C. 11 Epidemiology Research Unit, Future Farming Systems Research Group, Scotland’s Rural College, Kings Building, West Mains Road, Edinburgh, United Kingdom2 Animal and Plant Health Agency, New Haw, Addlestone, Weybridge, Surrey, United Kingdom3 Animal and Plant Health Agency, Rougham Hill, Bury St Edmunds, Suffolk, United KingdomHeneberg Petr EditorCharles University in Prague, CZECH REPUBLICCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: CCG RPS SCT. Data curation: CCG RPS JIE. Formal analysis: CCG RPS JIE MKH. Funding acquisition: CCG SCT SW. Investigation: CCG RPS. Methodology: CCG RPS JIE. Project administration: CCG SCT. Resources: GJG SCT. Supervision: CCG RPS SCT. Writing – original draft: CCG RPS JIE SCT. Writing – review & editing: CCG RPS JIE MKH GJG SW SCT. * E-mail: carla.gomes@sruc.ac.uk26 8 2016 2016 11 8 e01619901 4 2016 16 8 2016 © 2016 Correia-Gomes et al2016Correia-Gomes et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Statutory recording of carcass lesions at the abattoir may have significant potential as a resource for surveillance of livestock populations. Food Standards Agency (FSA) data in Great Britain are not currently used for surveillance purposes. There are concerns that the sensitivity of detection, combined with other issues, may make the outputs unreliable. In this study we postulate that FSA data could be used for surveillance purposes. To test this we compared FSA data with BPHS (a targeted surveillance system of slaughtered pigs) and laboratory diagnostic scanning surveillance (FarmFile) data, from mid-2008 to mid-2012, for respiratory conditions and tail bite lesions in pigs at population level. We also evaluated the agreement/correlation at batch level between FSA and BPHS inspections in four field trials during 2013. Temporal trends and regional differences at population level were described and compared using logistic regression models. Population temporal analysis showed an increase in respiratory disease in all datasets but with regional differences. For tail bite, the temporal trend and monthly patterns were completely different between the datasets. The field trials were run in three abattoirs and included 322 batches. Pearson’s correlation and Cohen’s kappa tests were used to assess correlation/agreement between inspections systems. It was moderate to strong for high prevalence conditions but slight for low prevalence conditions. We conclude that there is potential to use FSA data as a component of a surveillance system to monitor temporal trends and regional differences of chosen indicators at population level. At producer level and for low prevalence conditions it needs further improvement. Overall a number of issues still need to be addressed in order to provide the pig industry with the confidence to base their decisions on these FSA inspection data. Similar conclusions, at national level, may apply to other livestock sectors but require further evaluation of the inspection and data collection processes. Agricultural and Horticultural Development Board - PorkAn effective methodology for monitoring health and welfare status in the pig industry and to establish the current baseline health and welfare status of pigs in EnglandTongue Sue C Agricultural and Horticultural Development Board - PorkBPHS Data analysisCorreia-Gomes Carla S Scottish GovernmentRERAD funded programme Food, Land and People (Programme 2)-Theme 6Gunn George J This work was funded by the Agricultural and Horticultural Development Board (AHDB)-Pork (formerly BPEX) under the project “An effective methodology for monitoring health and welfare status in the pig industry and to establish the current baseline health and welfare status of pigs in England” and the project “BPHS Data analysis”. This work was underpinned by Theme 6 work funded by the Scottish Government (RERAD funded programme ―Food, Land and People (Programme 2)). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityData used in this research has not been generated by the authors but was obtained from other third parties. The data are available for researchers who meet the criteria for access which is specified by the respective data providing organisations. Data are obtainable by interested researchers from: a. British Pig Health Scheme (England and Wales) – Martin Smith, Martin.Smith@ahdb.org.uk, AHDB-Pork, Stoneleigh Park, Kenilworth, Warwickshire, CV8 2TL b. Food Standards Agency (FSA) – Gemma Forrest Gemma.Forrest@foodstandards.gsi.gov.uk, Service Development Lead, Business Transformation and Intervention Team, Food Standards Agency, Foss House, Peasholme Green, York YO1 7PR c. Farmfile—Rob Ready, Rob.Ready@defra.gsi.gov.uk, Weybridge IT Unit—Laboratory Information Management System (LIMS), APHA – Weybridge, Woodham Lane, New Haw, Addlestone, Surrey KT15 3NB.Data Availability Data used in this research has not been generated by the authors but was obtained from other third parties. The data are available for researchers who meet the criteria for access which is specified by the respective data providing organisations. Data are obtainable by interested researchers from: a. British Pig Health Scheme (England and Wales) – Martin Smith, Martin.Smith@ahdb.org.uk, AHDB-Pork, Stoneleigh Park, Kenilworth, Warwickshire, CV8 2TL b. Food Standards Agency (FSA) – Gemma Forrest Gemma.Forrest@foodstandards.gsi.gov.uk, Service Development Lead, Business Transformation and Intervention Team, Food Standards Agency, Foss House, Peasholme Green, York YO1 7PR c. Farmfile—Rob Ready, Rob.Ready@defra.gsi.gov.uk, Weybridge IT Unit—Laboratory Information Management System (LIMS), APHA – Weybridge, Woodham Lane, New Haw, Addlestone, Surrey KT15 3NB. ==== Body Introduction Endemic diseases are a particular concern to all livestock industry sectors, causing production losses and welfare issues. Surveillance can play an important role in their control by the provision of estimates of the frequency, such as prevalence or incidence, of disease and/or welfare conditions. These estimates can be monitored over time and significant changes detected [1]. Detection of change may lead to the initiation of appropriate interventions with continued surveillance enabling the evaluation of their impact. The units of analysis may vary from the population level to the individual animal and the outputs may be similarly targeted; from population (industry/government/region) to an individual producer level. Each purpose, or surveillance objective, may ideally need the data collection and/or analysis to be designed differently to produce the required output [2,3]. Asking a surveillance system to meet multiple objectives will require multiple components and, potentially, compromises will need to be made. When resources are limited, cost effective surveillance methods are also important; hence, existing data sources are being explored to determine if they can be used for surveillance purposes [4,5,6,7]. In Great Britain, probably the most robust data sources [8] that are currently used, or have the potential to be used, for endemic disease surveillance in livestock species are the Animal and Plant Health Agency (APHA)’s FarmFile system and statutorily collected abattoir inspection data [8]. Furthermore the pig sector also has robust health schemes [9], which were purposively designed for specific surveillance purposes. S1 Table describes these three systems in detail for pigs. Briefly abattoir inspection started in the late 1800s [10], with the objective to identify and discard carcasses infected with major zoonotic pathogens and parasitic infections in order to reduce public health risks. Since then, abattoir inspection has expanded and now has several purposes associated with public health, animal health and welfare and meat quality. In Great Britain, abattoir inspection has been audited and controlled by the Food Standards Agency (FSA). It is performed by the official auxiliaries (commonly known as meat hygiene inspectors (MHIs)) and official veterinarians. The data for the conditions/lesions observed and recorded can be provided to the farmer and the farmer’s private veterinary surgeon, to allow them to take action on-farm to improve animal health and welfare i.e. it has the potential to be used at a producer level. However, the abattoir inspection data are not currently being used for surveillance purposes at population level, e.g. to monitor trends of specific conditions or to detect significant changes in endemic disease trends. The pig health schemes in Great Britain were designed to provide information about specific endemic conditions of economic concern for producers, with relevant outputs provided both to individual producers and at a population (industry) level [11]. In Great Britain there are two schemes: Wholesome Pigs Scotland, in Scotland, and the British Pig Health Scheme (BPHS), in England and Wales. BPHS started in 2005 and provides frequent feedback of benchmarked results from targeted abattoir inspections to the participating producers and their herd veterinarians [8], helping to increase their awareness of the occurrence of subclinical diseases in their farms [11]. Since 1999, farm and laboratory data collected from voluntary laboratory submissions made across Great Britain, to APHA Regional Laboratories and Scottish Agricultural College’s Consultancy Division’s Disease Surveillance Centres has been aggregated in the FarmFile database [12]. Carcass and non-carcass submissions are submitted through private veterinary surgeons for laboratory testing and diagnostic investigation. Individual results are sent back to the private veterinary surgeon and the data contribute to surveillance for the detection of new and emerging threats, including significant changes in endemic disease trends. Denmark and the Netherlands have pioneered the use of abattoir inspection data as a means of animal health surveillance, followed by countries such as Sweden, Norway and Germany [10]. Advantages include: increased coverage using an existing data collection system and data on many health and welfare conditions. Such data could be used to help identify potential emerging animal health threats, or to target slaughtered animals entering the food chain that present a direct risk to people from zoonotic organisms. However, difficulties have been identified with FSA data in the past and there is a lack of trust in the system. Some of the issues that have been identified with the FSA include a lack of sensitivity (ability to detect an affected carcass); a lack of standardisation with multiple operators and conditions; poor data quality and coverage of only those healthy pigs sent to slaughter [13]. The aim of this study was to test if abattoir inspection data has potential for surveillance purposes: firstly at a population level, to measure how much disease is present? and how this changes over time, and secondly at a producer level, to provide information to direct individual action, using the pig industry and respiratory conditions and tail bite lesions, as an example. Respiratory conditions were chosen as they are one of the major diseases affecting pigs worldwide [14]. Most of the respiratory lesions in pigs in Great Britain are lung consolidation likely to be caused by enzootic pneumonia (EP) and pleurisy [9]. Tail bite lesions are considered a welfare problem, especially in Great Britain, where routine tail docking of pigs is banned. These lesions may indicate stress, adverse environmental conditions and intercurrent disease among other factors [15]. To achieve this aim the FSA data were compared with data from the two existing data sources commonly used to monitor trends in endemic disease: FarmFile and BPHS. The specific objectives were a) to compare the prevalence (over the time period and spatial distribution) of FSA data with BPHS and FarmFile for respiratory conditions and tail bite lesions, looking for similarities and differences, and b) to examine the agreement between BPHS and FSA data–the question here being, could FSA replace BPHS data? Materials and Methods Population level comparison Data description The data sources used for the population level analysis were FSA, BPHS and FarmFile data. The details of these systems and type of data collected are explained in detail in S1 Table. Datasets were collected to cover a 48 month period from June 2008 to May 2012. However, the FSA dataset only started at August 2009, following the implementation of a new electronic format for the data. Data management and analyses To assess regional prevalences, the postcode of the farm of origin of the carcass, or submission, was used to link a record to a geographical region, based on British NUTS 1 regions (Nomenclature of Units for Territorial Statistics [16]). Where full postcodes were missing, county definitions were used to classify which region the record belonged to. The FSA dataset did not include postcode information, so other information available to the researchers was used to link a postcode to each herd/slap mark in the dataset. Slap mark is the herd mark that is tattooed on a pig and identifies farm of origin. It is a legally required official reference for each pig farm. In some cases a farm can have more than one slap mark. It was not possible to identify those cases in our analysis so it was assumed that each slap mark will approximate to a farm. Two conditions were considered: respiratory conditions and tail bite lesions. As each data source has different ways of assessing some of the same lesions/diseases, the data were recategorised into “lesions/diseases” (S2 Table). The FSA data include two inspection types: ante mortem (AM) and post mortem (PM). Within the PM inspection, conditions were recorded as present in the carcass, offal, or whether they were generalised conditions. For each of the datasets, the date of assessment was used to produce variables for the year and month of assessment. Regional prevalences were calculated and plotted in a map. Epidemiological modelling was used to detect statistically significant differences (P<0.05) between the results. The data from each data source was modelled separately. Respiratory lesions All three datasets were used for the analysis of respiratory conditions. In the FSA data tuberculosis (TB)-like lesions and suspected generalised tuberculosis were included, although it should be noted that these are not necessarily respiratory conditions. For the FarmFile system interpretation was required to decipher multiple diagnoses from the same submission to clarify if a respiratory condition was present. A submission may relate to multiple animals being ill on a holding with samples being collected from single or multiple pigs, but typically a diagnosis for a single pig is produced and so for the purposes of this analysis we have assumed that a submission is related to a single pig with a diagnosed condition. For the FSA and BPHS dataset the number of animals assessed and the number of individual animals that had any positive result for any of the respiratory conditions were summarised per batch. For the FarmFile dataset a binary score was created for each submission, based on the presence/absence of any respiratory conditions, and the results summarised per submission. As each of the data sources contained multiple samples from the same farm, this contravenes the assumption of independence between records. Generalised linear mixed models (GLMM) were used to account for this. The unique farm identifier (slap mark) recorded in each data source was applied as a random effect in the GLMM to account for the potential similarity of results between samples from the same farm. GLMM also adequately corrects for the differences in sample size between batches by incorporating the number of animals sampled and the number of positive cases in the model. This adjustment is necessary in order to reduce bias in prevalence estimation due to varying sample size. In the multivariable models, month and year variables were used as fixed effects to model the temporal trend effects. The model used to analyse the FSA data for the respiratory lesions outcome did not converge within 24 hours. A simpler and faster method was used instead, in which a robust cluster function for the unique farm identifier was added to a logistic model. This method provides robust standard errors but does not model the effect of each farm effect separately. This method was then used for all of the respiratory condition models to aid comparison of results. The logistic regression analysis was completed using Stata 12 (StataCorp. 2011). Tail bite lesions For the tail bite lesions only the BPHS and FSA data were used because this condition is not recorded in FarmFile. As for the respiratory conditions, the number of animals assessed and the number of individual animals that had any positive result for tail bite lesions were summarised per batch. GLMM were used to model the results for tail bite lesions, as described above for respiratory lesions. GLMM analyses were performed with R version 2.12.1 from R Foundation for Statistical Computing. http://www.r_project.org. using the following packages: nortest [17] and lme4 [18]. Field trials comparison Data description Four field trials were held in three abattoirs during 2013. These field trials were not part of normal BPHS operations and were carried out especially for the purpose of comparing FSA and BPHS data. Data were recorded at animal level by the BPHS assessors and at batch level by the MHIs, in accordance with their normal practices. During the field trials an external observer was present to guarantee that the same pigs were being scored by BPHS assessors and MHIs. The MHIs were informed about the aim of the field trials before they were carried out and the results of the comparison were discussed with them afterwards. In one abattoir two field trials were carried out at an interval of almost three months. Data management and analyses Different ways of assessing the same lesions are used, so the data were recategorised into “Conditions” (S3 Table). The data collected were compared at batch level. A batch was defined as pigs belonging to the same slap mark that were slaughtered in the same abattoir on the same date. For the conditions that were found to have a normal distribution (using the Anderson-Darling test), a Pearson’s correlation test was used to test for correlations between the two proportions (FSA vs BPHS). The conditions that did not follow a normal distribution (right skewed due to the high number of batches with zero prevalence) were categorised according to absence or presence of the condition for each batch. The data were then analysed as binary variables. The Cohen’s Kappa test, which measures the agreement between two methods, was also used. Due to the number of analyses performed in the comparisons (n = 8) any differences were considered statistically significant when P<0.00625, using Bonferroni correction. For the interpretation of the kappa values the following categories, as described by Shoukri and Pause [19] were used: 0 –“poor”, 0.01–0.20 –“slight”, 0.21–0.40 –“fair”, 0.41–0.60 –“moderate”, 0.61–0.80 –“substantial” and 0.81–1.0 –“almost perfect”. The Pearson’s correlation values were interpreted using the following categories, as suggested by Evans [20]: 00–0.19 - “very weak”, 0.20–0.39 - “weak”, 0.40–0.59 - “moderate”, 0.60–0.79 - “strong”,0.80–1.0 - “very strong”. The data were also reanalysed by comparing the batches with high and low discrepancy, in terms of the number of animals assessed. These two categories were defined as follows: Low discrepancy batches: at least 80% of the animals in the batch were assessed by the BPHS assessor, assuming that the MHI assessed the full batch. High discrepancy batches: less than 80% of the animals in the batch were assessed by the BPHS assessor, assuming that the MHI assessed the full batch. This was done because it was assumed that a big difference in the number of animals assessed between these two systems may influence the results. The cut-off of 80% was arbitrarily chosen. The field trial comparison analyses were performed using the vcd [21] package with R version 2.12.1. Results Population level comparison The FSA dataset had the largest number of unique holdings and had the greatest number of pigs assessed compared to the other data sources (Table 1). 10.1371/journal.pone.0161990.t001Table 1 Summary of the three data sources (FSA, BPHS, FarmFile) used to analyse trends.   FSA BPHS FarmFile Number of pigs assessed 19,534,728 933,771 6,012* Batches 306,004 21,929 - Holdings 31,578 2,543 2,699 Identifier Slap mark, postcode Slap mark CPH/ postcode/ unique submission reference Holdings with postcodes 18,783 2,076 2,076 Number of pigs (%) that could be allocated to regions 17,278,953 (88.5%) 900,840 (96.5%) 5,574 (92.7%) * Assumed to be a single pig–see Materials and Methods text, Population level comparison–Respiratory lesions, CPH–county parish holding number (identification number for farm or business, which relates to the location of the land). Respiratory conditions The BPHS data suggested a possible increase in the prevalence of respiratory conditions over the years, with the prevalence of cases rising from 32.2% in 2009 (95%CI: 32.0%-32.4%) to 40.0% in 2012 (95%CI: 39.7%-40.3%) (Fig 1A). The monthly summaries showed that the highest prevalence occurred from November to January, whereas the lowest prevalence occurred in July (Fig 1B). The FSA and FarmFile data also suggested an increased prevalence over the years (Fig 1A). However, the monthly results differed from BPHS. The FSA and FarmFile data had peaks in prevalence occurring in spring and a low prevalence in autumn (Fig 1B). For the FSA data the highest prevalence occurred between March to May, whereas the lowest prevalence occurred in September and October. This was complemented by FarmFile, which had the highest prevalence in April and May and lowest in September (Fig 1B). However, due to the small number of records in this dataset and the overlapping 95% confidence intervals, the monthly prevalence estimates for FarmFile should not be over-interpreted. 10.1371/journal.pone.0161990.g001Fig 1 Respiratory disease conditions in the three datasets (FSA, BPHS and FarmFile). Prevalence estimate (dots) and 95% confidence interval (bars) from 2008 to 2012 (A) and from January (1) to December (12) (B). Logistic regression odds ratio (OR) (dots) and 95% confidence interval (bars) from 2008 to 2012 (C) and from January (1) to December (12) (D). In the multivariable temporal analyses, the BPHS model showed significant increasing odds of a pig having respiratory conditions in 2011 and 2012 compared to 2008 (Fig 1C) and that the months July and March had the lowest odds (Fig 1D). The FarmFile model showed that the odds of a respiratory case increased over the years with significant increases in years 2010 and 2011 (Fig 1C). In this model April had the highest odds and September had the lowest (Fig 1D). Finally, the FSA model showed that the odds increased over the years and that 2010 and 2012 were significantly higher than 2009 (Fig 1C). The results for month showed that there were two peaks, with high odds in the spring months and in December (Fig 1D). The prevalence for the BPHS data was comparable across the geographical regions (range 30–39.8%). It was highest in the East of England and Wales, and lowest in South East and North West England (Fig 2). For the FarmFile data, the prevalence was more variable across the regions (8.1–20.5%), being lowest in Scotland and Northern Ireland (where few samples were collected) and highest in the North West and North East. In contrast, for the FSA data, the prevalence was comparable to BPHS data in Scotland, where the prevalence was highest (37.7% [BPHS, 35.8%]) and lower across England and Wales (<20%), being lowest in the South East, North East and East of England (<12%, Fig 2). 10.1371/journal.pone.0161990.g002Fig 2 Prevalence of respiratory lesions by UK NUTS1 during the studied period for the three datasets (FSA, BPHS and FarmFile). England is indicated by the regions shown in pink. The 95% confidence interval for the prevalence is within brackets. Tail bite lesions The analysis of the BPHS data suggests a general decrease in the prevalence of tail biting from 0.44% in 2008 to 0.35% in 2012. However, year by year comparison shows that the prevalence was highest in 2010 (Fig 3A). There was no clear average monthly pattern (Fig 3B). The FSA data analysis produced contrary results suggesting a general increase in the prevalence from 2009 to 2012 (Fig 3A). The average monthly pattern indicates that the prevalence was highest in February and lowest in August (Fig 3B). 10.1371/journal.pone.0161990.g003Fig 3 Tail bite lesions in the two datasets (BPHS, FSA). Prevalence estimate (dots) and 95% confidence interval (bars) from 2008 to 2012 (A) and from January (1) to December (12) (B). Generalised Linear Mixed Model odds ratio (OR) (dots) and 95% confidence interval (bars) from 2008 to 2012 (C) and from January (1) to December (12) (D). Results from the BPHS data indicate that odds of tial bite lesions declined over the years (Fig 3C). Odds were lower in 2011 and 2012 when compared with 2008. However, the highest odds was in 2010. The odds of tail biting were higher in July and lower in December compared to January (Fig 3D). Temporal patterns differed for FSA: the odds overall the years increased with a similar peak in 2010, however the continued decline in 2012—seen in the BPHS—was not observed. The patterns across the year by month were similar for the first half of the year (January to June) and then differed, with specific differences seen between the datasets in July and December. The odds were higher in May than January and lower in September and August compared to January. However, the 95% confidence intervals for the BPHS dataset are wide and overlap considerably. Descriptive analysis of tail bite lesions by region (Fig 4) using BPHS data suggests that prevalence was highest in the South East and lowest in Scotland (0.68% and 0.17%, respectively). In Scotland it was comparable to the estimate from the FSA data (0.18%). Otherwise the FSA estimates were generally lower than the BPHS ones, with a different distribution; prevalence was highest in the North West (0.31%) and lowest in South East (0.08%) and Wales (0.05%). 10.1371/journal.pone.0161990.g004Fig 4 Prevalence of tail bite lesions by UK NUTS1 during the studied period for the two datasets (BPHS and FSA). England is indicated by the regions shown in pink. The 95% confidence interval for the prevalence is within brackets. Field trials comparison Batch level FSA and BPHS data were obtained from a total of 53,479 and 18,748 pigs, respectively, from 332 batches (Table 2). Two batches, from trial 1, were not included in this study due to incorrect recording of the pig slap mark by the MHIs and mixing of conditions data from pigs of different slap marks into the same batch of pigs. 10.1371/journal.pone.0161990.t002Table 2 Dates of the field trials with the number of slap marks, batches and pigs assessed for the three different abattoirs that participated in the field trials for the two data sources (BPHS, FSA). Abattoirs Trial number Dates (2013) Number of slap marks Number of batches Number of pigs assessed A 1: week 22 to 26/04 35 50 2,479 (BPHS), 8,805 (FSA) 1: days 21/02, 8/05, 13/05, 14/05 23 34 1,685 (BPHS), 5,895 (FSA) 3: week 8 to 12/07 46 56 3,374 (BPHS), 12,729 (FSA) B 2: week 17 to 21/06 87 101 4,408 (BPHS), 13,104 (FSA) 2: days 28/06, 2/07, 8/07 46 49 3,389 (BPHS), 8,249 (FSA) C 4: week 30/09 to 4/10 22 24 2,044 (BPHS), 2,765 (FSA) 4: days 7/10, 18/10, 22/10 17 18 1,369 (BPHS), 1,932 (FSA) The overall results show a moderate correlation between the inspection methods for pericarditis and pleurisy and a strong correlation for pneumonia. The agreement is moderate for milk spots and slight for tail bite, peritonitis and pyaemia (Table 3). The results at abattoir C (which had a low line speed) showed better agreement for milk spots, pyaemia and abscesses in the lungs compared to the other abattoirs although not statistically significant (Table 3). Abattoir C also had a moderate correlation for pericarditis and a moderate to strong correlation for pneumonia (Table 3). 10.1371/journal.pone.0161990.t003Table 3 Summary of the correlation/agreement results between the two datasets (BPHS, FSA) for pericarditis, pleurisy, pneumonia, milk spots, tail bite, pyaemia, peritonitis and abscesses in the lung for each of the week field trials and all data of the field trials, with significance levels. Condition Abattoir A: Week trial 1 Abattoir B: Week trial 2 Abattoir A: Week trial 3 Abattoir C: Week trial 4 All data (week plus days trials) Pearson’s correlation coefficients Pericarditis 0.511*** 0.612*** 0.304* 0.531** 0.415*** Pleurisy 0.162 0.545*** 0.507*** 0.393 0.473*** Pneumonia—v1 0.837*** 0.571*** 0.684*** 0.769*** 0.650*** Pneumonia—v2 0.883*** 0.572*** 0.798** 0.622** 0.687*** Pneumonia—v3 0.826*** 0.597*** 0.746*** 0.405* 0.653*** Cohen’s kappa agreement values Milk spots 0.32 0.36 0.325** 0.408 0.414** Tail bite 0.14** 0.10*** 0*** 0.023*** 0.038*** Pyaemia 0.23* 0 0*** 0.333 0.069*** Peritonitis 0.33* 0.18 0.062*** 0.167** 0.175*** Abscesses in the lung 0.18 0.19*** 0.103*** 0.250* 0.131 *p<0.05 ** p<0.01 *** p<0.001 Pneumonia–v1: number of animals with enzootic pneumonia-like lesions score >0, or/and Viral-like pneumonia or/and pleuropneumonia lesions; Pneumonia–v2: number of animals with enzootic pneumonia-like lesions score >5, or/and Viral-like pneumonia or/and pleuropneumonia lesions; Pneumonia–v3: number of animals with enzootic pneumonia-like lesions score >10, or/and Viral-like pneumonia or/and pleuropneumonia lesions When batches with low or high discrepancy in the number of pigs assessed were compared (Table 4), the correlation for pericarditis was higher for the batches with low discrepancy compared to the batches with a high discrepancy. For pneumonia the opposite was observed and no major difference was seen for pleurisy (Table 4). In general, the agreement for peritonitis was better for the batches with low discrepancy compared to the ones with high discrepancy (Table 4). 10.1371/journal.pone.0161990.t004Table 4 Summary of the correlation/agreement results between the two datasets (BPHS, FSA) for pericarditis, pleurisy, pneumonia, milk spots, tail bite, pyaemia, peritonitis and abscesses in the lung for low and high discrepancy batches, with significance levels. Condition Low discrepancy High discrepancy Pearson’s correlation coefficients Pericarditis 0.702*** 0.364*** Pleurisy 0.488** 0.467*** Pneumonia–v1 0.483** 0.676*** Cohen’s kappa agreement values Milk spots 0.648 0.372** Tail bite NC 0.097*** Pyaemia NC 0.091*** Peritonitis 0.325** 0.172*** Abscesses in the lung 0.155 0.139 Low discrepancy batches: at least 80% of the animals in the batch were assessed by the BPHS assessor, assuming that the MHI assessed the full batch, high discrepancy batches: less than 80% of the animals in the batch were assessed by the BPHS assessor, assuming that the MHI assessed the full batch ** p<0.01 *** p<0.001 NC: Not possible to estimate due to zero results in both positive categories, Pneumonia–v1: number of animals with enzootic pneumonia-like lesions score >0, or/and Viral-like pneumonia or/and pleuropneumonia lesions. Discussion The aim of this study was to investigate if pig abattoir inspection data could be used for surveillance purposes. To achieve this we compared the prevalence (over the time period and regions) of two conditions at population level and investigated the agreement at batch level between FSA and BPHS data. Population level comparison The major difference between data sources was in the values of the prevalence estimates generated. These were higher for respiratory lesions in BPHS in comparison to both FSA and FarmFile, in which the estimates were similar. In addition, even though the prevalence of tail bite lesions is approximately 100 times less, the estimates were equivalently higher in the BPHS dataset compared to the FSA dataset. While some difference would be expected due to the differences in the populations sampled and in the criteria for inclusion in the respiratory lesion category of each dataset, this consistent difference between the FSA and BPHS dataset, even when the condition is classified in a similar way (tail bite lesions), suggests that something else is happening. A similar finding was observed in another study that compared routine abattoir inspection findings with systematic health monitoring in pigs in Denmark [22]. This seems to suggest that there are differences in recording sensitivities between the two systems i.e. the BPHS assessors detect a higher proportion of carcasses affected with respiratory conditions compared to the MHIs; similarly with a low prevalence condition such as tail bite lesions. The FarmFile data is a very different subset of the population, relating to more overtly diseased animals and diagnostic requests (a minority of which were for monitoring purposes). Why this should have similar prevalence estimate values for respiratory lesions to those in the FSA dataset is not entirely clear, unless it is related to the sensitivity of diagnostic uncertainty. For respiratory conditions the yearly and monthly patterns were similar between the different data sources. For tail bite lesions the two data sources showed differences in both temporal patterns. The major difference was a year effect in 2012 and variations in the second half (June–December) of the year. In both sets of analyses the patterns in the FSA data are smoother. This would be expected from a larger, continuously collected dataset that encompasses the wider population. The FSA and FarmFile data complement each other for respiratory conditions, which was expected; as subclinical conditions detected in the wider population may either be an indicative lead into an increase in clinical observations–and subsequent submission for diagnosis–or, as a sequela. Similarly one would expect the FarmFile and BPHS data to complement each other (clinical v. subclinical) but for both to be more variable than the FSA data, as they are from smaller but potentially more investigatory minded systems than the general abattoir inspection. BPHS is known to cover a smaller population of farms, with members representing only 75% of the commercial units [23]. Furthermore, the intermittent (voluntary and quarterly) nature of the data collection will have an effect on the variability in prevalence estimates, particularly on a monthly basis. All of these aspects may contribute to the differences observed between the BPHS and FSA data for tail bite lesions. Additionally, the age of the assessed pigs in the datasets differs. BPHS pigs are slaughter age, finished pigs with an average age of 5.7 months [24], as are the vast majority of those contributing to the FSA dataset. FarmFile, however, can include data from pigs of any age, although those pigs being submitted for respiratory disease are usually younger than typical slaughter age pigs. This would have the consequence that, in the temporal analysis, an increase in prevalence of respiratory disease detected in weaners or growers, recorded by FarmFile, might not be expected to be reflected in lesions observed at slaughter until 6–16 weeks later i.e. the FSA and BPHS data would be expected to ‘lag’ FarmFile The spatial differences between the datasets reflect the overall prevalence difference, with BPHS regional estimates being consistently higher than the FSA estimates for both lesions in all regions, with the noticeable exception of Scotland. Here the BPHS and FSA prevalence estimates concur, however the number of Scottish farms that send pigs to slaughter in England is very small and not representative of the Scottish situation. For each of the two conditions, the variation in the regional spatial distribution within each dataset is not the same. It is not possible, therefore, to tell whether there is true regional variation in the conditions themselves without considerable further investigatory work into other potential contributory factors. For example, regional differences due to assessor differences should be minimised in the BPHS dataset as standardisation exercises are held, whereas these are not common practice between MHIs and abattoirs. There are other data limitations that might influence the results. Firstly, BPHS and FSA record non-specific lesions, which has the advantage of high sensitivity but low specificity. In contrast, FarmFile records specific diagnoses with high specificity but low sensitivity due to submission bias. Secondly, the data, even within recategorised conditions, are not directly comparable due to the different definitions and criteria for recording (e.g. combination of multiple respiratory conditions in FarmFile data). This has resulted in a loss of sensitivity for patterns of specific respiratory conditions. Thirdly, there is also potential for misclassification; in the FarmFile data the main presenting condition would have been recorded, but in some cases other secondary conditions were not recorded. This is a known hazard in the FSA inspection system, too [13]. In addition, for the FSA data, the conditions details were summarised into the total number of animals that had body parts rejected due to a condition. As data on individual animals was not provided, it was possible that some double counting of animals with conditions was included e.g. a batch with one record of tail bite and a record of fight/bite may have been the same animal and would have been recorded as two animals instead of one. An improvement, for future analyses of FarmFile data, would be to analyse only diagnostic samples for which respiratory clinical signs had been noted by the submitting veterinary practice. A further refinement would be to split carcass and non-carcass diagnostic submissions, as carcass submissions would be indicators of more severe incidents of respiratory conditions. Field trial comparison The BPHS and FSA data were also compared in terms of agreement per batch. Overall, the prevalence of respiratory conditions was higher within BPHS data than the FSA (data not shown) and the results of the field trials showed a moderate correlation between data sources for pericarditis and pleurisy and a strong correlation for pneumonia. This is similar to findings by Nielsen and colleagues [22], except for pleurisy where the correlation was stronger than the one observed in this study. The agreement between data sources was moderate for milk spots and slight for tail bite, peritonitis and pyaemia. For some conditions (pericarditis and peritonitis) with low prevalence, the agreement between data sources improved when all, or almost all, animals in the batch were assessed and not only a sample of the batch. This suggests that assessing a subset of animals per batch (as BPHS currently does) is not enough to account for the variability associated with low prevalence conditions (e.g. for detecting a disease with 1% prevalence in a group of 100 animals with 95% CI, 95 animals should be tested) [25]. One limitation of this comparison was the power of the statistical analysis of the field trials. This was low because only three abattoirs were involved, with a low number of batches assessed. However, it was a resource intensive exercise and the abattoirs were chosen to represent the situation of English abattoirs (e.g. abattoir A and B were high output abattoirs while abattoir C was a low output abattoir and therefore had a different line speed). The MHIs were aware of the aim of the field trials before they were carried out, as were the BPHS assessors, and the results were discussed with all involved in the field trials at the end of each day. This potentially may have affected the outcome of the field trials. While a blinded process would have been preferable that was not possible and both sets of assessors had the opportunity to raise their game on the day. One abattoir (A) had two field trials. This meant that all involved in the trials in that abattoir were aware, in the second field trial, of previous pitfalls identified and therefore could have changed their procedures. However, an improvement in agreement for the different conditions assessed from week 1 to week 3 for abattoir A was not observed. The comparison between outcomes is not one between equivalent inspections; it was a comparison between two types of inspection. Another potential limitation of abattoir inspection data could be its differing ability to detect conditions. The sensitivity (probability of detecting conditions actually present in the diseased group) and specificity (probability of correctly identifying healthy carcasses in a healthy group) of routine abattoir inspection for parasitic, intestinal and heart disorders was considered low in a study that compared two groups of observers (regular meat inspectors and two veterinary researchers) [26]. In addition, in a recent Danish study [22], pericarditis, pleuritic and lung lesions data collected through routine abattoir inspection were compared to data collected through a systematic health monitoring system (voluntary scheme for farmers having clinical problems) for the same batches. The results suggested that the correlation was moderate for pleuritic and lung lesions, but poor for pericarditis and the authors concluded that caution should be used whenever routine abattoir inspection data are used for purposes other than those for which they were originally intended. Potential for animal health surveillance purposes The results of our work suggest that the FSA data could, potentially, be used to measure the apparent prevalence of conditions and monitor trends over time at the population level. This holds, as long as the system remains stable in denominator population (number and type of pigs slaughtered; management types, ages, etc.), methodology and application. This, however, is not the case: the current FSA system has been subjected to challenges, such as the transition to visual inspection from June 2015 and the retraining of MHIs. These changes will have affected the types of conditions that can be recorded and could affect the performance of the system over time, respectively. Neither aspect was evaluated in this study. Nielsen and colleagues [22] have suggested that, at least for pericarditis, some cases will be missed at visual inspection, which can result in a loss of overall sensitivity. The industry would also need to accept that the prevalence in this population is different than the estimates acquired via the BPHS because the two inspection systems are different. While useful for temporal trends, further work would be required to investigate whether there is the potential to detect emerging or re-emerging diseases. The continuous nature of the data collection and the large numbers inspected make it easier to detect a statistically significant change in the overall population in a timely manner. The loss of sensitivity for some conditions and the larger scale may, however, mask more localised fluctuations that are of clinical significance within specified populations or locations. Alternatively, it could draw a number of these fluctuations together into a coherent indication of a problem that would not otherwise be recognised. Furthermore, it is expected that emerging and re-emerging endemic diseases will usually be detected before slaughter; for example by production and clinical monitoring at farm-level. Abattoir inspection could act as an ultimate ‘fail-safe’ in the case of failure at this level, particularly when faced with an insidious, sub-clinical, situation. It enables an alarm to be raised when the other components fail [5,27]. Elsewhere, several studies [4,5,6,7,28,29,30,31] have evaluated the use of abattoir data for integration in syndromic surveillance systems, for the early detection of emerging diseases in livestock animals or to measure animal health and welfare. Some of the advantages identified were the good coverage and availability of syndromic indicators; while some of the limitations were the lag of time between reporting and occurrence of disease in live animals and the influence on the results of abattoir-related factors. The potential utility of the FSA data for the provision of evidence to inform decision making and individual action at the producer level (i.e. could FSA data replace BPHS data?) is hampered by a number of issues; this is illustrated by the field trial batch level results. The major issues are the recording of correct slap mark for each pig and the double counting of animals with multiple conditions. Standardisation of MHIs in terms of assessment and recording of conditions is also required to give confidence to producers and their veterinarians. The lack of agreement between the two systems and the higher prevalence of lesions recorded in the BPHS system (especially for pneumonia and pleurisy) may suggest that, for the measurement and monitoring of endemic disease and welfare conditions, the detection of change and the evaluation of the effects of any control measures that are implemented at producer level, the BPHS system performs better than the FSA currently could. This probably occurs because BPHS assessors are focused only on a limited number of standardised specific animal health related lesions, while FSA ante and post mortem inspection has the wider primary remit to protect public health. However, given the reduced ability to detect conditions with low prevalence in large batches and the quarterly nature of the BPHS inspection, it may be less suited than the FSA data to the early detection of emerging or re-emerging diseases, at a producer level. The data sources investigated in this paper are considered to be three of the most robust data sources currently available for the British pig industry. The analyses highlighted that no single source could fulfil all the purposes that a surveillance system may be asked to meet; however, they each have useful contributions to make as separate components of an animal health surveillance system. As each of the data sources differs in terms of coverage and representativeness of the pig industry, amalgamating all the data together could be attempted, as could integration through statistical analyses. Such approaches were not within the remit of this study. The FSA data could be interpreted in parallel with FarmFile and BPHS data, or with FarmFile data alone to provide monitoring of trends of the chosen indicators at population (i.e. national) level. Stärk and colleagues [5] drew a similar conclusion that the combination of abattoir inspection with other surveillance components can provide more information than abattoir inspection alone. The FSA data requires improvement to be effective for use at the producer level. Recently AHDB–Pork began working with the FSA and abattoirs to improve the FSA system. The aim is to bring it to a position where it can provide similar information on pig health and welfare issues from post mortem data, to that which is currently delivered through the BPHS. When such a system is in place further comparisons should be done. In this paper, the abattoir inspection data for the pig industry could be assessed because of the existence of the purposively designed abattoir based health scheme data. For the other livestock industries in Great Britain no such comparative schemes exist. However, it is likely that the same conclusions can be made: the FSA data can be interpreted in parallel with FarmFile to provide monitoring of trends of any chosen indicators at population (i.e. national) level but requires improvement to be valued at producer level. Standardised collection and recording of the core data requirements, including those variables that enable linkages to be made with other data sources (slap mark, CPH and postcode) is vital. This and the provision of a background standardised description and evaluation of the data sources is essential for any attempt to use multiple data sources for surveillance purposes. Similar conclusions may apply to other countries where statutory abattoir inspection data are collected; especially within the EU, as data collection is regulated by the same legislation. It would, however, require an evaluation of the inspection and data collection processes before robust comparisons can be drawn between countries. Conclusion There is potential to use abattoir inspection data, as a component of an animal health and welfare surveillance system, in order to monitor temporal trends and regional differences of chosen indicators at population (i.e. national) level. Further work is required to evaluate its potential for the early detection of emerging or re-emerging diseases at a population level; however, there is the potential to do this with low prevalence conditions at producer level (i.e. within herd/batch). To inform decision making at producer level, the lack of agreement is greater and therefore improvements need to be made if the FSA system is to replace the BPHS system. Supporting Information S1 Table Description of characteristics of the three data sources (FSA, BPHS, FarmFile) for pigs in Great Britain. (DOCX) Click here for additional data file. S2 Table The information within each dataset (FSA, BPHS and FarmFile) and how it was recategorised into ‘conditions’ for use in the population-level analysis. (DOCX) Click here for additional data file. S3 Table The information recorded by each set of assessors/inspectors and how it was recategorised into ‘conditions’ for use in the batch-level comparison. (DOCX) Click here for additional data file. We thank ADHB-Pork (formerly BPEX), APHA and FSA for the data provided for this analysis, and BPHS veterinary assessors and meat inspectors for their participation. ==== Refs References 1 Hoinville L.J. , Alban L. , Drewe J.A. , Gibbens J.C. , Gustafson L. , Häsler B. et al Proposed terms and concepts for describing and evaluating animal-health surveillance systems . Prev Vet Med 2013 , 112 , 1 –2 . 10.1016/j.prevetmed.2013.06.006 23906392 2 Grosbois V. , Häsler B. , Peyre M. , Hiep D.T. , Vergne T. A rationale to unify measurements of effectiveness for animal health surveillance . Prev Vet Med 2015 , 120 (1 ), 70 –85 . 10.1016/j.prevetmed.2014.12.014 25623970 3 Rodríguez-Prieto V. , Vicente-Rubiano M. , Sánchez-Matamoros A. , Rubio-Guerri C. , Melero M. , Martínez-López B. , et al Systematic review of surveillance systems and methods for early detection of exotic, new and re-emerging diseases in animal populations . Epidemiol Infect 2015 , 143 (10 ), 2018 –2042 . 10.1017/S095026881400212X 25353252 4 Dupuy C. , Morignat E. , Maugey X. , Vinard J-L. , Hendrikx P. , Ducrot C. , et al Defining syndromes using cattle meat inspection data for syndromic surveillance purposes: a statistical approach with the 2005–2010 data from ten French slaughterhouses . BMC Vet Res 2013 , 9 :88 10.1186/1746-6148-9-88 23628140 5 Stärk K.D.C. , Alonso S. , Dadios N. , Dupuy C. , Ellerbroek L. , Georgiev M. , et al Strengths and weakness of meat inspection as a contribution to animal health and welfare surveillance . Food Control 2014 , 39 , 154 –162 . 6 Thomas-Bachli A.L. , Pearl D.P. , Frienship R.M. , Berke O. Suitability and limitations of portion-specific abattoir data as a part of an early warning system for emerging disease of swine in Ontario , BMC Vet Res 2012 , 8 :3 10.1186/1746-6148-8-3 22225910 7 Vial F. , Reist M. Evaluation of Swiss slaughterhouse data for integration in a syndromic surveillance system . BMC Vet Res 2014 , 10 :33 10.1186/1746-6148-10-33 24479844 8 Stärk K.D.C. , Nevel A. Strengths, weaknesses, opportunities and threats of the pig health monitoring systems used in England . Vet Rec 2009 , 165 ,461 –465 . 19850852 9 Eze JI , Correia-Gomes C , Borobia-Belsue J , Tucker AW , Sparrow D , Strachan DW , et al Comparison of Respiratory Disease Prevalence among Voluntary Monitoring Systems for Pig Health and Welfare in the UK . PloS One 2015 , Volume 10 , Issue 5 , http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0128137 10 Harley S , More S , Boyle L , O' Connell N , Hanlon A . Good animal welfare makes economic sense: potential of pig abattoir meat inspection as a welfare surveillance tool . Ir Vet J 2012 , 65 , http://www.irishvetjournal.org/content/65/1/11. 11 Sanchez-Vazquez MJ , Nielen M , Edwards SA , Gunn GJ , Lewis FI . Identifying associations between pig pathologies using a multi-dimensional machine learning methodology . BMC Vet Res 2012 , 8 :151 10.1186/1746-6148-8-151 22937883 12 Gibbens JC , Robertson S , Willmington J , Milnes A , Ryan JB , Wilesmith JW , et al Use of laboratory data to reduce the time taken to detect new diseases: VIDA to FarmFile . Vet Rec 2008 , 162 (24 ):771 –776 . 18552327 13 Watson E. , Marier E. , Weston J. MC1001: Review of historic ante mortem and post mortem inspection data–Final Report Veterinary Laboratory Agency 2011 , pp 1 –118 . Sponsored by Food Standards Agency. Available: https://www.food.gov.uk/science/research/choiceandstandardsresearch/meatcontrolsprojects/mc1001 14 Brockmeier SL , Halbur PG , Thacker EL . Chapter 13: Porcine Respiratory Disease Complex In: Brogden KA , Guthmiller JM , editors. Polymicrobial Disease . Washington (DC) : ASM Press , 2002 . 15 Schrøder-Petersen D.L. and Simonsen H. B. Tail Biting in Pigs . The Veterinary Journal 2001 , 162 , 196 –210 . 11681870 16 European Commission. NUTS—Nomenclature of territorial units for statistics 2015. Available: http://ec.europa.eu/eurostat/web/nuts/overview. Accessed 9 October 2015. 17 Gross, J. and Ligges, U. nortest: Tests for Normality. R package version 1.0–4 2015. http://CRAN.R-project.org/package=nortest 18 Bates D , Maechler M , Bolker B and Walker S . Fitting Linear Mixed-Effects Models Using lme4 . J Stat Softw 2015 , 67 (1 ), 1 –48 . 19 Shoukri, M.M. and Pause, C.A. Chapter 3: Statistical analysis of cross classified data In: CRC Editors. Statistical Methods for Health Sciences, 2nd Edition, 1999. pp 93. 20 Evans JD . Straightforward Statistics for the Behavioural Sciences . Brooks/Cole Publishing ; Pacific Grove, Calif ,1996 . 21 Meyer, D., Zeileis, A., and Hornik, K. vcd: Visualizing Categorical Data. R package version 1.4–1 2015, https://cran.r-project.org/web/packages/vcd/index.html. 22 Nielsen S. , Nielsen G.B. , Denwood M.J. , Haugegaard J. , Houe H. Comparison of recording of pericarditis and lungs disorders at routine meat inspection with findings at systematic health monitoring in Danish finisher pigs . Acta Vet Scand 2015 , 57 :18 10.1186/s13028-015-0109-z 25887329 23 Sanchez-Vazquez MJ , Strachan WD , Armstrong D , Nielen M , Gunn GJ . The British pig health schemes: integrated systems for large-scale pig abattoir lesion monitoring . Vet Rec 2011 , 169 : 413 –U93 . 10.1136/vr.d4814 21881022 24 BPEX . The BPEX Yearbook 2013–2014: key industry statistics, pig performance data and details of knowledge transfer, research and development activity Agricultural and Horticulture Development Board , 2014 Available: http://pork.ahdb.org.uk/media/2420/the_bpex_yearbook_2013-2014.pdf 25 Dohoo, I., Martin, W., Stryhn, H. Chapter 2: Sampling. In: VER inc. Editors. Veterinary Epidemiologic Research, 2nd Edition, Charlottetown, 2009. pp 54. 26 Bonde M. , Toft N. , Thomsen P.T. , Sørensen J.T. Evaluation of sensitivity and specificity of routine meat inspection of Danish slaughter pigs using Latent Class Analysis . Prev Vet Med 2010 , 94 ,165 –169 . 10.1016/j.prevetmed.2010.01.009 20132995 27 Huneau-Salaün A. , Stärk K.D.C , Mateus A. , Lupo C. , Lindberg A. , Le Bouquin-Leneveu S. Contribution of meat inspection to the surveillance of poultry health and welfare in the European Union . Epidemiol Infect 2015 , 143 , 2459 –2472 . 10.1017/S0950268814003379 25521240 28 Enøe C. , Christensen G. , Andersen S. , Willeberg P. The need for built-in validation of surveillance data so that changes in diagnostic performance of post-mortem meat inspection can be detected . Prev Vet Med 2003 , 57 ,117 –125 . 12581595 29 Alton G. , Pearl D.L. , Bateman K.G. , McNab B. , Berke O. Comparison of covariate adjustment methods using space-time scan statistics for food animal syndromic surveillance . BMC Vet Res 2013 , 9 :231 10.1186/1746-6148-9-231 24246040 30 Byrne A.W. , McBride S. , Lahuerta-Marin A. , Guelbenzu M. , McNair J. , Skuce R.A. , McDowell S.W.J. Liver fluke (Fasciola hepatica) infection in cattle in Northern Ireland: a large-scale epidemiological investigation utilising surveillance data . Parasit Vectors 2016 , 9 :209 10.1186/s13071-016-1489-2 27079910 31 Qekwana DN , Oguttu JW , Venter D , Odoi A . Disparities in Beef Tapeworm Identification Rates in the Abattoirs of Gauteng Province, South Africa: A Descriptive Epidemiologic Study . PLoSOne 2016 , 11 (3 ): e0151725 .

PMC005xxxxxx/PMC5001631.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756440510.1371/journal.pone.0161866PONE-D-16-18849Research ArticleBiology and Life SciencesEvolutionary BiologyPopulation GeneticsHaplotypesBiology and Life SciencesGeneticsPopulation GeneticsHaplotypesBiology and Life SciencesPopulation BiologyPopulation GeneticsHaplotypesBiology and Life SciencesEvolutionary BiologyPopulation GeneticsBiology and Life SciencesGeneticsPopulation GeneticsBiology and Life SciencesPopulation BiologyPopulation GeneticsBiology and Life SciencesMolecular BiologyMolecular Biology TechniquesMolecular Biology Assays and Analysis TechniquesPhylogenetic AnalysisResearch and Analysis MethodsMolecular Biology TechniquesMolecular Biology Assays and Analysis TechniquesPhylogenetic AnalysisBiology and Life SciencesEcologyEcosystemsForestsEcology and Environmental SciencesEcologyEcosystemsForestsEcology and Environmental SciencesTerrestrial EnvironmentsForestsBiology and life sciencesMolecular biologyMolecular biology techniquesSequencing techniquesSequence analysisDNA sequence analysisResearch and analysis methodsMolecular biology techniquesSequencing techniquesSequence analysisDNA sequence analysisBiology and Life SciencesBiogeographyPhylogeographyEcology and Environmental SciencesBiogeographyPhylogeographyEarth SciencesGeographyBiogeographyPhylogeographyBiology and Life SciencesEvolutionary BiologyPopulation GeneticsPhylogeographyBiology and Life SciencesGeneticsPopulation GeneticsPhylogeographyBiology and Life SciencesPopulation BiologyPopulation GeneticsPhylogeographyBiology and life sciencesGeneticsDNAForms of DNAMitochondrial DNABiology and life sciencesBiochemistryNucleic acidsDNAForms of DNAMitochondrial DNABiology and Life SciencesMolecular BiologyMolecular Biology TechniquesArtificial Gene Amplification and ExtensionPolymerase Chain ReactionResearch and Analysis MethodsMolecular Biology TechniquesArtificial Gene Amplification and ExtensionPolymerase Chain ReactionGenetic Diversity and Structure among Isolated Populations of the Endangered Gees Golden Langur in Assam, India Genetic Studies in Golden LangurRam Muthuvarmadam S. 1Kittur Sagar M. 1Biswas Jihosuo 2Nag Sudipta 2Shil Joydeep 23Umapathy Govindhaswamy 1*1 Laboratory for the Conservation of Endangered Species, CSIR-Centre for Cellular and Molecular Biology, Uppal Road, Hyderabad 500007, India2 Primate Research Centre NE India, H/N 4, Byelane 3, Ananda Nagar, Pandu, Guwahati 781012, India3 Sálim Ali Centre for Ornithology and Natural History, Anaikatty, Coimbatore 641108, IndiaChiang Tzen-Yuh EditorNational Cheng Kung University, TAIWANCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: GU JB. Data curation: GU MSR SMK. Formal analysis: GU MSR SMK. Funding acquisition: JB. Investigation: MSR SMK SN JS. Methodology: GU JB MSR. Project administration: GU JB. Resources: GU JB. Supervision: GU JB. Validation: MSR SMK. Writing – original draft: GU JB MSR SMK. Writing – review & editing: GU JB MSR. * E-mail: guma@ccmb.res.in26 8 2016 2016 11 8 e016186617 5 2016 13 8 2016 © 2016 Ram et al2016Ram et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Gee’s golden langur (Trachypithecus geei) is an endangered colobine primate, endemic to the semi-evergreen and mixed-deciduous forests of Indo-Bhutan border. During the last few decades, extensive fragmentation has caused severe population decline and local extinction of golden langur from several fragments. However, no studies are available on the impact of habitat fragmentation and the genetic diversity of golden langur in the fragmented habitats. The present study aimed to estimate the genetic diversity in the Indian population of golden langur. We sequenced and analyzed around 500 bases of the mitochondrial DNA (mtDNA) hypervariable region-I from 59 fecal samples of wild langur collected from nine forest fragments. Overall, genetic diversity was high (h = 0.934, π = 0.0244) and comparable with other colobines. Populations in smaller fragments showed lower nucleotide diversity compared to the larger forest fragments. The median-joining network of haplotypes revealed a genetic structure that corresponded with the geographical distribution. The Aie and Champabati Rivers were found to be a barrier to gene flow between golden langur populations. In addition, it also established that T. geei is monophyletic but revealed possible hybridization with capped langur, T. pileatus, in the wild. It is hoped that these findings would result in a more scientific approach towards managing the fragmented populations of this enigmatic species. Department of Science and Technology, Government of IndiaSERB project # SR/SO/ AS - 17/2012Biswas Jihosuo This work was supported by the Department of Science and Technology, Government of India, Science and Engineering Research Board, grant # SR/SO/ AS - 17/2012 to JB. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll sequences are available from the GenBank database (accession number(s) KX189494-KX189516).Data Availability All sequences are available from the GenBank database (accession number(s) KX189494-KX189516). ==== Body Introduction Gee’s golden langur (Trachypithecus geei), a colobine primate endemic to Indo-Bhutan border, was discovered in the late 1950s based on morphological differences with the capped langur (T. pileatus) [1,2]. It has been listed as endangered species in the IUCN Red List [3] and Schedule-I species in Indian Wildlife Protection Act (1972). Golden langur is found in the sub-tropical, monsoon-fed, semi-evergreen and mixed deciduous forests of western Assam in India and south-central Bhutan. It is considered to be one of the most restricted-range primates of South Asia [4–6]. Its distribution in India is confined between the rivers Manas in the east, Sankosh in the west, Brahmaputra in the south and high mountain ridges upto 2300 m altitude in the north, and in Bhutan it is found between Sankosh-Chamke-Mangde-Manas river complex [4–11]. The golden langur shares its habitat with other non-human primates such as the slow loris (Nycticebus bengalensis), Assamese macaque (Macaca assamensis) and rhesus macaque (Macaca mulatta). However, its distribution range rarely overlaps with that of the capped langur (Trachypithecus pileatus) but hybrids between golden and capped langur have been reported in Bhutan [11]. In Bhutan, the golden langur is mostly found in three Protected Areas viz. Black Mountain National Park (NP) (1,730 km2), Royal Manas NP (1,033 km2) and Phibsoo Wildlife Sanctuary (WLS) (266 km2) [12], whereas in India, except Manas NP (500 km2) and Chakrashila WLS (45.58 Km2), the major populations are found in Reserve Forests (RFs), Proposed Reserve Forests (PRFs) and Unclassified State Forests with a little or no protection. In the last few decades due to forest fragmentation and degradation, its habitat in India has been reduced by more than 30% leading to severe population fragmentation [4,5]. Of the estimated 6,500 individuals of Indian populations, 93% occur in three large Reserve Forests (Chirang, Manas and Ripu) and the western part of Manas NP, and the remaining occur in several small isolated fragments [13–17]. Recently, local population extinctions have been reported from seven isolated fragments [6]. Habitat fragmentation consistently has large, negative effects on measures of biodiversity such as population size, distribution and genetic diversity, especially in habitat specialist species [18–22]. Low genetic variation, in turn, results in reduction of survival and reproduction [23,24] which increases the probability of extinction [25–27]. To estimate genetic variation and structure within wild populations of primates, many studies have focussed on the use of mtDNA control region (D-loop), particularly the hypervariable regions I and II (HVRI and HVRII) because of their high rate of mutation and resistance to selection pressures [28–40]. Moreover, non-invasive samples such as feces and hair usually yield low-quality DNA and working with mtDNA is easier and more reliable due to their availability in high copy number in cells. The hypervariable region has also been used to study the evolution of pelage coloration in colobine species [41]. Genetic data generated so far on golden langur was intended to examine its taxonomic placement within the langur group in the subfamily Colobinae using mitochondrial cytochrome b gene (cytb) [10,42,43]. Interestingly, none of the above studies has used samples from wild populations of golden langur in India. Its placement in the colobine tree is of particular importance because of its presence where the ranges of the genera Semnopithecus and Trachypithecus meet. Furthermore, no information is available on the genetic diversity of fragmented populations of golden langur. The present study aims to estimate the genetic diversity in the Indian populations of golden langur using HVRI. It also aims to examine the phylogenetic position of Indian golden langur with respect to the Bhutanese population and closely related Trachypithecus spp. using cytb. It is hoped that this study would serve as a platform for a more scientific approach towards managing the wild population of this enigmatic species. Methods Study area Eight forest fragments in Assam that vary in size, time since isolation and disturbance level (Table 1) were part of the study, which include Chirang RF (CR), Manas NP (MN), Chakrashila WLS (CS), Bamungaon PRF (BG), Nayekgaon PRF (NY), Nadangiri RF (ND), Bhairabi Hill PRF (BH) and Kakoijana RF (KJ) (Fig 1). The fragment size ranged from eight km2 (NY) to 592.5 km2 (CR) and are inhabited by as few as 25 golden langur individuals (BG) to more than 1450 individuals (CR) [13–15]. Apart from these eight forest fragments, the study included the temple group in Umananda (UM), a small river island in Brahmaputra River which lies outside the natural range of golden langur. The temple group was founded by two individuals that are believed to be brought from Bhutan [17,44]. 10.1371/journal.pone.0161866.g001Fig 1 Study area and sampling locations. Current and past distribution of golden langur in Assam, India based on previous reports [4] and sampling locations in the present study. Umananda (UM) is not shown in this map. Modified and reprinted from Choudhury et al 2002 [4] under a CC BY license, with permission from Zoos’ Print Journal, Coimbatore, India, original copyright 2002. 10.1371/journal.pone.0161866.t001Table 1 Study sites and their isolation status. Sl no. Fragment Area (km2) GPS Coordinates Population Isolation status Disturbance 1 Chirang (CR) 592.5 26°37’39.4”N 90°17’21.5”E >1450 Isolated since 1990s after losing about 30% of the habitat. Little or no disturbance 2 Manas (MN) 352 26°47’26.2”N 90°57’27.7”E >200 Isolated since 1990s and is contiguous with Royal Manas NP, Bhutan. Little or no disturbance 3 Chakrashila (CS) 45.6 26°16’50.7”N 90°20’44.1”E >500 Isolated since 1990s from adjoining Nadangiri and Nayekgaon PRF Felling in the fringe areas 4 Bhairabi Hill (BH) 36 26°18’27.8”N 90°32’06.6”E 26 Isolated since 1970s from adjoining Nakkati RF Heavily degraded 5 Kakoijana (KJ) 17 26°25’41.9”N 90°39’15.9”E 144 Isolated since 1970s Heavily degraded 6 Bamungaon (BG) 10.5 26°21’34.8”N 90°40’16.7”E 25 Isolated since 1970s Heavily degraded 7 Nadangiri (ND) 10.2 26°25’08.5”N 90°21’36.0”E >20 Isolated since 1990s from adjoining Nayekgaon PRF and Chakrashila Partly degraded 8 Nayekgaon (NY) <8 26°21’49.3”N 90°23’10.3”E 112 Isolated since 1990s from adjoining Nadangiri and Chakrashila Partly degraded 9 Umananda (UM) 0.049 26°11’47.8”N 91°44’43.4”E 7 Founder individuals introduced in 1993 - Sample Collection Sample collection was performed without direct interaction with the golden langurs and without causing any disturbance to their habitat. Samples were collected only from government-owned lands. Permission to collect fecal samples from Assam state were obtained from the Principal Chief Conservator of Forests (Wildlife), Assam Forest Department (Order No. 336, dated 16th March, 2013) and Assam State Biodiversity Board (Letter No. ABB/Permission/2012/469, dated 31st August 2015). Fresh fecal samples were collected between 29/06/2014 and 16/12/2015 from wild langurs after watching them defecate. Care was taken not to collect multiple samples from the same individual. A total of 71 samples were collected from CR (N = 15), CS (N = 12), KJ (N = 10), ND (N = 9), NY (N = 6), BH (N = 6), UM (N = 5), MN (N = 4) and BG (N = 4). Approximately 5g of the samples were stored in small plastic bottles using the two-step method [39]. In short, the samples were first soaked in 90% ethanol for 24 hours, dried, stored with silica gel and transported to the lab as soon as possible. DNA isolation, amplification and sequencing Fecal material was dried overnight in hot air oven at 50°C to remove any moisture. 0.2g of the completely dried fecal material was scrapped from the surface and used for extraction. Genomic DNA was extracted from the samples using Qiagen Stool Kit following the manufacturer’s protocol. The extracted DNA was stored in the elution buffer provided in the extraction kit. DNA quantification was done using Nanodrop-Spectrometer. To avoid amplification of nuclear inserts of mitochondrial DNA fragments (numts), a PCR was performed with universal vertebrate mitochondrial cytochrome b (cytb) forward primer [45] and 16022R reverse primer [46] which generated a 1500 bp PCR product. A nested PCR with PresLoopF [46] and 16220R primers was then performed, which generated PCR products of 690 bp of partial control region, specifically the hypervariable region-I (HVRI), partial tRNA-Thr, and complete tRNA-Pro. Additionally, 450 bp of partial cytb was also amplified using the universal cytb primers [45]. PCRs were carried out with initial denaturation at 95°C for 5 min, 35 cycles each with initial denaturation at 95°C for 20 sec, annealing at 52°C for 30 sec and extension at 72°C for 30 sec followed by final extension at 72°C for 10 min. The total reaction volume was 15 μl with 1X BSA, 1X PCR buffer, 0.25 mM of each DNTP, 2.5 mM MgCl2, 0.25 μM each of forward and reverse primers and 0.75 units of Taq polymerase (ExTaq HS DNA polymerase, Takara Bio Inc.). Utmost precautions were taken while performing PCR to avoid contamination from an external source of DNA. All PCR reactions were carried out with a negative control. Pre- and post-PCR works were carried out at separate places and using separate micropipettes. The PCR products were visualised on a 2% agarose gel. DNA was eluted from the gel using Gel Elution Kit (Bioserve India). DNA sequencing was carried out using BigDye Terminator cycle sequencing Kit (Applied Biosystems) in ABI 3730XL sequencer. Sequencing was repeated twice in both the directions. A maximum 639 bases of a mitochondrial DNA fragment containing the partial tRNA-Thr sequence, complete tRNA-Pro sequence and 543 bases of HVRI were generated. Cytb sequences of a maximum 521 bases were generated from eight samples. Genetic analysis The sequences were assembled and manually edited to check for base-calling errors using CodonCode Aligner v.2.0.6 (http://www.codoncode.com/). The primer sequences along with low quality bases at the ends which were removed. CLUSTAL W with default settings as implemented in MEGA5 was used to realign the edited sequences. Diversity estimations were done using DnaSP v.5.10 [47]. Genetic diversity was estimated in terms of average haplotype diversity (h) and nucleotide diversity (π) for each fragment. Population pairwise FST were calculated from pairwise nucleotide differences between fragments and their statistical significance was checked using 10,000 permutations as implemented in Arlequin v.3.5.1.2 [48]. To represent the geographical distribution of the haplotypes and the mutational relationships of the haplotypes among the isolated populations, the software NETWORK v.5.0.0.0 (available at http://www.fluxus-engineering.com) was used to construct a median-joining haplotype network [49]. For determination of intra-specific genetic structure, all unique HVRI haplotype sequences of T. geei were used. Four orthologous sequences of other Trachypithecus spp. truncated from mitogenome sequences from GenBank nucleotide database were part of this 529 base dataset: two sequences of T. shortridgei (accession numbers HQ149048.1 and KP834334.1) and one sequence each of T. pileatus (acc. no. KF680163.1), and T. obscurus (acc. no. EU004477.1). Two sequences of Nasalis larvatus (acc. nos. DQ355298.1 and KM889667.1) were used as the outgroup. In addition to HVRI sequences, cytb sequences were used for ascertaining the phylogenetic position of Indian T. geei sequences with respect to T. geei from Bhutan, T. pileatus and T. shortridgei. For this purpose, apart from the aforementioned mitochondrial sequences, three partial cytb sequences of T. geei (acc. nos. EU519220.1, EU526384.1 and EU526385.1) and one of T. pileatus (acc. no. EU526386.1) of Bhutanese origin were also included in the analysis. The software jModelTest v.2.1.3 [50] was used to determine the simplest model of sequence evolution that best explains the nucleotide variations in the dataset. HKY+G model was selected based on interpretation of likelihood scores using Bayesian Information Criterion and Decision Theory. Bayesian Inference and Maximum Likelihood trees were reconstructed using Mr.Bayes v.3.1.2 [51] and raxmlGUI [52], respectively. For Bayesian Inference tree reconstructions, two parallel MCMC runs of two million generations were performed with three heated chains and one cold chain each. Trees were sampled every 100 generations. The split frequency of standard deviation, that was calculated once every 1,000 generations, was used to assess the convergence of the two runs. The first 25% of generations were discarded as burn-in. The uncorrelated potential scale reduction factor (PSRF) values for all parameters were confirmed to be approximately equal to one. The trees were summarized to give a consensus tree with posterior probability values for each branch. Maximum Likelihood (ML) tree reconstruction was carried out in raxmlGUI with 10,000 rapid bootstrap replications using GTRGAMMA as the substitution model. Results Genetic diversity The possibility of amplifying numts was greatly reduced by the fact the all our DNA was extracted from fecal samples, which are naturally enriched with mtDNA molecules due to their small size, high copy number and lower vulnerability to degradation compared to nuclear DNA [53]. Moreover, all our PCR products were of the expected length and no extremely variant sequences were detected. Furthermore, during an initial tree construction with all available Trachypithecus sequences in GenBank, our sequences clustered with authentic mtDNA sequences instead of confirmed pseudogenes. No stop codons were detected in the partial cytb sequences. Out of 71 samples collected, 59 samples gave successful amplification for HVRI and 518 bases were sequenced and aligned. Nineteen haplotypes were revealed based on 47 mutations (45 transitions and two transversions) spread over 46 segregating sites, out of which 43 were parsimony-informative. Haplotype sequences were deposited to GenBank (accession numbers KX189494-KX189512). The overall haplotype diversity was high (h = 0.938) but the nucleotide diversity was low (π = 0.02443). The haplotypes were assigned an alphabetical code (from GL-A to GL-S) based on the chronology of discovery. Haplotype GL-C was the most frequent, represented by eight individuals from two fragments, followed by GL-A represented by seven individuals. More than half of the haplotypes represented fewer than three individuals. Haplotype and nucleotide diversities were calculated for each population (Table 2). CS and BH showed the highest number of haplotypes (four each) and CR showed the second highest (three haplotypes). Consequently, CS and BH also had the highest haplotype diversities (0.76 and 0.8 respectively). Five fragments (KJ, NY, ND, MN and BG) showed two haplotypes while UM showed no variation and KJ, ND and BG showed the least number of polymorphic sites among their haplotypes. Surprisingly, NY, which had only two haplotypes, showed the most number of polymorphic sites (17) for any fragment, giving it the highest nucleotide diversity. It was followed by CR (14), CS (7), MN (7) and BH (6). KJ had the lowest nucleotide diversity (π = 0.000483). 10.1371/journal.pone.0161866.t002Table 2 Fragment-wise estimates of genetic diversity in the golden langur. Sl. no. Fragment name Samples collected Sequences analyzed Haplotypes Polymorphic sites Haplotype diversity h (SD) Nucleotide diversity π (SD) 1 Chirang (CR) 9 7 3 14 0.524 (0.209) 0.00919 (0.00436) 2 Manas (MN) 4 4 2 7 0.5 (0.265) 0.00676 (0.00517) 3 Chakrashila (CS) 12 11 4 7 0.764 (0.099) 0.00527 (0.0007) 4 Bhairabi Hill (BH) 6 6 4 6 0.8 (0.172) 0.00541 (0.00156) 5 Kakoijana (KJ) 10 8 2 1 0.25 (0.18) 0.00048 (0.00035) 6 Bamungaon (BG) 4 3 2 1 0.667 (0.314) 0.00129 (0.00061) 7 Nadangiri (ND) 9 5 2 1 0.6 (0.175) 0.00116 (0.00034) 8 Nayekgaon (NY) 6 6 2 17 0.6 (0.0129) 0.01969 (0.00424) 9 Umananda (UM) 5 5 1 0 0 0 Genetic structure A median-joining haplotype network was drawn based on the HVRI sequence data (Fig 2). The 19 haplotypes differed from each other by between 1 and 42 bases. Three out of the 19 haplotypes were shared between utmost two fragments- haplotype GL-B was shared between NY and ND, GL-C was shared between NY and UM, and GL-D was shared between CS and BG. The haplotype network revealed no clear-cut structure, but there are at least four distinct groups- (i) all CS haplotypes along with two out of the three CR haplotypes and two other haplotypes from BG and NY/UM, (ii) Two ND haplotypes and the remaining CR haplotype, (iii) all BH and KJ haplotypes and (iv) all MN haplotypes (Fig 2). 10.1371/journal.pone.0161866.g002Fig 2 Median-joining network. Median-joining network of mtDNA HVRI haplotypes of golden langurs. The size of each circle represents the frequency of each haplotype (N = 59) and the colours represent each fragment. Each bar on the lines connecting two haplotypes represents one mutational step. The pairwise FST were calculated between each fragment and statistical significance assessed with 10,000 permutation steps (Table 3). NY was not significantly different from ND, UM and CR, and CS and BG were not significantly different from each other. Of the groups that were significantly different, NY and BG were most similar, followed by CS and UM, and CS and NY. The most dissimilar groups were KJ and UM. 10.1371/journal.pone.0161866.t003Table 3 Pairwise FST values between fragments (below diagonal) and their significance (above diagonal) after 10,000 permutations. CR MN CS BH KJ BG ND NY UM CR 0 + + + + + + - + MN 0.76278 0 + + + + + + + CS 0.54884 0.90687 0 + + - + + + BH 0.51405 0.89296 0.81496 0 + + + + + KJ 0.67172 0.95875 0.87668 0.50222 0 + + + + BG 0.52859 0.92501 0.27654 0.89701 0.97673 0 + + + ND 0.45262 0.92393 0.86979 0.84401 0.97244 0.95869 0 - + NY 0.1784 0.72321 0.40284 0.5 0.66239 0.27427 0.32999 0 - UM 0.60446 0.95098 0.39934 0.89701 0.98876 0.91477 0.98171 0.3617 0 ‘+’ indicates significant FST (p<0.05) and ‘-’ indicates not significant. The 19 HVRI haplotypes of T. geei along with GenBank sequences of T. pileatus, T. shortridgei and T. obscurus, and Nasalis larvatus as the outgroup were used to reconstruct phylogenetic trees. Both Bayesian Inference and Maximum Likelihood trees gave almost the same tree topology and similar statistical support for most clades (Fig 3). In both trees, T. geei emerged as a monophyletic clade with very high posterior probability/bootstrap value and showed most recent divergence from the lineage leading to T. pileatus and T. shortridgei. Golden langurs fell into two distinct clades, one containing the two haplotypes from MN and the other containing all the remaining haplotypes. Both these clades were statistically well-supported. In the second clade, haplotypes GL-B (NY and ND) and GL-M (ND) were basal to the rest of the 15 haplotypes of which 14 fell into two clades of high statistical support, and one (GL-O from CR) remained unresolved. The BH-KJ clade showed a typical stepping stone-like progression from BH haplotypes to the KJ haplotypes, with high support for all nodes but one. On the other hand, the CR-CS clade also showed a similar progression from CR haplotypes to the rest. However, the basal node had low support and the derived sequences fanned out instead of being strictly dichotomous, but the nodes had high support. 10.1371/journal.pone.0161866.g003Fig 3 Bayesian Inference tree of hypervariable region-I. Bayesian inference tree reconstructed from 518 bases of HVRI. Leaves are labelled with the species and accession codes in parentheses and, for golden langurs, haplotype codes followed by the fragment(s) of origin in parentheses. The numbers at each node are the Bayesian posterior probabilities followed by bootstrap values of the corresponding nodes (where present) in the maximum likelihood tree. We also found that the cytb sequences obtained from T. geei formed four haplotypes (named GL-A, GL-D, GL-N and GL-R, according to their corresponding HVRI haplotypes). Haplotype sequences were deposited to GenBank (accession numbers KX189513-KX189516). To check for the taxonomic placement of T. geei, a dataset containing 392 bases of partial cytb sequences was used in phylogenetic tree reconstruction. T. geei, T. pileatus and T. shortridgei clustered together with very high support, as expected, but the relationship among the three remained unresolved (Fig 4). Our sequences of T. geei clustered with the three sequences of Bhutanese T. geei. The haplotype GL-R from Manas had the same sequence as the Bhutanese T. geei. Curiously, one of the T. pileatus sequences (sampled from Bhutan by Wangchuk et al., 2008 [43]) also clustered with this T. geei clade (posterior probability of 1), whereas the other T. pileatus sequence [54] separated out. 10.1371/journal.pone.0161866.g004Fig 4 Bayesian Inference tree of cytochrome b. Bayesian inference tree reconstructed from 392 bases of cytb. Leaves are labelled with the species and accession codes in parentheses and, for golden langurs, haplotype codes. The numbers at each node are the Bayesian posterior probability values. Sequences generated in the present study and by other authors have been indicated. Discussion Genetic diversity in the fragmented forests While all previous genetic work carried out on golden langur were aimed at addressing questions regarding its taxonomic placement among congeneric species [10,42,43], the present study aimed to estimate its genetic diversity. To achieve this, 59 golden langur individuals were sampled from nine sites of varying sizes and times since isolation, and a 518 base segment of HVRI was sequenced. Golden langur showed a very high haplotype diversity (h = 0.938) and a moderately high nucleotide diversity (π = 0.0244) compared to other Asian colobines. For example, the critically endangered white-headed langur, T. leucocephalus (h = 0.57, π = 0.00323) and Guizhou snub-nosed monkey, Rhinopithecus brelichi (h = 0.457, π = 0.014) showed much less genetic diversity even when a much higher percentage of their population was sampled [36,40]. These species have severely depleted populations due to habitat loss and have an extremely small geographical distribution. In spite of golden langur’s small geographical footprint, its overall genetic diversity is comparable to other endangered Asian colobines with wider distribution like Sichuan snub-nosed monkey, R. roxellana (h = 0.845, π = 0.034) [55], Yunnan snub-nosed monkey, R. bieti (h = 0.944, π = 0.034) [31,56] and proboscis monkey, Nasalis larvatus (h = 0.9, π = 0.022) [35]. Notably though, the nucleotide diversity in golden langur is comparatively low, perhaps because of a more recent divergence from its sister species compared to Rhinopithecus spp. [36,43]. Golden langur, once found in large populations, now occurs in several isolated populations. The highest haplotype diversities were observed in Bhairabi Hill and Chakrashila fragments (30–50 km2), whereas lowest was in Kakoijana (10 km2). Interestingly two of the largest (>350 km2) fragments showed lower haplotype diversities. This might be due to over-sampling of major haplotypes. Nucleotide diversities showed greater correspondence to the fragment areas with larger fragments showing higher nucleotide diversities compared to smaller fragments, with the exception of Nayekgaon, thus reflecting the greater gene flow expected in large, continuous forests. Nayekgaon’s unusually high nucleotide diversity may be due to the presence of individuals from divergent populations which originated from adjoining fragments which it was connected with until the 1990s. Kakoijana, although not the smallest fragment, has the lowest haplotype and nucleotide diversities, which could be due to relatively large geographical and temporal isolation from adjoining forest (Fig 1, Table 1). Umananda population, being very recently founded by two individuals that were donated by devotees, showed no variation. Overall nucleotide diversity increased with increases of population and area of the fragment, while haplotype diversity did not, however more samples from more groups in larger fragments are required to ascertain the relationship. Genetic structure in the golden langur To examine the relationship among fragmented populations, we constructed a median-joining haplotype network (Fig 2) and phylogenetic trees (Fig 3) using HVRI haplotypes. The haplotypes fall broadly into four groups and two of them (MN and BH-KJ) are well-defined. The Aie River, one of the largest rivers to occur in the distributional range of golden langur (Fig 1), separates the Manas RF and Manas NP from the other forests. Therefore, it should not come as a surprise that haplotypes from MN are the most distant. The perennial Champabati River (Fig 1) also creates a barrier to westward movement of golden langurs from BH and KJ, meaning that any movement from this area to CS, NY, ND or CR can be achieved only by circumventing it. The UM temple group shared its haplotype with CS indicating that the founder individuals were from CS. In general, the haplotype network corresponded well with the geographical origins of the samples. However, there were some discrepancies. Firstly, NY contains two haplotypes that differ by a massive 17 nucleotide bases. While one of these is shared with ND and falls in CR group, the other is shared with UM and falls in the CS group (Fig 2). This is evidence of past connectivity of CR with NY, ND and CS and gene flow among them. This argument is bolstered by the fact that there are no natural barriers that occur between CS-NY-ND and CR. It is likely, therefore, that recent anthropogenic modification of the habitat has isolated CS-NY-ND from CR. Secondly, BG which is the easternmost small isolated fragment shares haplotypes with the distant CS fragment. This might be due to introduction of individuals from CS, but this needs further detailed investigation. The phylogenetic trees reconstructed with HVRI haplotypes (Fig 3) revealed an early divergence of golden langur into two clades separated by the Aie River (MN to the east and the rest to the west of Aie). In addition, it was found that in the west clade, NY, ND and CR haplotypes were basal to the remaining 14 haplotypes which form a clade. The area encompassing NY, ND and CR could be considered as the putative place of origin of the west clade. In late 1960s the area was more or less interconnected with each other but now fragmented in to Gaourang RF, Nadangiri RF and Nayekgaon PRF. The barrier (most probably the Champabati River) between CS-NY-ND-CR and BH-KJ seems to have been formed after an early expansion of golden langurs from near ND. After the barrier was formed, there seems to have been a range expansion through migration from ND-CR to CS and from BH to KJ, as evidenced from the stepping-stone-like progression. Golden langur phylogeny and hybridization Gee’s golden langur appears to be monophyletic with respect to its sister species (Fig 3). The cytb tree (Fig 4), however, throws some light on the possibility of hybridization with T. pileatus, especially in Bhutan. A previous report from Bhutan did not find sequence variation between golden and capped langur [43]. This might be due to sampling of hybrid individuals. Golden langur is often confused with capped langur owing to their variable coat morphology [1], seasonal variations in coat colour [2] and difficulty in observing the individuals through thick canopies. Moreover, natural hybrids between golden langur and capped langur have also been observed in their range peripheries [11]. Ironically, though, our study also supports Wangchuk et al.’s [43] conclusion that the golden and capped langur individuals of Indian zoos sampled by Karanth [42] are hybrids with Semnopithecus entellus. However, future studies should use nuclear markers to determine the genetic structure of the paternal component in golden langur. Management recommendations Although a majority of the fragments showed moderate to high genetic diversity for small, isolated populations, some fragments showed very low genetic diversity, particularly Kakoijana which has a population of about 144 golden langurs. Kakoijana was connected to Nakkati and Bhairabi Hill until the 1970s. Steps should be taken to re-establish this link through canopy corridors to facilitate gene flow between Kakoijana and neighbouring fragments. The group in Bamungaon (25 individuals) seems to be founded by individuals from or near Chakrashila. This has to be ascertained to prevent any inadvertent genetic intermixing during future repopulations. Most forest fragments lost their connectivity between 1970s and 1990s. These fragments should be protected from further degradation and steps may be taken to connect them to maintain genetic diversity. Golden langurs are monophyletic, and it appears that they are capable of hybridizing with capped langurs in the wild and Hanuman langurs in captivity. In fact, some of the individuals kept in Indian zoos may be hybrids. Therefore, the genetic affiliation of captive golden langurs has to be ascertained before they are used in captive breeding programs and reintroductions. We are thankful to the Department of Environment and Forest, Govt of Assam particularly PCCF, Wildlife Mr. S. Chand, Council Head of the Department, Forest etc., BTC Mr. G.C. Basumatary, Director, Manas TR Mr. A. Sargiary and others for providing necessary permission and other logistic support. We specially thank Prof. S.M. Mohnot, Dr. Nabajit Das, Mr. Pankaj Milli, Mr. Dharmeswar Rabha, Mr. Sunil Basumatary and Mr. Ratneswar Mushahary for their constant encouragement and support in the field. We also indebted to all members of Primate Research Centre, NE India for their constant support and kind help during the survey and sample collection. This study was supported by DST (SERB project # SR/SO/ AS—17/2012) to Dr. Jihosuo Biswas ==== Refs References 1 Khajuria H . A new langur (Primates: Colobidae) from Goalpara District, Assam . Annals and Magazine of Natural History . 1956 ;12 (9 ):86 –8 . 10.1080/00222935608655728 2 Gee EP . The distribution and feeding habit of the golden langur, Presbytis geei Gee (Khajuria, 1956) . Journal of the Bombay Natural History Society . 1961 ;58 (1 ):1 –12 . 3 The IUCN Red List of Threatened Species. Version 2015–4. 4 Srivastava A , Biswas J , Das J , Bujarbarua P . Status and distribution of golden langurs (Trachypithecus geei) in Assam, India . American Journal of Primatology . 2001 ;55 :15 –23 . 10.1002/ajp.1035 11536313 5 Biswas J. Ecology and social behaviour of golden langur (Trachypithecus geei) Khajuria, 1956 [dissertation]. Gauhati: Gauhati University; 2002. 6 Choudhury AU . Golden langur, Trachypithecus geei threatened by habitat fragmentation . Zoo’s Print Journal . 2002 ;17 (2 ):699 –703 . 10.11609/JoTT.ZPJ.17.2.699-703 7 Mukherjee RP. Distribution and present status of the Golden langur, Presbytis geei, Khajuria, in some forests of Assam, India. Proceeding of Wildlife Workshop; 1980. p. 129–32. 8 Subba PB . The status and conservation of golden langur (Presbytis geei Khajuria, 1956) in the Manas National Parks, Bhutan . Tiger Paper . 1989 ;16 :16 –8 . 9 Wangchuk T , Kawamto Y , Aimi M . Discovery of a contact zone between golden langurs (Trachypithecus geei) and capped langurs (Trachypithecus pileatus) in Bhutan . American Journal of Primatology . 2001 ;54 Suppl 1 :40 . 10 Wangchuk T. Report on the status and distribution of golden langur (Trachypithecus geei) in Bhutan based on surveys between 1994–2000. Unpublished report to Nature Conservation Division, Royal Government of Bhutan, Thimphu; 2003. 11 Choudhury AU . Primates of Bhutan and observations of hybrid langurs . Primate Conservation . 2008 ;23 :66 –73 . 10.1896/052.023.0107 12 Wangchuk T . A census and the biogeography of golden langurs (Presbytis geei) in Bhutan Tigerpaper (FAO) 1995 . 13 Ghosh S. Report on the distribution and population status of golden langur (Trachypithecus geei) in Bodoland Territorial Council, Assam, India; 2009. 14 Biswas J, Ghose S, Sangma A, Ray PC. Conserving through communities: Adopting an inclusive approach of research, population monitoring, conservation education and extensionism on Golden langur in Manas Tiger cum Biosphere Reserve, India. Final Report. Primate Research Centre NE India, CEPF / ATREE, Kokrajhar WL Division; 2010 BTC Collaborative Project. No. PRCNE/Tecr-14. 15 Ghosh S , Ramesh C . Management Plan of Chakrashilla Wildlife Santuary Part-I & Part-II, 2009–2010 to 2018–2019. Department of Environment & Forest, Govt. Of Assam ; 2011 . 16 Horwich RH , Das R , Bose A . Conservation and the Current Status of the Golden Langur in Assam, India, with Reference to Bhutan . Primate Conservation . 2013 ;27 (1 ):77 –83 . 10.1896/052.027.0104 17 Medhi R , Chetry D , Bhattacharjee PC , Patiri BN . Status of Trachypithecus geei in a rubber plantation in Western Assam, India , 2004 ;25 (6 ):1331 –7 . 10.1023/B:IJOP.0000043965.38722.63 18 Andrén H . Effects of habitat fragmentation on birds and mammals in landscapes with different proportions of suitable habitat: a review . Oikos . 1994 ;71 :355 –66 . 10.2307/3545823 19 Bender DJ , Contreras TA , Fahrig L . Habitat loss and population decline: A meta-analysis of the patch size effect . Ecology . 1998 ;79 (2 ):517 –33 . 10.1890/0012-9658(1998)079[0517:HLAPDA]2.0.CO;2 20 Fahrig L . Relative effects of habitat loss and fragmentation on population extinction . The Journal of Wildlife Management . 1999 ;61 (3 ):603 –10 . 21 Fahrig L . Effects of habitat fragmentation on biodiversity . Annual Review of Ecology, Evolution, and Systematics . 2003 ;34 (1 ):487 –515 . 10.1146/annurev.ecolsys.34.011802.132419 22 Gibbs JP , Stanton EJ . Habitat fragmentation and arthropod community change: carrion beetles, phoretic mites, and flies . Ecological Applications . 2001 ;11 (1 ):79 –85 . 23 Frankham R . Inbreeding and extinction: a threshold effect . Conservation Biology . 1995 ;9 (4 ):792 –9 . 24 Reed DH , Frankham R . Correlation between fitness and genetic diversity . Conservation Biology . 2003 ;17 (1 ):230 –7 . 25 Saccheri I , Kuussaari M , Kankare M , Vikman P , Hanski I . Inbreeding and extinction in a butterfly metapopulation . Nature . 1998 ;392 (6675 ):491 –4 . 26 Westemeier RL , Brawn JD , Simpson SA , Esker TL , Jansen RW , Walk JW , et al Tracking the long-term decline and recovery of an isolated population . Science . 1998 ;282 (5394 ):1695 –8 . 10.1126/science.282.5394.1695 9831558 27 Dixon JD , Oli MK , Wooten MC , Eason TH , McCown JW , Cunningham MW . Genetic consequences of habitat fragmentation and loss: the case of the Florida black bear (Ursus americanus floridanus) . Conservation Genetics . 2007 ;8 (2 ):455 –64 . 10.1007/s10592-006-9184-z 28 Chu JH , Lin YS , Wu HY . Mitochondrial DNA Diversity in two populations of Taiwanese macaque (Macaca cyclopis) . Conservation Genetics . 2005 ;6 (1 ):101 –9 . 10.1007/s10592-004-7749-2 29 Mondolo L , Salzburger W , Martin RD . Phylogeography of Barbary macaques (Macaca sylvanus) and the origin of the Gibraltar colony . Proceedings of the National Academy of Sciences of the United States of America . 2005 ;102 (20 ):7392 –7 . 10.1073/pnas.0502186102 15870193 30 Smith DG , Mcdonough J . Mitochondrial DNA variation in Chinese and Indian Rhesus macaques (Macaca mulatta) . American Journal of Primatology . 2005 ;65 (1 ):1 –25 . 10.1002/ajp.20094 15645455 31 Liu Z , Ren B , Wei F , Long Y , Hao Y , Li M . Phylogeography and population structure of the Yunnan snub-nosed monkey (Rhinopithecus bieti) inferred from mitochondrial control region DNA sequence analysis . Molecular Ecology . 2007 ;16 (16 ):3334 –49 . 10.1111/j.1365-294X.2007.03383.x 17688537 32 Monda K , Simmons RE , Kressirer P , Su B , Woodruff DS . Mitochondrial DNA hypervariable region-1 sequence variation and phylogeny of the concolor gibbons, Nomascus . American Journal of Primatology . 2007 ;69 (11 ):1285 –1306 . 17455231 33 Blancher A , Bonhomme M ., Crouau-Roy B , Terao K , Kitano T , Saitou N . Mitochondrial DNA sequence phylogeny of 4 populations of the widely distributed cynomolgus macaque (Macaca fascicularis fascicularis) . Journal of Heredity . 2008 ;99 (3 ):254 –64 . 10.1093/jhered/esn003 18334507 34 Chaves PB , Alvarenga CS , Possamai CB , Dias LG , Boubli JP , Strier KB , et al Genetic diversity and population history of a critically endangered primate, the northern muriqui (Brachyteles hypoxanthus) . PloS One . 2011 ;6 (6 ):e20722 10.1371/journal.pone.0020722 21694757 35 Munshi-South J , Bernard H . Genetic diversity and distinctiveness of the proboscis monkeys (Nasalis larvatus) of the Klias Peninsula, Sabah, Malaysia . Journal of Heredity . 2011 ;102 (3 ):342 –6 . 10.1093/jhered/esr013 21414965 36 Yang M , Yang Y , Cui D , Fickenscher G , Zinner D , Roos C , Brameier, M. Population genetic structure of Guizhou snub-nosed monkeys (Rhinopithecus brelichi) as inferred from mitochondrial control region sequences, and comparison with R. roxellana and R. bieti . American Journal of Physical Anthropology . 2012 ;147 (1 ):1 –10 . 10.1002/ajpa.21618 21953032 37 Hagell S , Whipple AV , Chambers CL . Population genetic patterns among social groups of the endangered Central American spider monkey (Ateles geoffroyi) in a human-dominated landscape . Ecology and Evolution . 2013 ;3 (5 ):1388 –99 . 10.1002/ece3.547 23762523 38 Badhan A , Eichstaedt CA , Almond NM , Knapp LA , Rose NJ . Analysis of full-length mitochondrial DNA D-loop sequences from Macaca fascicularis of different geographical origin reveals novel haplotypes . Journal of Medical Primatology . 2015 ; 44 (3 ):125 –36 . 10.1111/jmp.12163 25707924 39 Ram MS , Marne M , Gaur A , Kumara HN , Singh M , Kumar A , Umapathy G . Pre-historic and recent vicariance events shape genetic structure and diversity in endangered lion-tailed macaque in the Western Ghats: Implications for Conservation . Plos One . 2015 ;10 (11 ):e0142597 10.1371/journal.pone.0142597 26561307 40 Wang W , Qiao Y , Pan W , Yao M . Low genetic diversity and strong geographical structure of the critically endangered white-headed langur (Trachypithecus leucocephalus) inferred from mitochondrial DNA control region sequences . PloS One . 2015 ;10 (6 ):e0129782 10.1371/journal.pone.0129782 26057239 41 Liu Z , Wang B , Nadler T , Liu G , Sun T , Huang C , et al Relatively recent evolution of pelage coloration in Colobinae: phylogeny and phylogeography of three closely related langur species . PloS One . 2013 ;8 (4 ):e61659 10.1371/journal.pone.0061659 23613895 42 Karanth P. Molecular systematics and phylogeography of the langurs of Asia [dissertation]. Albany (NY): University at Albany; 2000. 43 Wangchuk T , Inouye DW , Hare MP . The Emergence of an Endangered Species: Evolution and Phylogeny of the Trachypithecus geei of Bhutan . International Journal of Primatology . 2008 ;29 (3 ):565 –82 . 10.1007/s10764-008-9258-5 44 Menon V . First record of commensalism in golden langur . The Journal of Bombay Natural History Society . 1996 ;93 :284 –5 . 45 Verma SK , Singh L . Novel universal primers establish identity of an enormous number of animal species for forensic application . Molecular Ecology Notes . 2003 ;3 (1 ):28 –31 . 46 Meyer D , Rinaldi ID , Ramlee H , Perwitasari-Farajallah D , Hodges JK , Roos C . Mitochondrial phylogeny of leaf monkeys (genus Presbytis, Eschscholtz, 1821) with implications for taxonomy and conservation . Molecular Phylogenetics and Evolution . 2011 ;59 (2 ):311 –9 . 10.1016/j.ympev.2011.02.015 21333742 47 Librado P , Rozas J . DnaSP v5: a software for comprehensive analysis of DNA polymorphism data . Bioinformatics . 2009 ;25 (11 ):1451 –2 . 10.1093/bioinformatics/btp187 19346325 48 Excoffier L , Lischer HEL . Arlequin suite ver 3.5: a new series of programs to perform population genetics analyses under Linux and Windows . 2010 ;564 –7 . 10.1111/j.1755-0998.2010.02847.x 49 Bandelt HJ , Forster P , Röhl A . Median-joining networks for inferring intraspecific phylogenies . Molecular Biology and Evolution . 1999 ;16 (1 ):37 –48 . 10.1093/oxfordjournals.molbev.a026036 10331250 50 Posada D . jModelTest: Phylogenetic model averaging . Molecular Biology and Evolution . 2008 ;25 (7 ):1253 –6 . 10.1093/molbev/msn083 18397919 51 Ronquist F , Huelsenbeck JP . MrBayes 3: Bayesian phylogenetic inference under mixed models . Bioinformatics . 2003 ;19 (12 ):1572 –4 . 10.1093/bioinformatics/btg180 12912839 52 Silvestro D , Michalak I . RaxmlGUI: A graphical front-end for RAxML . Organisms Diversity and Evolution . 2012 ;12 (4 ):335 –337 . 10.1007/s13127-011-0056-0 53 Frantzen MAJ , Silk JB , Ferguson JW , Wayne RK , Kohn MH . Empirical evaluation of preservation methods for fecal DNA . Molecular Ecology . 1998 ;7 :1423 –8 . 9787450 54 Shi FL , Wang BS , Pan HJ . Complete mitochondrial genome of the capped langur (Trachypithecus pileatus) . Mitochondrial DNA: The Journal of DNA Mapping, Sequencing, and Analysis . 2015 ;26 (5 ):809 –10 . 10.3109/19401736.2013.855904 55 Li M , Liu Z , Gou J , Ren B , Pan R , Su Y , Funk SM , Wei F . Phylogeography and population structure of the golden monkeys (Rhinopithecus roxellana): inferred from mitochondrial DNA sequences . American journal of primatology . 2007 ;69 (11 ):1195 –209 . 17385699 56 Xiao W , Ding W , Cui LW , Zhou RL , Zhao QK . Habitat degradation of Rhinopithecus bieti in Yunnan, China . International Journal of Primatology . 2003 ;24 (2 ):389 –98 .

PMC005xxxxxx/PMC5001632.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756485710.1371/journal.pone.0160906PONE-D-16-03978Research ArticleBiology and Life SciencesBiomechanicsBiological LocomotionWalkingBiology and Life SciencesPhysiologyBiological LocomotionWalkingMedicine and Health SciencesPhysiologyBiological LocomotionWalkingMedicine and Health SciencesGeriatricsGeriatric RehabilitationMedicine and Health SciencesRehabilitation MedicineGeriatric RehabilitationEngineering and TechnologyElectronicsAccelerometersPeople and PlacesPopulation GroupingsAge GroupsElderlyMedicine and Health SciencesRehabilitation MedicineGait RehabilitationEngineering and TechnologyEquipmentMeasurement EquipmentMedicine and Health SciencesPublic and Occupational HealthPhysical ActivityPhysical FitnessExerciseMedicine and Health SciencesSports and Exercise MedicineExerciseBiology and Life SciencesSports ScienceSports and Exercise MedicineExerciseMedicine and Health SciencesPublic and Occupational HealthPhysical ActivityPromoting Activity in Geriatric Rehabilitation: A Randomized Controlled Trial of Accelerometry Promoting Activity in Geriatric Rehabilitationhttp://orcid.org/0000-0002-3038-8881Peel Nancye M. 1‡*Paul Sanjoy K. 2‡Cameron Ian D. 3Crotty Maria 4Kurrle Susan E. 5Gray Leonard C. 11 Centre for Research in Geriatric Medicine, The University of Queensland, Brisbane, Australia2 Clinical Trials & Biostatistics Unit, QIMR Berghofer Medical Research Institute, Brisbane, Australia3 John Walsh Centre for Rehabilitation Research, University of Sydney, Sydney, Australia4 Department of Rehabilitation and Aged Care, Flinders University, Adelaide, Australia5 Division of Rehabilitation and Aged Care, Hornsby Ku-ring-gai Hospital, Sydney, AustraliaTaheri Shahrad EditorWeill Cornell Medical College in Qatar, QATARCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: NMP SKP IDC MC SEK LCG. Data curation: NMP SKP. Formal analysis: NMP SKP. Funding acquisition: NMP SKP IDC MC SEK LCG. Investigation: NMP SKP IDC MC SEK LCG. Methodology: NMP SKP IDC MC SEK LCG. Project administration: NMP. Resources: NMP SKP LCG. Software: SKP. Supervision: NMP IDC MC LCG. Visualization: NMP SKP. Writing – original draft: NMP. Writing – review & editing: NMP SKP IDC MC SEK LCG. ‡ These authors are joint senior authors on this work. * E-mail: n.peel@uq.edu.au26 8 2016 2016 11 8 e01609066 2 2016 26 7 2016 © 2016 Peel et al2016Peel et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Background Low activity levels in inpatient rehabilitation are associated with adverse outcomes. The study aimed to test whether activity levels can be increased by the provision of monitored activity data to patients and clinicians in the context of explicit goal setting. Methods A randomized controlled trial in three sites in Australia included 255 inpatients aged 60 and older who had a rehabilitation goal to become ambulant. The primary outcome was patients’ walking time measured by accelerometers during the rehabilitation admission. Walking times from accelerometry were made available daily to treating therapists and intervention participants to motivate patients to improve incidental activity levels and reach set goals. For the control group, ‘usual care’ was followed, including the setting of mobility goals; however, for this group, neither staff nor patients received data on walking times to aid the setting of daily walking time targets. Results The median daily walking time in the intervention group increased from 10.3 minutes at baseline to 32.1 minutes at day 28, compared with an increase from 9.5 to 26.5 minutes per day in the control group. Subjects in the intervention group had significantly higher non-therapy walking time by about 7 minutes [mean (95% CI): 24.6 (21.7, 27.4)] compared to those in the control group [mean(95% CI): 17.3 (14.4, 20.3)] (p = 0.001). Conclusions Daily feedback to patients and therapists using an accelerometer increased walking times during rehabilitation admissions. The results of this study suggest objective monitoring of activity levels could provide clinicians with information on clinically important, mobility-related activities to assist goal setting. Trial Registration Australian New Zealand Clinical Trials Registry ACTRN12611000034932 http://www.ANZCTR.org.au/ Australian National Health and Medical Research CouncilAPP1007886Gray Leonard C Australian National Health and Medical Research CouncilPractitioner FellowshipCameron Ian D This work was supported by Australian National Health and Medical Research Council (NHMRC) Grant (APP1007886), https://www.nhmrc.gov.au/, LCG NMP SKP IDC SEK MC. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityData are available from the Queensland Clinical Trials and Biostatistics Unit, QIMR Berghofer Medical Research Institute, Brisbane, Australia 4006 for researchers who meet the criteria for access to confidential data. Potential exists to identify participants so that ethical and legal considerations restrict access to the data. Requests for the data may be sent to Sanjoy Paul (sanjoy.paul@qimrberghofer.edu.au).Data Availability Data are available from the Queensland Clinical Trials and Biostatistics Unit, QIMR Berghofer Medical Research Institute, Brisbane, Australia 4006 for researchers who meet the criteria for access to confidential data. Potential exists to identify participants so that ethical and legal considerations restrict access to the data. Requests for the data may be sent to Sanjoy Paul (sanjoy.paul@qimrberghofer.edu.au). ==== Body Introduction For frail older people, low levels of mobility during hospitalization are associated with functional decline and deconditioning (decline in muscle strength and bulk as a result of physical inactivity), leading to increased length of stay, post discharge readmission or transfer to permanent residential care [1]. Observational studies [2–4] have shown that activity levels in older people undergoing inpatient rehabilitation are low, with only 7% to 9% of a monitored eight hour period spent walking [3] and 65% of daytime hours either asleep or completely inactive [2]. The factors that contribute to low activity in rehabilitation have not been well studied. One survey of older patients in a post-acute hospital setting found that, whilst attitudes to exercise were generally positive, patients over-estimated the adequacy of their activity levels. They were unsure if they should be doing more exercise and only 11% recalled having been advised to exercise regularly by a health professional [5]. Subjective clinician assessment of the level of patient activity in a rehabilitation setting has also been shown to be inaccurate [6]. Rehabilitation for older people should have specific goals, set in conjunction with the patient, family members and multi-disciplinary team [7]. Goal-setting enhances both the process and outcome of clinical care and is a core practice in rehabilitation [8]. To restore physical function and independence in frail and deconditioned patients, it is important to set measurable, attainable mobility goals and to monitor progress carefully [9]. Monitoring and feedback are essential tenets in health promotion strategies to change behavior, since exercise is significantly influenced by self-efficacy (confidence in one's ability to exercise) and exercise outcome expectations [10]. The setting of activity targets (such as achieving a particular heart rate or step count) has been used to increase levels of physical activity in healthy older people [9]. However, measuring and targeting activity in older patients in rehabilitation settings using models of exercise prescription and monitoring developed for healthy individuals may not be appropriate or accurate [9]. Pedometers, widely used to promote physical activity [11] are inaccurate when assessing step counts as a measure of activity in elderly populations with varying levels of physical dysfunction and gait anomalies [12]. Advances in technology over the past decade have led to the development of wearable devices, such as accelerometers, with potential for continuous ambulatory activity monitoring of older adults in clinical settings [13–16]. Such devices have been used in research studies of gait and balance analysis for falls risk assessment and to detect posture changes for activity monitoring in groups such as amputees, medical and surgical patients, and those with Parkinson’s disease, diabetes, stroke and multiple sclerosis [14]. While the importance of physical activity in the functional recovery of older rehabilitation patients has been recognized [17], questions remain about the optimal methods to promote and monitor activity for frail older people in post-acute hospital settings. Based on the findings of a small feasibility study [18], this study was developed to evaluate the efficacy of using accelerometry to promote activity for older inpatients in rehabilitation settings. The primary aim was to test whether incidental activity levels can be increased by the provision of objectively measured activity data to patients and clinicians in the context of explicit goal setting. Secondary aims were to explore the effects of increased walking activity (if achieved) on patient outcomes. Materials and Methods Study Design and Setting The study design was a parallel group randomized controlled trial, complying with the recommendations from the CONSORT statement [19]. Before selecting this approach, a variety of designs, including matched controlled, before-after and cluster randomized trials were extensively considered, against a set of criteria which included feasibility, degree of contamination, cost and power. The setting was post-acute care Geriatric Rehabilitation Units or Geriatric Evaluation and Management Units with at least a 40 bed capacity, at three Australian sites. The trial was registered with the Australian New Zealand Clinical Trials Registry (registration number ACTRN12611000034932). Participants Patients admitted to post-acute care rehabilitation who were (1) aged 60 years and older; (2) able to ambulate independently or with supervision/assistance and had a rehabilitation goal to become ambulant within the context of the current admission and (3) expected to have a length of stay of at least two weeks, were eligible to participate. Exclusion criteria were those (1) with lower limb amputation; (2) with delirium or agitated dementia, as documented by the geriatric treating team; or (3) not expected to walk within four weeks of admission. The study was approved by the Human Research Ethics Committees at each site. Patients gave informed written consent to participate; in the event of incapacity to consent, assent for participation was sought from next of kin or carers. Study Protocol Accelerometers were used to monitor patients’ activity in both the intervention and control groups. For the intervention group, accelerometer data was downloaded daily. Feedback was provided to the intervention participant and their therapists of the previous day’s walking time in numerical and graphical form, showing walking time outside therapy sessions compared with walking target. The treating therapist, in consultation with the patient, set mobility goals, including provisional targets for daily walking time. These goals were reviewed weekly and modified, informed by the accelerometer data, to motivate the patient to improve incidental activity levels outside of therapy sessions and reach set targets. Walking times over a week were summarized in chart form and made available at the weekly case conference. All staff were trained in the use of accelerometry data and asked to encourage patients to meet their activity goals. For the control group, ‘usual care’ was followed, including the setting of mobility goals. However, neither staff nor patients received data on walking times to aid the setting of walking time targets. The monitoring period was four weeks from date of study entry, unless in the interim, the patient was discharged or unable to continue by virtue of a sudden change in condition which precluded mobility. Data Collection and Measures The primary outcome measure was walking time per day in minutes. Walking time was divided into time within and outside of therapy sessions, based on recorded therapy session times. Secondary outcome measures included lower extremity function and functional status, quality of life, length of monitoring period, and discharge destination measured at study exit and hospital readmissions at 28 day follow-up post study exit. Walking time was downloaded daily from the accelerometer by the research assistant responsible for activity monitoring. A variation in protocol occurred with change in the activity monitoring device. The accelerometer devices initially used in this study were triaxial ALIVE Heart and Activity Monitors, manufactured by Alive Technologies Pty. Ltd, Ashmore, Queensland, Australia. The device was fitted to a band worn around the waist. Based on validated algorithms [18], the accelerometer measured daily sitting, standing and walking times. A record was kept of periods when the accelerometer was not being worn, mainly at night, and calibrated with times when the signal indicated no activity. Due to difficulties with supply and servicing, the devices were changed part way through the study to ActivPal TM (PAL Technologies LTD, Glasgow, UK), a validated device [20] which classifies an individual's activity into periods spent sitting or lying, standing and walking. The ActivPal is attached to the front of the mid-thigh with waterproof tape and is capable of continuous monitoring for three to seven days. A comprehensive geriatric assessment supported by the interRAI Acute Care Post-Acute Care (AC-PAC) instrument was administered by a research nurse within three days of entry into the study, at the 14th day and at discharge or exit from the study. The interRAI instrument [21] measures a comprehensive set of items including patient demographics, cognition, mood, functional and mobility status, diagnoses and social support. A number of scales imbedded in interRAI instruments combine single items belonging to a domain, such as personal and instrumental activities of daily living (ADL, IADL), which were used to describe the presence and extent of deficits in that domain [21]. Lower extremity function was assessed by the research nurse at entry into the study, on the 14th day and at discharge or exit from the study, using the Short Physical Performance Battery (SPPB) [22]. The SPPB is a brief, quantitative estimate of future risk for hospitalization and decline in health and function, validated in clinical populations of older adults [23]. The SPPB score is based on timed measures of standing balance, walking speed, and ability to rise from a chair. Each of the three performance measures was assigned a score ranging from 0 to 4, with 4 indicating the highest level of performance and 0 the inability to complete the test. A summary score (range 0–12) is subsequently calculated by adding the three scores [24]. Health-related quality of life was measured using the EQ-5D three level, a concise measure widely used internationally across many different diseases and health states [25]. The output of the instrument is in the form of a utility value (between zero and one) representing quality of life at that point in time. These utilities can be used to estimate quality adjusted life years (QALYs), a common measure of health gain or benefit attributable to an intervention. The EQ-5D was administered by the research nurse at study entry and exit. A telephone follow-up by the research nurse at 28 days post exit from the study assessed current living arrangements (community, institutional care, died) and adverse outcomes such as readmissions to hospital. According to the approved research protocol, this information could be obtained from the carer, in the event of participant incapacity to respond. All data collection instruments were administered by trained assessors. Data Management An Electronic Case Report Form (eCRF) was developed and validated following Good Clinical Practice standards on Food and Drug Administration approved OpenClinica (www.openclinica.org), a clinical trial software platform for Electronic Data Capture (EDC). The eCRF development and online data capture was managed by a dedicated data manager from the Queensland Clinical Trials and Biostatistics Unit. Recruitment and Randomization Procedure At each site, at admission to post-acute rehabilitation, potentially eligible patients were identified by clinical staff and their names provided to the research assistant who obtained informed consent for participation for those who met eligibility criteria. A random number sequence was generated for the order of group allocation at each site. The randomization codes were generated by the Queensland Clinical Trials and Biostatistics Unit and placed in sealed envelopes. Patients who consented to participate were allocated a unique identification number (ID) and the research assistant opened the sealed envelope for that ID, which contained the randomization code for group allocation. Once opened, the envelope was dated, signed, securely stored and accessed only by the research assistant who was responsible for activity monitoring. A research nurse, trained in the use of assessment instruments and blinded to group allocation, conducted assessments. Rehabilitation staff were not be able to be blinded to the patient’s allocation. Statistical Methods The power analysis for this study was based on accelerometry pilot data of 60 patients [3] with a mean (Standard Deviation (SD)) daily walking time of 45 (51) minutes, correlation coefficient of 0.20 between baseline and 14 day follow-up measurements and a correlation coefficient of 0.60 between the follow-up measurements. Comparative power analyses were conducted for 14 to 20 days of possible repeated measures of daily activities. To observe an increase of activity by at least 15 minutes (33%) in the intervention group, 108 patients were required in each group with 14 days of measurements and 105 patients in each group with 20 days of measurements, with 80% power at two-sided 5% level of significance. During the conduct of the study, activity data were collected for a maximum of 28 days on 255 subjects, thereby significantly increasing the power of the study. The basic statistics on study parameters were presented by number (%), mean (SD) or median (Interquartile Range (IQR)), as appropriate. Intervention and control groups were compared using Chi-square or Fisher’s Exact tests for categorical variables and t-tests or Mann-Whitney U tests for continuous variables, depending on the distribution of the data, with p value <0.05 taken as the level of significance. The longitudinal trajectories of the daily in-therapy and non-therapy walking time over 28 days were compared between the treatment groups using the generalized estimating equation (GEE) approach, with normal distribution and identity link. The bootstrapped standard errors were obtained, and hence the bootstrapped 95% CI of walking time were presented. The GEE regression model based analyses were weighted by the use of two devices. The changes in the secondary outcome measures at discharge from admission were presented by mean and 95% confidence interval (CI), compared between the treatment groups. All analyses were conducted following the intention-to-treat approach. Analyses were performed using IBM SPSS Version 23.0 (Armonk, NY) and Stata Statistical Software, Release 14 (College Station, TX). Results A total of 270 subjects (90 from each of the three sites) were randomized equally between the intervention and control group. A total of 128 intervention and 127 control subjects received and continued to follow the study protocol during the course of the study (Fig 1). The median (IQR) length of the monitoring period was 14 (11–18) days, with 6% of patients completing 28 days of monitoring. There was no significant difference in accelerometer device used by treatment group with 141 (55%) of the participants using the ActivPal (71 intervention; 70 control). The daily monitoring period of ActivPal users was 24 hours, as the device was worn continuously. For the ALIVE Heart and Activity Monitor users, the median (IQR) daily hours of monitoring was 8.3 (6.7, 9.6) for the intervention group and 7.3 (6.0, 9.0) for the control group, with no significant difference between treatment groups (p = 0.07). 10.1371/journal.pone.0160906.g001Fig 1 Recruitment Flow Diagram. The mean (SD) age of subjects was 81 (8) years, 42% were male, 16% had BMI above 30 kg/m2, and the mean (SD) number of co-morbidities was 8 (4) at randomization. At admission, 84% (n = 215) subjects required supervision or person assistance for walking, and only 16% (n = 40) could walk unsupervised. Gait speed mean (SD) of 0.32 (0.25) m/sec was slow, characteristic of a population in a sub-acute setting with functional dependence and mobility disability [26]. The primary diagnosis in 88 (34.5%) patients was a fracture, most frequently hip fracture (n = 47). Other primary diagnoses were infections (n = 47), including pneumonia and urinary tract infections, neurological conditions (n = 29) including stroke, and cardio-pulmonary conditions (n = 25). There were no significant differences between intervention and control groups on baseline characteristics as shown in Table 1. 10.1371/journal.pone.0160906.t001Table 1 Baseline characteristics of the study subjects by treatment group. Intervention Control n = 128 n = 127 Age (years) mean (SD) 81 (9) 82 (8) Male n(%) 50 (39) 57 (45) BMI (kg/m2) mean (SD) 25.6 (6.7) 24.5 (5.2) BMI ≥ 30 kg/m2 n(%) 25 (20) 16 (13) Walking- Supervised or Person Assist n(%) 111 (87) 104 (82) Walking Without Aids n(%) 4 (3) 11 (9) SPPB median (IQR)a 2 (1, 4) 3 (1, 5) • Gait speed m/s mean (SD) 0.31 (0.24) 0.33 (0.27) Cognitive Function median (IQR) b 1 (0, 2) 1 (0, 2) Cognitive Function Score < 2 n(%)b 89 (70) 86 (69) ADL Scale median (IQR)c 10 (5, 13) 9 (5, 13) Number of Co-morbidities mean (SD) 8 (4) 8 (4) Primary Diagnosis n(%) • Fractures 46 (36) 33) • Infections 24 (19) 18) • Neurological 14 (11) 12) • Cardiopulmonary 16 (12) 9 (7) Notes: Abbreviations: BMI—Body Mass Index; SPPB- Short Physical Performance Battery; ADL—Activities of Daily Living; SD- Standard Deviation; IQR- Interquartile Range a Based on Short Physical Performance Battery range 0–12 with higher scores indicating better performance b Based on Cognitive Performance Scale range 0–6 with higher scores indicating greater incapacity c Based on Activities of Daily Living Scale (Long Form) range 0–28 with higher scores indicating greater dependence The average in-therapy and non-therapy walking time during the 28-day measurement period are presented in Figs 2 and 3. Subjects in the intervention groups had significantly higher non-therapy walking time by about 7 minutes [mean (95% CI): 24.6 (21.7, 27.4)] compared to those in the control group [mean (95% CI): 17.3 (14.4, 20.3)] (p = 0.001, Table 2). The separation of the non-therapy walking time between treatment groups was evident from day 3 post study initiation (Fig 2). There was an observed significant difference in average walking time during therapy (although this is unlikely to be clinically important) (intervention: 4.4 minutes, control: 3.7 minutes, p = 0.021) (Table 2). The median daily walking time in the intervention group increased from 10.3 minutes at baseline to 32.1 minutes at day 28, compared with increase in median walking times from 9.5 to 26.5 minutes per day in the control group. 10.1371/journal.pone.0160906.g002Fig 2 Average daily measures of non-therapy walking time by treatment groups. 10.1371/journal.pone.0160906.g003Fig 3 Average daily measures of in-therapy walking time by treatment groups. The measurements are based on GEE regression models, as described in the method section. 10.1371/journal.pone.0160906.t002Table 2 Mean (95% CI) of walking time by study group. Non-Therapy Walking Time In-Therapy Walking Time Mean (95% CI) p Mean (95% CI) p Intervention 24.6 (21.7, 27.4) 0.001 4.4 (4.0, 4.9) 0.021 Control 17.3 (14.4, 20.3) 3.7 (3.2, 4.1) Notes: Data are in minutes. Abbreviations: CI- Confidence Interval In regard to secondary outcomes (Table 3), the median (IQR) length of stay was 23 days (16, 35) and was not different between the treatment groups (p = 0.52). Overall, 16% (n = 41) were discharged to a higher level of care and 7.5% (n = 19) subjects were readmitted within 28 days of discharge. The proportions were not significantly different between the treatment groups (p = 0.63 and p = 0.46 respectively). The improvements in the short physical performance battery (SPPB) and the ADL scale between study entry and exit were not significantly different between the treatment groups. The average levels of individual components of EQ-5D scores were similar between the treatment groups. 10.1371/journal.pone.0160906.t003Table 3 Secondary outcomes at study exit. Outcome Intervention Control p Change in SPPBa mean (95% CI) 1.76 (1.33, 2.20) 1.64 (1.21, 2.07) 0.69 • Change in Gait Speed m/s mean (SD) 0.22 (0.21) 0.18 (0.25) 0.12 Change in ADL scaleb mean (95% CI) -5.59 (-6.42, -4.76) -4.69 (-5.50, -3.88) 0.13 EQ-5D Scoresc median (IQR) • Mobility 2 (1, 2) 2 (1, 2) 1.00 • Personal Care 1 (1, 2) 1 (1, 2) 1.00 • Usual Activities 2 (1, 2) 2 (1, 2) 1.00 • Pain Discomfort 2 (1, 2) 2 (1, 2) 1.00 • Anxiety / Depression 1 (1, 2) 1 (1, 2) 1.00 Length of monitoring period (days) median (IQR) 15 (12, 21) 14 (11, 17) 0.20 Discharged to higher level of cared n (%) 22 (17) 19 (15) 0.63 Readmitted within 28 days n (%) 8 (6) 11 (9) 0.46 Notes a Based on Short Physical Performance Battery range 0–12 with higher scores indicating better performance b Based on Activities of Daily Living Scale (Long Form) range 0–28 with higher scores indicating greater dependence c Based on EQ-5D Items are scored from 1 to 3, corresponding to 3 levels: no problems, some problems, extreme problems respectively. d Discharged to a higher level of care (eg admitted from community and discharged to residential care) With the exception of one patient from the control group who withdrew from the study after two hours because they could not tolerate wearing the accelerometer belt, no adverse events were recorded from the use of accelerometers. Discussion Patients in the intervention arm achieved significantly higher non-therapy walking time by 7.6 min/day on average, compared to the control group. Whether this increase in walking time confers a clinically meaningful benefit is debatable, although it is possible that any increase in walking relative to individualised baseline values could confer important health benefits [27]. The prognostic value of physical activity measures such as walking time has not yet been established in a rehabilitation setting, although correlations with physical performance measures such as gait speed have been demonstrated [3, 28]. Remote monitoring using wireless technology is seen as a possible supplementary measure to assess outcomes. No significant differences were found between treatment groups on secondary outcome measures, although this may be a reflection of the short timeframe of intervention (maximum four weeks) and follow-up (28 days post discharge). Lack of sensitivity to change could also be a factor in not observing differences in the secondary outcome measures, and because a number of physical capacity measures including components of the Short Physical Performance Battery (such as gait speed, chair rise time and balance tests) have floor effects [28]. To our knowledge, this is the first randomized trial that utilized accelerometers as a strategy to improve walking activity for the general population of older patients in a geriatric rehabilitation setting. Two trials to increase walking activity through monitoring and feedback in rehabilitation of stroke patients have recently been published [29, 30]. In contrast to our study, no significant increase in walking time in the intervention group was reported. It was suggested that the rehabilitation environment, patient fatigue and time allocated for other priorities limited opportunities to ambulate more frequently [29, 30]. Comparability of walking times achieved in our study with those in previous studies monitoring activity in geriatric rehabilitation using wearable devices [28–33] are problematic because of differences in population characteristics (including primary diagnosis and acuity), setting (acute, sub-acute or post-acute rehabilitation), period of monitoring and activity measured (eg step count, walking, or ‘uptime’ which includes standing and walking). For example, compared with a baseline median daily walking time of 10 minutes in the current study, daily walking activity varied from 4 minutes in a study of a comparable population referred for rehabilitation therapy [33], to 7 minutes in a study of patients rehabilitated following hip fracture [28] and 23 minutes in moderately impaired elderly stroke patients [31]. Increase in walking time over the duration of rehabilitation shows similar wide variation, depending on diagnostic group and period of monitoring. While it is widely believed that bed rest and inactivity in hospital are detrimental for mobility and function [34], there are currently no definitive clinical guidelines on optimum physical activity levels for older adults to guide clinicians in the management for older people admitted to rehabilitation [35]. Barriers to mobility during hospitalization of older patients that need to be taken into account in planning successful strategies include health problems, especially weakness, pain, and fatigue; being attached to a medical device such as intravenous drip or catheter; being concerned about falls; and lack of staff to assist with out-of-bed activity [36, 37]. Low mobility among hospitalized older adults has also been attributed to lack of patient motivation [36] and environmental factors such as hospital traffic, noise and clutter that present physical barriers to ambulation [37] and lack of places to go in hospital environs (i.e. patients are not motivated to move) [33]. Strengths and Limitations Compared to previous studies, the sample size was large and sufficiently powered to detect changes in the primary outcome measure of difference in walking time between intervention and control groups. Because physical therapy sessions could reflect on time spent walking [33], the walking time in this study was measured for in-therapy and non-therapy times, since the aim of the study was to motivate patients to increase incidental walking activity. Analyses included adjustment for monitoring device. Accelerometry devices may find difficulty differentiating between walking and standing during very slow walking. This may cause an underestimation of the time spent walking. The absolute percentage error for the ActivPalTM when discriminating between standing and walking is <1% at walking speeds from 0.67–1.56 m/s [38] and 3.5% at 0.45 m/s [39]. Since the participants in this study had a mean gait speed (as measured by the SPPB) of 0.32 m/sec at admission and 0.51 m/sec at discharge, the time spent walking may well be underestimated, and, for this reason, measures of upright time may be more accurate. Changing accelerometry devices during the study could potentially have influenced results, although there was no significant difference in the proportions of participants using each device by treatment group. Contamination cannot be discounted, since blinding of both therapists and patients was not possible. Both intervention and control groups were aware that activity was being monitored. Such motivation may have minimized possible differences. Conclusions Objective monitoring for the amount of physical activity per day in hospitalized older adults could provide clinicians with information on clinically important, mobility-related activities to assist goal setting. Activity prescription could then be a routine component of the care plan, with professional staff and the patient contributing and responding to the activity plan. Supporting Information S1 File Consort Checklist. (DOCX) Click here for additional data file. S2 File Accelerometry Trial Study Protocol. (PDF) Click here for additional data file. The authors wish to acknowledge the staff and patients in the Geriatric Rehabilitation Units in Princess Alexandra Hospital Brisbane, Hornsby Ku-ring-gai Hospital Sydney and the Repatriation Hospital Adelaide for their support and participation in this study. ==== Refs References 1 Brown CJ , Friedkin RJ , Inouye SK . Prevalence and outcomes of low mobility in hospitalized older patients . J Am Geriatr Soc . 2004 ;52 (8 ):1263 –1270 . 15271112 2 Patterson F , Blair V , Currie A , Reid W . An investigation into activity levels of older people on a rehabilitation ward: an observational study . Physiotherapy . 2005 ;91 (1 ):28 –34 . 3 Peel NM , Kuys SS . Walking activity of older patients in rehabilitation: a prospective study . J Frailty Aging . 2013 ;2 (1 ):22 –26 . 10.14283/jfa.2013.4 27070454 4 Smith P , Galea M , Woodward M , Said C , Dorevitch M . Physical activity by elderly patients undergoing inpatient rehabilitation is low: an observational study . Aust J Physiother . 2008 ;54 (3 ):209 –213 . 18721125 5 Buttery AK , Martin FC . Knowledge, attitudes and intentions about participation in physical activity of older post-acute hospital inpatients . Physiotherapy . 2009 ;95 (3 ):192 –198 . 10.1016/j.physio.2009.03.002 19635339 6 Cheung VH , Salih SA , Crouch A , Karunanithi MK , Gray L . Monitoring ambulation of patients in geriatric rehabilitation wards: the accuracy of clinicians' prediction of patients' walking time . Int J Rehabil Res . 2012 ;35 (4 ):375 –377 . 10.1097/MRR.0b013e32835a23e8 23060084 7 Cameron ID , Kurrle SE . 1: Rehabilitation and older people . Med J Aust . 2002 ;177 (7 ):387 –391 . 12358585 8 Wade DT . Goal setting in rehabilitation: an overview of what, why and how . Clin Rehabil . 2009 ;23 (4 ):291 –295 . 10.1177/0269215509103551 19293289 9 Dawes H . The role of exercise in rehabilitation . Clin Rehabil . 2008 ;22 (10–11 ):867 –870 . 10.1177/0269215508098423 18955418 10 Killey B , Watt E . The effect of extra walking on the mobility, independence and exercise self-efficacy of elderly hospital in-patients: a pilot study . Contemp Nurse . 2006 ;22 (1 ):120 –133 . 16863419 11 Bravata DM , Smith-Spangler C , Sundaram V , Gienger AL , Lin N , Lewis R , et al Using pedometers to increase physical activity and improve health: a systematic review . JAMA . 2007 ;298 (19 ):2296 –2304 . 18029834 12 de Bruin ED , Hartmann A , Uebelhart D , Murer K , Zijlstra W . Wearable systems for monitoring mobility-related activities in older people: a systematic review . Clin Rehabil . 2008 ;22 (10–11 ):878 –895 . 10.1177/0269215508090675 18955420 13 Cheung VH , Gray L , Karunanithi M . Review of accelerometry for determining daily activity among elderly patients . Arch Phys Med Rehabil . 2011 ;92 (6 ):998 –1014 . 10.1016/j.apmr.2010.12.040 21621676 14 Culhane KM , O'Connor M , Lyons D , Lyons GM . Accelerometers in rehabilitation medicine for older adults . Age Ageing . 2005 ;34 (6 ):556 –560 . 16267178 15 Dakin LE , Gray LC , Peel NM , Salih SA , Cheung VH . Promoting walking amongst older patients in rehabilitation: are accelerometers the answer? J Nutr Health Aging . 2010 ;14 (10 ):863 –865 . 21125206 16 Matthews CE , Hagstromer M , Pober DM , Bowles HR . Best practices for using physical activity monitors in population-based research . Med Sci Sports Exerc . 2012 ;44 (1 Suppl 1):S68 –76 . 10.1249/MSS.0b013e3182399e5b 22157777 17 Peiris CL , Taylor NF , Shields N . Extra physical therapy reduces patient length of stay and improves functional outcomes and quality of life in people with acute or subacute conditions: a systematic review . Arch Phys Med Rehabil . 2011 ;92 (9 ):1490 –1500 . 10.1016/j.apmr.2011.04.005 21878220 18 Salih SA , Peel NM . Clinical utility of accelerometry in geriatric rehabilitation . Geriatr Gerontol Int . 2012 ;12 (4 ):753 –754 . 10.1111/j.1447-0594.2012.00844.x 22998379 19 Moher D , Schulz KF , Altman DG . The CONSORT statement: revised recommendations for improving the quality of reports of parallel-group randomised trials . Clin Oral Investig . 2003 ;7 (1 ):2 –7 . 12673431 20 Taraldsen K , Askim T , Sletvold O , Einarsen EK , Bjastad KG , Indredavik B , et al Evaluation of a body-worn sensor system to measure physical activity in older people with impaired function . Phys Ther . 2011 ;91 (2 ):277 –285 . 10.2522/ptj.20100159 21212377 21 Gray LC , Berg K , Fries BE , Henrard JC , Hirdes JP , Steel K , et al Sharing clinical information across care settings: the birth of an integrated assessment system . BMC Health Serv Res . 2009 ;9 :71 10.1186/1472-6963-9-71 19402891 22 Guralnik JM , Simonsick EM , Ferrucci L , Glynn RJ , Berkman LF , Blazer DG , et al A short physical performance battery assessing lower extremity function: association with self-reported disability and prediction of mortality and nursing home admission . J Gerontol . 1994 ;49 (2 ):M85 –94 . 8126356 23 Studenski S , Perera S , Wallace D , Chandler JM , Duncan PW , Rooney E , et al Physical performance measures in the clinical setting . J Am Geriatr Soc . 2003 ;51 (3 ):314 –322 . 12588574 24 Gawel J , Vengrow D , Collins J , Brown S . The short physical performance battery as a predictor for long term disability or institutionalization in the community dwelling population aged 65 years old or older . Phys Ther Rev . 2012 ;17 (1 ):37 –44 . 25 Holland R , Smith RD , Harvey I , Swift L , Lenaghan E . Assessing quality of life in the elderly: a direct comparison of the EQ-5D and AQoL . Health Econ . 2004 ;13 (8 ):793 –805 . 15322991 26 Peel NM , Kuys SS , Klein K . Gait speed as a measure in geriatric assessment in clinical settings: a systematic review . J Gerontol A Biol Sci Med Sci . 2013 ;68 (1 ):39 –46 . 10.1093/gerona/gls174 22923430 27 Tudor-Locke C , Craig CL , Aoyagi Y , Bell RC , Croteau KA , De Bourdeaudhuij I , et al How many steps/day are enough? For older adults and special populations . Int J Behav Nutr Phys Act . 2011 ;8 :80 10.1186/1479-5868-8-80 21798044 28 Benzinger P , Lindemann U , Becker C , Aminian K , Jamour M , Flick SE . Geriatric rehabilitation after hip fracture. Role of body-fixed sensor measurements of physical activity . Z Gerontol Geriatr . 2014 ;47 (3 ):236 –242 . 10.1007/s00391-013-0477-9 23780628 29 Dorsch AK , Thomas S , Xu X , Kaiser W , Dobkin BH , Sirract investigators. SIRRACT: An international randomized clinical trial of activity feedback during inpatient stroke rehabilitation enabled by wireless sensing . Neurorehabil Neural Repair . 2015 ;29 (5 ):407 –415 . 10.1177/1545968314550369 25261154 30 Mansfield A , Wong JS , Bryce J , Brunton K , Inness EL , Knorr S , et al Use of accelerometer-based feedback of walking activity for appraising progress with walking-related goals in inpatient stroke rehabilitation: a randomized controlled trial . Neurorehabil Neural Repair . 2015 ;29 (9 ):847 –857 . 10.1177/1545968314567968 25605632 31 Lindemann U , Jamour M , Nicolai SE , Benzinger P , Klenk J , Aminian K , et al Physical activity of moderately impaired elderly stroke patients during rehabilitation . Physiol Meas . 2012 ;33 (11 ):1923 –1930 . 10.1088/0967-3334/33/11/1923 23111341 32 Taraldsen K , Sletvold O , Thingstad P , Saltvedt I , Granat MH , Lydersen S , et al Physical behavior and function early after hip fracture surgery in patients receiving comprehensive geriatric care or orthopedic care—a randomized controlled trial . J Gerontol A Biol Sci Med Sci . 2014 ;69 (3 ):338 –345 . 10.1093/gerona/glt097 23902933 33 Villumsen M , Jorgensen MG , Andreasen J , Rathleff MS , Molgaard CM . Very low levels of physical activity in older patients during hospitalization at an acute geriatric ward—a prospective cohort study . J Aging Phys Act . 2015 ;23 :542 –549 . 10.1123/japa.2014-0115 25415513 34 McCusker J , Kakuma R , Abrahamowicz M . Predictors of functional decline in hospitalized elderly patients: a systematic review . J Gerontol A Biol Sci Med Sci . 2002 ;57 (9 ):M569 –577 . 12196493 35 Said CM , Morris ME , McGinley JL , Szoeke C , Workman B , Liew D , et al Evaluating the effects of increasing physical activity to optimize rehabilitation outcomes in hospitalized older adults (MOVE Trial): study protocol for a randomized controlled trial . Trials . 2015 ;16 (1 ):13 .25588907 36 Brown CJ , Williams BR , Woodby LL , Davis LL , Allman RM . Barriers to mobility during hospitalization from the perspectives of older patients and their nurses and physicians . J Hosp Med . 2007 ;2 (5 ):305 –313 . 17935241 37 Callen BL , Mahoney JE , Grieves CB , Wells TJ , Enloe M . Frequency of hallway ambulation by hospitalized older adults on medical units of an academic hospital . Geriatr Nurs . 2004 ;25 (4 ):212 –217 . 15311196 38 Grant PM , Dall PM , Mitchell SL , Granat MH . Activity-monitor accuracy in measuring step number and cadence in community-dwelling older adults . J Aging Phys Act . 2008 ;16 (2 ):201 –214 . 18483442 39 Kanoun N . Validation of the ActivPAL Activity Monitor as a Measure of Walking at Pre-determined Slow Walking Speeds in a Healthy Population in a Controlled Setting . Reinvention: a Journal of Undergraduate Research [Internet]. 2009 ; 2 (2 ).[Accessed 14 June 2015]. Available from: http://www.warwick.ac.uk/go/reinventionjournal/issues/volume2issue2/kanoun

PMC005xxxxxx/PMC5001633.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756454210.1371/journal.pone.0161868PONE-D-16-25307Research ArticlePhysical SciencesPhysicsThermodynamicsFree EnergyPhysical SciencesPhysicsCondensed Matter PhysicsSolid State PhysicsCrystallographyCrystal StructurePhysical SciencesChemistryChemical ElementsOxygenPhysical SciencesChemistryPhysical ChemistryChemical BondingHydrogen BondingPhysical SciencesChemistryPhysical ChemistryChemical BondingElectrostatic BondingPhysical SciencesChemistryComputational ChemistryMolecular DynamicsPhysical SciencesChemistryPhysical ChemistryIonsAnionsPhysical SciencesChemistryChemical CompoundsOrganic CompoundsAminesCatecholaminesPhysical SciencesChemistryOrganic ChemistryOrganic CompoundsAminesCatecholaminesBiology and Life SciencesBiochemistryNeurochemistryNeurotransmittersBiogenic AminesCatecholaminesBiology and Life SciencesNeuroscienceNeurochemistryNeurotransmittersBiogenic AminesCatecholaminesBiology and Life SciencesBiochemistryHormonesCatecholaminesComputational Investigation of the Interplay of Substrate Positioning and Reactivity in Catechol O-Methyltransferase Substrate Positioning and Reactivity in COMTPatra Niladri Ioannidis Efthymios I. http://orcid.org/0000-0001-9342-0191Kulik Heather J. *Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge, Massachusetts, 02139, United States of AmericaSoares Claudio M EditorUniversidade Nova de Lisboa Instituto de Tecnologia Quimica e Biologica, PORTUGALCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: HJK. Data curation: HJK. Formal analysis: HJK. Funding acquisition: HJK. Investigation: HJK NP. Methodology: HJK. Project administration: HJK. Supervision: HJK. Validation: HJK. Visualization: HJK EII. Writing – original draft: HJK. Writing – review & editing: HJK. * E-mail: hjkulik@mit.edu26 8 2016 2016 11 8 e016186823 6 2016 13 8 2016 © 2016 Patra et al2016Patra et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Catechol O-methyltransferase (COMT) is a SAM- and Mg2+-dependent methyltransferase that regulates neurotransmitters through methylation. Simulations and experiments have identified divergent catecholamine substrate orientations in the COMT active site: molecular dynamics simulations have favored a monodentate coordination of catecholate substrates to the active site Mg2+, and crystal structures instead preserve bidentate coordination along with short (2.65 Å) methyl donor-acceptor distances. We carry out longer dynamics (up to 350 ns) to quantify interconversion between bidentate and monodentate binding poses. We provide a systematic determination of the relative free energy of the monodentate and bidentate structures in order to identify whether structural differences alter the nature of the methyl transfer mechanism and source of enzymatic rate enhancement. We demonstrate that the bidentate and monodentate binding modes are close in energy but separated by a 7 kcal/mol free energy barrier. Analysis of interactions in the two binding modes reveals that the driving force for monodentate catecholate orientations in classical molecular dynamics simulations is derived from stronger electrostatic stabilization afforded by alternate Mg2+ coordination with strongly charged active site carboxylates. Mixed semi-empirical-classical (SQM/MM) substrate C-O distances (2.7 Å) for the bidentate case are in excellent agreement with COMT X-ray crystal structures, as long as charge transfer between the substrates, Mg2+, and surrounding ligands is permitted. SQM/MM free energy barriers for methyl transfer from bidentate and monodentate catecholate configurations are comparable at around 21–22 kcal/mol, in good agreement with experiment (18–19 kcal/mol). Overall, the work suggests that both binding poses are viable for methyl transfer, and accurate descriptions of charge transfer and electrostatics are needed to provide balanced relative barriers when multiple binding poses are accessible, for example in other transferases. Burroughs Wellcome Fund (US)Career Award at the Scientific Interfacehttp://orcid.org/0000-0001-9342-0191Kulik Heather J. H.J.K. holds a Career Award at the Scientific Interface from the Burroughs Wellcome Fund, which supported both H.J.K. and N.P. This work was carried out in part using computational resources from the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1053575. This work was also carried out in part using the computational resources of the Center for Nanoscale Materials (Carbon cluster), an Office of Science user facility, which was supported by the U. S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under Contract No. DE-AC02-06CH11357. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction Quantum-mechanical/molecular-mechanics (QM/MM) simulation[1–8] has taken a central role in unraveling enzyme mechanism. Challenges remain in fully enumerating sources of enzymatic rate enhancement even for fundamental reactions such as methyl transfer in neurotransmitter[9] and gene regulation[10, 11]. Catechol O-methyltransferase (COMT) is an S-Adenosyl-L-methionine (SAM)- and Mg2+-dependent methyltransferase (MTase)[9] that reacts with an array of catecholamine substrates (e.g., dopamine neurotransmitters). In order to form an active Michaelis complex in COMT, SAM binds first, followed by Mg2+, and the catecholamine substrate binds last in a bidentate fashion to Mg2+ at the solvent-exposed active site[12]. Catechol deprotonation, mediated by a lysine or a histidine in a recently discovered variant[13], has been thought to be an important intermediate step in the catalytic cycle since the catecholate anion is expected to be more reactive[14, 15]. The rate-determining[16], direct SN2 methyl transfer[17] from SAM[18] occurs primarily at the meta position[18, 19] of substituted catecholamines, with kinetic studies[12, 20–22] providing a free energy barrier estimate of 18.1[21]-19.2[22] kcal/mol for the soluble, human form of the enzyme. It is believed that Mg2+ plays a critical role in bringing the substrates together[18] because complete inhibition of COMT is achieved[9] when Mg2+ is replaced by Ca2+, and Mg2+ may provide a large component of the estimated 1016-fold rate enhancement[23] over solution. Alanine mutagenesis of an active site Tyr68 increases the barrier to 21 kcal/mol[22], and combined computational-experimental studies of COMT mutants have suggested[24] enhanced flexibility with respect to the wildtype enzyme is responsible for the reduction in rates. Atomistic simulation can provide valuable insight into the mechanism by which COMT achieves an estimated 1016-fold rate enhancement[23] over solution. Classical molecular dynamics studies should capture motions both at the active site and in the overall protein[25–27], and quantum mechanical studies on model systems of COMT can begin to provide insight into reactivity, typically from static structures[28] due to higher computational cost. Some computational studies have employed classical molecular dynamics (MD) of COMT in apo[29, 30], intermediate[31], and holo[25–27] forms. Long, well-equilibrated MD simulations have only been carried out on apo-COMT[29, 30], and earlier studies on holo-COMT[25–27] were limited by then-available computational power to much shorter 1 ns simulations. Early holo-COMT studies fixed bidentate catecholate coordination to Mg2+ observed in X-ray crystal structures[15] with explicit Mg2+-O bonds[25, 26]. Although qualitative catecholate-Mg2+ coordination was held consistent with experiment, non-bonded SAM methyl-catecholate oxygen (C-O) distances sampled during dynamics averaged 3.55 Å in poor agreement with the approximately 2.6 Å C-O distance observed in crystal structures. Other MD studies[27, 32] in which bidentate catecholate coordination to Mg2+ was not enforced instead have resulted in reorientation to form a monodentate catecholate characterized by a single Mg2+-O- coordination, an intramolecular hydrogen bond, and a compensating sixth interaction with Mg2+ derived from an active site carboxylate. Mixed semi-empirical/classical (SQM/MM) calculations have suggested an even weaker interaction between catecholate and Mg2+ with only a single coordination to the neutral hydroxyl of catecholate with a bond elongated by as little as 0.5 Å[33] or as much as 2–5 Å[34] (i.e., catecholate does not coordinate Mg2+) with respect to typical Mg2+-O bonds. These distinct poses have never before been directly compared, and the substrate's suitability as a methyl acceptor may depend on its position in the active site. For instance, substrate placement has been shown to strongly influence reactivity and branching ratios in metalloenzymes[35, 36]. First-principles simulation of the methyl transfer barrier requires careful selection of which portion of the enzyme will be treated quantum mechanically since QM methods are typically higher scaling and more computationally expensive than classical methods. SAM, catecholamine substrate, and the Mg2+ alone are 64 atoms, and the Mg2+ coordination sphere enlarges this system size to over 100 atoms. For efficient sampling and calculation, studies have leveraged partial models of SAM and reactants[28] and have often treated Mg2+ classically. Despite these approximations, a number of DFT[37, 38] and semi-empirical[32, 39, 40] computational studies[27, 32, 37–40] have produced a wide range of free energy barriers (3–30 kcal/mol) that are sometimes in good agreement with experimental barriers (18–19 kcal/mol), especially after corrections for some approximations. Some of us have identified[24] that large-scale QM/MM treatments (ca. 500 atoms) of the COMT active site 500 atoms in the COMT active site are beneficial for explaining and interpreting the effects of active site mutations. Despite the many approaches that have been proposed in the literature to balance accuracy and efficiency in evaluating methyl transfer barriers, there have not yet been any comparative studies between how differing binding poses accessible in MD may be more or less suitable methyl group acceptors in methyl transfer reactions. We thus use comprehensive classical and quantum mechanical (both semi-empirical and density functional theory) methods to address open questions in the structure and function of COMT. We quantify the free energy landscape of substrate dynamics, identify the driving force and interactions in differing substrate poses in the active site, and determine the extent to which substrate placement alters the methyl transfer reaction coordinate. We additionally demonstrate that charge transfer between substrates and the active site is required to reproduce experimental crystal structure geometries and methyl transfer barrier heights. The structure of this article is as follows. First, we describe the Computational Details of all simulations carried out in this work. In the Results and Discussion, we first evaluate the structure, dynamics, and binding free energies of catecholate in differing binding poses and identify the effect of binding pose on free energy barriers of the rate-determining step (RDS). Finally, we provide our Conclusions. Methods Classical Molecular Dynamics Classical MD simulations of COMT were carried out using the GPU-accelerated version[41, 42] of the AMBER 14 software package[41]. The starting structure was obtained from the COMT crystal structure (PDB ID: 3BWM[43]), which had bound SAM, dinitrocatecholate (DNC) inhibitor, and an Mg2+ cation. Three resolved water molecules in the X-ray crystal structure that were buried in the active site were kept in the simulations while external waters were replaced during protein solvation. For catecholate structures, nitro groups were removed from the DNC structure. The protein was described by the AMBER ff12SB[44] force field, which is derived from the ff99SB[45] force field with updates to backbone torsional parameters. For SAM and catecholate or DNC substrates, we employ the generalized AMBER force field (GAFF)[46] with partial charges assigned from restrained electrostatic potential (RESP) charges[47] obtained with GAMESS-US[48] at the Hartree-Fock level using a 6-31G*[49] basis set, as implemented by the R.E.D.S. web server[50–52]. Thoroughly tested parameters for Mg2+ were obtained from Ref. [53] (S1 Table and S1 Fig). The charge of protein residues was assigned with the H++ webserver[54–57] assuming a pH of 7.0 to yield -8 (-6) for the apoprotein (holoprotein). H++ assigns the protonation state of neutral histidine residues based on van der Waals’ contacts and hydrogen bonding distances, which results in His12, His142, and His192 being protonated at the δ position and His16, His57, and His182 being protonated at the ε position. The protein was solvated with a 10 Å buffer of TIP3P[58] water on all sides (a total average size of around 62x69x71 Å during the NPT production runs) and neutralized with 6 Na+ ions. Several multistage equilibration protocols were employed, which differed only by the extent and nature of restraints that enforced the crystal structure active site coordination (S1 Text, S2–S6 Figs and S2 Table). These protocols included restrained and/or unrestrained minimizations, a quick NVT heating stage, NPT equilibration, and NPT production dynamics at T = 300 K and p = 1 bar. We used a Langevin dynamics thermostat with a collision frequency of 1.0 ps-1 and a random seed to avoid synchronization artifacts. For constant pressure dynamics, a Berendsen barostat with a pressure relaxation time of 1 ps was used. The SHAKE algorithm[59] was applied to fix all bonds involving hydrogen, permitting a 2 fs timestep to be used for all MD. For the long-range electrostatics, the particle mesh Ewald method was used with a 10-Å electrostatic cutoff. Free energy surfaces Two-dimensional MD free energy surfaces (FES) were obtained by applying the weighted-histogram analysis method (WHAM)[60, 61] using the Grossfield lab software package[62] to unbias distributions obtained with umbrella sampling[63]. Equally spaced windows were obtained at 0.1 Å spacing over the range of 1.8–4.5 Å in the Mg2+-OH bond distance and 5° spacing over the 0–180° range in the C-C-O-H dihedral (see Fig 1). Force constants were 500 kcal/(mol.Å2) for the bond distance restraint and 200 kcal/(mol.rad2) for the dihedral restraint. Each window was equilibrated for 100 ps before 200 ps of production dynamics was carried out over which sampled distances and dihedrals were collected. During the WHAM fitting, the iterative solution of the free energy weights was converged with a 1x10-8 threshold, and the final FES was described by 76 C-C-O-H dihedral bins of 2.5° width and 104 Mg2+-OH distance bins of 0.025 Å width. 10.1371/journal.pone.0161868.g001Fig 1 Model of key active site residues in COMT. a) Bidentate and b) monodentate configurations of COMT, including Mg2+ and its coordinating residues, catecholate and its interactions with Lys144 and Glu199, and the SAM substrate. The C-C-O-H dihedral angle is highlighted in green. Quantum Mechanical FESs Hybrid semi-empirical quantum-mechanical/molecular mechanics (SQM/MM) dynamics employed the internal AMBER[41, 42] SQM routines. These SQM/MM calculations include electronic embedding of SQM atoms and hydrogen capping of all cleaved covalent bonds that span the SQM/MM boundary. For the SQM approach, both AM1[64] with explicit treatment of d electrons and PM6[65] were employed. These dynamics were carried out with spherical boundary conditions in the largest sphere afforded by the MD box from which they were extracted with no electrostatic cutoff, a 0.5 fs timestep, and constant temperature (T = 300 K) enforced by a Langevin thermostat as in the MD runs. No atoms were held fixed, but a 1.5 kcal/(mol.Å2) restraining potential kept water molecules from evaporating from the sphere. Initial configurations across the methyl transfer coordinate were obtained from a quick steered molecular dynamics (SMD) run. Snapshots from the SMD runs were extracted for each of the ten windows in umbrella sampling for 10 ps equilibration and 20 ps production SQM/MM. The reaction coordinate in both SMD and umbrella sampling was defined as an antisymmetric linear combination of distances (LCOD) between the S-C and C-O bonds, which break and form, respectively, during methyl transfer. Variable force constants ranging from 10 kcal/(mol.Å2) in low energy regions to 240 kcal/(mol.Å2) in high energy regions were employed to minimize the number of windows required while maximizing overlap over the -1.5 to 2.2 Å LCOD range (details of force constants, window targets, and window widths are provided in S3 Table). WHAM[60, 61] software[62] was used to reconstruct one-dimensional free energy curves with 0.02 Å bin widths. In order to validate PM6 for FESs, additional geometry optimizations were carried out both at the hybrid density functional theory (DFT, B3LYP[66–68]) and PM6[65] levels of theory (S2 Text). Energy Decomposition Analysis Binding free energy calculations employed the AMBER MMPBSA.py[69] utility, which follows protocols outlined in Ref. [70]. New 5 ns MD trajectories were generated starting from representative bidentate, monodentate, and local max geometries for a single trajectory protocol. The local max geometry represents a high-energy point on the transition from bidentate to monodentate (see Results and Discussion) and was sampled with distance restraints, as described above. For this method, implicit solvent calculations within the Poisson-Boltzmann (PB)[71] or Generalized Born (GB)[72] approximations are carried out on snapshots obtained from MD both with a noncovalent ligand present and rigidly removed. In our simulations, the rigid binding free energies were averaged from configurations extracted every 8 ps for a total of 625 snapshots. MMPBSA total binding free energies of Mg2+ and catecholate as well as the individual residue contribution to the binding energy of catecholate were obtained. Full energy decomposition analysis with MMPBSA is computationally intensive, and pairwise residue interactions were computed instead with MMGBSA using the "OBC1" model[73], motivated by recent benchmarks[74]. In both MMPBSA and MMGBSA cases, the internal dielectric was set to 1, and the salt concentration was set to 0.1 M. Entropic contributions to binding computed within the quasi-harmonic approximation were not found to vary across points being compared and were therefore neglected. More description of contributions to the MMPBSA binding free energies is provided in S3 Text. Results and Discussion Structure and Dynamics in the Active Site At least ten experimental crystal structures[15, 43, 75–79] of COMT have been solved with bound inhibitors ranging from dinitro to coumarine in nature along with SAM and Mg2+ in the active site at resolutions ranging from 1.3 to 2.4 Å. On average, the bidentate substrate analogue in these structures has two Mg2+-O bond distances averaging around 2.16 Å that are comparable to the 2.12 Å average distance for the remaining species in the active site (Asp141, Asp169, Asn170, and H2O) that coordinate Mg2+. It has been proposed[14, 15] that the inhibitor molecule was bound in a monoanionic form. Based on expected pKas, it is believed that Glu199 forms a hydrogen bond with the hydroxyl of one catechol, and the other oxygen proximal to the methyl group of SAM is deprotonated by Lys144 (Fig 1A). Some simulations[27, 32, 80] have identified alternate active site configurations for catecholate in which a monodentate (m) structure with an intramolecular catecholate hydrogen bond is formed. The intramolecular hydrogen bond is known to be stable in gas phase structures of catecholate[81]. The newly available sixth coordination site on Mg2+ is either filled by monodentate coordination to Glu199[80] or bidentate coordination with Asp141[27, 32] (Fig 1B). The bidentate Asp141 configuration is likely an artifact of force field parameter choice, as bidentate coordination of Mg2+ is exceedingly rare under physiological conditions[82]. An alternate m configuration (m-alt)[33] has been proposed in which the neutral hydroxyl oxygen of catecholate weakly binds Mg2+ with a > 2.5 Å Mg2+-O(H) bond and the sixth Mg2+ coordination site is instead occupied by Glu199. This binding orientation necessitates rearrangement in the active site or proton transfer along a path that crosses very close to Mg2+, and the weak catecholate-Mg2+ interaction is unlikely to be sustained during the relatively slow catalytic cycle of COMT (turnover frequency of 12–24 min-1 [21, 22]). When we sampled neutral catechol or a species in which only the neutral hydroxyl oxygen of catecholate was directly coordinated to Mg2+, we observed extremely short (i.e., sub-ns) lifetimes for that species. In an attempt to stabilize m-alt, we also carried out restrained MD in the configuration proposed in Ref. [33] for 50 ns before release and were able to stabilize this configuration for up to 50 ns before it rearranged to the standard m configuration. As with previous work[27, 32, 80], we also observe rapid bidentate (b) to monodentate (m) rearrangement (Fig 2) of catecholate-Mg2+ coordination (Fig 1), depending upon the equilibration protocol employed (S2–S6 Figs and S1 Text). The most rapid rearrangement is observed in protocols in which no restraints are applied to the protein or ligand environment during equilibration. Use of harmonic restraints on Mg2+-O distances during NVT heating and NPT equilibration leads to b structures that are stabilized for at least 80 ns before rearrangement to a m structure (Fig 2, see S1 Text for a full 350 ns bidentate run when dihedral restraints are employed). The long lifetime of the b species departs from unrestrained equilibration protocols in earlier[27, 32, 80] short timescale MD and our own MD. Analysis of bond distance changes reveals that the dynamical rearrangement is induced by the destabilization of the catecholate hydroxyl-Glu199 hydrogen bond via Glu199 sidechain rotation. This fluctuation encourages the catecholate hydroxyl to reorient, forming an intramolecular hydrogen bond and monodentate coordination to Mg2+. Shortly thereafter, Glu199 is observed to coordinate directly to Mg2+. In all simulations, no reverse transition from m to b coordination is observed on the 100-ns timescale. 10.1371/journal.pone.0161868.g002Fig 2 Active site bond distances from a 160 ns molecular dynamics trajectory. Distances include Mg2+ coordination with catecholate anion (red line) and hydroxyl oxygen atoms (magenta line); catecholate intramolecular hydrogen bond distance (green line); and catecholate hydrogen bond distance to Glu199 carboxylate oxygen atoms (blue and gray lines). The electrostatic driving force for the catecholate rearrangement during classical MD simulations is clear: oxygen partial charges on catecholate are weaker than those on Glu199 (S7 Fig). However, a key question is whether the overall free energy of the m configuration is substantially lower than b since the lack of m-to-b rearrangement during MD suggests a higher barrier for the backward than the forward transformation and thus possibly a lower free energy for the m configuration. The two-dimensional (2D) free energy surface (FES) spanning Mg2+-OH distances and C-C-O-H dihedral values confirms observations from MD (Fig 3). The minimum free energy path involves Mg-OH bond elongation followed by formation of the intramolecular hydrogen bond (C-C-O-H dihedral of 180°, see Fig 1 for dihedral location on catecholate). The free energy barrier for the Mg2+-OH bond elongation step of around 6.5 kcal/mol is consistent with observations of a long-lived b structure during MD simulations. 10.1371/journal.pone.0161868.g003Fig 3 Catecholate-Mg2+ coordination FES in COMT with respect to C-C-O-H dihedral and Mg-OH distance. Color bar at right shown in kcal/mol with high free energy regions in excess of 15 kcal/mol shown in white. A second free energy minimum that is equienergetic with the bidentate structure corresponds to elongated Mg2+-OH distances in a monodentate configuration without the intramolecular catecholate hydrogen bond. Here, a hydrogen bond with Glu199 stabilizes the hydroxyl of monodentate catecholate instead (see Fig 1). Rearrangement to a m structure with an intramolecular hydrogen bond has a low free energy barrier of around 3 kcal/mol. Overall, the m-catecholate (m-CAT) with an intramolecular hydrogen bond is stabilized by around 4 kcal/mol with respect to the other two free energy basins. The m-alt configuration employed in Ref. [33], on the other hand, was found to be 11 kcal/mol higher in energy than a bidentate reference, and interconversion to m-alt is prohibitive with a free energy barrier of 24 kcal/mol (S8 Fig). During MD sampling of m-CAT, we observe occasional interconversion between the intramolecular H-bond monodentate (C-C-O-H dihedral = 180°) and the extended hydroxyl case (C-C-O-H dihedral = 0°) consistent with the features of the free energy surface. Although no rearrangement back to b-CAT is observed, the barrier for conversion from the m-CAT, intramolecular H-bond catecholate to b-CAT is predicted to be 10.5 kcal/mol, corresponding to an interconversion frequency of 16 μs-1. This exchange frequency is within an order of magnitude of experimental 0.67 μs-1 exchange rates at room temperature (1.5 μs-1 at 37°C) of neutral ligands such as H2O around Mg2+[83] and should occur multiple times between the slower RDS methyl transfer steps. Representative structures from b-CAT and m-CAT free energy basins reveal key differences in orientations (Fig 4). Since X-ray structures are usually solved in the presence of a dinitrocatecholate (DNC) inhibitor, we also carried out MD simulation with DNC bound and confirmed that the b-to-m rearrangement was comparable to CAT (S9 Fig). 10.1371/journal.pone.0161868.g004Fig 4 Representative bidentate (top) and monodentate (bottom) catecholate (CAT) substrate configurations at the COMT active site. Substrates are shown in orange, protein residues in blue, and key distances are shown (in Å), except for D141-Mg2+, D169-Mg2+, and Mg2+-H2O, which are unchanged. In both CAT and DNC MD simulations, a short hydrogen-Glu199 oxygen distance of around 1.7 Å in the b-CAT configuration is replaced by a 2.1 Å intramolecular hydrogen bond in the m-CAT case (S4 Table). Whereas the CAT oxygen anion coordination distance to Mg2+ is always predicted to be shorter than crystal structure values (1.94 Å in m-CAT, 1.97 Å in b-CAT), the longer DNC distances (2.08–2.16 Å, S9 Fig) are consistent with experimental crystal structure distances. Overall distances between both substrate and inhibitor are comparable, with short 1.90–1.94 Å distances for Mg2+ coordination to carboxylates of Asp141 and Asp169 and longer 2.02–2.13 Å bond distances with Asn170 and the axial water ligand. The neutral catechol hydroxyl coordination distance to Mg2+ is consistently longer by 0.15–0.25 Å in both CAT and DNC with respect to the compensating interaction that is formed with the Glu199 carboxylate, consistent with analysis of the charges of each residue in the active site (S7 Fig and S4 Table). Interestingly, distances of the transferring methyl carbon on SAM to the nearest oxygen on catechol vary in the b and m MD configurations. Crystal structures have unusually short SAM C–CAT O distances (average: 2.65 Å, range: 2.45–2.81 Å), and it has been hypothesized that unusually short reactant distances may be a source of catalytic rate enhancement[25]. The GAFF repulsive van der Waals terms for SAM and CAT prevent extensive sampling of C-O distances observed in X-ray structures. This repulsion is slightly balanced by weak electrostatic attraction between methyl hydrogen atoms (q = +0.12 e-) and the anionic oxygen of the substrate (q = -0.55–0.80 e-). Weaker partial charges on the DNC oxygen atoms lengthen the C-O distances even further (S4 Table). There are still distinguishable differences in sampled SAM-CAT distances for the two CAT m and b binding orientations (Fig 5). The b-CAT configuration samples distances shorter than 3.2 Å (or 2.8 Å) between a methyl C and O- 42% (4%) of the time, whereas m-CAT only samples those distances for 11% (0.3%) of the trajectory. Both the b-CAT and m-CAT C-O distances may be fit to normal distributions with means of 3.22 Å and 3.44 Å, standard deviations of 0.175 Å and 0.236 Å, and minimum distances of 2.65 Å and 2.75 Å, respectively. Semi-empirical AM1 QM/MM geometry optimizations have indicated a C-O distance of 2.92 Å in an alternate, weakly-coordinating monodentate configuration (m-alt)[33], but the closest comparison to our room temperature MD result from the same level of theory instead indicated a 3.25–3.50 Å[84] preferred C-O distance in that orientation. Classical MD on the m-alt structure reveals a comparable C-O distribution for the Mg2+-coordinated hydroxyl oxygen (3.44 Å) and a broader and longer C-O distance for the free anionic oxygen (centered about 4.0 Å). This longer distance indicates that the anionic oxygen is seldom available to form a catalytically competent geometry and thus this configuration is not considered further (S10 Fig). We will revisit QM/MM equilibrium bond distances from QM/MM MD in the next section, where the possible effect of the different starting structures on the methyl transfer RDS will now be considered in more detail. 10.1371/journal.pone.0161868.g005Fig 5 Histograms of C(SAM)-O-(catecholate) distances for bidentate (black lines) and monodentate (red lines) catecholate-Mg2+ coordination in COMT. Dashed lines are the best-fit normal distributions. The range of C-O distances in X-ray crystal structures of COMT is indicated by two gray vertical dotted lines. Substrate-position-dependent Free Energy Barriers We now investigate how substrate-positioning differences affect predicted free energy barriers for the rate-determining methyl transfer step. Although a number of computational studies have made a range of predictions for the COMT methyl transfer free energy barrier[27, 32, 33, 37, 85] (3–30 kcal/mol), a systematic comparison of methyl transfer barriers between the distinct bidentate and monodentate substrate configurations has not been carried out. Recent work by some of us[24] has indicated that properties of COMT converge only with inclusion of 100s of atoms in the QM region in QM/MM calculations. It appears that accurate free energy calculations should be carried out with charge transfer (CT) permitted between the substrates, Mg2+, and the protein environment. In order to enable sufficient sampling but maintain a sufficient description of the substrate electronic environment, we turn to computationally efficient semi-empirical methods and validate them with hybrid DFT. Semi-empirical models from Stewart and coworkers have been demonstrated[86] to describe Mg2+ hydration free energies well. The central importance of Mg2+ in this protein motivated our selection of the PM6 semi-empirical Hamiltonian as well as its availability in the AMBER code. Three SQM region sizes were considered: catecholate and SAM substrates only (S region), substrates with Mg2+ cation (SMg), and the substrates, Mg2+, and Mg2+ ligands (Asp141, Asp169, Asn170, Glu199, and water) as well as Lys144 (SMgL). Recall that in b-CAT, Glu199 hydrogen bonds with CAT whereas in m-CAT Glu199 directly coordinates Mg2+. Lys144 is included in order to allow for proton transfer or hydrogen bonding with CAT. The largest model selected here (SMgL) contains most of the critical residues used in previous large-scale QM/MM models[24]. Umbrella sampling was carried out to obtain the free energy for methyl transfer at the PM6/MM level of theory with an LCOD coordinate that was the difference (Δ) between the donor (S-C) bond and the acceptor (C-O) bond (sampled values illustrated in top panel of Fig 6). We observe a significant decrease in the methyl transfer free energy barrier as the SQM region size is increased (Fig 6), with a 6.3 kcal/mol drop from SMg to SMgL SQM regions. Only with the full SMgL region do we recover the prediction of an exothermic methyl transfer step, whereas the other two regions correspond to strongly endothermic reactions (see Fig 6 and S5 Table). Overall prediction of a 21.8 kcal/mol free energy barrier for b-CAT is in good agreement with the experimental value of 18.1–19.2 kcal/mol[21, 22]. Continued QM region enlargement, which is currently computationally prohibitive, could improve agreement further. 10.1371/journal.pone.0161868.g006Fig 6 Bidentate methyl transfer free energy curves. (bottom) Methyl transfer free energy curves from PM6/MM (in kcal/mol) for three models of increasing size: catecholate and SAM substrate-only (S, black circles), with Mg2+ (SMg, red circles), and with Mg2+ coordination ligands (SMgL, green circles) plotted against the difference of S-C and C-O bond distances. (top) Absolute, average and range of S-C (yellow line) and C-O (gray line) distances (in Å). In addition to the methyl transfer barrier, our sampling of the difference (Δ) between S-C and C-O distances provides information about the FES of the Michaelis complex. For all SQM regions the free energy surface is relatively flat over the range of Δ = -1.5 to -0.75 Å, but the free energy minimum shifts significantly from S to SMgL. The equilibrium C-O distance in b-CAT S SQM/MM is 3.12 Å, around the same as observed in MD, but this value decreases to 2.98 and 2.72 Å for SMg and SMgL SQM/MM, respectively (Fig 6, Table 1 and S6 Table). The CT-mediated minimum free energy C-O non-bonded distance in the SMgL SQM/MM Michaelis complex is consistent with the range of distances observed in crystal structures (2.45–2.81 Å). We note absolute bond distances are subject to the errors inherent in PM6 at around 0.03–0.10 Å[65], but trends in distances with QM-region likely benefit from cancellation of errors. 10.1371/journal.pone.0161868.t001Table 1 Geometries of the enzyme-substrate (ES) complex and transition state (TS) for m-CAT and b-CAT configurations for the SMgL PM6 SQM region. b-CAT ES m-CAT ES b-CAT TS m-CAT TS S-C (Å) 1.76 1.76 2.21 2.19 C-O (Å) 2.72 3.02 1.91 1.90 Δ (Å) -0.96 -1.26 0.32 0.29 ∠S-C-O (°) 158.7 162.7 173.2 173.3 The CT-mediated C-O distance shortening appears strongly correlated to the reduction in free energy barriers. In fact, the three SQM regions can be fit (R2 = 0.99) to the expression: ΔG‡=25.5×d(C-O)−47.8,(1) where d(C-O) is the C-O distance in Å of the Michaelis complex and the units are in kcal/mol (S5 Table). This observation suggests that within QM treatments of COMT, shorter distances are correlated to greater recovery of electronic contributions to enzymatic rate enhancement. However, we have not yet considered the effect the QM region size has on the highest energy point, which we refer to approximately as the transition state (TS) of the reaction coordinate. Examining the b-CAT methyl transfer TS geometries (see Table 1 and S6 Table) reveals some S-C distance shortening as the SQM region is increased (from 2.28 to 2.21 Å for S to SMgL) but no significant difference in C-O distances (1.88 Å to 1.91 Å). However, this reduction in S-C distance pushes the transition state earlier on the Δ coordinate from 0.40 to 0.32 Å. Our SQM/MM results are consistent with or correspond to a slightly tighter transition state compared to distances in previous semi-local DFT/MM studies of 2.24 Å and 2.07 Å for S-C and C-O, respectively[85]. Differences in the S-C-O angle reveal some region dependence with the SMgL SQM/MM calculations showing the closest correspondence between the Michaelis complex and the TS of a little under 15°. Therefore, in the SMgL simulations, the effective barrier height reduction is arising because the TS and MC resemble each other more closely than in the smallest S SQM/MM calculations. We now compute the region dependence of the free energy barriers for methyl transfer to m-CAT (S11 Fig, S5 and S6 Tables). We again observe a decrease in the free energy barrier as the SQM region is increased along with a shortening of the C-O distance in the Michaelis complex. Here, the S SQM/MM C-O distance of 3.38 Å is comparable to the classical MD m-CAT free energy minimum, and it shortens to 3.02 Å in the SMgL SQM/MM simulation. The distance shortening with SQM region increase is somewhat smaller in m-CAT than b-CAT (0.36 versus 0.40 Å), consistent with reduced CT and interaction with Mg2+ for m-CAT vs. b-CAT. Although the distance decrease is smaller, the free energy barrier reduction from S to SMgL SQM/MM is only slightly reduced (10.3 kcal/mol for bidentate to 8.7 kcal/mol for monodentate, see Table 2 and S5 Table). The resulting correlation between the free energy barrier and Michaelis complex is of similar magnitude to that for b-CAT, albeit with a reduced correlation coefficient of 0.85. 10.1371/journal.pone.0161868.t002Table 2 Free energy barriers (ΔG‡) and reaction free energies (ΔGRxn) for methyl transfer from the monodentate and bidentate catecholate configurations with the SMgL SQM region in SQM/MM calculations. ΔG‡ (kcal/mol) ΔGRxn (kcal/mol) bidentate 21.8 -0.9 monodentate 22.1 -2.4 expt. 18.1–19.2a — aRefs. [21, 22]. Overall comparison of the bidentate and monodentate methyl transfer for the largest SQM region considered reveals nearly identical methyl transfer energetics (Fig 7 and Table 2). The S-C and C-O distances of the m and b TSes are also nearly identical (see Table 1). It has been suggested[34] that m-CAT orientation, especially with the catecholate oxygen anion uncoordinated to Mg2+, might be necessary in order to make the catecholate oxygen anion a more suitable nucleophile for the SN2 methyl transfer and that the b-CAT form in X-ray structures is an inactive state of the enzyme. The nearly identical methyl transfer barriers of 21.8 and 22.1 kcal/mol for b-CAT and m-CAT, which are both slight overestimates of the experimental range[21, 22] (18.1–19.2 kcal/mol), however, suggest the oxygen atoms of both coordination geometries are equally suitable nucleophiles. One could draw a different conclusion that the m-CAT structure was preferred if one used the CT-restrictive S SQM region results. The m-CAT reaction is predicted to be slightly more exothermic, as a consequence of a smaller effect on Mg2+ stabilization by CAT when the methylated m-CAT product is formed (see inset of Fig 7). [87]A key distinction for future study is that the b-CAT Michaelis complex C-O distance is dramatically shorter by nearly 0.4 Å, which may enhance reaction probability by impacting recrossing from that configuration and would need to be addressed by considering a generalized transmission coefficient.[87] 10.1371/journal.pone.0161868.g007Fig 7 Bidentate and monodentate methyl transfer free energy curves. Methyl transfer free energy curves (in kcal/mol) for the SMgL model for bidentate (blue circles) and monodentate (gray circles) catecholate-Mg2+ coordination plotted against the difference in S-C and C-O distances (in Å). Representative product geometries are shown in the inset. We also considered the extent to which the choice of the PM6 semi-empirical method has impacted predictions of the methyl transfer free energy barriers. Earlier[33] AM1/MM calculations had used an MP2-derived correction in order to add > 10 kcal/mol to the semi-empirical free energy barriers. However, two major differences in those calculations from the current work were the exclusion of Mg2+ from the SQM region and the use of a reactant reference in which Mg2+ was only weakly coordinated by the neutral hydroxyl oxygen of catechol and average C-O distances ranged from 2.95[33] to 3.25[84] Å. We compared our AM1/d/MM and PM6/MM free energy barriers, where the AM1/d semi-empirical approach incorporates parameters for Mg2+ and found the AM1/d/MM results to produce comparable energy barriers (S12 Fig). In the previous work[33], a very weakly coordinating CAT reference was used as the reactant, likely causing the lower computed barrier for methyl transfer (see Fig 3). In addition to semi-empirical methods, we compared enthalpies obtained with hybrid DFT using the B3LYP functional with the 6–311++G* basis set. Using a number of techniques outlined in the Supporting Information including model calculations in gas phase and implicit solvent as well as clusters cut directly from the SQM/MM free energy curves, we confirmed consistency in barrier estimates between PM6 and hybrid DFT (S7 Table and S13 and S14 Figs). Overall, these results suggest that the electronic environment of COMT enforces a shortened C-O distance and lowers the free energy barrier for methyl transfer. Interconversion to a monodentate structure may occur on rapid timescales, but this structure likely is less, rather than more, reactive for methyl transfer. Differences in Substrate-Protein Interactions Interactions between CAT, SAM, Mg2+, and protein appear distinct between b-CAT and m-CAT configurations. The electrostatic driving force for b-CAT to m-CAT rearrangement has been identified as the higher point charges for Glu199 than for the neutral catechol hydroxyl (S7 Fig). Nevertheless, electrostatic interactions are only part of free energy differences, and we now carry out binding free energy calculations and energy decomposition analysis (EDA) with MMPBSA in order to further identify differences in binding modes (S3 Text). The MMPBSA approach has been used before to analyze COMT inhibitor binding[27, 88, 89], but we leverage it for the first time to identify differences in b-CAT and m-CAT configurations. We first computed total binding free energy components for CAT in a b-CAT, m-CAT, and local maximum (local max) configuration. The local max configuration corresponds to the highest free energy structure for a C-C-O-H dihedral around 0° on the 2D-FES (Fig 3) for the b-to-m transition. For these 3 orientations, we computed the CAT binding energy with a SAM- and Mg2+-containing receptor protein, consistent with experimental observations of the order of binding of substrates in COMT[12]. In addition to CAT binding, we considered the artificial case of treating Mg2+ as the ligand and the SAM-, CAT-containing protein as the receptor. Although not necessarily physical, this analysis enables us to answer the question of whether the overall stabilization of m-CAT is derived from stronger binding of Mg2+ or CAT to the protein. Total and component binding free energies for CAT and Mg2+ reveal that Mg2+ binding is enhanced and CAT binding weakened from the b-CAT to m-CAT configurations (Fig 8). In CAT, a 59% reduction in the electrostatic (elst) stabilization is observed from b-CAT to m-CAT, but this reduction is counteracted by other components to give an overall decrease in binding free energy of 28%. These components are the reduction in cost to desolvate m-CAT, indicated primarily through the less repulsive Poisson-Boltzmann (PB) polar solvation term as well as a shift from weakly repulsive van der Waals (vdW) substrate-protein interactions to weakly attractive ones. Other components of the binding free energy are relatively unchanged. The increase in binding free energy of Mg2+ (see Fig 8) in the b-to-m transition is primarily (92% of total free energy difference) in the reduction of PB solvation penalty. The direct difference in gas phase elst of binding provides a comparatively smaller (5%) contribution. Local max elst become unfavorable, which is a likely component of the free energy barrier observed earlier on the 2D-FES. These MMPBSA results suggest the lower free energy of m-CAT on the 2D-FES is due to a stabilization of Mg2+ and destabilization of CAT. 10.1371/journal.pone.0161868.g008Fig 8 Binding free energy components. Binding components for Mg2+ (top) and catecholate (CAT, bottom) from MMPBSA for the bidentate (left), monodentate (right) and local max (center) configurations. Contributions of each bar for both graphs are shown in the legend (bottom right). Lines represent the total (squares), gas phase (triangles), and solvent (circles) contributions. Decomposition of individual residue contributions to the binding free energy also provides insight into how protein-CAT interactions shift between the b-CAT and m-CAT configurations. We identified the residues or substrates that had at least 1 kcal/mol binding free energy difference between b-CAT and m-CAT: Mg2+, SAM, CAT, Glu199, Asp141, Lys46, Asp169, and Lys144 (Fig 9). Glu199 had been previously identified (see Fig 4) as forming a hydrogen bond to b-CAT or alternately coordinating Mg2+ and repelling m-CAT, consistent with a 7 kcal/mol more repulsive contribution to m-CAT binding due to increasing elst binding penalty. The Mg2+-coordinating residues Asp141 and Asp169 are similarly more repulsive (3 and 2 kcal/mol, respectively), but, unlike Glu199, these differences are derived from the gas phase vdW and polar solvation (pol) contributions as a result of m-CAT's greater proximity to those residues. The increased electrostatic attraction of Lys144 and binding free energy contribution change (-1 kcal/mol) is due to increased hydrogen bonding between m-CAT and the NH3+ group of Lys144. Similarly, the protonated Lys46 is positioned on the neutral hydroxyl side of CAT. The b-to-m hydroxyl reorientation increases (-2 kcal/mol) favorable interactions between m-CAT OH and Lys46. 10.1371/journal.pone.0161868.g009Fig 9 Residue contributions to catecholate binding. Residue contributions to binding in bidentate (green circle, left bar in clustered bars) and monodentate (red circle, right bar) binding free energy as well as elst (blue bar) and pol+vDW (gray bar) contributions. The remaining b-to-m shifts in residue contributions to binding energies are derived from SAM, CAT, and Mg2+ and recapitulate earlier observations of Mg2+ increasing binding and CAT decreasing binding (see Fig 8). This difference may be rationalized by an increase in solvent exposure of Mg2+ in the m-CAT configuration as well as a more stabilizing coordination sphere from Glu199 even in the absence of m-CAT (i.e. in the receptor alone). CAT gas phase electrostatic binding is greatly reduced (20 kcal/mol), although this effect is counterbalanced slightly by a reduction in the repulsive pol+vdW contributions for overall binding (17 kcal/mol). Changes in SAM contribution to binding are weakly attractive by -1 kcal/mol, likely due to reorientation of the CAT O-. Finally, we consider the shift in residue-by-residue interactions for b-CAT and m-CAT using MMGBSA. We compare the absolute energies of the b-CAT and m-CAT complexes and identify a large number of residue pairs that contribute more than 1 kcal/mol difference in pairwise-residue total interaction energies. The map shown in Fig 10 is a 2D projection of the center of mass of each residue in the protein that has pairwise-residue interaction energy shifts indicated by a line connecting residues (red for increasing, blue for decreasing). Most of the strongest pairwise changes are adjacent to CAT in the active site and consistent with earlier EDA. From b-CAT to m-CAT, the CAT-Mg2+ interactions are weakened, Lys46 interactions increase with CAT and decrease with Mg2+, and Mg2+ interactions with nearly all other directly coordinating residues (Glu199, Asp169, Asp141, and Asn170) are strengthened. Other strong b-to-m shifts are strengthened binding of SAM to the protein via a SAM carboxylate-Asn41 sidechain interaction and hydrogen bonding to the Glu90 sidechain. 10.1371/journal.pone.0161868.g010Fig 10 Network graph of interaction shifts. Graph (full left and inset right) of difference in total pairwise residue interaction energies between bidentate and monodentate catecholate binding (greater than 1 kcal/mol). Line color indicates strengthening (red) or weakening (blue) with saturation of colors at +/- 15 kcal/mol. Subtler shifts in pairwise interactions highlight key residues that have been identified in earlier work but also suggest new interactions that may warrant further study. In m-CAT, the "gatekeeper" residues[12] Trp38 and Pro174 have strengthened interactions with Glu199 and CAT, respectively. Tyr68 interactions with Glu6 strengthen while interactions with Met40 weaken, suggesting a slight movement of Glu6 and Tyr68 away from the active site in the monodentate structure. COMT Tyr68Ala mutants have been shown experimentally and computationally[22, 24] to be less catalytically competent, strongly implicating Tyr68 as a key residue to mediate methyl transfer. Although Tyr200 has been identified as a component of the substrate binding pocket[25], one strong pairwise interaction shift involving residues not highlighted in previous work is the Asp30-Tyr200 pairwise interaction. The change in sidechain hydrogen bonding between the solvent- and active-site-facing loop Tyr200 and α-helix Asp30 disrupts the loop orientation. This observation is consistent with experiments that have identified significant loop movements are occur in COMT when comparing an ensemble of structures ranging from apo to holo forms[90, 91]. Another important cluster of interaction shifts is the strengthening of Arg161 pairwise interactions with Asp136 and Lys162 along with Asp136 and Ser60 pairwise interaction weakening from b to m configurations. These residues are maximally distant from the active site, adjacent to the well-studied[22, 24, 29, 30, 43, 92] Val108Met polymorph site and could play a similar role in altering the solvent accessibility of the active site and stability of the protein as was suggested for Val108Met. More proximally, the α-helical residues Tyr71 and Arg75 have a strengthened interaction adjacent to a number of active site residues. These results motivate future extensions to use this energy decomposition analysis to guide computational mutagenesis. Conclusions We have used classical and semi-empirical quantum mechanical methods to investigate multiple substrate binding modes observed in computation and experiment and how they influence mechanistic predictions in COMT. At the classical level, a monodentate CAT structure in which only a single oxygen anion is coordinated to Mg2+ is preferred by 4 kcal/mol over the experimental structure in which both CAT oxygen atoms are coordinated to Mg2+. However, the free energy barrier for transfer between these two basins of around 7 kcal/mol leads to 350 ns or longer stabilization of the bidentate structure in MD trajectories after careful equilibration of the protein. Although the barrier for substrate interconversion is substantially higher than kBT, it is still lower than the methyl transfer RDS, which is known to experimentally to have a free energy barrier around 18–19 kcal/mol. Although both binding modes have been remarked on, we have presented the first systematic determination of the free energy of these two binding modes. Our work suggests the importance of charge transfer in methyltransferase modeling: as charge transfer is permitted between Mg2+ and the catecholate substrate in the increasingly large SQM regions, i) the free energy barrier is reduced by around 10 kcal/mol to bring it into consistency with experiment and ii) the non-bonded C-O distance of the Michaelis complex (2.7 Å) simultaneously comes into agreement with short distances (2.5–2.8 Å) observed experimentally in COMT crystal structures for the bidentate configuration. Future work will be aimed at using accelerated, fully first-principles methods to quantify differences in stability and reactivity of diverse substrate binding poses and protonation states in MTase active sites. Supporting Information S1 Fig Bond distance distribution using different Mg2+ parameters. (TIF) Click here for additional data file. S2 Fig Bond distances from protocol 1. (TIF) Click here for additional data file. S3 Fig Bond distances from protocol 2. (TIF) Click here for additional data file. S4 Fig Bond distances from protocol 3. (TIF) Click here for additional data file. S5 Fig Bond distances from protocol 4. (TIF) Click here for additional data file. S6 Fig Dihedral angle distribution in restrained simulations. (TIF) Click here for additional data file. S7 Fig Oxygen charges and coordination environments for magnesium in bidentate and monodentate catecholcate (CAT) and dinitrocatecholate (DNC). Charges for CAT/DNC are obtained from RESP (HF/6-31G*) calculations as described in the text while the other charges are from the TIP3P or Amber ff12SB force fields. (TIF) Click here for additional data file. S8 Fig Free energy of rearrangement (in kcal/mol) of catecholate from bidentate to alternate monodentate configuration. Error bars shown are from Monte Carlo analysis. (TIF) Click here for additional data file. S9 Fig Representative, average bidentate (left) and monodentate (right) dinitrocatecholate (DNC) substrate configurations at the COMT active site. Substrates are shown in orange and labeled in brown while protein residues are shown in blue and labeled in dark blue. Key distances are labeled (in Å), except for D141-Mg2+, D1619-Mg2+, and Mg2+-H2O, which are omitted for clarity. (TIF) Click here for additional data file. S10 Fig C-O distance distribution of alternate monodentate anionic oxygen and hydroxyl oxygen compared to standard monodentate and bidentate configurations. (TIF) Click here for additional data file. S11 Fig SQM region dependence of monodentate methyl transfer barriers. (TIF) Click here for additional data file. S12 Fig AM1/d and PM6 SMg free energy barrier comparison. (TIF) Click here for additional data file. S13 Fig Unconstrained SAM-CAT methyltransfer reaction coordinate computed with B3LYP and PM6. The pathway is obtained at the B3LYP/6-311++G* level of theory with nudged elastic band compared to single point energies obtained at the PM6 level of theory all treated with the COSMO implicit solvent model (ε = 10). (TIF) Click here for additional data file. S14 Fig Constrained methyl transfer reaction coordinate for PM6 vs B3LYP. The (d(S-C)-d(C-O) in Å) reaction coordinate is obtained for PM6 and B3LYP/6–311++G* in COSMO implicit solvent (ε = 10). (TIF) Click here for additional data file. S1 Table Mg2+ ion force field parameters. (DOCX) Click here for additional data file. S2 Table Dihedral angle and bond distances in restraints simulations with different force constant. (DOCX) Click here for additional data file. S3 Table Target d(S-C)-d(C-O) values for each window in umbrella sampling along with maximum and minimum values sampled for the bidentate configuration and SMgL semi-empirical model. (DOCX) Click here for additional data file. S4 Table Catecholate and dinitrocatecholate active site distances. (DOCX) Click here for additional data file. S5 Table SQM region dependence of monodentate and bidentate methyl transfer. (DOCX) Click here for additional data file. S6 Table SQM region dependence of geometric properties for the enzyme-substrate complex (ES) and transition state (TS). (DOCX) Click here for additional data file. S7 Table Comparison of reaction energies for methyltransfer to catecholate (CAT) from S-Adenosyl methionine (SAM) or trimethyl sulfonium (TMS) in the gas phase and with the COSMO solvation model (ε = 10 or ε = 78.4) computed with PM6 or B3LYP/6-311++G*. (DOCX) Click here for additional data file. S1 Text Extended discussion of equilibration protocols and force field parameters. (PDF) Click here for additional data file. S2 Text Discussion of PM6 vs. B3LYP reaction energies. (PDF) Click here for additional data file. S3 Text Description of MMPBSA. (PDF) Click here for additional data file. The authors acknowledge helpful conversations with Michael Funk and Adam H. Steeves. This work was carried out in part using computational resources from the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1053575. This work was also carried out in part using the computational resources of the Center for Nanoscale Materials (Carbon cluster), an Office of Science user facility, was supported by the U. S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under Contract No. DE-AC02-06CH11357. ==== Refs References 1 Senn HM , Thiel W . QM/MM methods for biomolecular systems . Angewandte Chemie International Edition . 2009 ;48 (7 ):1198 –229 .19173328 2 Carvalho ATP , Barrozo A , Doron D , Kilshtain AV , Major DT , Kamerlin SCL . Challenges in computational studies of enzyme structure, function and dynamics . J Mol Graphics Modell . 2014 ;54 :62 –79 . 3 Gao J , Truhlar DG . Quantum mechanical methods for enzyme kinetics . Annual Review of Physical Chemistry . 2002 ;53 (1 ):467 –505 . 4 Martí S , Roca M , Andrés J , Moliner V , Silla E , Tuñón I , et al Theoretical insights in enzyme catalysis . Chemical Society Reviews . 2004 ;33 (2 ):98 –107 . 14767505 5 Warshel A , Levitt M . Theoretical studies of enzymic reactions: dielectric, electrostatic and steric stabilization of the carbonium ion in the reaction of lysozyme . J Mol Biol . 1976 ;103 (2 ):227 –49 . 985660 6 Hu H , Yang W . Free energies of chemical reactions in solution and in enzymes with ab initio QM/MM methods . Annual review of physical chemistry . 2008 ;59 :573 10.1146/annurev.physchem.59.032607.093618 18393679 7 Xu D , Cui Q , Guo H . Quantum mechanical/molecular mechanical studies of zinc hydrolases . International Reviews in Physical Chemistry . 2014 ;33 (1 ):1 –41 . 8 van der Kamp MW , Mulholland AJ . Combined quantum mechanics/molecular mechanics (QM/MM) methods in computational enzymology . Biochemistry . 2013 ;52 (16 ):2708 –28 . 10.1021/bi400215w 23557014 9 Axelrod J , Tomchick R . Enzymatic O-methylation of epinephrine and other catechols . J Biol Chem . 1958 ;233 (3 ):702 –5 . 13575440 10 Cheng X , Kumar S , Posfai J , Pflugrath JW , Roberts RJ . Crystal structure of the Hhal DNA methyltransferase complexed with S-adenosyl-l-methionine . Cell . 1993 ;74 (2 ):299 –307 . 8343957 11 Labahn J , Granzin J , Schluckebier G , Robinson DP , Jack WE , Schildkraut I , et al Three-dimensional structure of the adenine-specific DNA methyltransferase M.Taq I in complex with the cofactor S-adenosylmethionine . Proc Natl Acad Sci U S A . 1994 ;91 (23 ):10957 –61 . 7971991 12 Lotta T , Vidgren J , Tilgmann C , Ulmanen I , Melen K , Julkunen I , et al Kinetics of Human Soluble and Membrane-Bound Catechol O-Methyltransferase: A Revised Mechanism and Description of the Thermolabile Variant of the Enzyme . Biochemistry . 1995 ;34 (13 ):4202 –10 . 7703232 13 Du X , Schwander M , Moresco EMY , Viviani P , Haller C , Hildebrand MS , et al A catechol-O-methyltransferase that is essential for auditory function in mice and humans . Proc Natl Acad Sci U S A . 2008 ;105 (38 ):14609 –14 . 10.1073/pnas.0807219105 18794526 14 Coward JK , Slisz EP , Wu FYH . Kinetic studies on catechol O-methyltransferase, Product inhibition and the nature of the catechol binding site . Biochemistry . 1973 ;12 (12 ):2291 –7 . 4736330 15 Vidgren J , Svensson LA , Liljas A . Crystal structure of catechol O-methyltransferase . Nature . 1994 ;368 (6469 ):354 –8 . 8127373 16 Hegazi MF , Borchardt RT , Schowen RL .alpha.-Deuterium and carbon-13 isotope effects for methyl transfer catalyzed by catechol O-methyltransferase. SN2-like transition state . J Am Chem Soc . 1979 ;101 (15 ):4359 –65 . 17 Woodard RW , Tsai MD , Floss HG , Crooks PA , Coward JK . Stereochemical course of the transmethylation catalyzed by catechol O-methyltransferase . J Biol Chem . 1980 ;255 (19 ):9124 –7 . 6997310 18 Axelrod JJ . Methylation Reactions In The Formation And Metabolism Of Catecholamines And Other Biogenic Amines . Pharmacol Rev . 1966 ;18 (1 ):95 –113 . 5323771 19 Creveling CR , Dalgard N , Shimizu H , Daly JW . Catechol O-Methyltransferase: III. m- and p-O-Methylation of Catecholamines and Their Metabolites . Mol Pharmacol . 1970 ;6 (6 ):691 –6 . 5497718 20 Schultz E , Nissinen E . Inhibition of rat liver and duodenum soluble catechol-O-methyltransferase by a tight-binding inhibitor OR-462 . Biochem Pharmacol . 1989 ;38 (22 ):3953 –6 . 2597177 21 Lautala P , Ulmanen I , Taskinen J . Molecular Mechanisms Controlling the Rate and Specificity of Catechol O-Methylation by Human Soluble Catechol O-Methyltransferase . Mol Pharmacol . 2001 ;59 (2 ):393 –402 . 11160877 22 Zhang J , Klinman JP . Enzymatic Methyl Transfer: Role of an Active Site Residue in Generating Active Site Compaction That Correlates with Catalytic Efficiency . J Am Chem Soc . 2011 ;133 (43 ):17134 –7 . 10.1021/ja207467d 21958159 23 Mihel I , Knipe JO , Coward JK , Schowen RL .alpha.-Deuterium isotope effects and transition-state structure in an intramolecular model system for methyl-transfer enzymes . J Am Chem Soc . 1979 ;101 (15 ):4349 –51 . 24 Zhang J , Kulik HJ , Martinez TJ , Klinman JP . Mediation of donor–acceptor distance in an enzymatic methyl transfer reaction . Proc Natl Acad Sci U S A . 2015 ;112 (26 ):7954 –9 . 10.1073/pnas.1506792112 26080432 25 Lau EY , Bruice TC . Importance of correlated motions in forming highly reactive near attack conformations in catechol O-methyltransferase . J Am Chem Soc . 1998 ;120 (48 ):12387 –94 . 26 Lau EY , Bruice TC . Comparison of the Dynamics for Ground-State and Transition-State Structures in the Active Site of Catechol O-Methyltransferase . J Am Chem Soc . 2000 ;122 (30 ):7165 –71 . 27 Kuhn B , Kollman PA . QM−FE and Molecular Dynamics Calculations on Catechol O-Methyltransferase:  Free Energy of Activation in the Enzyme and in Aqueous Solution and Regioselectivity of the Enzyme-Catalyzed Reaction . J Am Chem Soc . 2000 ;122 (11 ):2586 –96 . 28 Zheng YJ , Bruice TC . A theoretical examination of the factors controlling the catalytic efficiency of a transmethylation enzyme: Catechol O-methyltransferase . J Am Chem Soc . 1997 ;119 (35 ):8137 –45 . 29 Rutherford K , Bennion BJ , Parson WW , Daggett V . The 108M Polymorph of Human Catechol O-Methyltransferase Is Prone to Deformation at Physiological Temperatures† . Biochemistry . 2006 ;45 (7 ):2178 –88 . 16475806 30 Rutherford K , Daggett V . A Hotspot of Inactivation: The A22S and V108M Polymorphisms Individually Destabilize the Active Site Structure of Catechol O-Methyltransferase . Biochemistry . 2009 ;48 (27 ):6450 –60 . 10.1021/bi900174v 19435324 31 Tsao D , Diatchenko L , Dokholyan NV . Structural Mechanism of S-Adenosyl Methionine Binding to Catechol O-Methyltransferase . PLoS ONE . 2011 ;6 (8 ):e24287 10.1371/journal.pone.0024287 21904625 32 Lameira J , Bora RP , Chu ZT , Warshel A . Methyltransferases do not work by compression, cratic, or desolvation effects, but by electrostatic preorganization . Proteins: Struct, Funct, Bioinf . 2015 ;83 (2 ):318 –30 . 33 Roca M , Martí S , Andrés J , Moliner V , Tuñón I , Bertrán J , et al Theoretical modeling of enzyme catalytic power: analysis of “cratic” and electrostatic factors in catechol O-methyltransferase . J Am Chem Soc . 2003 ;125 (25 ):7726 –37 . 12812514 34 Kanaan N , Ruiz Pernia JJ , Williams IH . QM/MM simulations for methyl transfer in solution and catalysed by COMT: ensemble-averaging of kinetic isotope effects . Chem Commun (Cambridge, U K) . 2008 ;(46 ):6114 –6 . 35 Kulik HJ , Drennan CL . Substrate Placement Influences Reactivity in Non-heme Fe (II) Halogenases and Hydroxylases . J Biol Chem . 2013 ;288 :11233 –41 . 10.1074/jbc.M112.415570 23449977 36 Martinie RJ , Livada J , Chang W-c , Green MT , Krebs C , Bollinger JM , et al Experimental Correlation of Substrate Position with Reaction Outcome in the Aliphatic Halogenase, SyrB2 . J Am Chem Soc . 2015 . 37 Rod TH , Ryde U . Quantum Mechanical Free Energy Barrier for an Enzymatic Reaction . Phys Rev Lett . 2005 ;94 (13 ):138302 15904045 38 Rod TH , Ryde U . Accurate QM/MM Free Energy Calculations of Enzyme Reactions:  Methylation by Catechol O-Methyltransferase . J Chem Theory Comput . 2005 ;1 (6 ):1240 –51 . 10.1021/ct0501102 26631668 39 Roca M , Moliner V , Tuñón I , Hynes JT . Coupling between Protein and Reaction Dynamics in Enzymatic Processes:  Application of Grote−Hynes Theory to Catechol O-Methyltransferase . J Am Chem Soc . 2006 ;128 (18 ):6186 –93 . 16669689 40 Ruggiero GD , Williams IH , Roca M , Moliner V , Tuñón I . QM/MM Determination of Kinetic Isotope Effects for COMT-Catalyzed Methyl Transfer Does Not Support Compression Hypothesis . J Am Chem Soc . 2004 ;126 (28 ):8634 –5 . 15250699 41 Case DA , Berryman JT , Betz RM , Cerutti DS , Cheatham TE III, Darden TA , et al AMBER 2015, University of California , San Francisco 2015 . 42 Salomon-Ferrer R , Case DA , Walker RC . An overview of the Amber biomolecular simulation package . Wiley Interdiscip Rev: Comput Mol Sci . 2013 ;3 (2 ):198 –210 . 43 Rutherford K , Le Trong I , Stenkamp RE , Parson WW . Crystal Structures of Human 108V and 108M Catechol O-Methyltransferase . J Mol Biol . 2008 ;380 (1 ):120 –30 . 10.1016/j.jmb.2008.04.040 18486144 44 Maier JA , Martinez C , Kasavajhala K , Wickstrom L , Hauser KE , Simmerling C . ff14SB: Improving the Accuracy of Protein Side Chain and Backbone Parameters from ff99SB . J Chem Theory Comput . 2015 ;11 (8 ):3696 –713 . 10.1021/acs.jctc.5b00255 26574453 45 Hornak V , Abel R , Okur A , Strockbine B , Roitberg A , Simmerling C . Comparison of multiple Amber force fields and development of improved protein backbone parameters . Proteins: Struct, Funct, Bioinf . 2006 ;65 (3 ):712 –25 . 46 Wang J , Wolf RM , Caldwell JW , Kollman PA , Case DA . Development and testing of a general amber force field . J Comput Chem . 2004 ;25 (9 ):1157 –74 . 15116359 47 Bayly CI , Cieplak P , Cornell W , Kollman PA . A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model . J Phys Chem . 1993 ;97 (40 ):10269 –80 . 48 Gordon MS , Schmidt MW . Advances in electronic structure theory: GAMESS a decade later . Theory Appl Comput Chem: First Forty Years . 2005 :1167 –89 . 49 Harihara PC , Pople JA . Influence of Polarization Functions on Molecular-Orbital Hydrogenation Energies . Theor Chim Acta . 1973 ;28 (3 ):213 –22 . 50 Wang F, Becker J-P, Cieplak P, Dupradeau F-Y. R.E.D. Python: Object oriented programming for Amber force fields, Université de Picardie—Jules Verne, Sanford|Burnham Medical Research Institute, Nov. 2013 [6/16/15]. Available from: http://q4md-forcefieldtools.org/REDServer-Development/. 51 Vanquelef E , Simon S , Marquant G , Garcia E , Klimerak G , Delepine JC , et al R.E.D. Server: a web service for deriving RESP and ESP charges and building force field libraries for new molecules and molecular fragments . Nucleic Acids Res . 2011 ;39 (suppl 2 ):W511 –W7 .21609950 52 Dupradeau F-Y , Pigache A , Zaffran T , Savineau C , Lelong R , Grivel N , et al The R.E.D. tools: advances in RESP and ESP charge derivation and force field library building . Phys Chem Chem Phys . 2010 ;12 (28 ):7821 –39 . 10.1039/c0cp00111b 20574571 53 Allnér O , Nilsson L , Villa A . Magnesium Ion–Water Coordination and Exchange in Biomolecular Simulations . J Chem Theory Comput . 2012 ;8 (4 ):1493 –502 . 10.1021/ct3000734 26596759 54 H++ server v3.1. [6/16/2015]. Available from: http://biophysics.cs.vt.edu/H++. 55 Anandakrishnan R , Aguilar B , Onufriev AV . H++ 3.0: automating pK prediction and the preparation of biomolecular structures for atomistic molecular modeling and simulations . Nucleic Acids Res . 2012 ;40 (W1 ):W537 –W41 .22570416 56 Gordon JC , Myers JB , Folta T , Shoja V , Heath LS , Onufriev A . H++: a server for estimating pKas and adding missing hydrogens to macromolecules . Nucleic Acids Res . 2005 ;33 (suppl 2 ):W368 –W71 .15980491 57 Myers J , Grothaus G , Narayanan S , Onufriev A . A simple clustering algorithm can be accurate enough for use in calculations of pKs in macromolecules . Proteins: Struct, Funct, Bioinf . 2006 ;63 (4 ):928 –38 . 58 Jorgensen WL , Chandrasekhar J , Madura JD , Impey RW , Klein ML . Comparison of simple potential functions for simulating liquid water . J Chem Phys . 1983 ;79 (2 ):926 –35 . 59 Ryckaert J-P , Ciccotti G , Berendsen HJC . Numerical integration of the cartesian equations of motion of a system with constraints: molecular dynamics of n-alkanes . J Comput Phys . 1977 ;23 (3 ):327 –41 . 60 Souaille M , Roux Bt . Extension to the weighted histogram analysis method: combining umbrella sampling with free energy calculations . Comput Phys Commun . 2001 ;135 (1 ):40 –57 . 61 Kumar S , Rosenberg JM , Bouzida D , Swendsen RH , Kollman PA . THE weighted histogram analysis method for free-energy calculations on biomolecules. I. The method . J Comput Chem . 1992 ;13 (8 ):1011 –21 . 62 Grossfield A. "WHAM: the weighted histogram analysis method", version 2.0.9 [06/15/2015]. Available from: http://membrane.urmc.rochester.edu/content/wham. 63 Torrie GM , Valleau JP . Nonphysical sampling distributions in Monte Carlo free-energy estimation: Umbrella sampling . J Comput Phys . 1977 ;23 (2 ):187 –99 . 64 Dewar MJS , Zoebisch EG , Healy EF , Stewart JJP . Development and use of quantum mechanical molecular models. 76. AM1: a new general purpose quantum mechanical molecular model . J Am Chem Soc . 1985 ;107 (13 ):3902 –9 . 65 Stewart JJP . Optimization of parameters for semiempirical methods V: Modification of NDDO approximations and application to 70 elements . J Mol Model . 2007 ;13 (12 ):1173 –213 . 17828561 66 Lee C , Yang W , Parr RG . Development of the Colle-Salvetti correlation-energy formula into a functional of the electron density . Phys Rev B . 1988 ;37 (2 ):785 –9 . 67 Becke AD . Density‐functional thermochemistry. III. The role of exact exchange . J Chem Phys . 1993 ;98 (7 ):5648 –52 . 68 Stephens PJ , Devlin FJ , Chabalowski CF , Frisch MJ . Ab Initio Calculation of Vibrational Absorption and Circular Dichroism Spectra Using Density Functional Force Fields . J Phys Chem . 1994 ;98 (45 ):11623 –7 . 69 Miller BR III, McGee TD Jr, Swails JM , Homeyer N , Gohlke H , Roitberg AE . MMPBSA. py: an efficient program for end-state free energy calculations . J Chem Theory Comput . 2012 ;8 (9 ):3314 –21 . 10.1021/ct300418h 26605738 70 Massova I , Kollman PA . Combined molecular mechanical and continuum solvent approach (MM-PBSA/GBSA) to predict ligand binding . Perspect Drug Discovery Des . 2000 ;18 (1 ):113 –35 . 71 Sharp KA , Honig B . Calculating total electrostatic energies with the nonlinear Poisson-Boltzmann equation . J Phys Chem . 1990 ;94 (19 ):7684 –92 . 72 Tsui V , Case DA . Theory and applications of the generalized born solvation model in macromolecular simulations . Biopolymers . 2000 ;56 (4 ):275 –91 . 11754341 73 Onufriev A , Bashford D , Case DA . Exploring protein native states and large‐scale conformational changes with a modified generalized born model . Proteins: Struct, Funct, Bioinf . 2004 ;55 (2 ):383 –94 . 74 Hou T , Wang J , Li Y , Wang W . Assessing the Performance of the MM/PBSA and MM/GBSA Methods. 1. The Accuracy of Binding Free Energy Calculations Based on Molecular Dynamics Simulations . J Chem Inf Model . 2011 ;51 (1 ):69 –82 . 10.1021/ci100275a 21117705 75 Bonifácio MJ , Archer M , Rodrigues ML , Matias PM , Learmonth DA , Carrondo MA , et al Kinetics and Crystal Structure of Catechol-O-Methyltransferase Complex with Co-Substrate and a Novel Inhibitor with Potential Therapeutic Application . Mol Pharmacol . 2002 ;62 (4 ):795 –805 . 12237326 76 Palma PN , Rodrigues ML , Archer M , Bonifácio MJ , Loureiro AI , Learmonth DA , et al Comparative Study of ortho- and meta-Nitrated Inhibitors of Catechol-O-methyltransferase: Interactions with the Active Site and Regioselectivity of O-Methylation . Mol Pharmacol . 2006 ;70 (1 ):143 –53 . 16618795 77 Tsuji E , Okazaki K , Takeda K . Crystal structures of rat catechol-O-methyltransferase complexed with coumarine-based inhibitor . Biochem Biophys Res Commun . 2009 ;378 (3 ):494 –7 . 10.1016/j.bbrc.2008.11.085 19056347 78 Ellermann M , Lerner C , Burgy G , Ehler A , Bissantz C , Jakob-Roetne R , et al Catechol-O-methyltransferase in complex with substituted 3'-deoxyribose bisubstrate inhibitors . Acta Crystallogr Sect D-Biol Crystallogr . 2012 ;68 (3 ):253 –60 .22349227 79 Harrison ST , Poslusney MS , Mulhearn JJ , Zhao Z , Kett NR , Schubert JW , et al Synthesis and Evaluation of Heterocyclic Catechol Mimics as Inhibitors of Catechol-O-methyltransferase (COMT) . ACS Med Chem Lett . 2015 ;6 (3 ):318 –23 . 10.1021/ml500502d 25815153 80 Tsao D , Liu S , Dokholyan NV . Regioselectivity of catechol O-methyltransferase confers enhancement of catalytic activity . Chem Phys Lett . 2011 ;506 (4–6 ):135 –8 . 21731105 81 Gerhards M , Perl W , Schumm S , Henrichs U , Jacoby C , Kleinermanns K . Structure and vibrations of catechol and catechol⋅H2O(D2O) in the S0 and S1 state . J Chem Phys . 1996 ;104 (23 ):9362 –75 . 82 Dudev T , Lim C . Principles Governing Mg, Ca, and Zn Binding and Selectivity in Proteins . Chem Rev . 2003 ;103 (3 ):773 –88 . 12630852 83 Bleuzen A , Pittet PA , Helm L , Merbach AE . Water exchange on magnesium (II) in aqueous solution: a variable temperature and pressure 17O NMR study . Mag Reson Chem . 1997 ;35 (11 ):765 –73 . 84 Roca M , Andrés J , Moliner V , Tuñón I , Bertrán J . On the Nature of the Transition State in Catechol O-Methyltransferase. A Complementary Study Based on Molecular Dynamics and Potential Energy Surface Explorations . J Am Chem Soc . 2005 ;127 (30 ):10648 –55 . 16045352 85 Sparta M , Alexandrova AN . How Metal Substitution Affects the Enzymatic Activity of Catechol-O-Methyltransferase . PLoS ONE . 2012 ;7 (10 ):e47172 10.1371/journal.pone.0047172 23056605 86 Furuki T , Sakurai M , Inoue Y . An application of the reaction field theory to hydrated metal cations in the framework of the MNDO, AM1, and PM3 methods . J Comput Chem . 1995 ;16 (3 ):378 –84 . 87 Garcia-Viloca M , Gao J , Karplus M , Truhlar DG . How enzymes work: analysis by modern rate theory and computer simulations . Science . 2004 ;303 (5655 ):186 –95 . 14716003 88 Cao Y , Chen Z-J , Jiang H-D , Chen J-Z . Computational Studies of the Regioselectivities of COMT-Catalyzed Meta-/Para-O Methylations of Luteolin and Quercetin . J Phys Chem B . 2014 ;118 (2 ):470 –81 . 10.1021/jp410296s 24354565 89 Bai H-W , Shim J-Y , Yu J , Zhu BT . Biochemical and Molecular Modeling Studies of the O-Methylation of Various Endogenous and Exogenous Catechol Substrates Catalyzed by Recombinant Human Soluble and Membrane-Bound Catechol-O-Methyltransferases† . Chem Res Toxicol . 2007 ;20 (10 ):1409 –25 . 17880176 90 Tsuji E , Okazaki K , Isaji M , Takeda K . Crystal structures of the Apo and Holo form of rat catechol-O-methyltransferase . J Struct Biol . 2009 ;165 (3 ):133 –9 . 10.1016/j.jsb.2008.11.012 19111934 91 Ehler A , Benz J , Schlatter D , Rudolph MG . Mapping the conformational space accessible to catechol-O-methyltransferase . Acta Crystallogr Sect D-Biol Crystallogr . 2014 ;70 (8 ):2163 –74 .25084335 92 Rutherford K , Alphandéry E , McMillan A , Daggett V , Parson WW . The V108M mutation decreases the structural stability of catechol O-methyltransferase . Biochim Biophys Acta, Proteins Proteomics . 2008 ;1784 (7–8 ):1098 –105 .

PMC005xxxxxx/PMC5001634.txt

==== Front PLoS BiolPLoS BiolplosplosbiolPLoS Biology1544-91731545-7885Public Library of Science San Francisco, CA USA 2756485810.1371/journal.pbio.1002547PBIOLOGY-D-16-01004PerspectiveResearch and Analysis MethodsResearch AssessmentPeer ReviewPeople and PlacesPopulation GroupingsProfessionsScientistsBiology and Life SciencesCell BiologyCellular TypesAnimal CellsNeuronsBiology and Life SciencesNeuroscienceCellular NeuroscienceNeuronsResearch and Analysis MethodsResearch AssessmentReproducibilitySocial SciencesEconomicsExperimental EconomicsSocial SciencesEconomicsSocial SciencesPolitical SciencePublic PolicyTaxesSocial SciencesEconomicsHealth EconomicsHealth InsuranceMedicine and Health SciencesHealth CareHealth EconomicsHealth InsuranceTruth in Science Publishing: A Personal Perspective Südhof Thomas C. *Department of Molecular and Cellular Physiology and Howard Hughes Medical Institute, Stanford University Medical School, Stanford, California, United States of AmericaThe author has declared that no competing interests exist. * E-mail: tcs1@stanford.edu26 8 2016 8 2016 26 8 2016 14 8 e1002547© 2016 Thomas C. Südhof2016Thomas C. SüdhofThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Scientists, public servants, and patient advocates alike increasingly question the validity of published scientific results, endangering the public’s acceptance of science. Here, I argue that emerging flaws in the integrity of the peer review system are largely responsible. Distortions in peer review are driven by economic forces and enabled by a lack of accountability of journals, editors, and authors. One approach to restoring trust in the validity of published results may be to establish basic rules that render peer review more transparent, such as publishing the reviews (a practice already embraced by some journals) and monitoring not only the track records of authors but also of editors and journals. In this Perspective, Thomas C. Südhof describes some of the current challenges to the peer review system that have endangered public acceptance of science and discusses possible avenues to addressing these challenges. The author received no specific funding for this work. ==== Body As scientists, we are part of society. Thus, it may be expected that questions about reproducibility of published studies, which are, in effect, questions about the truthfulness of these studies, are fair game [1,2]. After all, much science is publicly funded, and the public has a right to ask whether their funds are well spent. What’s more, in a time when many publicly held and expressed opinions are patently false (although it is not always clear whether the people who express these opinions are aware that they are lies), it is important for an engaged citizenry to demand evidence in support of claims. For too long, elected officials in the United States have not been challenged when they claim, for example, that global warming does not exist, that tax cuts will increase tax revenues, and that government-run health insurance is less efficient than privately run health insurance, even though in each case, the evidence indicates the opposite. But scientists, unlike politicians, are supposed to be guided by facts. Scientists should be insulated from the patent disregard of facts in public discourse. Why, then, are increasing numbers of scientific studies becoming suspect? In my personal view, we as scientists must accept at least part of the blame for failing to ensure that studies report facts and not fantasy. The increasingly severe problem here stems not from outright fraud, which continues to be rare. Instead, the problem is due to biased interpretations of experimental results. These interpretations lead to exaggerated statements of fact (a.k.a. conclusions), which not infrequently are only distantly related to the actual data in a paper. As a concrete example, the last decade witnessed hundreds of neuroscience papers in which an animal’s behavior is analyzed after defined populations of neurons were excited or inhibited using optogenetic methods. These papers provide valuable information, but commonly conclude that the manipulated neurons physiologically perform the studied behaviors, which change during the manipulations. However, these manipulations induce massive changes in neuronal activity that do not replicate the normal operation of the affected neurons. Moreover, the large activity changes that are induced propagate throughout the brain and thus likely induce myriads of downstream effects. As a result, conclusions from such experiments about the normal functions of the manipulated neurons are difficult to sustain without complementary, independent evidence [3]. Indeed, sometimes ablating the same population of neurons whose optic manipulation produces major effects has no behavioral consequences [4]. Two checkpoints are meant to safeguard scientific truth and to prevent unjustified conclusions: peer review and reproducibility. Peer review examines the validity of the experiments and conclusions as presented, and reproducibility ensures that the experimental results and conclusions can be replicated. Both checkpoints are under threat primarily due to economic factors. Compromised Peer Review Peer review of scientific results occurs at multiple levels, among which the review of scientific manuscripts is arguably the most important because published papers provide the reference point for all other types of peer review (such as selection of grants, promotions, lecture invitations, and prizes). Given the vast amount of data produced, journals serve a valuable role in identifying results and conclusions that merit attention—nobody can possibly read all papers in a field! In recent times, the role of journals in selecting studies has become all-powerful. At some journals, editors who direct the selection process have become akin to high priests and priestesses of science, and here a whiff of ancient Egypt with pharaohs controlling access to wisdom can permeate the review process. As a result, three problems have emerged in peer review that have corrupted the process, decreasing its value. First, journals and their peer reviewers often have a conflict of interest that is hidden. Journal profits depend on the broadness of the readership and on advertisers, leading to geographical biases (articles from economically important countries are preferred) and content biases (articles on trendy subjects are selected), while reviewers may have other agendas (e.g., supporting friends or holding an economic or professional stake in the results). Sometimes journals and reviewers may not even be aware of the corrupting influence of commercial interests. Attacking this problem will require more than just declaring consulting and ownership relationships. In an ideal world, a journal should not be funded by advertisers or subscribers but by authors’ fees, and reviewers should recuse themselves in cases of commercial and personal conflicts. As argued below, at a minimum, journals should be held accountable not only by their owners for the money they make but also by the public for the value they provide—just as a drug company cannot simply sell any drug but has to show that the drug is safe and effective, a journal should not be allowed to “sell” its products without being accountable for its content. This brings me to the second problem related to peer review: there is little accountability for journals and reviewers. If a journal repeatedly publishes papers that draw untenable conclusions, eventually the authors of the papers may be blamed, but editors and reviewers who are arguably responsible for gross negligence are not held responsible. There are insufficient checks and balances in the publishing system; when high-ranked journals repeatedly publish papers that are later considered unreliable or even retracted, the journals seem to face no consequences—their premier status remains untouched. The third peer review problem, finally, is that there is no real competition between journals as the conduit for communicating science. Capitalism thrives and depends on competition. Just like in many other commercial domains nowadays, however, authors have no true choice between journals. The majority of high-profile journals are run by a few companies with significant profits, and it is very hard for newcomers to break into this system. The lack of journal competition means that authors have limited choices in selecting journals with better peer review, decreasing the economic pressure on journals to invest in high-quality peer review. Endangered Reproducibility The other pillar of scientific truth, reproducibility, means that another scientist can repeat an experiment and arrive at the same results or, conversely, show that the results are not reproducible. Just as for peer review, multiple problems increasingly imperil reproducibility. For example, it’s not uncommon for an initial high-profile study to report amazing results with a stunning conclusion. Then, when the experiments are repeated, only trends toward the same conclusion are observed with increasingly smaller effect sizes. This outcome neither contradicts nor confirms the original study but is a dead end, and the original paper is slowly forgotten. As discussed above, the problem is not that the initial paper is fraudulent, but that the results were “tweaked” or selected, or represented a statistical outlier, leading to a misleading conclusion. A second emerging reproducibility problem is that many experiments are by design impossible to repeat. As formalized by Karl Popper [5], scientific truth requires interpersonal reproducibility. Based on this postulate, any conclusion that cannot be falsified because the underlying experiment cannot be repeated in exactly the same way is not a scientific conclusion. Many current experiments are so complex that differences in outcome can always be attributed to differences in experimental conditions (as is the case for many recent neuroscience studies because of the complexity of the nervous system). If an experiment depends on multiple variables that cannot be reliably held constant, the scientific community should not accept the conclusions from such an experiment as true or false. Such conclusions are simply non-scientific, even if based on an experiment. A third reproducibility problem is validation of reagents and methods. Too often, papers in premier journals are published without sufficient experimental controls—they take up too much space in precious journal real estate!—or with reagents that have not been vetted after they were acquired. Added to these reproducibility problems is the near impossibility of actually publishing negative results, owing to the reluctance of journals—largely motivated by economic pressures—to devote precious space to such papers, and to the reluctance of authors to acknowledge mistakes. Towards More Reliable Results Thus, we as the scientific community face major problems in ensuring the legitimacy of science. Although correcting these problems will not be trivial, simple steps could increase scientific truthfulness. In my personal view, there is no alternative to journals—we need journals as a filter, now more than ever, and journals need to be economically sustainable. However, given the robust profit margins of many journals, I feel that it is reasonable to insist that scientific journals adhere to a minimum set of rules. For example, reviews should be published, not hidden. Editors should be named as part of the published reviews and should be held accountable if papers fail to meet basic quality and reproducibility standards. Papers should be evaluated historically, not by citations (which can be misleading), but by tracking the follow-up to these papers. At least, the more prominent journals should systematically monitor subsequent work (or lack thereof) emerging from important studies. Submitted papers should be assessed by a checklist that ensures that proper controls and reagent validations are present, and such validations should be required for the supplementary materials. It is amazing how many prominent papers show immunoblots in which the supposed target proteins have the wrong size! Editors need to have the qualifications for judging the overall technical validity of experiments even if they cannot assess specific details (which is the job of the reviewers). Moreover, editors need to have the time to carefully read the papers and to understand the methods and experiments, and they should be paid better for the vast amount of work they are asked to do. Most importantly, as reviewers, we should emphasize less how exciting a result is even though it may not be true and focus more on whether a result is actually solid (i.e., true). A more demanding but possibly necessary change to ensure scientific truthfulness is to demonstrate immediate reproducibility. A conclusion should not be based on a single type of measurement, but on multiple parallel approaches. Ideally, scientists should recruit other groups to independently reproduce key results. Most pressing, however, in ensuring validity of scientific studies may be what I would call the common sense rule: the more a paper arouses amazement because it appears to contravene common sense and/or because it arrives at conclusions that diametrically differ from previous studies (also referred to as “novelty”), the more evidence should be required. Occasionally, studies that most challenge credulity are published with the least actual experimental support because such studies exude excitement—however, these are the studies that require the most experimental support! Many of these ideas have been expressed multiple times before. Never, however, has the need for action been more urgent than now, when our entire society is increasingly threatened by untruthfulness, with science being only a tiny part of it. Because the driving factors behind the threats to scientific practice are economic and political, we should speak up and express our concerns. As “voluntary” action seems unlikely, we should demand rules that inject accountability into the system, as capitalism without rules appears to become self-destructive and leads to self-sustaining and self-serving monopolies that impede progress. Provenance: Not commissioned; not externally peer reviewed. ==== Refs References 1 Open Science Collaboration . Estimating the reproducibility of psychological science . Science . 2015 ; 349 10.1126/science.aac4716 2 Allison DB , Brown AW , George BJ , Kaiser KA . Reproducibility: A tragedy of errors . Nature . 2016 ; 530 : 27 –29 . 10.1038/530027a 26842041 3 Südhof TC . Reproducibility: Experimental mismatch in neural circuits . Nature . 2015 ; 528 : 338 –339 . 10.1038/nature16323 26649825 4 Otchy TM , Wolff SBE , Rhee JY , Pehlevan C , Kawai R , Kempf A , et al Acute off-target effects of neural circuit manipulations . Nature . 2015 ; 528 : 358 –363 . 10.1038/nature16442 26649821 5 Keuth Herbert (Ed.): Popper Karl . Logik der Forschung . Akademie-Verlag , Berlin 2004 , ISBN 3-05-004085-8.

PMC005xxxxxx/PMC5001635.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756409310.1371/journal.pone.0161149PONE-D-16-13823Research ArticleMedicine and Health SciencesPediatricsMedicine and Health SciencesCritical Care and Emergency MedicineMedicine and Health SciencesDiagnostic MedicineMedicine and Health SciencesHealth CareHealth Care ProvidersMedical DoctorsPhysiciansPeople and PlacesPopulation GroupingsProfessionsMedical DoctorsPhysiciansMedicine and Health SciencesSurgical and Invasive Medical ProceduresPediatric SurgerySocial SciencesEconomicsLabor EconomicsSalariesMedicine and Health SciencesPediatricsChild HealthMedicine and Health SciencesPublic and Occupational HealthChild HealthPeople and PlacesPopulation GroupingsAge GroupsChildrenPeople and PlacesPopulation GroupingsFamiliesChildrenDedicated Pediatricians in Emergency Department: Shorter Waiting Times and Lower Costs Dedicated Pediatricians in Emergency DepartmentMelo Manuel Rocha 12*Ferreira-Magalhães Manuel 13Flor-Lima Filipa 1Rodrigues Mariana 1Severo Milton 24Almeida-Santos Luis 1Caldas-Afonso Alberto 12Barros Pedro Pita 56Ferreira António 71 Department of Pediatrics, Integrated Pediatric Hospital, Centro Hospitalar de S. João, Porto, Portugal2 Institute of Public Health, University of Porto, Porto, Portugal3 Center for Health Technologies and Information Systems Research, Faculty of Medicine, University of Porto, Porto, Portugal4 Department of Clinical Epidemiology, Predictive Medicine and Public Health, Faculty of Medicine, University of Porto, Porto, Portugal5 Lisbon Nova University, Lisboa, Portugal6 Centre for Economic Policy Research, London, United Kingdom7 Centro Hospitalar de S. João, Porto, PortugalZhang Harry EditorOld Dominion University, UNITED STATESCompeting Interests: The authors have declared that no competing interests exist. Conceived and designed the experiments: MRM MFM ACA AF. Performed the experiments: MRM MFM FFL MR. Analyzed the data: MRM MFM MS. Contributed reagents/materials/analysis tools: MS LAS PPB. Wrote the paper: MRM MFM ACA PPB AF. * E-mail: manuel.melo@hsjoao.min-saude.pt26 8 2016 2016 11 8 e01611495 4 2016 1 8 2016 © 2016 Melo et al2016Melo et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Background Dedicated pediatricians in emergency departments (EDs) may be beneficial, though no previous studies have assessed the related costs and benefits/harms. We aimed to evaluate the net benefits and costs of dedicated emergency pediatricians in a pediatric ED. Methods Cost-consequences analysis of visits to a pediatric ED of a tertiary hospital. Two pediatric ED Medical Teams (MT) were compared: MT-A (May–September 2012), with general pediatrics physicians only; and MT-B (May–September 2013), with emergency dedicated pediatricians. The main outcomes analyzed were relevant clinical outcomes, patient throughput time and costs. Results We included 8,694 children in MT-A and 9,417 in MT-B. Medication use in the ED increased from 42.3% of the children in MT-A to 49.6% in MT-B; diagnostic tests decreased from 24.2% in MT-A to 14.3% in MT-B. Hospitalization increased from 1.3% in MT-A to 3.0% in MT-B; however, there was no significant difference in diagnosis-related group relative weight of hospitalized children in MT-A and MT-B (MT-A, 0.979; MT-B, 1.075). No differences were observed in ED readmissions or in patients leaving without being seen by a physician. The patient throughput time was significantly shorter in MT-B, with faster times to first medical observation. Within the cost domains analyzed, the total expenditures per children observed in the ED were 16% lower in MT-B: 37.87 euros in MT-A; 31.97 euros in MT-B. Conclusion The presence of dedicated emergency pediatricians in a pediatric ED was associated with significantly lower waiting times in the ED, reduced costs, and similar clinical outcomes. The authors received no specific funding for this work. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction There are different models of pediatric emergency care worldwide.[1–5] Differences in facilities, age of admission, triage, referral or physician’s qualification may influence the outcomes in pediatric emergency care.[6–8] Over the last decade, pediatric emergency medicine has been progressively recognized, [9–10] first in the United States, Canada, [11, 12] and Australia and later in Europe, [13, 14] where it is now regarded as a subspecialty by the European Academy of Pediatrics.[15] However, the practice still varies widely across different countries and regions, [4] and little is known about the benefits of having dedicated emergency pediatricians in a pediatric ED. Specific performance measures for pediatric emergency care were identified and should be used to assess efficiency of pediatric emergency care systems.[16, 17] Evidence-based guidelines are required to standardize pediatric emergency care models and clinical practice procedures. In this context, analysis of the net benefits and related costs of dedicated emergency pediatricians teams are of major importance. Cost-consequences analysis is a comprehensive economic evaluation that fits the society’s health values, [18] and can be applied to the assessment of dedicated emergency pediatricians in ED. Objective results of not only costs, but also consequences, will inform decision-makers and improve emergency care for children. The present study aimed to assess the net benefits and costs of dedicated emergency pediatricians in a pediatric ED. The specific aims were to analyse the following outcomes: clinical outcomes; patient throughput time; and costs of hospital medication, diagnostic tests, consumable materials, and medical staff. Patients and Methods Study Setting Portuguese National Health System is a Beveridge-based system, supported by taxes and, therefore, the access is universal and free for every citizen. Centro Hospitalar de São João (CHSJ) is a public, mixed adult-children, and university tertiary-level hospital located in Porto, Portugal. Pediatric ED occupies a separate place within the hospital facilities, the catchment area is approximately 700 000 children/adolescents (0–17 years-old, inclusive) and there are on average 80 000 visits per year. Patients may direct themselves to the ED or be referred by a health professional, and the admissions to this ED are universal and totally free of charge. The patients undergo triage by trained nurses in accordance with the Pediatric Canadian Triage and Acuity Scale (Paed CTAS).[19] Medical staffing working in this ED is organized in 8 medical teams composed by physicians from three different hospitals: four teams from CHSJ, three from Centro Hospitalar do Porto, and one from Matosinhos Hospital. Each team ensures coverage of 24 hours per week, divided into two shifts of 12 hours. Until 2013, each shift comprised a medical team with five ‘general pediatrics consultants’–MT-A. In May 2013, in accordance with national guidelines, [20] CHSJ reorganized the medical staff for its four pediatric medical teams. After this reorganization, three ‘pediatric consultants dedicated to the ED’ and two ‘general pediatrics consultants’ composed each shift—MT-B. The medical teams were the only aspect that changed between these two periods; every other aspect of pediatric ED remained the same: facilities, electronic medical systems, non-medical staff (nurses and auxiliary personnel), and triage procedures. ‘General pediatricians’ were consultant physicians with board certification in pediatrics. These general pediatricians work 40 hours per week in inpatient, outpatient clinics and ED; each of these consultants work 12 hours per week in the pediatric ED. ‘Pediatricians dedicated to the ED’ were also consultant physicians with board certification in pediatrics; however, these consultants had an additional multidisciplinary training in pediatric emergency medicine. This training was conducted in full-time regimen (40 hours per week) over a 6-month period, and included practice in: pediatric intensive care, neonatal intensive care, pediatric interhospital transport, orthopedics, otorhinolaryngology, general surgery, and anesthetics. After this training period, pediatricians dedicated to ED started to work 40 hours per week exclusively in pediatric ED and interhospital transport of critically ill children. Study Design and Participants We performed a cost-consequences analysis of the implementation of MT-B in our pediatric ED, using patients’ electronic medical records and hospital administrative databases. To examine the performance of MT-A and MT-B, we selected comparable annual periods in consecutive years: May 1 to September 30, 2012 for MT-A; and May 1 to September 30, 2013 for MT-B. We included all patients (0–17 years-old, inclusive) that visited the pediatric ED of CHSJ. The exclusion criteria were: patients evaluated by pediatric medical teams from Centro Hospitalar do Porto and Matosinhos Hospital; and patients referred directly to other specialties in triage. Patients who left the pediatric ED without being seen by a physician (dropouts) were excluded from the overall analysis, but the frequency was compared between the two periods. Confidentiality of participants was assured by patient records anonymization and de-identification prior to analysis. This study was reviewed and approved by the Ethics Committee of CHSJ, Porto, Portugal. Outcome Measures The main comparisons were relevant clinical outcomes, patient throughput times, and costs. We also analysed demographic characteristics. Discharge diagnoses were classified in accordance with the International Classification of Diseases, Ninth Revision (ICD-9).[21] The relevant clinical outcomes considered were: need for medication within the ED; need for diagnostic tests within the ED; number of hospitalisations; readmission to ED in the 72-hour period after discharge; death in the ED; and patients leaving without being seen by a physician. In order to address potential differences in hospitalisations, we assessed the complexity of the hospitalised patients with the diagnosis-related group relative weight.[22] Patient throughput time was defined as the time from patient arrival to time of discharge from pediatric ED. This time was further divided into the following durations between: arrival and triage; triage and first medical observation; and first medical observation and discharge. Cost analysis included four different cost domains: hospital medication given inside the ED; diagnostic tests; clinical consumables; and medical staff (regular salary and overtime remuneration). As stated before, besides differences in medical teams composition, no other change occurred between MT-A and MT-B; also, our analysis was focused in costs potentially influenced by medical staff management. So we did not account for the costs with nursing, security professionals, facilities, administrative consumables or other general costs. Statistical Analysis Categorical variables (clinical outcomes) were characterized by counts and proportions. Numerical variables (throughput times) that presented a right-skewed distribution were log transformed and described with geometric mean (meang) and standard deviation (SDg). To compare numerical variables after log transformation, we used two-independent sample t test; for categorical variables, we employed a chi-square test. We categorized the effect size of the comparisons between the two ED models according to Cohen.[23] Considering that we are working with large samples, measures of effect size were preferred over significance tests to remove the dependence on sample size and the associated high probability of significant differences. For continuous variables (or the respective log transformation) the effect size was calculated as the difference between two means (MT-A and MT-B) divided by the pooled standard deviation. For categorical variables we used the formula ∑i(p1i−p0i)2/p0i, where p0i is the proportion of the ith cell under H0 and p1i is the proportion of the ith cell under H1 (observed proportion). For comparisons of the categorical variables (clinical outcomes), the effect size of 0.1–0.3 was considered small, 0.3–0.5 medium, and >0.5 large. For comparisons of numerical variables (throughput times), the effect size of 0.2–0.5 was considered small, 0.5–0.8 medium, and >0.8 large. To measure the association between clinical outcomes and the main exposure, we used the odds ratio (OR) and respective 95% confidence interval (95%CI). We estimated the OR by means of an unconditional logistic regression. We performed the cost analysis using the total cost in each time period (sum of the four cost domains for MT-A and for MT-B) divided by the number of patients observed in each one. The results for the costs are presented in euros per patient observed in the pediatric ED. We also performed a multiple quantile regression for the costs of diagnostic tests since that was the only cost domain where all the costs were specifically allocated to each patient observed in the ED.[24] We compare the diagnostic tests costs in MT-B with MT-A adjusting for age, month of admission, and level of triage. In particular, a quantile regression model follows the changes in coefficients and can indicate heterogeneity in the direction and magnitude of the associations between predictor variable (MT-B vs. MT-A) and the linear dependent variable (costs of diagnostic tests) from the mean estimates. We conducted the statistical analysis using SPSS version 22 (SPSS IBM, New York, NY, USA). Quantile regression was estimated using the library quantreg from the software R 2.14.1.[25] The level of significance was fixed at 0.05. Results From a total of 56 298 visits to pediatric ED, we included 8 694 (15.4%) in MT-A and 9 417 (16.7%) in MT-B. The study flowchart is presented in Fig 1 and demographic data summarized in Table 1. In the final sample, children aged 1 to 5 years old were the most frequent patients visiting the pediatric ED (MT-A—n = 4 282, 49.3%; MT-B—n = 4 666, 49.5%). 10.1371/journal.pone.0161149.g001Fig 1 Study flowchart. Patients included in each medical team (MT-A and MT-B) are shown; exclusion criteria appear in the central boxes. 10.1371/journal.pone.0161149.t001Table 1 Sample characteristics. Total MT-A MT-B p value (n = 18 111) (n = 8 694) (n = 9 417) Gender, n (%)  Male 9 402 (51.9) 4 494 (51.7) 4 908 (52.1) 0.565  Female 8 709 (48.1) 4 200 (48.3) 4 509 (47.9) Age, n (%)  <12 months 2 306 (12.7) 1 004 (11.5) 1 302 (13.8)  12–24 months 2 836 (15.7) 1 367 (15.7) 1 469 (15.6)  1–5 years 6 112 (33.7) 2 915 (33.5) 3 197 (33.9) <0.001  6–10 years 3 306 (18.3) 1 654 (19.0) 1 652 (17.5)  11–15 years 2 316 (12.8) 1 147 (13.2) 1 169 (12.4)  ≥16 years 1 235 (6.8) 607 (7.0) 628 (6.7) Month of visit, n (%)  May 3 906 (21.6) 2 099 (24.1) 1 807 (19.2)  June 3 829 (21.1) 1 537 (17.7) 2 292 (24.3)  July 3 560 (19.7) 1 913 (22.0) 1 647 (17.5) <0.001  August 3 086 (17.0) 1 485 (17.1) 1 601 (17.0)  September 3 730 (20.6) 1 660 (19.1) 2 070 (22.0) Origin, n (%)  Primary care 818 (4.5) 412 (4,7) 406 (4.3)  Other hospital 153 (0.9) 82 (0.9) 71 (0.8)  Private clinic 105 (0.6) 53 (0.6) 52 (0.6) 0.115  Health-care call center 402 (2.2) 172 (2.0) 230 (2.4)  Outpatient department 22 (0.1) 9 (0.1) 13 (0.1)  Without referral 16 611 (91.7) 7 966 (91.7) 8 645 (91.8) Canadian triage level, n (%)  Level 1 and 2 954 (5.3) 426 (4.9) 528 (5.6)  Level 3 8 455 (46.7) 4 117 (47.4) 4 338 (46.1)  Level 4 8 027 (44.3) 3 845 (44.2) 4 182 (44.4) 0.049  Level 5 675 (3.7) 306 (3.5) 369 (3.9) Triage destination, n (%)  Pediatric resuscitation room 161 (0.9) 55 (0.6) 106 (1.1)  Observation ward 629 (3.5) 211 (2.4) 418 (4.4) <0.001  Waiting room 17 321 (95.6) 8 428 (97.0) 8 893 (94.5) Disorders according to ICD-9*, n (%)  Infectious 2 869 (16.0) 1 456 (16.9) 1 413 (15.1)  CNS and Sense organs 1 480 (8.2) 692 (8.0) 788 (8.4)  Respiratory system 3 477 (19.3) 1 460 (16.9) 2 017 (21.6)  Digestive system 1 226 (6.8) 613 (7.1) 613 (6.6) 0.213  Signs and symptoms 4 180 (23.2) 2 034 (23.6) 2 146 (22.9)  Injury and poisoning 2 062 (11.5) 975 (11.3) 1 087 (11.6)  Others 2 696 (15.0) 1 402 (16.2) 1 294 (13.8) ICD-9: International Classification of Diseases, Ninth Revision; CNS: Central Nervous System. Paed CTAS level 3 patients were the most common (MT-A—n = 4 117, 47.4%; MT-B—n = 4 338, 46.1%). MT-B had more level 1 and level 2 Paed CTAS episodes than MT-A (MT-A—n = 426, 4.9%; MT-B—n = 528, 5.6%; p = 0.049), as well as more resuscitation room (MT-A—n = 55, 0.6%; MT-B—n = 106, 1.1%; p<0.001) and observation ward admissions (MT-A—n = 211, 2.4%; MT-B—n = 418, 4.4%; p<0.001). There were no statistically significant differences regarding gender and the origin of children that visited pediatric ED. Clinical Outcomes The number of children that received medication increased from 42.3% in MT-A to 49.6% in MT-B. Children who underwent diagnostic tests decreased from 24.2% in MT-A to 14.3% in MT-B (Table 2). The number of children hospitalized increased from 1.3% in MT-A to 3.0% in MT-B. However, all the effect sizes of these comparisons were extremely low (Table 2). There was no significant difference on the diagnosis-related group relative weight of hospitalized children (MT-A, 0.979; MT-B, 1.075; p = 0.45). 10.1371/journal.pone.0161149.t002Table 2 Clinical outcomes: proportions, effect size, and odds ratio (crude and adjusted) for comparisons between the two medical teams (MT-A and MT-B). Total, n (%) MT-A, n (%) MT-B, n (%) Effect size Crude OR (95%CI) Adjusted OR (95%CI) Medication use 8 348 3 677 4 671 0.103 1.34 1.40 (46.1) (42.3) (49.6) (1.27–1.42) (1.31–1.49) Diagnostic test use 3 453 2 106 1 347 0.178 0.52 0.52 (19.1) (24.2) (14.3) (0.48–0.56) (0.49–0.57) Hospitalization 397 115 282 0.082 2.30 2.26 (2.2) (1.3) (3.0) (1.85–2.87) (1.81–2.83) Readmission 1 264 575 689 0.020 1.12 1.11 (7.0) (6.6) (7.3) (0.99–1.25) (0.99–1.25) Bold values: p<0.001 for comparisons between the MT-A and MT-B OR, odds ratio; CI, confidence interval Adjusted OR: adjusted to age, month of visit, and level of triage. There was no statistically significant difference regarding readmissions to pediatric ED in the 72-hour period after discharge. In the periods under study, there was only one death within the ED (MT-A). There were 73 (0.8%) patients who left without being seen by a physician in MT-A and 85 (0.9%) in MT-B (p = 0.652). Patient throughput time Patient throughput time in the pediatric ED was significantly shorter in MT-B, as well as all the subdivisions of this time (Table 3). The effect sizes of those differences were classified as small except for duration between patient arrival and first medical observation, which had a medium effect size. 10.1371/journal.pone.0161149.t003Table 3 Patient throughput time (hours) in pediatric emergency department—comparison between the two medical teams (MT-A and MT-B). Total, meang (SDg) MT-A, meang (SDg) MT-B, meang (SDg) Effect size Patient throughput time in ED (hours) 1.84 (2.47) 2.08 (2.43) 1.65 (2.47) 0.248  Duration between arrival and triage (hours) 0.12 (1.98) 0.13 (2.06) 0.11 (1.88) 0.220  Duration between triage and first medical observation (hours) 0.28 (3.79) 0.39 (3.32) 0.20 (3.91) 0.513  Duration between first medical observation and discharge (hours) 0.48 (2.41) 0.58 (2.41) 0.40 (2.33) 0.414 Bold values: p<0.001 for comparisons between MT-A and MT-B. SDg, geometric standard deviation. The meang duration between arrival and triage was less than 10 minutes in both MT-A and MT-B. The meang duration between triage and first medical observation was 23.4 minutes in MT-A and 12 minutes in MT-B (p<0.001; effect size 0.513) and was inversely proportional to the priority levels of triage (S1 Fig). Cost Analysis Within the cost domains analysed, the total expenditure per patient observed in the pediatric ED was 16% lower in MT-B (37.87 euros in MT-A and 31.97 euros in MT-B; Fig 2). The relative costs of diagnostic tests, clinical consumables, and physicians’ overtime pay were lower in MT-B (relative decrease of 47%, 9%, and 62%, respectively). The relative costs of medication and physicians’ regular salary were higher in MT-B (relative increase of 26% and 13%, respectively). 10.1371/journal.pone.0161149.g002Fig 2 Specific costs per patient observed in the pediatric emergency department in medical teams (MT-A and MT-B). In the multivariate quantile regression, we observed that the highest savings in diagnostic tests observed in MT-B were mainly driven by a lower use of the cheaper (38% lower) and the costliest (15% higher) diagnostic tests. The costs for medium expenditure with diagnostic tests were similar in the two models (S2 Fig). Discussion This is the first study that evaluates net benefits (clinical outcomes and patient throughput time) and costs of dedicated emergency medicine pediatricians in the pediatric ED. The cost-consequences approach allowed us to perform a comprehensive analysis, aiming to impact the future decision-making in pediatric emergency care. We were able to describe and compare the two models using most of the recommended performance indicators for pediatric ED with the highest utility score, as ranked by the Royal College of Physicians and Surgeons of Canada.[16] In our study, the differences between general pediatricians and pediatricians dedicated to ED were mainly two: 1) the training program, i.e. a 6-month extra training period in pediatric emergency medicine for pediatricians dedicated to ED; 2) clinical activity, i.e. pediatricians dedicated to ED have a full-time work (40 hours per week) in pediatric ED and interhospital transport of critically ill children, as opposed to general pediatricians that work 40 hours per week in several areas of pediatric department (inpatient, outpatient clinics and ED) and, of these, only 12 hours per week in ED. We propose that the differences in training and clinical practice may give a greater know-how in treating children in ED, leading to better outcomes in the emergency setting: higher confidence in managing medication in the ED and less diagnostic tests use may decrease waiting times in the ED and, altogether, decrease the costs in pediatric ED. Clinical Outcomes There is a lack of published studies about diagnostic tests and medication prescription within the pediatric ED. In our study, we observed a decrease in diagnostic testing in MT-B and an increase in medication use, even when adjusted for patient age, level of triage, and month of ED visit. According to the specificity of the training program and clinical activities, pediatricians dedicated to ED were probably more prone to have an integrated approach to the children admitted to ED. This may explain why pediatric emergency physicians have had more restrictive criteria for employing diagnostic testing in an emergency setting and, in turn, may have had greater confidence in managing medication. Further studies are needed to address this hypothesis. Nevertheless, those differences were considered clinically irrelevant (see the low effect size in Table 2) and may also be related to our large sample size and the power to detect small differences. This point also applies to the increase in hospitalisations observed in MT-B (see the low effect size in Table 2). Moreover, we did not observed any difference in diagnosis-related group relative weight between the two periods, indicating that physician’s clinical criteria on 14ospitalization remained the same in MT-A and MT-B. The overall rate of 14ospitalization observed is similar to that found in other pediatric ED settings.[26] The rates of readmission to the ED after 72 hours and of patients left without being seen were extremely low in both periods and no differences were found. Previous studies have reported higher proportions (7%-16%) of patients leaving without being seen.[27,28] Patient throughput time The total patient throughput time inside the pediatric ED, as well as all the subdivided times, were significantly shorter in MT-B, with acceptable effect sizes (small to medium effect sizes; Table 3). We believe that this reduction in duration may be partly a consequence of the fewer diagnostic tests performed and lack of need to wait for test results. These findings support the effectiveness of dedicated emergency pediatricians in reducing waiting times / length of stay in the pediatric ED. Keijzers et al. reported a reduction of the total ED length of stay (in an adult-child mixed ED) with a pediatric medical team, but they were unable to find an effect on the waiting time to see a doctor.[29] In our setting, both models were composed of pediatricians only (or pediatric residents); the difference was in ‘general pediatricians’ vs ‘dedicated emergency pediatricians’. We did not make satisfaction inquiries, however, previous reports have demonstrated the importance of shorter waiting times on carer satisfaction levels and in the likelihood of recommending the pediatric ED.[29] Costs MT-B showed a 16% reduction in the total costs analysed (medications, diagnostic tests, clinical consumables, and physicians’ salaries). The highest saving was in physicians’ overtime pay (62% reduction). That saving was accompanied by a slight increase in the physicians’ regular salaries. That change was anticipated since the emergency pediatricians had been hired, and consequently general pediatrics physicians worked less overtime. The cost domain that most affected the final cost reduction in MT-B was diagnostic tests. The 47% decrease amounted to a reduction of 7.2 euros per patient observed. The savings were greater among patients who underwent less expensive diagnostic tests (38% lower) and in those who received the most costly diagnostic tests (15% higher), meaning that fewer cheaper and costly diagnostic tests were performed (the number of diagnostic tests of medium prices remained similar). Strengths Portuguese National Health System access is universal and free for every citizen so, our study population is representative of all the children, without having the possible confounding effect of social factors like health insurance or socioeconomic status. MT-A and MT-B differed only in the composition of the pediatric medical teams; all the other features were identical. This allowed us to assess the effect of having dedicated emergency pediatricians in a pediatric ED. In addition, we focus our analysis on the children attended by pediatricians and excluded patients referred to other specialties (Fig 1). We also chose the same time of year in 2012 and 2013 to minimize possible bias related to the incidence of acute pathologies linked to different seasons. Finally, the study sample size allowed us to greatly reduce the probability of type II errors. Limitations This study had some limitations. There were small differences in some baseline characteristics between the two models (Table 1); however, we adjusted our estimates to those variables, and the differences in the baseline characteristics were probably detected because of the large sample. The comparison between two different periods may had also not account for the small changes in practices that could have been occurred; nevertheless, we did not identify any major change in practices in pediatric ED. Finally, our results are fitted to the population of the studied hospital, and extrapolation to other settings has to be made with caution. Conclusion The presence of dedicated emergency pediatricians in a pediatric ED was associated to significantly lower waiting times in the ED, less diagnostic tests use and reduced costs. Clinical outcomes were similar in the two models studied. More studies on utility, benefits and costs of specialized pediatric emergency teams in pediatric ED are needed in order to help decision makers to improve pediatric emergency care models, maximizing health results in this setting. Supporting Information S1 Fig Comparison of estimated marginal means of duration between triage and first medical observation in MT-A and MT-B. Blue line: MT-A. Green line: MT-B. (TIFF) Click here for additional data file. S2 Fig Multivariate quantile regression for costs with diagnostic tests in MT-B compared with MT-A, adjusted for age, month of admission, and level of triage. Slash-dotted line: mean difference of MT-B compared with MT-A. Gray zone: 95% confidence interval for the mean difference. Red solid line shows the mean difference of MT-B compared with MT-A using the usual least squares linear regression. (TIFF) Click here for additional data file. ==== Refs References 1 Van Veen M , Moll HA . Reliability and validity of triage systems in paediatric emergency care . Scand J Trauma Resusc Emerg Med . 2009 ;17 :38 10.1186/1757-7241-17-38 19712467 2 Fernández A , Pijoan JI , Ares MI , Mintegi S , Benito FJ . Canadian Paediatric Triage and Acuity Scale: assessment in a European paediatric emergency department . Emergencias . 2010 ;22 :355 –360 . 3 American Academy of Pediatrics , Committee on Pediatric Emergency Medicine ; American College of Emergency Physicians , Pediatric Committee ; Emergency Nurses Association Pediatric Committee . Joint policy statement—guidelines for care of children in the emergency department . Pediatrics . 2009 ;124 (4 ):1233 –1243 . 10.1542/peds.2009-1807 19770172 4 Mintegi S , Shavit I , Benito J , REPEM group (Research in European Paediatric Emergency Medicine) . Paediatric emergency care in europe: a descriptive survey of 53 tertiary medical centers . Pediatr Emerg Care . 2008 ;24 (6 ):359 –363 . 10.1097/PEC.0b013e318177a762 18562877 5 Models of Emergency Care. Sydney. NSW Ministry of Health; 2012. http://www.health.nsw.gov.au/Performance/Publications/ed-model-of-care-2012.pdf Accessed January 29, 2016. 6 Sacchetti A , Baren J , Carraccio C . The paradox of the nested pediatric emergency department . Acad Emerg Med . 2005 ;12 (12 ):1236 –1239 . 16293893 7 Davies F, Gaushe-Hill M, Chu S, Cheema B, Ang A, Caceres L, et al. 2012 International Standards of Care for Children in Emergency Departments. Melbourne. International Federation for Emergency Medicine—Paediatric Special Interest Group; 2012. http://www.ifem.cc/Resources/PoliciesandGuidelines.aspx Accessed September 24, 2015. 8 Dharmar M , Marcin JP , Romano PS , Andrada ER , Overly F , Valente JH , et al Quality of Care of Children in the Emergency Department: Association with Hospital Setting and Physician Training . J Pediatr . 2008 ;153 (6 ):783 –789 . 10.1016/j.jpeds.2008.05.025 18617191 9 Li M , Baker MD , Ropp LJ . Pediatric emergency medicine: a developing subspecialty . Pediatrics . 1989 ;84 (2 ):336 –42 . 2748264 10 Pou Fernández J , Benito Fernández J . Pediatría de urgencias: una nueva especialidad . An Esp Pediatr . 2002 ;56 (1 ):2 –4 . 11 Curriculum Subcommittee , Section of Emergency Medicine , American Academy of Pediatrics . Pediatric emergency medicine (PEM) fellowship curriculum statement . Pediatr Emerg Care . 1993 ;9 (1 ):60 –66 . 8488151 12 Babl FE , Weiner DL , Bhanji F , Davies F , Berry K , Barnett P . Advanced training in pediatric emergency medicine in the United States, Canada, United Kingdom, and Australia: An international comparison and resources guide . Ann Emerg Med . 2005 ;45 (3 ):269 –275 . 15726049 13 Benito J , Mintegi S , Ruddy RM , Gonzalez del Rey JA . Changing Clinical Practices and Education in Pediatric Emergency Medicine Through Global Health Partnerships . Clin Pediatr Emerg Med . 2012 ;13 (1 ):37 –43 . 14 Salter R , Maconochie IK . Implementation of recommendations for the care of children in UK emergency departments: national postal questionnaire survey . BMJ . 2005 ;330 (7482 ):73 –74 . 15579478 15 Benito J, Mintegi S, Luaces C, Liviana D, Chérón G, Martinot A, et al (Working Group on Paediatric Emergency Medicine). The European Syllabus for Paediatric Emergency Medicine. European Academy of Paediatrics; Paediatric Section of the UEMS; UEMS Section of the European Board of Emergency Medicine; 2011. http://www.simeup.com/doc/syllabus.pdf Accessed January 29, 2016. 16 Hung GR , Chalut D . A consensus-established set of important indicators of pediatric emergency department performance . Pediatr Emerg Care . 2008 ;24 (1 ):9 –15 . 10.1097/pec.0b013e31815f39a5 18165798 17 Alessandrini E , Varadarajan K , Alpern ER , Gorelick MH , Shaw K , Ruddy RM , et al Emergency department quality: An analysis of existing pediatric measures . Acad Emerg Med . 2011 ;18 (5 ):519 –526 . 10.1111/j.1553-2712.2011.01057.x 21569170 18 Coast J . Is economic evaluation in touch with society's health values? BMJ . 2004 ;329 (7476 ):1233 –6 . 15550430 19 Gouin S , Gravel J , Amre DK , Bergeron S . Evaluation of the Paediatric Canadian Triage and Acuity Scale in a paediatric ED . Am J Emerg Med . 2005 ;23 (3 ):243 –247 . 15915392 20 Paiva JA, Silva AM, Almeida AL, Seco CM, Gomes CM, Ribeiro E, et al. Reavaliação da Rede Nacional de Emergência e Urgência; 2012. http://www.portaldasaude.pt/NR/rdonlyres/0323CC90-45A4-40E4-AA7A-7ACBC8BF6C75/0/ReavaliacaoRedeNacionalEmergenciaUrgancia.pdf Accessed January 29, 2016. 21 International Classification of Diseases, Ninth Revision (ICD-9). World Health Organization; 2011. http://www.cdc.gov/nchs/icd/icd9cm.htm Accessed January 29, 2016. 22 Fetter RB , Shin Y , Freeman JL , Averill RF , Thompson JD . Case mix definition by diagnosis related groups . Med Care . 1980 ;18 (2 ):1 –53 . 23 Cohen J . A power primer . Psychol Bull . 1992 ;112 (1 ):155 –159 . 19565683 24 Koenker RW , Bassett QW . Regression quantiles . Econometrica . 1978 ;46 :33 –50 25 R Core Team (2013). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.org/. Accessed January 29, 2016. 26 Hampers LC , Faries SG , Poole SR . Regional after-hours urgent care provided by a tertiary children’s hospital . Pediatrics . 2002 ;110 (6 ):1117 –1124 . 12456908 27 Ng Y , Lewena S . Leaving the paediatric emergency department without being seen: understanding the patient and the risks . J Paediatr Child Health . 2012 ;48 (1 ):10 –15 . 10.1111/j.1440-1754.2011.02187.x 21988657 28 Gaucher N , Bailey B , Gravel J . Who Are the Children Leaving the Emergency Department Without Being Seen by a Physician? Acad Emerg Med . 2011 ;18 (2 ):152 –157 . 10.1111/j.1553-2712.2010.00989.x 21314774 29 Keijzers G , Crilly J , Walters B , Crawford R , Bell C . Does a dedicated pediatric team within a busy mixed emergency department make a difference in waiting times, satisfaction, and care transition? Pediatr Emerg Care . 2010 ;26 (4 ):274 –280 . 20401972

PMC005xxxxxx/PMC5001636.txt

==== Front PLoS GenetPLoS GenetplosplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, CA USA 2756444910.1371/journal.pgen.1006268PGENETICS-D-15-02670Research ArticleBiology and Life SciencesCell BiologyChromosome BiologyChromosomesChromosome Structure and FunctionTelomeresBiology and life sciencesBiochemistryProteinsPost-translational modificationSUMOylationBiology and Life SciencesCell BiologyChromosome BiologyChromosomesBiology and life sciencesGeneticsGene expressionDNA transcriptionBiology and life sciencesGeneticsDNADNA replicationBiology and life sciencesBiochemistryNucleic acidsDNADNA replicationBiology and Life SciencesComputational BiologyGenome AnalysisChromatin ImmunoprecipitationBiology and Life SciencesGeneticsGenomicsGenome AnalysisChromatin ImmunoprecipitationResearch and Analysis MethodsElectrophoretic TechniquesGel ElectrophoresisElectrophoretic BlottingSouthern BlotBiology and Life SciencesMolecular BiologyMolecular Biology TechniquesMolecular Probe TechniquesElectrophoretic BlottingSouthern BlotResearch and Analysis MethodsMolecular Biology TechniquesMolecular Probe TechniquesElectrophoretic BlottingSouthern BlotBiology and Life SciencesOrganismsFungiYeastSaccharomycesSaccharomyces CerevisiaeResearch and Analysis MethodsModel OrganismsYeast and Fungal ModelsSaccharomyces CerevisiaeSmc5/6 Is a Telomere-Associated Complex that Regulates Sir4 Binding and TPE Smc5/6 Maintains Transcriptional Silencing at Telomerehttp://orcid.org/0000-0002-5611-2680Moradi-Fard Sarah 1Sarthi Jessica 1Tittel-Elmer Mireille 1http://orcid.org/0000-0002-0495-8967Lalonde Maxime 2http://orcid.org/0000-0002-7683-8542Cusanelli Emilio 2¤Chartrand Pascal 2Cobb Jennifer A. 1*1 Departments of Biochemistry & Molecular Biology and Oncology, Robson DNA Science Centre, Arnie Charbonneau Cancer Institute, Cumming School of Medicine, University of Calgary, Calgary, Alberta, Canada2 Département de Biochimie, Université de Montréal, Montréal, Quebec, CanadaMcKinnon Peter EditorSt Jude Children's Research Hospital, UNITED STATESThe authors declare that they have no conflict of interest. Conceived and designed the experiments: SMF PC JC. Performed the experiments: SMF JS MTE ML EC. Analyzed the data: SMF PC JC. Contributed reagents/materials/analysis tools: PC JC. Wrote the paper: SMF JC. ¤ Current Address: Max F. Perutz Laboratories, Department of Chromosome Biology, University of Vienna, Vienna, Austria * E-mail: jcobb@ucalgary.ca26 8 2016 8 2016 12 8 e10062682 11 2015 28 7 2016 © 2016 Moradi-Fard et al2016Moradi-Fard et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.SMC proteins constitute the core members of the Smc5/6, cohesin and condensin complexes. We demonstrate that Smc5/6 is present at telomeres throughout the cell cycle and its association with chromosome ends is dependent on Nse3, a subcomponent of the complex. Cells harboring a temperature sensitive mutant, nse3-1, are defective in Smc5/6 localization to telomeres and have slightly shorter telomeres. Nse3 interacts physically and genetically with two Rap1-binding factors, Rif2 and Sir4. Reduction in telomere-associated Smc5/6 leads to defects in telomere clustering, dispersion of the silencing factor, Sir4, and a loss in transcriptional repression for sub-telomeric genes and non-coding telomeric repeat-containing RNA (TERRA). SIR4 recovery at telomeres is reduced in cells lacking Smc5/6 functionality and vice versa. However, nse3-1/ sir4 Δ double mutants show additive defects for telomere shortening and TPE indicating the contribution of Smc5/6 to telomere homeostasis is only in partial overlap with SIR factor silencing. These findings support a role for Smc5/6 in telomere maintenance that is separate from its canonical role(s) in HR-mediated events during replication and telomere elongation. Author Summary Structural Maintenance of Chromosome (SMC) complexes, include cohesin, condensin, and the Smc5/6 complex. These complexes are involved in many aspects of chromosome organization, with cohesin and condensin having relatively well-characterized roles. Cohesin holds newly replicated DNA strands together and condesin is critical for chromosome condensation and genome compaction as cells enter mitosis. However, a role for the Smc5/6 complex in higher-level chromosome organization has remained ill defined. The Smc5/6 complex is recovered at chromosome ends in all stages of the cell cycle and has a role in telomere biology. Smc5/6 integrity is necessary for Sir4 binding, telomere clustering, and transcriptional silencing. In all, our data suggest that Smc5/6 has a physiologically relevant role in chromatin maintenance at telomeres and telomere organization within the nucleus that are distinct of it functionality in homologous recombination. http://dx.doi.org/10.13039/501100000024Canadian Institutes of Health ResearchMOP- 82736 ans MOP-137062Cobb Jennifer A. http://dx.doi.org/10.13039/501100000038Natural Sciences and Engineering Research Council of Canada418122Cobb Jennifer A. http://dx.doi.org/10.13039/501100000024Canadian Institutes of Health ResearchMOP-89768Chartrand Pascal This work was supported by the Canadian Institutes of Health Research MOP-82736; MOP-137062 and Natural Sciences and Engineering Research Council of Canada 418122 awarded to JC, Canadian Institutes of Health Research MOP-89768 awarded to PC. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction Structural maintenance of chromosome (SMC) protein complexes facilitate chromosome structure and organization in eukaryotes. Three SMC complexes are found in eukaryotes and each has a unique role in chromosome dynamics and metabolism. Underscoring their importance and distinct functionality, all three complexes and their individual components are essential for cell viability. Cohesin regulates sister chromatid cohesion and condensin is important for chromosome compaction by tethering different regions of the same chromosome [1–3]. The third complex, Smc5/6, contains six non-SMC proteins in addition to Smc5 and 6 including Mms21/ non-Smc element 2 (Nse2), which is an E3 SUMO ligase (Fig 1A) [4–6]. As well, Nse1 and Nse3 bind to Nse4 to form a heterotrimer, which in turn interacts with the ATPase head domain generated by the N- and C-termini of Smc5 and Smc6 [7, 8]. Nse1 is a putative ubiquitin ligase and Nse3 is a MAGE (melanoma-associated antigen gene) domain containing protein that is important for loading the complex onto chromatin [9–11]. The Smc5/6 complex functions in homologous recombination (HR) and replication, and it localizes to repetitive elements such as the rDNA and telomeres presumably to promote and resolve HR-dependent intermediates [12–14]. 10.1371/journal.pgen.1006268.g001Fig 1 Smc5/6 is a telomere binding complex. (A) A schematic representation of the Smc5/6 complex showing the location of Nse3 as part of a trimeric sub-complex located at the head region where Smc5 and Smc6 meet. (B) Chromatin immunoprecipitation (ChIP) followed by qPCR was performed on Smc6FLAG (JC1594) at the indicated time points after release from α-factor. The fold enrichment at three native subtelomeres (Tel1L, Tel6R and Tel15L) compared to a control (ctrl) late replicating region on Chromosome V (469104–469177) is reported with the mean ± SD for n≥3 experiments performed in technical duplicate. (*) Indicates a statistically significant level of enrichment compared to the ctrl with p values < .05 by a two-tailed t-test. Smc6FLAG enrichment at Tel1L is higher at 0 and 15 minutes after release, but with p values = 0.08 and p = 0.06 respectively. The lower panels show flow cytometry on ChIP samples with an asynchronous culture shown in black at the 0 time point. (C) Drop assay of exponentially growing wild type (JC470) and nse3-1 (JC3607) cells that were grown for 48 hours at the indicated temperatures on YPAD and 1:5 serial dilutions. (D) Schematic diagram of Nse3. “MHD” represents Melanoma Homology Domain in Nse3 protein. Seven amino acid substitutions in Nse3-1 are shown in red. (E) Chromatin immunoprecipitation (ChIP) on Smc6FLAG in wild type (JC1594), nse3-1 (JC2630), mms21-11 (JC2075) and the non-tagged (nt) control strains for wild type (JC470), nse3-1 (JC3607), and mms21-11 (JC1879) in asynchronous cultures. The fold enrichment levels are relative to the late-replicating control region on Chr V for n = 3 experiments with the mean ± SD. All primers are listed in S2 Table. Enrichment levels for wild type and mutant cells with p values < .05 from a two-tailed t-test are indicated by (*). (F) Telomere length was determined as previously described [15]. Southern blot analysis was performed on 1μg XhoI-digested genomic DNA hybridized with a radiolabeled poly (GT/CA) probe in wild type (JC471), nse3-1 (JC3032), mms21-11 (JC1981), and smc6-9 (JC1358).In higher eukaryotes, telomeres are challenged by the continuous loss of DNA due to the end replication problem. However, in Saccharomyces cerevisiae, telomere length is maintained by the continued expression of telomerase, an enzyme containing a RNA subunit that serves as a template for de novo telomere synthesis [16]. After the 3’ end is extended by telomerase, the replicative DNA polymerase fills in the complementary strand. Both telomerase extension and semiconservative replication at telomeres are included in the final events of S phase (for review see [17]). In the absence of telomerase activity, telomeres shorten extensively, leading to senescence, however a small percentage of cells survive by extending their telomeres through the HR dependent alternative lengthening of telomeres (ALT) pathway [18–21]. A telomeric function for the Smc5/6 complex in ALT has been demonstrated in both human and yeast cells [22–24]. In human ALT cells, a knockdown of components in the Smc5/6 complex inhibits recombination at telomeres, resulting in telomere shortening and senescence [22]. As well, in telomerase negative yeast cells smc6-9 and mms21-11 mutant alleles exhibited accelerated senescence attributed to the accumulation of recombination intermediates, but also to an HR–independent mechanism involving the untimely termination of DNA replication [23, 24]. The Smc5/6 complex is enriched at telomeres in telomerase positive asynchronous cultures [12, 13], however its characterization outside the ALT pathway remains limited. In telomerase positive cells, the smc6-9 allele exhibited mis-segregation of repetitive elements at telomeres which is attributed to defects in HR [12] and the mms21-11 allele was shown to have defects in telomere clustering with increased telomere position effect (TPE) [4]. Subsequent to the initial characterization of mms21-11, mms21Δsl mutants showed a loss of TPE and SIR binding [25]. Thus, allele specific variations have complicated the understanding of Mms21 and SUMO mediated events in TPE [4, 25]. Further characterization of Smc5/6 in telomere homeostasis using a mutant allele of a distinct complex component will provide additional information about the functionality of Smc5/6 at telomeres. Telomeric DNA in S. cerevisiae contains tandem repeats of (AC1-3/TG1-3) n; n = 275–375 [26] along with two types of subtelomeric repeat elements called Y’ and X [27]. The Y’ sequence is located adjacent to the tandem repeats at many, but not all subtelomeres, whereas X-elements are found at the ends of all chromosomes [28]. Rap1 binds directly to the double-stranded TG1-3 DNA moiety and is a central regulator of telomere biology [29]. The C-terminal domain of Rap1 interacts with Rif1 and Rif2 and regulates telomere length via a counting system that involves their interaction with Rap1 [30, 31]. Telomeres are elongated in rif1Δ and rif2Δ cells via telomerase dependent and HR independent events [32, 33]. The C-terminal domain of Rap1 also binds the SIR complex, which is important for transcriptional silencing primarily via interactions with Sir4 [30, 32, 34, 35]. SIR proteins are important for telomere position effect (TPE) and the formation of heterochromatin, which nucleates at telomeres and then spreads several kilobases into subtelomeric regions [36, 37]. Subtelomeric heterochromatin is maintained by seemingly distinct events that are likely to be interrelated in vivo. For example, in budding yeast, 32 telomeres cluster together in 3–8 foci at the nuclear periphery, and this drives the sequestration of SIR complex sub-compartments within the nucleus, and promotes silencing [38]. Additionally, the SIR complex, along with Rif1 and Rif2, modulates the level of long non-coding telomeric repeat-containing RNA, TERRA, which is also an integral factor in heterochromatin formation [39–42]. TERRA levels have never been reportedly assessed in Smc5/6 compromised cells and a role for the complex in heterochromatin maintenance and transcription at telomeres remains to be clearly defined. Here we show that the Smc5/6 complex binds telomeres, not only during late S phase when telomeres are synthesized, but also throughout the cell cycle in telomerase positive cells. Telomere clustering and full Sir4 binding is indeed dependent on the SUMO ligase activity of Mms21, however in the course of characterizing a temperature sensitive (ts) mutant of NSE3, telomere defects were observed in cells harboring the nse3-1 allele, which have not been previously reported with other alleles having compromised Smc5/6 functionality. TPE and TERRA regulation, as well as telomere length defects in nse3-1 mutants were additive with the loss of SIR4. In all, our data support a model that extends the functionality of Smc5/6 at telomeres beyond its previously reported roles in homology-mediated events in the ALT pathway [22–24]. Results The Smc5/6 complex is constitutively bound to telomeres and reduced in nse3-1 mutant cells The Smc5/6 complex has been detected at telomeres [12, 13] and stalled and collapsed replication forks [43–48]. Given that telomeres are difficult to replicate sites and prone to fork stalling, we wanted to determine if the presence of Smc5/6 at chromosome ends coincided solely with telomere duplication or if it was present at telomeres independent of replication. We monitored Smc6FLAG enrichment as a marker for the complex and performed chromatin immuno-precipitation (ChIP)–qPCR at multiple time points after cells were synchronously released from G1 into S phase. Significant enrichment of Smc6FLAG was observed at three telomeres above a late-replicating control region on Chr V (Fig 1B) [49, 50], showing the Smc5/6 complex is constitutively present at telomeres and not only during the time of telomere replication in late S phase (Fig 1B). It was recently demonstrated that Nse3 in fission yeast is important for loading the Smc5/6 complex onto chromatin [11]. We wanted to determine the involvement of Nse3 in localizing Smc5/6 to its endogenous binding sites such as telomeres in budding yeast. As with all subcomponents of the complex (Fig 1A), NSE3 is essential precluding its deletion. Therefore, we utilized a mutant allele, nse3-1, which contains seven amino acid substitutions and was isolated from a screen for temperature sensitivity (ts) at 37°C [51] (Fig 1C and 1D). As nse3-1 mutant cells do not synchronize efficiently with α-factor (S1A Fig), we determined Smc5/6 localization in asynchronous cultures at the semi-permissive temperature 34°C. The enrichment of Smc6FLAG was significantly reduced in nse3-1 mutant cells at telomeres and other known sites of Smc5/6 binding (Fig 1E, S1B Fig). In contrast, the level of Smc6FLAG recovered at telomeres in mms21-11 mutant cells, which are HR and SUMO ligase deficient, was similar to wild type (Fig 1E, S1B Fig). One explanation for the loss of Smc6FLAG recovery is that the complex is unstable in nse3-1 mutant cells. To address this possibility, we performed co-immunoprecipitation with two subcomponents that do not directly interact with one another, Nse6 and Smc5, as previously described [48]. In nse3-1 mutant cells, Nse6 was recovered in Smc5 pull-downs at levels comparable to wild type cells (S1C Fig), suggesting the complex does not markedly dissociate in nse3-1 mutants. Telomeres were also slightly shorter in nse3-1 mutants compared to wild type and HR-defective smc6-9 mutant cells (Fig 1F). In contrast, slightly longer telomeres were observed in mms21-11 mutants (Fig 1F), which is consistent with its initial characterization showing that this allele had longer telomeres [4]. The changes are indeed subtle, however there is a noticeable difference in telomere length when comparing the nse3-1 to the other complex mutants, suggesting that the Smc5/6 complex might have a role at telomeres distinct from HR-mediated events. The Smc5/6 complex is important for telomere clustering Telomere clustering at the nuclear periphery in S. cerevisiae establishes sub-nuclear zones that sequester repressors of transcription [52, 53]. Clustering can be visualised by performing immunofluorescences and counting GFP-Rap1 foci. In haploid cells, it has been demonstrated that 32 telomeres cluster in limited number [54], and consistent with this, our quantification showed ~90% of wild type cells contained ≤ 6 foci in both G1 and S phases of the cell cycle at 34°C (Fig 2A–2C). In contrast, nse3-1 mutants had ≥ 6 foci in ~65% and ~80% of the cells in G1 and S phases respectively, with 10–20% having ≥ 9 foci (Fig 2A–2C). In a side-by-side comparison and in line with its initial characterization, a similar clustering defect was observed in mms21-11 mutants [4], but smc6-9 mutant cells were similar to wild type (Fig 2D). 10.1371/journal.pgen.1006268.g002Fig 2 Smc5/6 is critical for telomere clustering and Sir4 binding to telomeres. (A) Rap1-GFP foci in WT (JC1822) and nse3-1 (JC3041) cells counted as a measurement for telomere clustering with representative merged images of GFP and DIC channels. (B-C) The number of GFP-Rap1 foci was determined for cells within G1 (unbudded) or S (small budded cells) phases in at least 100 cells for each cell cycle stage, and (D) compared with mms21-11 (JC1827) and smc6-9 (JC2710). (E-F) Western blot analysis and immunofluorescence staining using α-Myc antibody (green in IF) to detect Sir4Myc in WT (JC3433), nse3-1 (JC3452), mms21-11 (JC3597), and smc6-9 (JC2907) cells with DAPI staining shown in blue. (G) ChIP was performed on Sir4Myc as in Fig 1E from asynchronous cultures and in more than one isogenic strain if available. The fold enrichment for each strain is calculated for n≥3 experiments with the mean ± SD at three native subtelomeres (Tel1L, Tel6R and Tel15L). The p values < 0.05 from a two-tailed t-test are indicated by (*) for wild type (JC2671 and JC3433), nse3-1 (JC3452 and JC3849), mms21-11 (JC3597), and smc6-9 (JC2907 and JC3087) and non-tagged (nt) control strains included wild type (JC470), nse3-1 (JC3607), mms21-11 (JC1879), and smc6-9 (JC1358). Defects in clustering coincide with a disruption in SIR proteins, [55, 56]. Sir4Myc is expressed at similar levels in all strains (Fig 2E), and as measured by immunofluorescence, Sir4Myc forms discrete punctate foci in wild type cells (Fig 2F). In contrast, Sir4Myc became relatively dispersed throughout the nucleus in nse3-1 mutant cells (Fig 2F). Dispersion was also observed in mms21-11 and smc6-9 alleles, but to a lesser extent than the level observed in nse3-1 mutants (Fig 2F). Foci, albeit with reduced intensity, remained in all mutants to varying degrees, therefore as a complement to immunofluorescence and to quantify changes at telomere, we performed ChIP with Sir4Myc. The level of Sir4Myc recovered at telomeres in both nse3-1 and mms21-11 mutants was reduced to ~40% that of wild type cells (Fig 2G). In smc6-9 mutant cells, the level of Sir4Myc at telomeres was not significantly different from the amount recovered in wild type (Fig 2G). Taken together, the alleles with defects in clustering, nse3-1 and mms21-11, also showed a reduction in the level of Sir4 bound at telomeres. Sir4 sumoylation by Siz2 was previously implicated in peripheral telomere position [57, 58]. Given that our results indicated Sir4 localization to be regulated by Mms21, we investigated if the SUMO status of Sir4 itself might provide a level of regulation. Similar to siz2Δ, the level of Sir4 sumoylation was reduced in mms21-11, however SUMO levels remained similar to WT, if not higher in nse3-1 mutants (S2 Fig). These data suggest that Sir4 localization to telomeres is not regulated by the SUMO status of Sir4 in nse3-1 cells. The Smc5/6 complex binding to telomeres is regulated by Sir4 and is important for TPE To further understand the relationship between Sir4 and the Smc5/6 complex we performed co-immunoprecipitation to see if we could detect a physical interaction. Upon Smc6FLAG immunoprecipitation, we recovered Sir4Myc (Fig 3A). We had variable results with the reciprocal IP, however we found that Nse3HA associated with Sir4Myc pull downs (Fig 3B), suggesting that the Smc5/6 and SIR complexes physically associate in vivo. 10.1371/journal.pgen.1006268.g003Fig 3 Smc5/6 physically associate with Sir4 and is important for TPE. (A) Co-immunoprecipitation (Co-IP) as described in the materials and methods section was performed in cells carrying Sir4Myc and Nse3HA (JC3736) with Nse3HA (JC2823) as control or (B) Sir4Myc and Smc6FLAG (JC3853) with Smc6FLAG (JC1594) as a control. (C) ChIP was performed on Smc5FLAG in wild type (JC3728) and sir4Δ (JC3720) and (D) Smc6FLAG in wild type (JC1594) and sir4Δ (JC3732) and non-tagged (nt) strains in wild type (JC470) and sir4 Δ (JC3737) as described in Fig 1E. The fold enrichment levels are relative to the late-replicating control region on Chr V for n≥3 experiments with the mean ± SD at three native subtelomeres (Tel1L, Tel6R and Tel15L) with p values < .05 from a two-tailed t-test indicated. (E) TPE was determined in strains with the URA3 reporter at the adh4 locus of Chromosome VIIL. Tenfold (1:10) serial dilutions of overnight cultures were spotted onto SC (complete medium) and SC + .1% 5-FOA plates at 25°C and 34°C in wild type (JC1991), sir4Δ (JC3818), nse3-1 (JC3860), mms21-11 (JC1080) and smc6-9 (JC1077) isogenic strains. The Smc5/6 complex influenced Sir4 recovery at telomeres and a physical interaction between the complexes was detected. Thus, the reverse was performed to determine if Sir4 levels impacted the localization of Smc5/6 at telomeres. ChIP was performed with Smc6FLAG and Smc5FLAG and recovery at telomeres was compared in sir4Δ and wild type cells (Fig 3C and 3D). The level of Smc6FLAG in cells lacking SIR4 decreased to ~60% the amount recovered in wild type cells (Fig 3C). Similarly, Smc5FLAG was reduced in sir4Δ mutants to ~25% that of wild type levels (Fig 3D). As Smc5 and Smc6 are present at stoichiometric levels in the complex [4, 59], the greater relative change with Smc5FLAG might result from IP variability. Nonetheless, there is a statistically significant decrease in both core factors of the Smc5/6 complex bound to telomeres in sir4Δ mutants compared to wild type cells (Fig 3C and 3D). Sir4 is a critical factor for TPE and in the maintenance of heterochromatin near telomeres [60]. As the Smc5/6 complex interacts with Sir4, and the presence of Smc5/6 is important for Sir4 recovery at telomeres and vice versa, we assessed a role for the complex in transcriptional gene silencing regulation. TPE was determined in reporter strains where URA3 was integrated at the left arm of telomere VII [61]. Consistent with previous reports, sir4Δ cells showed defects in TPE as measured by their compromised ability to form colonies on medium containing 5-fluoroorotic acid (5-FOA) (Fig 3E) [60]. For nse3-1 mutants, TPE was disrupted but not to the level observed with sir4Δ (Fig 3E). In contrast and consistent with previous reports, TPE in mms21-11 and smc6-9 mutant cells remained intact at 25°C and 34°C (Fig 3E; [4]). This data indicated that the loss of silencing in nse3-1 mutant cells could not be solely attributed to a defect in Sir4 recruitment. This is supported by the observation that both nse3-1 and mms21-11 mutants showed a comparable defect of Sir4 recovery at telomeres and this was sufficient to silence the reporter transgene in the mms21-11 allele. The Smc5/6 complex contributes to telomere homeostasis and interacts genetically with SIR4 and RIF2 To bring insight to the functionality of Smc5/6 in transcriptional silencing at telomeres the nse3-1 allele was combined with the loss of either SIR4 and/or RIF2. Utilizing the URA3 reporter assay (Fig 4A, S3 Fig), it was difficult to observe an additive defect in silencing for nse3-1 sir4Δ double mutants because the loss of silencing is so penetrant with the loss of SIR4. Therefore, two endogenous sub-telomeric sites, YR043C and CHA1, on Tel9R and Tel3L respectively were assessed [62, 63]. Gene transcription increased in nse3-1 sir4Δ double mutants compared to sir4Δ single mutant cells (Fig 4B). Moreover, a defect in silencing was also observed in nse3-1 mutants at VAC17, a gene adjacent to CHA1 and previously determined to be silenced independently of Sir4 (Fig 4B; [63]). An additive loss of silencing was not observed when smc6-9 was combined with sir4Δ (S4 Fig), suggesting that HR-regulated functions involving the Smc5/6 complex are separable from its function in transcriptional silencing. In rif2Δ cells, silencing remains and even increases presumably through increased binding of Sir4 to Rap1 at telomeres (Fig 4C) [30, 64]. The nse3-1 rif2Δ double mutants exhibited a loss of silencing that was similar to nse3-1 single mutant cells (Fig 4C), however this was difficult to observe when measuring TPE from the URA3 reporter unless cell concentrations were low (S3 Fig). 10.1371/journal.pgen.1006268.g004Fig 4 The nse3-1 allele exhibits genetic interactions with the loss of SIR4 and RIF2. (A) TPE was determined in strains with the URA3 reporter at the adh4 locus of Chromosome VIIL as in Fig 3E. Overnight cultures were spotted onto SC (complete medium) and SC + .1% 5-FOA plates at 34°C in wild type (JC1991), sir4Δ (JC3818), nse3-1(JC3860), nse3-1 sir4Δ (JC3870) isogenic strains. (B) Transcription levels in wild type (JC470), nse3-1 (JC3607), sir4Δ (JC3737), and nse3-1 sir4Δ (JC3741), and (C) rif2Δ (JC2992) and nse3-1 rif2Δ (JC3269) at sub-telomeric genes CHA1 and VAC17 on Tel3L and YR043C on Tel9R as described in [62, 63]. Expression values are mRNA levels relative to ACT1 and normalization to wild type cells. Error bars represent ± SD of n = 3 experiments with p values < .05 from a two-tailed t-test indicated by (*). (D) Chromatin immunoprecipitation (ChIP) was performed on Rif2Myc and showed similar levels of recovery in wild type (JC2380) and nse3-1 (JC3235) mutants. (E) ChIP on Smc6FLAG in wild type (JC1594), rif1Δ (JC2754) and rif2Δ (JC3074) cells with enrichment levels for untagged strains in wild type and mutants shown in S5C and S5D Fig. The mean ± SD of the fold enrichment at three native subtelomeres (Tel1L, Tel6R and Tel15L) relative to the control (ctrl) late replicating region on Chromosome V (469104–469177) is reported. In rif2Δ mutants the p values < .05 = 0.53 (Tel1L), 0.13 (Tel6R), and 0.15 (Tel15L) indicated that the difference was not significant from wild type. (F) Telomere length was determined as previously described [15]. Southern blot analysis was performed on 1μg XhoI-digested genomic DNA hybridized with a radiolabeled poly (GT/CA) probe in wild type (JC470), nse3-1 (JC3607), rif1Δ (JC3448), nse3-1 rif1Δ (JC3623), rif2Δ (JC2992), nse3-1 rif2Δ (JC3269), rad52Δ (JC1427), nse3-1 rad52Δ (JC3629), rif2Δ rad52Δ (JC3603), and nse3-1 rad52Δ rif2Δ (JC3627) strains. Rap1 binds both Sir4 and Rif1/2 [30, 32, 34, 65], and given the interactions nse3-1 had with these factors it was important to assess Rap1 binding to telomeres in nse3-1 mutants. By ChIP, we observed no significant difference in the level of Rap1Myc bound at telomeres in nse3-1 mutants compared to the levels in wild type cells (S5A Fig). These data also support the interpretation that the increased number of Rap1 foci we measured in nse3-1 cells resulted from a disruption in telomere clustering rather than a disruption of Rap1 binding to telomeres (Fig 2A–2C). Nse3 was previously reported to interact with Rif2 in a high-throughput yeast two-hybrid (Y2H) screen [66]. We verified the Rif2-Nse3 interaction and determined it was reduced when nse3-1 was expressed (S6 Fig), however, in contrast to Sir4Myc, the levels of Rif1Myc and Rif2Myc at telomeres in nse3-1 were similar to wild type (Fig 4D, S5B Fig), and no significant change with Smc6FLAG was measured at telomeres in cells lacking RIF1 or RIF2 (Fig 4E). In all, these data suggest that the physical association between Nse3 and Rif2 is not driving the recruitment of either factor/complex to telomeres. Cells carrying the nse3-1 allele exhibit slightly shorter telomeres (Figs 1F and 4F), which is opposite to cells lacking RIF1 or RIF2, which are negative regulators of telomerase [33]. Telomere length was determined when nse3-1 was combined with rif1Δ and rif2Δ. The nse3-1 rif2Δ double mutant cells exhibited a partial reversion in the telomere length phenotype (lanes 5 and 6; Fig 4F). However, when nse3-1 was combined with rif1Δ, telomere length looked indistinguishable from rif1Δ single mutants (lanes 3 and 4; Fig 4F). These data suggest the nse3-1 mutation does not counteract telomere elongation as a general mechanism per se and support the model that Rif1 and Rif2 having non-overlapping roles in telomere maintenance even though they interact with each other and with Rap1 [67–70]. As the Smc5/6 complex is implicated in HR and the ALT pathway, we also investigated if the partial reversion of long telomeres in nse3-1 rif2Δ was regulated by HR events. Upon disruption of RAD52, no detectable changes were observed, as telomeres for nse3-1 rad52Δ and nse3-1 rif2Δ rad52Δ mutants were similar in size to nse3-1 and nse3-1 rif2Δ mutants respectively (lanes 2 and 8; lanes 6 and 10; Fig 4F). Moreover, telomere shortening was not observed when the loss of RIF2 was combined with the HR-deficient smc6-9 allele (S7 Fig). Taken together, these data provide additional support for Smc5/6 having a role at telomeres distinct of its functionality in HR-mediate events. TERRA regulation is altered in nse3-1 mutants In addition to the transcription of gene-coding regions, RNA polymerase II also transcribes TERRA at telomeres [40]. There are reported correlations between non-physiological increases and decreases in TERRA levels with telomeric abnormalities [39, 71]. Moreover, TERRA expression was previously demonstrated to be regulated by Rap1, the SIR complex, and Rif1/2 proteins, with the role of Rif2 being minimal and only at a subset of telomeres [42]. As the nse3-1 mutation results in a loss of silencing at subtelomeric genes and showed interactions with Rif2 and Sir4 we measured TERRA expression in cell carrying the nse3-1 allele. Compared to wild type, there was a significant de-repression in TERRA expressed from both X only and Y’ telomeres in nse3-1 mutants at both 28°C and 34°C (red, Fig 5A and 5B, S8 Fig). Consistent with previous reports [42], sir4Δ mutants showed substantial TERRA expression from X only telomeres (purple, Fig 5A and 5B), and we observed no distinguishable increase in TERRA levels in cells lacking RIF2 at TEL1R, 6R, or Y’ (aqua, Fig 5A and 5B, S8 Fig). TERRA levels in nse3-1 and nse3-1 rif2Δ were similar and significantly higher than the level measured in rif2Δ mutant cells (red, green, and aqua; Figs 5A and 5B and S8). Interestingly, and consistent with the TPE reporter assay, TERRA levels in sir4Δ rif2Δ cells (light grey) were silenced to levels not statistically different from wild type (dark grey), and similar to rif2Δ (aqua, Fig 5A and 5B). There was a 2- and 4- fold increase in the level of TERRA from Y’ and X-only telomeres respectively in nse3-1 sir4Δ cells (blue) compared to cells lacking SIR4 (purple) at 28°C (Fig 5A, S8A Fig). The same trend was observed at 34°C, however variability between experiments resulted in p values > 0.05 (Fig 5B, S8B Fig). 10.1371/journal.pgen.1006268.g005Fig 5 Increases in TERRA and telomere shortening are additive in nse3-1 sir4Δ double mutant cells. (A and B) TERRA expression was determined by RT-qPCR for Tel1R and Tel6R, X only telomeres, at 28°C and 34°C in wild type (JC470), nse3-1 (JC3607), rif2Δ (JC2992), nse3-1 rif2Δ (JC3269), sir4Δ (JC3737), nse3-1 sir4Δ (JC3741), and sir4Δ rif2Δ (JC3738). TERRA expression from Y’ telomeres is shown in S8 Fig. Statistical significance with p values < .05 (*) or < .01(**) are reported from a two-tailed t-test. (C) Telomere length was determined as in Fig 1F by Southern blot analysis on 1μg XhoI-digested genomic DNA hybridized with a radiolabeled poly (GT/CA) probe in wild type (JC470), nse3-1 (JC3607), sir4Δ (JC3737), and nse3-1 sir4Δ (JC3741). (D) A model comparing telomere organization in wild type and nse3-1 mutants. The Smc5/6 complex localizes to telomeres but significantly decreases in nse3-1 mutants (Fig 1E). Moreover, nse3-1 alleles exhibit shorter telomeres, reduced telomere clustering, reduced Sir4 binding and defects in TPE. When nse3-1 is combined with the loss of SIR4, the resulting double mutant cells show additive defects in transcriptional repression and telomere shortening. Both nse3-1 and sir4Δ mutants have slightly shorter telomeres (Figs 1F and 5C) [53]. As well, transcription and TERRA levels increased in nse3-1 and these phenotypes were additive with sir4Δ. Given the correlations between increased TERRA levels and induced transcription with telomere shortening [40, 72] we proceeded to assess telomere length in nse3-1 sir4Δ double mutants. Telomeres shorten further in double mutants compared to cells harboring either nse3-1 or sir4Δ single mutant alone (Fig 5C). Highlighting the difference again between nse3-1 and smc6-9, the level of TERRA expression was not additive in smc6-9 sir4Δ double mutant cells (S9A and S9B Fig) and in contrast to nse3-1, telomere length in smc6-9 did not result in additive shortening when combined with sir4Δ. (S9C Fig). Taken together, our data support a model whereby Smc5/6 has a role in transcriptional silencing and telomere length maintenance that is different from its involvement in HR dependent events at telomeres and underscore the value of characterizing various ts alleles of the complex. Discussion We report a previously uncharacterized function for the Smc5/6 complex with links to transcriptional silencing and demonstrate a role for the complex in telomere homeostasis. In cells carrying the nse3-1 allele, Smc5/6 complex levels are markedly reduced at telomeres. This was true for cells grown at 25°C or 34°C, the temperature we used in many of our measurements, indicating that higher temperature did not introduce confounding defects to the complex in this mutant background (S10 Fig). Utilizing nse3-1, we show that Smc5/6 is critical for 1. Maintaining proper telomere length, 2. Telomere clustering, 3. SIR complex recovery at telomeres, 4. TPE, and 5. Regulating TERRA levels. Telomere defects involving mutations in the Smc5/6 complex were first reported with the mms21-11 allele; however, the loss of SUMO ligase activity did not appear to impact TPE, as expression from a URA3 reporter construct integrated at Tel5R remained silent [4]. Upon characterization of the nse3-1 allele, we also observed a loss of clustering, but unlike mms21-11, TPE was disrupted as shown by an increase in expression of sub-telomeric genes and URA3 reporter expression. Further characterization of nse3-1 and mms21-11 alleles demonstrated that a decrease in Sir4 binding at telomeres was common to both alleles (a summary of phenotypes can be found in S3 Table). In agreement with previous reports (Zhao & Blobel, 2005), and in side-by-side comparison with nse3-1 and wild type, we find silencing at sub-telomeres remained intact for mms21-11 and smc6-9 mutants (Fig 3E), suggesting that the partial reduction in Sir4 at telomeres in mms21-11 and nse3-1 mutants was not sufficient to abrogate silencing. These data also raise the possibility that the complex might have additional functions, which are disrupted in nse3-1, that are important for silencing. Our data also suggest a partial interdependency between the Smc5/6 complex and Sir proteins at telomeres. Indeed, a physical interaction is detected between the Smc5/6 complex and Sir4 (Fig 3A and 3B) and in the absence of SIR4 there is a moderate but statistically significant ~30% reduction in the levels of Smc6FLAG recovered at telomeres, however for comparison, Smc6FLAG was reduced further in nse3-1 mutant cells by ~60% the levels of wild type (S10B Fig). Even though Smc5/6 and Sir4 contribute to the stability of one another at telomeres, the defects in TPE and TERRA expression associated with the loss of Smc5/6 at telomeres are additive with the loss of SIR4. Live-cell imaging at the single-cell level demonstrated that when telomeres become critically short, TERRA is transcribed, and this recruits telomerase to the TERRA-expressing telomere to promote elongation [73]. Increased TERRA levels above physiologically important levels likely have an inhibitory affect on telomere length maintenance. TERRA levels in nse3-1 mutants are above wild type and when combined with sir4Δ, the double mutants show an even greater increase in TERRA compared to the levels measured in cells lacking SIR4 only. The elevated transcription and loss of TPE in nse3-1 is likely to have a direct effect on TERRA expression and supports the model that Smc5/6 functionality is important for silencing, and when deregulated, transcription lead to increases in TERRA and telomere loss [74]. Telomere shortening is additive in nse3-1 sir4Δ mutants. The robust expression of TERRA in nse3-1 sir4Δ cells possibly reinforces the shortening of telomeres, and vice versa. Indeed this explanation is consistent with previous work showing that when TERRA increases, telomeres shorten via telomerase inhibition [40], as well as disrupting the inhibitory effect of yKu70/80 on Exonuclease 1, leading to its increased activity at telomeres [75]. A more speculative model, that will require additional investigation, is that increases in TERRA expression might lead to increased RNA-DNA hybrids at telomeres and subsequently more aberrant replication fork structures that fail to be resolved by Smc5/6, and this results in telomere loss specifically in alleles deficient in silencing, as in nse3-1 and nse3-1 sir4Δ mutant cells. Lastly, an alterative model that we cannot exclude is that there is a more direct effect of nse3-1 on telomere length independent of TERRA, which might involve interactions of the Nse1-Nse3-Nse4 sub-complex within Smc5/6 that become altered in cells carrying the nse3-1 allele. We also assessed the SUMO status of Sir4 and determined that sumoylation was reduced in mms21-11 mutant cells to levels similar to those previously observed in cells lacking SIZ2 (S2 Fig) [57]. However, Sir4 sumoylation remained, and was slightly higher in cells harbouring the nse3-1 allele when silencing is reduced. This is consistent with previous work showing that increased levels of Siz2, and by extension elevated sumoylation, function antagonistically to silencing [58], and also suggests there is no direct correlation between Sir4 sumoylation in telomere clustering at the periphery. These data are also consistent with the observation that a SUMO-Sir4 fusion construct could not restore anchoring in siz2Δ mutants, which suggested that sumoylation of another target, besides Sir4, is important for telomere positioning at the periphery [57]. Telomere clustering and silencing are distinguishable functions [76, 77]. Our data indicates that Smc5/6 likely contributes to both and independently of HR as smc6-9 was not distinguishable from wild type in all measures, and that Mms21 sumoylation is important for clustering, but not silencing. Determining the role of Smc5/6 in clustering at the periphery will require further investigation. Organization of telomeres at the periphery is driven by partially redundant pathways involving Sir4 binding to membrane bound Esc1 and Yku70/80 [76, 78]. First, although Sir4 sumoylation does not control clustering we have not assessed if Esc1, which is also a target of sumoylation, regulates clustering in a pathway dependent on Mms21 activity [79, 80]. Secondly, unlike Sir4, we observed that the level of YKu70 at telomeres in nse3-1 mutant cells was not statistically different from wild type cells (S11 Fig). However, determining if Mms21 dependent sumoylation of yKu70 at telomeres is critical for Smc5/6 mediated anchoring will provide an additional level of understanding as both Yku70 and Yku80 sumoylation are important for perinuclear positioning [57], and while Yku80 sumoylation is markedly reduced in siz2Δ mutants, Yku70-sumoylation is primarily dependent on Mms21 [4, 57]. Our data support a model where the Smc5/6 complex, like other proteins involved in DNA repair, such as Tel1 and Mre11, contributes to transcriptional silencing via two pathways, one involving direct interactions with SIR factors and the other regulating nuclear position and association with the periphery [81]. The current study demonstrates a role for Smc5/6 complex in telomere maintenance that is distinct from its previously characterized functions in replication and HR. Our data show that the Smc5/6 complex is a bona fide telomere-binding factor that has reduced recovery in nse3-1 mutant cells (Fig 5D). Our study establishes Smc5/6 as having a physiological role in the structural maintenance of chromosome ends where its localization and integrity contribute to the stabilization of factors with well-established roles in telomere maintenance and metabolism. Consistent with a role in end protection, the localization of Smc5/6 to telomeres is critical for telomere clustering and transcriptional repression (Fig 5D). These roles for Smc5/6 together its involvement in the various aspects of HR-mediated DNA metabolism, such as replication and repair, perhaps contribute to the essential requirement of this complex for cell survival. Materials and Methods Yeast strains and plasmids All strains used in this study are listed in S1 Table. The nse3-1 mutant was a kind gift from Dr. P. Hieter at Michael Smith Laboratories. In all experiments exponentially growing cells were incubated at 34°C for 2hrs before harvesting, unless indicated otherwise. Drop assays were performed by growing cells overnight, and then performing 10-fold serial dilutions where 4μl of each dilution were plated on YPAD an incubated at the indicated temperature. For repression assays, 5-fold or 10-fold dilutions from overnight cultures were plated on SC or SC + 5-FOA as described [76, 82] at the indicated temperatures. Chromatin immunoprecipitation (ChIP) ChIP experiments performed as described previously [83], except that cells were incubated at 34°C for 2 hours before crosslinking with formaldehyde in media where the temperature was held a 25°C to allow efficient crosslinking. Immunoprecipitates were washed once with lysis buffer (50 mm HEPES, 140 mm NaCl, 1 mm EDTA, 1% Triton X-100, 1 mM PMSF and protease inhibitor pellet (Roche)) and twice with wash buffer (100 mM Tris (pH 8), 0.5% Nonidet P-40, 1 mM EDTA, 500 mM NaCl, 250 mM LiCl, 1 mM PMSF and protease inhibitor pellet (Roche)). Real-time qPCR reactions were carried on using SYBR green method. Results shown as fold enrichment at three native subtelomeres (Tel1L, Tel6R and Tel15L) compared to a control (ctrl) late replicating region on Chromosome V (469104–469177) [49, 50]. Primer sequences are listed in S2 Table. Microscopy For Rap1-GFP foci imaging, cell were grown to the 5x106 cells/ml at 34°C for 2 hours in synthetic complete (SC) media. Images were captured immediately in 21 Z-stacks of 0.2 μm using Zeiss Axiovert 200 microscope. GFP foci per nucleus were manually counted as a representation for telomere foci. For Sir4 immunofluorescence, cell cultures were grown to the 5x106 cells/ml at 34°C for 2 hours in synthetic complete (SC) media. Cells were immediately fixed using 3.7% formaldehyde and spheroplasted in SK (0.1M KPO4/1.2M sorbitol) buffer containing 0.4 mg/ml Zymolase (US, Biological). Spheroplasted cells were fixed on poly-lysine coated coverslips as described previously [84]. Coverslips were blocked in 1% BSA in PBS for 1 hour, then incubated with primary (αMyc, ab9106-100) followed by secondary (Alexa 488; Molecular Probes, Invitrogen) antibodies each for 30 minutes. Coverslips were mounted on microscope slides using vectashield-containing DAPI (Molecular Probes, Invitrogen). Images were taken in 21 Z-stacks of 0.2 μm using Zeiss Axiovert 200 microscope and Z-stack images were flattened and presented in the figures. ImageJ (NIH, USA) was used for adjusting background in both live and immunofluorescence imaging methods. Co-immunoprecipitation assay Strains carrying HA-tagged Nse6 and Myc-tagged Smc5 were grown to the log phase at room temperature and then incubated for 2 hours at 34°C in YPAD media. Cells were lysed with zirconia beads in lysis buffer (50 mm HEPES, 140 mm NaCl, 1 mm EDTA, 1% Triton X-100, 1 mM PMSF and protease inhibitor pellet (Roche)). Cell lysates were incubated with αMyc antibody-coupled Dynabeads (Invitrogen) for 2 hours at 4°C. Immunoprecipitates were washed once with lysis buffer and twice with wash buffer (100 mM Tris (pH 8), 0.5% Nonidet P-40, 1 mM EDTA, and 400 mM NaCl, 1 mM PMSF and protease inhibitor pellet (Roche)), each for 5 minutes. Beads were resuspended in SDS loading buffer and subjected to SDS gel electrophoresis followed by western blotting by αHA (Santa Cruz, F7) and αMyc (9E10) antibodies. The same procedure was performed for Sir4-Nse3 except that lysates were clarified with one round of centrifugation at 13200 rpm before incubating with Myc antibody-coupled beads and immunoprecipitates were washed once with lysis buffer and twice with wash buffer (100 mM Tris (pH 8), 0.5% Nonidet P-40, 1 mM EDTA, and 250 mM LiCl, 1 mM PMSF and protease inhibitor pellet (Roche)). The co-IP between Sir4 and Smc6 was performed in stationary phase cultures without a chromatin spin and with a wash buffer containing 250 mM NaCl rather than 250 mM LiCl. Telomere length analysis by Southern blotting Measurement of telomere length was performed as described in [15]. Cells were grown for 48 hours to stationary phase in liquid YPAD at 34°C and harvested for Southern blotting. Genomic DNA from each strain were digested with XhoI and then separated by 1% agarose gel electrophoresis. Denatured DNA was transferred to Amersham Hybond-XL (GE Healthcare Life Sciences) membrane and hybridized with radiolabeled telomeric repeat probe (TG1-3/C1-3A). Rediprime II DNA Labeling System used to radiolabel telomeric probe (GE). Gene expression analysis Exponentially growing WT and nse3-1 cells were incubated for 2 hours at 34°C prior to harvesting by centrifugation and snap freezing in liquid nitrogen. Cells were lysed and mRNA isolation was followed by reverse transcription Complementary DNA (cDNA) was amplified and quantified using the SYBR Green qPCR method. Primers are listed in S2 Table. Fold gene expression represents real time qPCR values relative to WT samples. Gene expression values were normalized to ACT1 expression as the internal control. RNA Extraction and RT-qPCR for TERRA expression analyses Total RNA was extracted as in [73]. 2 μg of RNA was treated with 4U of DNase I (Thermo-Fisher) for 4 hr at 37°C and then purified by phenol/chloroform extraction. 1μg of DNase-treated RNA was reverse transcribed by using RevertAid Reverse Transcriptase (Thermo Fisher) at 42°C for 1 hr. 0,5 μmol of a C-rich primer (CACCACACCCACACACCACACCCACA) and 0,5 μg of a poly(dT) primer was used for the reverse transcription reaction (RT). 20ng of cDNA was used for the qPCR, which was performed using the qPCR master mix SsoFAST EvaGreen Supermix from Bio-Rad. qPCR was carried out on a Roche LightCycler96. TERRA expression was normalized against ACT1 mRNA expression using the delta Ct method and than normalized against the WT yeast strain. Supporting Information S1 Table Strains used in this study. (PDF) Click here for additional data file. S2 Table qPCR primers used in this study. (PDF) Click here for additional data file. S3 Table Summary of mutant phenotypes. (PDF) Click here for additional data file. S1 Fig The nse3-1 mutants do not synchronize properly, however components of the Smc5/6 complex still interact. (A) Flow cytometry was performed as described in Fig 1. (B) The fold enrichment levels are relative to the late-replicating control region on Chr V for n = 3 experiments with the mean ± SD at the silent mating type locus (HMR) and two regions in the rDNA (NTS1) and (NTS2) [12]. All primers are listed in S2 Table. (C) Co-immunoprecipitation assay was performed by immunoprecipitating Smc5Myc using α-Myc antibody in WT (JC2229), nse3-1 (JC2677) and smc6-9 (JC2232) cells. Beads were washed in 400mM NaCl, followed by western blotting for Smc5Myc and Nse6Ha components. (TIFF) Click here for additional data file. S2 Fig Sir4 Sumoylation in mutant backgrounds. Sumoylated proteins were isolated by Ni-NTA affinity purification of His-Smt3 as described previously [48, 57, 80, 85] followed by western blotting with αMyc antibodies to visualize sumoylated proteins in cells containing Myc-tagged Sir4 with un-tagged Smt3 wild type (JC3433), or His8-tagged Smt3 in wild type (JC3823), siz2Δ (JC3822) nse5-ts1 (JC3851) and mms21-11 (JC3824). (TIFF) Click here for additional data file. S3 Fig TPE measurements from the URA3 reporter at Telomere VII L. TPE was determined in strains with the URA3 reporter at the adh4 locus of Chromosome VIIL. Overnight cultures were spotted onto SC (complete medium) and SC + .1% 5-FOA plates and photographed after incubation at 25C and 34C in wild type (JC1991), sir4Δ (JC3818), nse3-1(JC3860), nse3-1 sir4Δ (JC3870), rif2Δ (JC3852), sir4Δ rif2Δ (JC3872), nse3-1 rif2Δ (JC3861), nse3-1 rif2Δ sir4Δ (JC3871) isogenic strains. (TIFF) Click here for additional data file. S4 Fig Transcription at sub-telomeric genes in smc6-9 mutants. Levels of transcription were compared at sub-telomeric genes CHA1, VAC17 and YR043C as described in Fig 4 in wild type (JC470), sir4Δ (JC3737), smc6-9 (JC3039), and sir4Δ smc6-9 (JC3925). Expression values are mRNA levels relative to ACT1 and normalization to wild type cells. Error bars represent ± SD of n = 3 experiments. (TIFF) Click here for additional data file. S5 Fig ChIP performed on Rap1Myc and Rif1Myc and in non-tagged (nt) strains. ChIP was perform with Chromatin immunoprecipitation (ChIP) was performed on (A) Rap1Myc in wild type (JC2381) and nse3-1 (JC3272), (B) Rif1Myc in wild type (JC3277) and nse3-1 (JC3295), (C) α Myc in non-tagged wild type (JC470) and nse3-1 (JC3607) cells and (D) α FLAG in non-tagged wild type (JC470), rif1Δ (JC3448), and rif2Δ (JC2992) cells. The mean ± SD of the fold enrichment at three native subtelomeres (Tel1L, Tel6R and Tel15L) are normalized to the negative ctrl region described in Fig 1F. No statistically significant differences were calculated after a two-tailed t-test for Rap1Myc ChIP between wild type and nse3-1, the p values < .05 = 0.47 (Tel1), 0.28 (Tel6R), and 0.35 (Tel15L), or for Rif1Myc. (TIFF) Click here for additional data file. S6 Fig Rif1, Rif2 and Smc6 recruitment at native telomeres in various mutant cells. (A) Yeast-two Hybrid analysis was performed as previously described [48]. NSE3 full-length, nse3(1–150)—N-terminal end, nse3(150–300)—C-terminal end, or the nse3-1 mutant were cloned into bait plasmid (pEG202) and RIF2 into prey plasmid (pJG4-6) [86]. Plasmids containing bait and prey along with pSH18034 (LacZ reporter plasmid) were transformed into JC1280 and grown overnight in selective media containing 2% raffinose. Overnight cultures were then divided and growth continued in either 2% galactose or 2% glucose for 6 hours at 30°C. β-galactosidase activity was then measured in permeabilized cells as previously described [48, 87]. (B) Western blots with a-HA and a-LexA shows the expression levels of Rif2HA, Nse3LexA full-length, N (Nse3(1–150), C-terminal Nse3(150–300) and Nse3-1 peptides from Y2H vectors (TIFF) Click here for additional data file. S7 Fig The nse3-1 allele, but not the smc6-9 allele shortens the long telomeres in cells lacking RIF2. Telomere length is determined for the indicated strains by performing southern blot analysis using radiolabeled poly GT/CA probe as explained in Fig 1F and in the experimental procedures section for wild type (JC470), rif2Δ (JC2992), smc6-9 (JC3039), and smc6-9 rif2Δ (JC-2993). (TIFF) Click here for additional data file. S8 Fig TERRA expression levels in rif2Δ and nse3-1 mutants. (A and B) TERRA expression was determined for Y’ at 28°C and 34°C in wild type (JC470), nse3-1 (JC3607), rif2Δ (JC2992), nse3-1 rif2Δ (JC3269), sir4Δ (JC3737), nse3-1 sir4Δ (JC3741), and sir4Δ rif2Δ (JC3738). Statistical significance with p values < .05 (*) or < .01 (**) are reported from a two-tailed t-test. The Y’ primers detect TERRA expressed from these telomeres: 8L / 8R / 12L-YP1 / 12R-YP2 / 13L / 15R. The arms of chromosome XII contains two short telomeric Y’ elements, YP1 is more end-proximal and YP2 is more centromere-proximal [75]. (TIFF) Click here for additional data file. S9 Fig TERRA expression and telomeres length in smc6-9 mutants. (A and B) TERRA expression was determined by RT-qPCR for Tel1R and Tel6R, X only telomeres, at 28C (A) and 34C (B). Statistical significance with p values < .05 (*) or < .01(**) are reported from a two-tailed t-test. (C) Telomere length was determined as in Fig 1F by Southern blot analysis on 1μg XhoI-digested genomic DNA hybridized with a radiolabeled poly (GT/CA) probe in wild type (JC470), sir4Δ (JC3737), smc6-9 (JC3039), and smc6-9 sir4Δ (JC3925). (TIFF) Click here for additional data file. S10 Fig Comparison of ChIP levels for Smc6 at telomeres in sir4Δ and nse3-1 mutants and wild type cells. (A) Chromatin immunoprecipitation (ChIP) on Smc6FLAG in wild type (JC1594) and nse3-1 (JC2630) at 25°C. (B) ChIP comparison of Smc6FLAG in wild type (JC1594), sir4Δ (JC3732), nse3-1 (JC2630). The enrichment at three native subtelomeres (Tel1L, Tel6R and Tel15L) normalized to the negative control region as described in Fig 1B. The levels of Smc6 are reduced further in nse3-1 mutants than sir4Δ mutants. (TIFF) Click here for additional data file. S11 Fig ChIP of yKu70 at telomeres in nse3-1 mutant and wild type cells. Chromatin immunoprecipitation (ChIP) was performed on yKu70Myc in wild type (JC1352) and nse3-1 (JC3392). The enrichment at three native subtelomeres (Tel1L, Tel6R and Tel15L) normalized to the negative control region as described in Fig 1B. (TIFF) Click here for additional data file. We would like to thank Drs. Charlie Boone, Susan Gasser, Philip Hieter, David Shore, Virginia Zakian, and Xiaolan Zhao for strains. We thank all members of the Cobb Lab for helpful discussions. ==== Refs References 1 Nasmyth K , Haering CH . The structure and function of SMC and kleisin complexes . Annu Rev Biochem . 2005 ;74 :595 –648 . 10.1146/annurev.biochem.74.082803.133219 .15952899 2 Hirano T . At the heart of the chromosome: SMC proteins in action . Nat Rev Mol Cell Biol . 2006 ;7 (5 ):311 –22 . 10.1038/nrm1909 .16633335 3 Jeppsson K , Kanno T , Shirahige K , Sjogren C . The maintenance of chromosome structure: positioning and functioning of SMC complexes . Nat Rev Mol Cell Biol . 2014 ;15 (9 ):601 –14 . 10.1038/nrm3857 .25145851 4 Zhao X , Blobel G . A SUMO ligase is part of a nuclear multiprotein complex that affects DNA repair and chromosomal organization . Proc Natl Acad Sci U S A . 2005 ;102 (13 ):4777 –82 . 10.1073/pnas.0500537102 15738391 5 Potts PR , Yu H . Human MMS21/NSE2 is a SUMO ligase required for DNA repair . Mol Cell Biol . 2005 ;25 (16 ):7021 –32 . 10.1128/MCB.25.16.7021–7032.2005 16055714 6 Andrews EA , Palecek J , Sergeant J , Taylor E , Lehmann AR , Watts FZ . Nse2, a component of the Smc5-6 complex, is a SUMO ligase required for the response to DNA damage . Mol Cell Biol . 2005 ;25 (1 ):185 –96 . 10.1128/MCB.25.1.185–196.2005 15601841 7 Pebernard S , McDonald WH , Pavlova Y , Yates JR , 3rd, Boddy MN. Nse1, Nse2, and a novel subunit of the Smc5-Smc6 complex, Nse3, play a crucial role in meiosis . Mol Biol Cell . 2004 ;15 (11 ):4866 –76 . 10.1091/mbc.E04-05-0436 15331764 8 Palecek J , Vidot S , Feng M , Doherty AJ , Lehmann AR . The Smc5-Smc6 DNA repair complex. bridging of the Smc5-Smc6 heads by the KLEISIN, Nse4, and non-Kleisin subunits . J Biol Chem . 2006 ;281 (48 ):36952 –9 . 10.1074/jbc.M608004200 .17005570 9 Pebernard S , Perry JJ , Tainer JA , Boddy MN . Nse1 RING-like domain supports functions of the Smc5-Smc6 holocomplex in genome stability . Mol Biol Cell . 2008 ;19 (10 ):4099 –109 . 10.1091/mbc.E08-02-0226 18667531 10 Doyle JM , Gao J , Wang J , Yang M , Potts PR . MAGE-RING protein complexes comprise a family of E3 ubiquitin ligases . Mol Cell . 2010 ;39 (6 ):963 –74 . 10.1016/j.molcel.2010.08.029 20864041 11 Zabrady K , Adamus M , Vondrova L , Liao C , Skoupilova H , Novakova M , et al Chromatin association of the SMC5/6 complex is dependent on binding of its NSE3 subunit to DNA . Nucleic Acids Res . 2015 10.1093/nar/gkv1021 .26446992 12 Torres-Rosell J , Machin F , Farmer S , Jarmuz A , Eydmann T , Dalgaard JZ , et al SMC5 and SMC6 genes are required for the segregation of repetitive chromosome regions . Nat Cell Biol . 2005 ;7 (4 ):412 –9 . 10.1038/ncb1239 .15793567 13 Lindroos HB , Strom L , Itoh T , Katou Y , Shirahige K , Sjogren C . Chromosomal association of the Smc5/6 complex reveals that it functions in differently regulated pathways . Mol Cell . 2006 ;22 (6 ):755 –67 . 10.1016/j.molcel.2006.05.014 .16793545 14 Murray JM , Carr AM . Smc5/6: a link between DNA repair and unidirectional replication? Nat Rev Mol Cell Biol . 2008 ;9 (2 ):177 –82 . 10.1038/nrm2309 .18059412 15 Greider CW , Blackburn EH . The telomere terminal transferase of Tetrahymena is a ribonucleoprotein enzyme with two kinds of primer specificity . Cell . 1987 ;51 (6 ):887 –98 . .3319189 16 Wellinger RJ , Zakian VA . Everything you ever wanted to know about Saccharomyces cerevisiae telomeres: beginning to end . Genetics . 2012 ;191 (4 ):1073 –105 . 10.1534/genetics.111.137851 22879408 17 Lundblad V , Blackburn EH . An alternative pathway for yeast telomere maintenance rescues est1- senescence . Cell . 1993 ;73 (2 ):347 –60 . .8477448 18 Singer MS , Gottschling DE . TLC1: template RNA component of Saccharomyces cerevisiae telomerase . Science . 1994 ;266 (5184 ):404 –9 . 7545955 19 Le S , Moore JK , Haber JE , Greider CW . RAD50 and RAD51 define two pathways that collaborate to maintain telomeres in the absence of telomerase . Genetics . 1999 ;152 (1 ):143 –52 . 10224249 20 Teng SC , Zakian VA . Telomere-telomere recombination is an efficient bypass pathway for telomere maintenance in Saccharomyces cerevisiae . Mol Cell Biol . 1999 ;19 (12 ):8083 –93 . 10567534 21 Potts PR , Yu H . The SMC5/6 complex maintains telomere length in ALT cancer cells through SUMOylation of telomere-binding proteins . Nat Struct Mol Biol . 2007 ;14 (7 ):581 –90 . 10.1038/nsmb1259 .17589526 22 Chavez A , George V , Agrawal V , Johnson FB . Sumoylation and the structural maintenance of chromosomes (Smc) 5/6 complex slow senescence through recombination intermediate resolution . J Biol Chem . 2010 ;285 (16 ):11922 –30 . 10.1074/jbc.M109.041277 20159973 23 Noel JF , Wellinger RJ . Abrupt telomere losses and reduced end-resection can explain accelerated senescence of Smc5/6 mutants lacking telomerase . DNA Repair (Amst) . 2011 ;10 (3 ):271 –82 . 10.1016/j.dnarep.2010.11.010 .21190904 24 Wan Y , Zuo X , Zhuo Y , Zhu M , Danziger SA , Zhou Z . The functional role of SUMO E3 ligase Mms21p in the maintenance of subtelomeric silencing in budding yeast . Biochem Biophys Res Commun . 2013 ;438 (4 ):746 –52 . 10.1016/j.bbrc.2013.07.096 .23911609 25 Louis EJ . The chromosome ends of Saccharomyces cerevisiae . Yeast . 1995 ;11 (16 ):1553 –73 . 8720065 26 Louis EJ , Haber JE . The subtelomeric Y' repeat family in Saccharomyces cerevisiae: an experimental system for repeated sequence evolution . Genetics . 1990 ;124 (3 ):533 –45 . 2179052 27 Louis EJ , Haber JE . The structure and evolution of subtelomeric Y' repeats in Saccharomyces cerevisiae . Genetics . 1992 ;131 (3 ):559 –74 . 1628806 28 Longtine MS , Wilson NM , Petracek ME , Berman J . A yeast telomere binding activity binds to two related telomere sequence motifs and is indistinguishable from RAP1 . Curr Genet . 1989 ;16 (4 ):225 –39 . 2697465 29 Wotton D , Shore D . A novel Rap1p-interacting factor, Rif2p, cooperates with Rif1p to regulate telomere length in Saccharomyces cerevisiae . Genes Dev . 1997 ;11 (6 ):748 –60 . 9087429 30 Levy DL , Blackburn EH . Counting of Rif1p and Rif2p on Saccharomyces cerevisiae telomeres regulates telomere length . Mol Cell Biol . 2004 ;24 (24 ):10857 –67 . 10.1128/MCB.24.24.10857–10867.2004 15572688 31 Hardy CF , Sussel L , Shore D . A RAP1-interacting protein involved in transcriptional silencing and telomere length regulation . Genes Dev . 1992 ;6 (5 ):801 –14 . 1577274 32 Teng SC , Chang J , McCowan B , Zakian VA . Telomerase-independent lengthening of yeast telomeres occurs by an abrupt Rad50p-dependent, Rif-inhibited recombinational process . Mol Cell . 2000 ;6 (4 ):947 –52 . 11090632 33 Moretti P , Freeman K , Coodly L , Shore D . Evidence that a complex of SIR proteins interacts with the silencer and telomere-binding protein RAP1 . Genes Dev . 1994 ;8 (19 ):2257 –69 . 7958893 34 Buck SW , Shore D . Action of a RAP1 carboxy-terminal silencing domain reveals an underlying competition between HMR and telomeres in yeast . Genes Dev . 1995 ;9 (3 ):370 –84 . 7867933 35 Hecht A , Laroche T , Strahl-Bolsinger S , Gasser SM , Grunstein M . Histone H3 and H4 N-termini interact with SIR3 and SIR4 proteins: a molecular model for the formation of heterochromatin in yeast . Cell . 1995 ;80 (4 ):583 –92 . 7867066 36 Strahl-Bolsinger S , Hecht A , Luo K , Grunstein M . SIR2 and SIR4 interactions differ in core and extended telomeric heterochromatin in yeast . Genes Dev . 1997 ;11 (1 ):83 –93 . 9000052 37 Maillet L , Boscheron C , Gotta M , Marcand S , Gilson E , Gasser SM . Evidence for silencing compartments within the yeast nucleus: a role for telomere proximity and Sir protein concentration in silencer-mediated repression . Genes Dev . 1996 ;10 (14 ):1796 –811 . 8698239 38 Azzalin CM , Reichenbach P , Khoriauli L , Giulotto E , Lingner J . Telomeric repeat containing RNA and RNA surveillance factors at mammalian chromosome ends . Science . 2007 ;318 (5851 ):798 –801 . 10.1126/science.1147182 .17916692 39 Luke B , Panza A , Redon S , Iglesias N , Li Z , Lingner J . The Rat1p 5' to 3' exonuclease degrades telomeric repeat-containing RNA and promotes telomere elongation in Saccharomyces cerevisiae . Mol Cell . 2008 ;32 (4 ):465 –77 . 10.1016/j.molcel.2008.10.019 .19026778 40 Schoeftner S , Blasco MA . Developmentally regulated transcription of mammalian telomeres by DNA-dependent RNA polymerase II . Nat Cell Biol . 2008 ;10 (2 ):228 –36 . 10.1038/ncb1685 .18157120 41 Iglesias N , Redon S , Pfeiffer V , Dees M , Lingner J , Luke B . Subtelomeric repetitive elements determine TERRA regulation by Rap1/Rif and Rap1/Sir complexes in yeast . EMBO Rep . 2011 ;12 (6 ):587 –93 . 10.1038/embor.2011.73 21525956 42 Branzei D , Sollier J , Liberi G , Zhao X , Maeda D , Seki M , et al Ubc9- and mms21-mediated sumoylation counteracts recombinogenic events at damaged replication forks . Cell . 2006 ;127 (3 ):509 –22 . 10.1016/j.cell.2006.08.050 .17081974 43 Ampatzidou E , Irmisch A , O'Connell MJ , Murray JM . Smc5/6 is required for repair at collapsed replication forks . Mol Cell Biol . 2006 ;26 (24 ):9387 –401 . 10.1128/MCB.01335-06 17030601 44 Irmisch A , Ampatzidou E , Mizuno K , O'Connell MJ , Murray JM . Smc5/6 maintains stalled replication forks in a recombination-competent conformation . EMBO J . 2009 ;28 (2 ):144 –55 . 10.1038/emboj.2008.273 19158664 45 Sollier J , Driscoll R , Castellucci F , Foiani M , Jackson SP , Branzei D . The Saccharomyces cerevisiae Esc2 and Smc5-6 proteins promote sister chromatid junction-mediated intra-S repair . Mol Biol Cell . 2009 ;20 (6 ):1671 –82 . 10.1091/mbc.E08-08-0875 19158389 46 Choi K , Szakal B , Chen YH , Branzei D , Zhao X . The Smc5/6 complex and Esc2 influence multiple replication-associated recombination processes in Saccharomyces cerevisiae . Mol Biol Cell . 2010 ;21 (13 ):2306 –14 . 10.1091/mbc.E10-01-0050 20444977 47 Bustard DE , Menolfi D , Jeppsson K , Ball LG , Dewey SC , Shirahige K , et al During replication stress, non-SMC element 5 (NSE5) is required for Smc5/6 protein complex functionality at stalled forks . J Biol Chem . 2012 ;287 (14 ):11374 –83 . 10.1074/jbc.M111.336263 22303010 48 Tittel-Elmer M , Lengronne A , Davidson MB , Bacal J , Francois P , Hohl M , et al Cohesin association to replication sites depends on rad50 and promotes fork restart . Mol Cell . 2012 ;48 (1 ):98 –108 . 10.1016/j.molcel.2012.07.004 22885006 49 Crabbe L , Thomas A , Pantesco V , De Vos J , Pasero P , Lengronne A . Analysis of replication profiles reveals key role of RFC-Ctf18 in yeast replication stress response . Nat Struct Mol Biol . 2010 ;17 (11 ):1391 –7 . 10.1038/nsmb.1932 .20972444 50 Ben-Aroya S , Coombes C , Kwok T , O'Donnell KA , Boeke JD , Hieter P . Toward a comprehensive temperature-sensitive mutant repository of the essential genes of Saccharomyces cerevisiae . Mol Cell . 2008 ;30 (2 ):248 –58 . 10.1016/j.molcel.2008.02.021 18439903 51 Shore D . RAP1: a protean regulator in yeast . Trends Genet . 1994 ;10 (11 ):408 –12 . 7809947 52 Palladino F , Laroche T , Gilson E , Axelrod A , Pillus L , Gasser SM . SIR3 and SIR4 proteins are required for the positioning and integrity of yeast telomeres . Cell . 1993 ;75 (3 ):543 –55 . 8221893 53 Gotta M , Laroche T , Formenton A , Maillet L , Scherthan H , Gasser SM . The clustering of telomeres and colocalization with Rap1, Sir3, and Sir4 proteins in wild-type Saccharomyces cerevisiae . J Cell Biol . 1996 ;134 (6 ):1349 –63 . 8830766 54 Laroche T , Martin SG , Gotta M , Gorham HC , Pryde FE , Louis EJ , et al Mutation of yeast Ku genes disrupts the subnuclear organization of telomeres . Curr Biol . 1998 ;8 :653 –6 . 9635192 55 Taddei A , Van Houwe G , Nagai S , Erb I , van Nimwegen E , Gasser SM . The functional importance of telomere clustering: global changes in gene expression result from SIR factor dispersion . Genome research . 2009 ;19 (4 ):611 –25 . 10.1101/gr.083881.108 19179643 56 Ferreira HC , Luke B , Schober H , Kalck V , Lingner J , Gasser SM . The PIAS homologue Siz2 regulates perinuclear telomere position and telomerase activity in budding yeast . Nat Cell Biol . 2011 ;13 (7 ):867 –74 . 10.1038/ncb2263 .21666682 57 Pasupala N , Easwaran S , Hannan A , Shore D , Mishra K . The SUMO E3 ligase Siz2 exerts a locus-dependent effect on gene silencing in Saccharomyces cerevisiae . Eukaryot Cell . 2012 ;11 (4 ):452 –62 . 10.1128/EC.05243-11 22345352 58 Duan X , Yang Y , Chen YH , Arenz J , Rangi GK , Zhao X , et al Architecture of the Smc5/6 Complex of Saccharomyces cerevisiae Reveals a Unique Interaction between the Nse5-6 Subcomplex and the Hinge Regions of Smc5 and Smc6 . J Biol Chem . 2009 ;284 (13 ):8507 –15 . 10.1074/jbc.M809139200 19141609 59 Aparicio OM , Billington BL , Gottschling DE . Modifiers of position effect are shared between telomeric and silent mating-type loci in S. cerevisiae . Cell . 1991 ;66 (6 ):1279 –87 . 1913809 60 Gottschling DE , Aparicio OM , Billington BL , Zakian VA . Position effect at S. cerevisiae telomeres: reversible repression of Pol II transcription . Cell . 1990 ;63 (4 ):751 –62 . 2225075 61 Kueng S , Tsai-Pflugfelder M , Oppikofer M , Ferreira HC , Roberts E , Tsai C , et al Regulating repression: roles for the sir4 N-terminus in linker DNA protection and stabilization of epigenetic states . PLoS genetics . 2012 ;8 (5 ):e1002727 10.1371/journal.pgen.1002727 22654676 62 Ellahi A , Thurtle DM , Rine J . The Chromatin and Transcriptional Landscape of Native Saccharomyces cerevisiae Telomeres and Subtelomeric Domains . Genetics . 2015 ;200 (2 ):505 –21 . 10.1534/genetics.115.175711 25823445 63 Shi T , Bunker RD , Mattarocci S , Ribeyre C , Faty M , Gut H , et al Rif1 and Rif2 shape telomere function and architecture through multivalent Rap1 interactions . Cell . 2013 ;153 (6 ):1340 –53 . 10.1016/j.cell.2013.05.007 .23746845 64 Cockell M , Palladino F , Laroche T , Kyrion G , Liu C , Lustig AJ , et al The carboxy termini of Sir4 and Rap1 affect Sir3 localization: evidence for a multicomponent complex required for yeast telomeric silencing . J Cell Biol . 1995 ;129 (4 ):909 –24 . 7744964 65 Hazbun TR , Malmstrom L , Anderson S , Graczyk BJ , Fox B , Riffle M , et al Assigning function to yeast proteins by integration of technologies . Mol Cell . 2003 ;12 (6 ):1353 –65 . .14690591 66 Craven RJ , Petes TD . Dependence of the regulation of telomere length on the type of subtelomeric repeat in the yeast Saccharomyces cerevisiae . Genetics . 1999 ;152 (4 ):1531 –41 . 10430581 67 Bonetti D , Clerici M , Anbalagan S , Martina M , Lucchini G , Longhese MP . Shelterin-like proteins and Yku inhibit nucleolytic processing of Saccharomyces cerevisiae telomeres . PLoS genetics . 2010 ;6 (5 ):e1000966 10.1371/journal.pgen.1000966 20523746 68 Marcand S , Pardo B , Gratias A , Cahun S , Callebaut I . Multiple pathways inhibit NHEJ at telomeres . Genes Dev . 2008 ;22 (9 ):1153 –8 . 10.1101/gad.455108 18451106 69 Anbalagan S , Bonetti D , Lucchini G , Longhese MP . Rif1 supports the function of the CST complex in yeast telomere capping . PLoS genetics . 2011 ;7 (3 ):e1002024 10.1371/journal.pgen.1002024 21437267 70 Deng Z , Norseen J , Wiedmer A , Riethman H , Lieberman PM . TERRA RNA binding to TRF2 facilitates heterochromatin formation and ORC recruitment at telomeres . Mol Cell . 2009 ;35 (4 ):403 –13 . 10.1016/j.molcel.2009.06.025 19716786 71 Sandell LL , Gottschling DE , Zakian VA . Transcription of a yeast telomere alleviates telomere position effect without affecting chromosome stability . Proc Natl Acad Sci U S A . 1994 ;91 (25 ):12061 –5 . 7991584 72 Cusanelli E , Romero CA , Chartrand P . Telomeric noncoding RNA TERRA is induced by telomere shortening to nucleate telomerase molecules at short telomeres . Mol Cell . 2013 ;51 (6 ):780 –91 . 10.1016/j.molcel.2013.08.029 .24074956 73 Maicher A , Kastner L , Dees M , Luke B . Deregulated telomere transcription causes replication-dependent telomere shortening and promotes cellular senescence . Nucleic Acids Res . 2012 ;40 (14 ):6649 –59 . 10.1093/nar/gks358 22553368 74 Pfeiffer V , Lingner J . TERRA promotes telomere shortening through exonuclease 1-mediated resection of chromosome ends . PLoS genetics . 2012 ;8 (6 ):e1002747 10.1371/journal.pgen.1002747 22719262 75 Taddei A , Hediger F , Neumann FR , Bauer C , Gasser SM . Separation of silencing from perinuclear anchoring functions in yeast Ku80 Sir4 and Esc1 proteins . EMBO J . 2004 ;25 :in press. 76 Mondoux MA , Scaife JG , Zakian VA . Differential nuclear localization does not determine the silencing status of Saccharomyces cerevisiae telomeres . Genetics . 2007 ;177 (4 ):2019 –29 . 10.1534/genetics.107.079848 18073421 77 Hediger F , Neumann FR , Van Houwe G , Dubrana K , Gasser SM . Live Imaging of Telomeres. yKu and Sir Proteins Define Redundant Telomere-Anchoring Pathways in Yeast . Curr Biol . 2002 ;12 (24 ):2076 –89 . .12498682 78 Hannich JT , Lewis A , Kroetz MB , Li SJ , Heide H , Emili A , et al Defining the SUMO-modified proteome by multiple approaches in Saccharomyces cerevisiae . J Biol Chem . 2005 ;280 (6 ):4102 –10 . 10.1074/jbc.M413209200 .15590687 79 Wohlschlegel JA , Johnson ES , Reed SI , Yates JR 3rd. Global analysis of protein sumoylation in Saccharomyces cerevisiae . J Biol Chem . 2004 ;279 (44 ):45662 –8 . 10.1074/jbc.M409203200 .15326169 80 Kirkland JG , Peterson MR , Still CD 2nd, Brueggeman L , Dhillon N , Kamakaka RT . Heterochromatin formation via recruitment of DNA repair proteins . Mol Biol Cell . 2015 ;26 (7 ):1395 –410 . 10.1091/mbc.E14-09-1413 25631822 81 Gotta M , Palladino F , Gasser SM . Functional characterization of the N terminus of Sir3p . Mol Cell Biol . 1998 ;18 (10 ):6110 –20 . 9742128 82 Tittel-Elmer M , Alabert C , Pasero P , Cobb JA . The MRX complex stabilizes the replisome independently of the S phase checkpoint during replication stress . EMBO J . 2009 ;28 (8 ):1142 –56 . 10.1038/emboj.2009.60 19279665 83 Gotta M , Laroche T , Gasser SM . Analysis of nuclear organization in Saccharomyces cerevisiae . Methods Enzymol . 1999 ;304 :663 –72 . 10372389 84 Puglisi A , Bianchi A , Lemmens L , Damay P , Shore D . Distinct roles for yeast Stn1 in telomere capping and telomerase inhibition . EMBO J . 2008 ;27 (17 ):2328 –39 . 10.1038/emboj.2008.158 19172739 85 Psakhye I , Jentsch S . Protein group modification and synergy in the SUMO pathway as exemplified in DNA repair . Cell . 2012 ;151 (4 ):807 –20 . 10.1016/j.cell.2012.10.021 .23122649 86 Aushubel FM , Brent R , Kinston R , Moore D , Seidman JJ , Smith J , et al Current Protocols in Molecular Biology . New York : John Wiley and Sons ; 1994 . 87 Adams A , Gottschling DE , Kaiser CA , Stearns T . Methods in Yeast Genetics . New York : Cold Spring Harbor Laboratory Press ; 1997 .

PMC005xxxxxx/PMC5001637.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756441010.1371/journal.pone.0161595PONE-D-16-02989Research ArticleMedicine and Health SciencesPulmonologyChronic Obstructive Pulmonary DiseaseMedicine and Health SciencesHealth CarePatientsPeople and PlacesPopulation GroupingsEthnicitiesDutch PeopleMedicine and Health SciencesPublic and Occupational HealthBehavioral and Social Aspects of HealthMedicine and Health SciencesPharmaceuticsDrug TherapyMedicine and Health SciencesMental Health and PsychiatryResearch and Analysis MethodsMathematical and Statistical TechniquesStatistical MethodsFactor AnalysisPhysical SciencesMathematicsStatistics (Mathematics)Statistical MethodsFactor AnalysisBiology and Life SciencesPsychologyPsychometricsSocial SciencesPsychologyPsychometricsConstruct Validity of the Dutch Version of the 12-Item Partners in Health Scale: Measuring Patient Self-Management Behaviour and Knowledge in Patients with Chronic Obstructive Pulmonary Disease Construct Validity of the Dutch Partners in Health ScaleLenferink Anke 123*Effing Tanja 34Harvey Peter 5Battersby Malcolm 5Frith Peter 34van Beurden Wendy 1van der Palen Job 12Paap Muirne C. S. 61 Department of Pulmonary Medicine, Medisch Spectrum Twente, Enschede, The Netherlands2 Department of Research Methodology, Measurement, and Data-Analysis, Faculty of Behavioural, Management and Social sciences, University of Twente, Enschede, The Netherlands3 School of Medicine, Flinders University, Adelaide, South Australia, Australia4 Department of Respiratory Medicine, Repatriation General Hospital, Adelaide, South Australia, Australia5 Flinders Human Behaviour and Health Research Unit, Flinders University, Adelaide, Australia6 Centre for Educational Measurement at the University of Oslo (CEMO), Faculty of Educational Sciences, University of Oslo, Oslo, NorwayBorrow Ray EditorPublic Health England, UNITED KINGDOMCompeting Interests: The authors have read the journal's policy and have the following competing interests: PH and MB have competing interests as the developers of PIH, but both have no personal financial interests. The other authors have declared that no competing interests exist. This does not alter the authors' adherence to PLOS ONE policies on sharing data and materials. Conceptualization: AL TE PH MB PF JP MP. Data curation: AL. Formal analysis: AL MP. Funding acquisition: TE JP. Investigation: AL TE PH MB PF WB JP MP. Methodology: AL TE JP MP. Project administration: AL TE JP. Resources: AL WB. Supervision: TE JP MP. Validation: AL MP. Writing – original draft: AL TE PH MB PF WB JP MP. Writing – review & editing: AL TE PH MB PF WB JP MP. * E-mail: a.lenferink@mst.nl26 8 2016 2016 11 8 e016159522 1 2016 8 8 2016 © 2016 Lenferink et al2016Lenferink et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Objective The 12-item Partners in Health scale (PIH) was developed in Australia to measure self-management behaviour and knowledge in patients with chronic diseases, and has undergone several changes. Our aim was to assess the construct validity and reliability of the latest PIH version in Dutch COPD patients. Methods The 12 items of the PIH, scored on a self-rated 9-point Likert scale, are used to calculate total and subscale scores (knowledge; coping; recognition and management of symptoms; and adherence to treatment). We used forward-backward translation of the latest version of the Australian PIH to define a Dutch PIH (PIH(Du)). Mokken Scale Analysis and common Factor Analysis were performed on data from a Dutch COPD sample to investigate the psychometric properties of the Dutch PIH; and to determine whether the four-subscale solution previously found for the original Australian PIH could be replicated for the Dutch PIH. Results Two subscales were found for the Dutch PIH data (n = 118); 1) knowledge and coping; 2) recognition and management of symptoms, adherence to treatment. The correlation between the two Dutch subscales was 0.43. The lower-bound of the reliability of the total scale equalled 0.84. Factor analysis indicated that the first two factors explained a larger percentage of common variance (39.4% and 19.9%) than could be expected when using random data (17.5% and 15.1%). Conclusion We recommend using two PIH subscale scores when assessing self-management in Dutch COPD patients. Our results did not support the four-subscale structure as previously reported for the original Australian PIH. This study was supported by the Lung Foundation Netherlandsgrant number 3.4.11.061van der Palen Job This study was supported by the Lung Foundation Netherlands (grant number 3.4.11.061). URL: https://www.longfonds.nl/. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper and its Supporting Information files. All data are from the COPE-III study whose authors may also be contacted at Medisch Spectrum Twente, Department of Pulmonary Disease, Enschede, the Netherlands.Data Availability All relevant data are within the paper and its Supporting Information files. All data are from the COPE-III study whose authors may also be contacted at Medisch Spectrum Twente, Department of Pulmonary Disease, Enschede, the Netherlands. ==== Body Introduction Self-management interventions aim to improve the health behaviour and self-management skills of patients with chronic and complex health conditions in order to improve the physical health and well-being of these patients [1,2]. Problem solving, decision making, resource utilisation, forming patient-provider partnerships, and patient-tailored action planning are essential parts of self-management [2]. As patient self-management skills develop, increased confidence in their own health management becomes a powerful factor in inducing and sustaining behaviours that provide perceived benefits [2,3]. This is especially important in patients with Chronic Obstructive Pulmonary Disease (COPD) who are responsible for their day-to-day disease management [2]. COPD self-management interventions aim to e.g., instil the confidence to recognise COPD exacerbations [1] and to take appropriate actions when COPD symptoms deteriorate. The most recent Cochrane review regarding COPD self-management interventions showed that COPD self-management interventions are associated with improved health-related quality of life (HRQoL), a reduction in the number of hospitalisations, and improved dyspnoea [4]. In COPD patients, assessments have traditionally involved objective parameters (e.g., lung function). More recently, patient-reported outcomes (PROs) have become increasingly popular. Using PROs, it is not only possible to evaluate outcomes such as COPD-specific HRQoL [5] (e.g., St. George’s Respiratory Questionnaire (SGRQ)) [6] and COPD self-efficacy [7], but also perceived health outcomes. Little is known, however, about perceived health outcomes such as self-management behaviour and knowledge in COPD patients. To facilitate the measurement of self-management behaviour and self-management knowledge of patients with chronic diseases the 12-item Partners in Health scale (PIH) was developed by an Australian research group [8]. The Australian 12-item PIH was intended to provide a first step of assessing a patient’s self-management in developing a collaborative patient-clinician self-management care plan. It was designed to assist patients with chronic and complex conditions in learning how to participate more effectively in the management of their condition and to improve their self-management skills, because previous research indicated that providing coordinated care for people with chronic conditions was predominantly based on their self-management capabilities rather than on the severity and/or complexity of their illness [9]. The Australian 12-item PIH was therefore introduced as a generic self-rated clinical PRO tool suitable for: 1) assessing the effects of self-management interventions in populations with different chronic conditions; 2) comparing populations; and 3) determining changes in patient self-management knowledge and behaviour over time [8]. Subsequently, it was found to be a valid measure of patient competency in relation to the self-management of their chronic conditions [8]. Four subscales were reported based on Principal Component Analysis (PCA): knowledge, coping, recognition and management of symptoms, and adherence to treatment [8]. Hitherto, the Australian PIH has been successfully used to evaluate (self-) management strategies for chronic disease prevention and management [10]. In addition, the PIH has also been used as a screening tool to identify patients who would most benefit from a self-management care plan [11]. The PIH has been translated into Spanish and validated among healthcare users (patients with diabetes, hypertension and cancer) of primary care in Mexico [12]. Three subscales were reported for the Spanish PIH based on exploratory factor analysis (FA) [12]. Having greater insight into COPD patient behaviour and knowledge would facilitate the identification of key COPD self-management skills that could be improved. This could help inform further improvement of patient-tailored COPD self-management interventions and may reduce the high disease burden, hospitalisations and healthcare cost in COPD patients [13,14]. The PIH has, however, not been validated for use in patients with COPD nor has it been validated in the Dutch language. The aim of the current study was, therefore, to assess the construct validity and reliability of a Dutch translation of the latest PIH version in Dutch patients with COPD. More specifically, we assessed the underlying dimensionality of the Dutch PIH using data from a Dutch COPD sample participating in the COPE-III self-management intervention study [15] to determine whether the same four-subscale solution of self-management for the original Australian PIH as proposed by Petkov et al. [8] could be found for the Dutch PIH. Materials and Methods Measures Partners in Health scale The original PIH consists of 12 items (PIHv1), scored on a self-rated 9-point Likert scale with 0 indicating the worst and 8 the best possible patient self-management [8]. Both a total sum score and four subscale scores can be calculated for the PIHv1: knowledge (items 1, 2, 4, 8); coping (items 10–12); recognition and management of symptoms (items 6, 7, 9); adherence to treatment (items 3, 5). Reliability (estimated using Cronbach’s Alpha) equalled 0.82 for the total scale [8]. The 12-item PIHv1 is based on six key principles essential for effective self-management that were transformed into 12 items assessing how well persons were self-managing. It was revised by splitting two double-barrelled items into two questions each; for instance emotional and social impacts of the condition(s) became items 10 and 11 in PIHv2. The resulting 14-item PIH version was used clinically for several years and was also included in a RCT aimed at improving patient self-management competencies [16]. After a national project to determine a consensus definition of self-management the 14-item PIH was further revised [17], which allowed the number of items to be reduced and the time to administer and score the tool minimized, balanced against retention of items that were clinically relevant. Therefore, item 5 from PIHv1 (‘arranging and attend appointments’) was changed into item 6 ‘attend appointments’ in PIHv2. Two questions on monitoring and managing symptoms (item 6 and 8) were removed from PIHv1. In addition, an item on ability to access culturally appropriate services was added (item 5). The result was the current 12-item PIHv2 from which the Dutch version was derived. A copy of PIHv2 can be obtained from Flinders University, Australia. Development of the Dutch PIH For use in a Dutch speaking population the PIHv2 was translated into Dutch then back-translated into English by an independent translator (guidelines Guillemin et al. [18,19]). A Dutch PIHv2 (PIH(Du)) was defined (see S1 Table) and pre-tested in a qualitative evaluation with a small group of Dutch COPD patients who did not participate in the COPE-III self-management study [15], which is an ongoing RCT regarding self-management in COPD patients with comorbidities. Sampling of patients for the qualitative evaluation was continued until saturation of information was achieved. Comments on the wording, layout of the 9-point Likert scale, and issues encountered during the self-administration process were collected using the three-step test interview (TSTI) [20]. Respondents completed the PIH and concurrently verbalised their thoughts (‘think aloud technique’). Subsequently, they answered probes about terms or phrases in the PIH. A predefined cognitive testing protocol [21] was used for this second step. The third step elicited experiences and opinions of patients [20,21]. Non-verbal communications were documented and all verbalisations were audio recorded for further analysis. Data from the TSTI were analysed using content analysis approach [22], in which coding categories are derived directly from the text data. Patients We used baseline data from Dutch COPD patients with comorbidities participating in the COPE-III study for the psychometric analyses [15]. The patient eligibility criteria have been previously described [15] and can be summarised as follows: a clinical diagnosis of COPD [23]; clinically stable at the time of inclusion; at least one clinically relevant comorbidity (ischemic heart disease, heart failure, diabetes, anxiety and/or depression); at least three COPD exacerbations and/or one hospitalisation for respiratory problems in the two years preceding study entry; and adequate Dutch language proficiency. All procedures performed in the current study were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. The study protocol was approved by the Medical Ethical Committee at Medisch Spectrum Twente and by the Southern Adelaide Clinical Human Research Ethics Committee. The study is registered in the public Australian New Zealand Clinical Trials Registry (ACTRN12612000514808). Written informed consent was obtained from all individual participants prior to participation in this study. Statistical analyses Descriptive statistics were calculated using SPSS v20.0 [24]. Both scale structure and item properties were analysed. The analytic strategy was defined prior to viewing the dataset. Following Paap et al. (2015) [25], we used two complementary statistical methods to evaluate the dimensionality of the PIH(Du): 1) Mokken Scale Analysis (MSA; a non-parametric technique); and 2) common FA. In recent years, MSA has increased in popularity in health research [26–31]. MSA identifies scales that allow an ordering of individuals on an underlying scale using unweighted sum scores [32,33]. In order to ascertain which items co-vary and form a scale, scalability coefficients are calculated on three levels: item-pairs (Hij), items (Hi), and scale (H). H is based on Hi and reflects the degree to which the scale can be used to reliably order persons on the latent trait using their sum score. A scale is considered acceptable if 0.3≤H<0.4, good if 0.4≤H<0.5, and strong if H≥0.5 [32,33]. MSA can be used in both a confirmatory and exploratory manner. The exploratory procedure follows a bottom-up, iterative approach. First, a start set of items is identified in one of two ways: 1) the item pair with the highest Hij value is chosen (default), or 2) the researcher specifies the start set manually. Subsequently, the relationship (in terms of H coefficients) of each remaining item with the start set is evaluated one item at a time. At each step, the item that maximises H is added, but only if a) it has a positive relationship (in terms of Hij) with the set of items in the current scale, and b) adding the item results in an Hi value higher than a predefined user-specified constant c (typically 0.3). When no more items can be added, a second subscale is formed. The procedure stops when no items are left, or when no other items can be assigned to subscales anymore. For more detailed information on MSA, we refer to Paap et al. (2013; online supplement [25]). MSA was applied using the R [34] package Mokken [35]. We ran the exploratory analysis several times in a row, each time increasing the lower bound scalability coefficient c [33]. The outcomes indicate whether the data set is one-dimensional or multidimensional [33]. We used Parallel Analysis (PA) based on Minimum Rank Factor Analysis (MRFA); this method will be abbreviated as PA-MRFA [36]. MRFA is a common FA method that allows one to find the “most-unidimensional” solution [37]. In PA-MRFA, for each factor the empirical value of the proportion of explained common variance (ECV) is compared to corresponding factors ECV derived from random data [36]. The random data are generated based on the sample size of the real data assuming independence among items [38]. Typically, a large number of random datasets are generated, resulting in a sampling distribution of ECV-values for each factor. To determine the optimal number of factors, for each successive factor the observed ECV can be compared to the mean or the 95th percentile of the sampling distribution associated with the respective factor. We used the software package FACTOR [39] to perform the PA-MRFA analyses. We used the standard configuration for PA-MRFA: 500 random correlation matrices were generated based on “random permutation of sample values” [36]. Usually, it is advised to use polychoric correlation-based common FA in the case of ordinal data (with five or fewer answering categories). Although the PIH items were scored with nine response options (eligible to be treated as continuous), we had to collapse categories for all items prior to the analyses, in order to ensure adequate coverage (at least 10–15 observations per item-category combination). Polychoric correlation based models would, therefore, be more appropriate. However, they are known to be more prone to convergence issues when small sample sizes are involved. It was therefore decided to run two sets of analyses; one based on polychoric correlations and one based on Pearson correlations. The 95th percentile threshold was used for the polychoric analysis and the mean threshold for the Pearson analysis [36]. Since both sets of models converged and resulted in similar factor solutions, we will only report the findings based on the polychoric correlations. An oblique factor rotation (Promin) was used to facilitate interpretation of the factors [40]. Results Qualitative evaluation of the PIH(Du) Qualitative data were gathered during interviews with four Dutch COPD patients. In general, the instructions were found to be clear and patients indicated that the PIH(Du) was a proper, readable, synoptic, complete and clear instrument. Critical notes were: use of long sentences; information on a time period that fits with the completion of the instrument was lacking; and it could be more COPD-specific. In addition, more specific comments on the individual items and the clarity of wordings were provided for the items 5–12 (see Table 1). Patients’ suggestions for improvements were, for instance, adding a definition of a ‘healthcare professional’ and ‘blood glucoses level’. Other suggestions were: delete ‘culture, value and beliefs’ from item 5 (“You could leave out the last part of this question (culture, values and beliefs)”); add ‘life style’ and rephrase item 9; and split item 12 into different items for the different healthy life styles (e.g., ‘I manage to live a healthy life with no smoking’, ‘I manage to live a healthy life with moderate alcohol use’). The horizontal axis of the 9-point Likert scale was found acceptable and familiar (“This is quite similar to what they ask in connection with the pain threshold”). However, patients also indicated that a PIH(Du) item score of zero (lowest possible self-management) will most likely only be used by patients with an end-stage disease. Suggested improvements for the 9-point Likert scale were using fewer response options and visualising response options (“You could use it like a traffic light”). 10.1371/journal.pone.0161595.t001Table 1 Results of the qualitative evaluation of the 12-item PIH(Du) in four Dutch COPD patients. Item Interpretation Comments (e.g., on clarity of wordings) Improvements 1: Knowledge of illness “What I know in general about my health conditions.” “How much you know yourself about your illness.” “What the health reasons are.” “Whether I have lung issues.” “Whether you are well informed about your own health conditions.” - - 2: Knowledge of treatment “Whether I do know what the treatments and medications are for my conditions.” “It is about what I know in general about the medicines I use.” “The treatment with medication changes so quickly. I think, regarding the information about medicines, that it could be done better.”“And I have pointed that out a few times about my treatment.” - - 3: Taking prescribed medication “Just whether to take the medicines and to follow the treatment instructions.”“Regarding those medicines….nothing is ever said about it or how to use it.” “That you take what is prescribed, as has been agreed with your healthcare provider.” - - 4: Decision sharing “In principle, I always take decisions together with my doctor or healthcare provider.” “Actually, I haven’t been informed about that yet, about what’s wrong—or not wrong—with me.” “I don’t know what, what, what…where I always stand.” “I should talk about it with the doctor or healthcare provider then, shouldn’t I?“Whether you take decisions if you do experience symptoms.” - - 5: Services fit with culture/value/beliefs “Because I do occasionally discuss this with my doctor.” “Should I also arrange for a health professional? That‘s what it seems to say.” “That is self-evident that a healthcare provider should adapt to someone with a different cultural background.” “Yes, and just what does it all mean?” “I don’t understand it very well.” “But this has nothing to do with the kind of healthcare you need, I think.”“The most important thing is that you are able to arrange your healthcare as much as possible yourself.” “You could leave out the last part of this question (culture, values and beliefs).” 6: Arrange and attend appointments “Then you need to go to a doctor or health professional.” “An appointment where I need to go.” “I’ve never had contact with a health professional. Then I don’t know what this health professional is supposed to do.” “What do you mean by that, a health professional?” “So I’d think this word [health professional] is not appropriate in this questionnaire. “Add a definition of health professional.” 7: Track of symptoms “I understand my symptoms.” “Then you need to indicate how and what then. The same goes for your medicines. If I’m breathless or something.” “To act in time if you are not feeling well.” “That you need to know your body well yourself.”“I recognise the symptoms, but I don’t take action.” “I think that this is a good question.” “This is a very long sentence.” “This is not applicable to me, but I do understand it.” “I cannot fill in fairly well or very well, since I don’t know what that is: peak flow.” “Peak flow? What do they mean by that?” “For instance blood sugar levels and peak flows. I don’t know what that is.” “I don’t know to what extent blood glucose levels, peak flows, weight and sleeping problems are related to COPD. I don’t know that as a layperson, do I?” “Add a description of peak flow and blood glucoses level to this question.”“Shorten this question.” “Change this question into: ‘For instance, I watch my symptoms or early warning signs, such as breathlessness’, which makes this more relevant for COPD.” 8: Take action when symptoms deteriorate “Well, then I always tell the doctor when the symptoms get worse.” “Whether I do take action when there are warning signs” “I never take action when I have symptoms or something.” “Yes, well, yes, I do take action. But quite late, usually.” “Usually I contact the pulmonary physician then.” “Because I also think that many people will not understand this…symptoms and all those kinds of words.” “If you want to make it easier to understand for everyone, then you could simplify it.” “Make it more concrete.” 9: Dealing with effects on physical activity “How you function yourself.” “What is possible and what is not possible.” “That I have everything under control, such as performing household chores and walking.” “If I do those activities, how my health will develop.” “If someone leads a regular life, then you will have control over your lungs, over your walking, won’t you.” “Rather a mouthful, in my opinion. And that question really depends on how your complaints are at that moment.” “Short term or long term?” “Because that depends on how your physical condition is at that moment.” “So I think this question is very difficult defined.” “The effects will come later.” “I think this it is a little bit hard to answer.” “The effect of health conditions, I think that yes, that depends on the severity of your conditions, of course.” “Maybe add life style.” “So, I would describe it more, like ‘I can control my physical activities such as household chores, walking, in a normal way.’” “And you could put it in an even simpler way, like: ‘I have control over my health conditions and over my daily activities myself. For example, walking and household chores.’” 10: Dealing with effects on emotional wellbeing “Well, whether I have my emotions under control and that I mentally…That all is well mentally.” “Whether I have control over the effects on my emotional wellbeing.” “Whether I can keep my emotions under control, when I have problems.” “This question is not applicable to me. Actually, I’m always in a good mood.” “Very long sentences. It’s almost like two questions in one.” [reads first half of question out loud] “(…) the effect of my health condition, I think that is very incomprehensible for many people.” “I think the word ‘effect’ will be filled in differently than what is meant.” “You need to turn it around. What or with a question: ‘what is the effect of my health…ehm…condition on your own emotions and whether you have it under control?’” “Start this question with ‘I have insight into my health condition’, because that is easier to understand.” 11: Dealing with effects on social life “I often have things that I think I love to do this or that.” “How I behave and everything.” “Whether I can cope with my health issues.” “I’m not very sociable; I don’t need to be around a lot of people. So I’ll never visit a crowded place.” “It does not have any effect when my symptoms change.” “Also very broad.” “I think this is more about like a character trait.” “It is a general list. I have trouble relating it to lung problems.” “Just like before, start this question with ‘I have insight into (…)’.” 12: Manage to live a healthy life “Whether I am smoking, using alcohol or doing a lot of physical exercise.” “There are several things incorporated that I think are very difficult to answer.” “It can be difficult to indicate whether you eat healthy, I don’t know that.” “Everything has been added to this question.” “I cannot answer this question by giving one answer, since this question contains different things of a healthy life.” “Split this question into different questions for the different healthy life styles, e.g., smoking behaviour, alcohol use, sports etc.” Patient characteristics Patient characteristics for the Dutch COPD sample used for psychometric analysis can be found in Table 2. The PIH(Du) (see S2 Table) was completed by 118 COPD patients (65.3% male, mean age 67.6, 19.5% smoker) diagnosed with at least one clinically relevant comorbidity (71.2% cardiovascular disease, 40.7% diabetes, 19.5% anxiety, 16.9% depression). 10.1371/journal.pone.0161595.t002Table 2 Characteristics of Dutch COPD patients with comorbidities who completed the 12-item Dutch Partners in Health scale. Patient characteristics Total (n = 118) age in years; mean (SD) 67.6 (8.9) male; n (%) 77 (65.3) smoker; n (%) 23 (19.5) mMRC dyspnoea score, range 0–4; mean (SD) 1.99 (0.91) health literacy*, range 1–5; mean (SD) 2.56 (0.92) lung function parameters; mean (SD)  FEV1% predicted post-bronchodilator 52.4 (14.7)  FEV1/FVC post-bronchodilator 51.3 (12.9) diagnosed disease; n (%)  COPD 118 (100)  cardiovascular 84 (71.2)  diabetes 48 (40.7)  depression 20 (16.9)  anxiety 23 (19.5) 12-item PIH(Du) total score; mean (SD) 78.1 (9.7)  PIH(Du) subscale 1**; mean (SD) 35.2 (6.9)  PIH(Du) subscale 2***; mean (SD) 42.9 (4.3) FEV1: Forced Expiratory Volume in one second as percent predicted for age, gender and height; FVC: Forced (expiratory) Vital Capacity; mMRC: modified Medical Research Council; PIH(Du): Dutch Partners in Health scale; SD: Standard Deviation *Health literacy was measured by asking patients for their confidence in completing medical forms by themselves with higher scores indicating lower confidence. **Subscale 1 was tentatively labelled as ‘knowledge and coping’; ***Subscale 2 was tentatively labelled as ‘recognition and management of symptoms, adherence to treatment’. Dimensionality and reliability analyses Running exploratory MSA indicated a two-dimensional pattern for the PIH(Du) (see Table 3). The two PIH(Du) subscales were tentatively labelled as: 1) knowledge and coping (items 1, 2, 8–12) and 2) recognition and management of symptoms, adherence to treatment (items 3–7). The H-values of the two subscales based on the Dutch data were good (0.43, subscale 1) and acceptable (0.38, subscale 2). The correlation between the two subscales was 0.43. The lower-bound of the reliability (estimated using Cronbach’s Alpha) for the total scale equalled 0.84. Cronbach’s Alpha was 0.80 and 0.72 for the PIH(Du) subscales 1 and 2, respectively. 10.1371/journal.pone.0161595.t003Table 3 Scale solutions for the 12-item Dutch Partners in Health scale. 12-item Dutch Partners in Health scale MSA PA-MRFA Item 1: Knowledge of illness 1 1 Item 2: Knowledge of treatment of illness 1 1 Item 3: Taking medication as prescribed 2 2 Item 4: Decision sharing 2 2 Item 5: Services fit with culture/value/beliefs 2 2 Item 6: Arrange and attend appointments 2 2 Item 7: Track of symptoms 2 2 Item 8: Take action when symptoms deteriorate 2 1 Item 9: Dealing with effects on physical activity 1 1 Item 10: Dealing with effects on emotional wellbeing 1 1 Item 11: Dealing with effects on social life 1 1 Item 12: Manage to live a healthy life 1 1 MSA: Mokken Scale Analysis; PA-MRFA: Parallel Analysis based on Minimum Rank Factor Analysis; Note The last two columns indicate whether the item was assigned to the Dutch Partners in Health subscale 1 or 2. Subscale 1 was tentatively labelled as ‘knowledge and coping’, subscale 2 was tentatively labelled as ‘recognition and management of symptoms, adherence to treatment’. The factor analyses resulted in a very similar scale solution to the MSA analyses (see Table 3). The polychoric correlations matrix can be found in Table 4. The first two factors explained a larger percentage of common variance (39.4% and 19.9% for factor 1 and 2, respectively) than could be expected when using random data (see Table 5). The estimated correlation between the factors extracted from the Dutch data was 0.41. The factor analyses for the two PIH(Du) subscales showed that the newly added item 5 showed similar factor loadings for both subscales; 0.39 for subscale 1 and 0.48 for subscale 2 (see Table 6). 10.1371/journal.pone.0161595.t004Table 4 Polychoric correlations matrix for the 12-item Dutch Partners in Health scale. I1 I2 I3 I4 I5 I6 I7 I8 I9 I10 I11 I12 I1 1.00 I2 0.60 1.00 I3 0.03 0.16 1.00 I4 0.27 0.26 0.73 1.00 I5 0.40 0.38 0.34 0.61 1.0 I6 0.00 0.14 0.70 0.46 0.22 1.00 I7 0.12 0.26 0.42 0.39 0.44 0.20 1.00 I8 0.34 0.31 0.23 0.24 0.50 0.07 0.56 1.00 I9 0.25 0.28 -0.20 -0.05 0.24 -0.04 0.33 0.32 1.00 I10 0.32 0.26 -0.06 0.11 0.40 -0.01 0.22 0.31 0.58 1.00 I11 0.38 0.35 0.20 0.23 0.36 0.21 0.34 0.28 0.47 0.64 1.00 I12 0.20 0.32 0.17 0.23 0.36 0.19 0.34 0.38 0.41 0.60 0.51 1.00 10.1371/journal.pone.0161595.t005Table 5 Results of Minimum Rank Factor Analysis Dutch Partners in Health scale. Factor % ECV real data Mean % ECV random data 95th percentile % ECV random data Eigenvalue* 1 39.4 17.5 20.1 4.17 2 19.9 15.1 16.7 2.16 3 9.6 13.4 14.9 0.98 4 8.9 11.8 12.9 0.78 5 6.2 10.3 11.4 0.51 6 5.0 8.9 9.9 0.29 7 3.9 7.5 8.6 0.20 8 3.2 6.1 7.2 0.19 9 2.4 4.6 6.0 0.11 10 0.9 3.2 4.6 0.07 11 0.6 1.8 3.1 0.00 12 0.0 0.0 0.0 0.00 ECV: explained common variance *Based on reduced correlation matrix Note: Standardized Cronbach’s Alpha (total scale) = 0.84 10.1371/journal.pone.0161595.t006Table 6 Factor loadings of the Dutch Partners in Health scale based on Minimum Rank Factor Analysis. PIH(Du) subscale 1: ‘knowledge and coping’ PIH(Du) subscale 2: ‘recognition and management of symptoms, adherence to treatment’ Item 1: Knowledge of illness 0.57 0.07 Item 2: Knowledge of treatment of illness 0.47 0.19 Item 3: Taking medication as prescribed -0.39 1.05 Item 4: Decision sharing -0.13 0.93 Item 5: Services fit with culture/value/beliefs 0.39 0.48 Item 6: Arrange and attend appointments -0.26 0.74 Item 7: Track of symptoms 0.30 0.45 Item 8: Take action when symptoms deteriorate 0.49 0.26 Item 9: Dealing with effects on physical activity 0.80 -0.27 Item 10: Dealing with effects on emotional wellbeing 0.89 -0.17 Item 11: Dealing with effects on social life 0.65 0.12 Item 12: Manage to live a healthy life 0.60 0.13 PIH(Du): Dutch Partners in Health scale. Note: To aid interpretation, the factor loadings higher than 0.40 are printed in bold. Discussion Our dimensionality analyses showed a two-subscale solution for the PIH(Du): 1) knowledge and coping; 2) recognition and management of symptoms, adherence to treatment. Our results therefore did not support the four-subscale structure as previously reported for the original Australian PIH [8]. It is of interest that a Spanish version of the PIH was found to have a three-subscale solution [12]. Several possible explanations have been put forward to account for different findings in factorial solutions across studies: differences in statistical methods and target populations, sample size, number of items per factor, number of factors in the model, and the size of the communalities (proportion of the variance of an item that is accounted for by the common factors in the model) [31,41,42]. At the time of the original Australian PIH development [8], its dimensionality was evaluated by using a two-stage procedure: an exploratory PCA (data reduction technique to group items into a set of new variables) and a confirmatory common FA (a mathematical model to estimate the relationship between items and latent variables [43]) was subsequently used to “validate” the structure identified by the exploratory analysis. However, PCA and common FA will only produce similar results under very specific circumstances [38]. We favoured using exploratory IRT and common FA models over PCA in this study, because they are suitable for ordinal data [44] and result in meaningful scales (e.g., Borsboom et al. [45]). It is unclear which exploratory FA was performed for the Spanish PIH validation [12]. We were therefore unable to compare our results with the three-subscale solution for the Spanish PIH. The MRFA criteria used in our study require less interpretation in determining dimensionality and allows one to find the “most-unidimensional” solution [37], in comparison with conclusions based on a PCA. Petkov et al. used a Cattell’s Scree plot [46] as a graphical representation of the eigenvalues and suggested a cut-off of three components as defined by the ‘elbow’. This choice is somewhat arbitrary and the plot can be interpreted in different ways, since the slope has flattened from two components onwards and, therefore, the cut-off point could also be at two or one component. It has been shown that the Scree test has a tendency to overestimate the number of subscales [47] and it should be used and interpreted with care. Kaiser’s criterion to retain factors with eigenvalues greater than one for interpretation is the best known and most utilised method in practice [48]. Despite its simplicity, though, this method may also lead to arbitrary decisions and be inefficient in determining the number of subscales [48]. There is no consensus about a decision rule for the minimal sample size requirements in dimensionality analyses. In the current study, our sample size of 118 COPD patients is of a small to moderate size, with a correlation between the two PIH(Du) subscales of 0.43 and H-values of 0.43 and 0.38. According to the guidelines of Straat et al. (2014) [49] the sample size should be 50 to 250 to obtain 90 to 99% correct item assignment and adequate to good Per Element Accuracy in MSA. For MSA analyses the required minimal sample size is mainly dependent on the correlations between the latent variables and the H-values of the items [49]. Based on the correlations and H-values we found in the current study, our sample size should be sufficient to obtain 94–99% correct item assignment [49]. For FA the minimally required sample size depends on a complex interplay of many aspects, e.g., the estimated factor loadings and communalities [50]. When communalities are high, sample size tends to have less influence on the quality of factor solutions compared to when communalities are low [50]. In case of relatively low communalities, a larger sample size and number of items per factor are needed to obtain stable results in FA [41]. Conversely, in case of a relatively small sample size, a higher number of items per factor (≥ 4 items per factor [42]), a small number of factors and moderate to high communalities are needed to estimate a model that will give a good representation of the population factors [41]. Since the factorial solutions we found consist of a small number of well-identified factors with moderate to high communalities, we feel confident that our low-dimensional solutions for the PIH(Du) will be easy to replicate. Cross-cultural differences and adjustments made after publication of the original PIH may also have contributed to the discrepancy in dimensionality between the original Australian PIH and the PIH(Du). For instance, item 5 (‘dealing with health professionals to get services that fit with culture, values and beliefs’), which is unique to the PIHv2, was difficult to interpret for Dutch patients and most patients felt the item was not applicable to them. In addition, item 5 showed high factor loadings on both of the Dutch subscales, making it difficult to assign the item to either scale. We therefore suggest removing this item. Item 10 (‘manage the effect of health condition(s) on emotional wellbeing’) has recently been added by the PIH authors in an attempt to show the psychological/emotional impact of the disease(s). Their clinical experiences so far suggest that the item is powerful in ‘breaking open the case’ to uncover factors that can interfere with self-management. However, this item was poorly-received by patients completing the PIH(Du); patients indicated the item was too lengthy, the formulation too complex and it was unclear what the reference time period was. We therefore suggest specifying a recall period in the PIH. Differences in heterogeneity between the Australian and Dutch samples may also have contributed to the difference in the number of subscales found. Studies on other self-report instruments, such as the SCL-90, have indicated that the number of dimensions found can be related to for example disease severity [31]. Whereas the original Australian PIH was completed by patients with different kinds of chronic diseases, including respiratory problems, the PIH(Du) was administered exclusively to COPD patients, albeit with comorbidities and different COPD severity scores. Patients may provide different responses if multiple chronic conditions are present. For instance, ‘health condition(s)’, as used in the items 1, 2, 4, 9, 10 and 11 from the PIH(Du) is a broad definition and can be interpreted in different ways. Patients completing the PIH may only have considered those health conditions for which they have recently experienced symptom deterioration. Therefore, when multiple chronic conditions are present, the specific contribution and effects of each chronic condition cannot be assessed by the PIH scores. However, PIH scores were developed to enable assessment of the knowledge and behaviour of patients in general to improve self-management interventions. Based on our findings, we feel confident that the PIH is a useful tool in assessing self-management behaviour and knowledge in COPD patients, but we do recommend some minor changes to the instrument. Obviously, the PIH requires translation if used in other than the source language, which is often the case in international research [51–53]. However, when, besides translation, other changes are made over time to further improve measurement instruments, this may negatively impact its interpretation for use in research or clinical practice. First, with regard to changes made to the Australian PIH version, clear guidelines are needed before translation and validation of the instrument for use in other settings and countries can be continued. Second, we recommend introducing a recall period. Third, we suggest avoiding the use of terms with multiple meanings and composite items (e.g., it is difficult to respond unequivocally to the question ‘‘I take medications or carry out the treatments” if patients do take their medication, but do not carry out the treatments as asked by the doctor). Furthermore, none of the Dutch patients used all nine response options. Simplifying the PIH by using fewer response options could therefore be considered, although any such change would of course require re-validation. As a next step in our validation process, we plan to investigate the clinical relevance of the two-subscale solution by assessing the ability of both subscale scores to discriminate between patients who received benefit from the COPD self-management intervention (e.g. better self-treatment adherence, higher quality of life scores, fewer hospitalisations and fewer exacerbation days) and those who did not, and who demonstrated a poor self-management capacity. We will also assess the associations between the subscale scores and e.g. quality of life. In addition, we have planned to assess the responsiveness of the PIH, and whether response shift occurs in COPD patients. A study by Harvey and colleagues showed that self-reported Australian PIH scores improved significantly over time when patients with chronic diseases were involved in peer-led self-management education programs [54]. Their results indicated that patients had improved understanding of their condition and the ability to manage and deal with their symptoms resulting in a positive effect on self-management skills, confidence and health-related behaviour [54]. Our ongoing RCT regarding self-management in COPD patients [15] will allow us to assess the responsiveness of the PIH in more detail. Conclusion This is the first time that a translated Dutch PIH was validated in a sample of Dutch COPD patients. Our findings indicate that most items are well-received by patients and show favourable psychometric properties. We recommend making minor changes and refinements. More importantly, however, there is need for (international) consensus on a final version of the PIH which can be validated in several settings and populations. Nevertheless, the PIH shows great promise to facilitate the identification of self-management skills needing improvement in COPD patients with other comorbid conditions. PIH scores could be used to tailor COPD self-management interventions to the patient’s needs and capabilities, facilitating appropriate self-management of COPD exacerbations and a reduction of hospitalisations. For use in Dutch COPD patients, we recommend using two PIH subscale scores when assessing self-management knowledge and behaviour. More research is needed to evaluate whether this two-subscale solution is optimal for other populations as well. Supporting Information S1 Table Dutch translated 12-item Partners in Health scale (PIH(Du)). (PDF) Click here for additional data file. S2 Table Observed scores of the Dutch 12-item Partners in Health scale (PIH(Du)). (PDF) Click here for additional data file. We would like to thank the COPD patients who participated in this study. We would also like to thank Talencentrum Maastricht University for the Dutch translation of the PIH. We thank Mitzi Paap, Bachelor of Arts in English language and culture, for translating the citations of patients, and helpful discussions. This study was supported by the Lung Foundation Netherlands (grant number 3.4.11.061). ==== Refs References 1 Bourbeau J , van der Palen J . Promoting effective self-management programmes to improve COPD . Eur Respir J 2009 3 ;33 (3 ):461 –3 . 10.1183/09031936.00001309 19251792 2 Lorig KR , Holman H . Self-management education: history, definition, outcomes, and mechanisms . Ann Behav Med 2003 8 ;26 (1 ):1 –7 . 12867348 3 Bourbeau J , Nault D , Dang-Tan T . Self-management and behaviour modification in COPD . Patient Educ Couns 2004 3 ;52 (3 ):271 –7 . 14998597 4 Zwerink M , Brusse-Keizer M , van der Valk PD , Zielhuis GA , Monninkhof EM , van der Palen J , et al Self management for patients with chronic obstructive pulmonary disease . Cochrane Database Syst Rev 2014 3 ;19 ;3 :CD002990 . 5 Guyatt GH , Berman LB , Townsend M , Pugsley SO , Chambers LW . A measure of quality of life for clinical trials in chronic lung disease . Thorax 1987 10 ;42 (10 ):773 –8 . 3321537 6 Jones PW , Quirk FH , Baveystock CM , Littlejohns P . A self-complete measure of health status for chronic airflow limitation. The St. George's Respiratory Questionnaire . Am Rev Respir Dis 1992 6 ;145 (6 ):1321 –7 . 1595997 7 Wigal JK , Creer TL , Kotses H . The COPD Self-Efficacy Scale . Chest 1991 5 ;99 (5 ):1193 –6 . 2019177 8 Petkov J , Harvey P , Battersby M . The internal consistency and construct validity of the partners in health scale: validation of a patient rated chronic condition self-management measure . Qual Life Res 2010 9 ;19 (7 ):1079 –85 . 10.1007/s11136-010-9661-1 20437206 9 Battersby M , Harvey P , Mills PD , Kalucy E , Pols RG , Frith PA , et al SA HealthPlus: a controlled trial of a statewide application of a generic model of chronic illness care . Milbank Q 2007 ;85 (1 ):37 –67 . 17319806 10 Walters J , Cameron-Tucker H , Wills K , Schüz N , Scott J , Robinson A , et al Effects of telephone health mentoring in community-recruited chronic obstructive pulmonary disease on self-management capacity, quality of life and psychological morbidity: a randomised controlled trial . BMJ Open 2013 9 6 ;3 (9 ):e003097 10.1136/bmjopen-2013-003097 24014482 11 Battersby M , Harris M , Smith D , Reed R , Woodman R . A pragmatic randomized controlled trial of the Flinders Program of chronic condition management in community health care services . Patient Educ Couns 2015 . 12 Peñarrieta-de Córdova I , Barrios FF , Gutierrez-Gomes T , Piñonez-Martinez Mdel S , Quintero-Valle LM , Castaneda-Hidalgo H . Self-management in chronic conditions: partners in health scale instrument validation . Nurs Manag (Harrow) 2014 3 ;20 (10 ):32 –7 .24571163 13 Effing T , Kerstjens H , van der Valk , Zielhuis G , van der Palen J . (Cost)-effectiveness of self-treatment of exacerbations on the severity of exacerbations in patients with COPD: the COPE II study . Thorax 2009 11 ;64 (11 ):956 –62 . 10.1136/thx.2008.112243 19736179 14 Rice KL , Dewan N , Bloomfield HE , Grill J , Schult TM , Nelson DB , et al Disease Management Program for Chronic Obstructive Pulmonary Disease: A Randomized Controlled Trial . Am J Respir Crit Care Med 2010 1 14 . 15 Lenferink A , Frith P , van der Valk P , Buckman J , Sladek R , Cafarella P , et al A self-management approach using self-initiated action plans for symptoms with ongoing nurse support in patients with Chronic Obstructive Pulmonary Disease (COPD) and comorbidities: The COPE-III study protocol . Contemp Clin Trials 2013 9 ;36 (1 ):81 –9 . 10.1016/j.cct.2013.06.003 23770110 16 Battersby MW , Harris M , Reed RL , Harvey PW , Woodman R.J. , Frith P . A randomised trial of the Flinders Program to improve patient self-management compentencies in a range of chronic conditions: study rationale and protocol . Australasian Medical Journal AMJ 2010 ;1 (3 ):198 –204 . 17 Battersby M , Lawn S . Capabilities for supporting prevention and chronic condition self-management: A resource for educators of primary health care professionals . Flinders Human Behaviour and Health Research Unit, Flinders University , Bedford Park, SA ; 2009 . 18 Guillemin F , Bombardier C , Beaton D . Cross-cultural adaptation of health-related quality of life measures: literature review and proposed guidelines . J Clin Epidemiol 1993 ;46 (12 ):1417 –32 . 8263569 19 Guillemin F . Cross-cultural adaptation and validation of health status measures . Scand J Rheumatol 1995 ;24 (2 ):61 –3 . 7747144 20 Hak T , van der Veer K , Jansen H . The Three-Step Test-Interview (TSTI): An observation-based method for pretesting self-completion questionnaires . Journal of the European Survey Research Association 2008 ;2 (3 ):143 –50 . 21 Willis GB . Cognitive interviewing: A tool for improving questionnaire design . Thousand Oaks (CA) : SAGE ; 2005 . 22 Hsieh HF , Shannon SE . Three approaches to qualitative content analysis . Qual Health Res 2005 11 ;15 (9 ):1277 –88 . 16204405 23 Rabe KF , Hurd S , Anzueto A , Barnes PJ , Buist SA , Calverley P , et al Global strategy for the diagnosis, management, and prevention of chronic obstructive pulmonary disease: GOLD executive summary . Am J Respir Crit Care Med 2007 9 15 ;176 (6 ):532 –55 . 17507545 24 IBM SPSS Statistics for Windows, Version 20.0. Armonk, NY: IBM Corp. [computer program]. Version 20.0. Armonk, NY: IBM Corp.; 2011. 25 Paap MCS , Brouwer D , Glas CAW , Monninkhof EM , Forstreuter B , Pieterse ME , et al The St George's Respiratory Questionnaire revisited: a psychometric evaluation . Qual Life Res 2015 . 26 Watson R , van der Ark LA , Lin LC , Fieo R , Deary IJ , Meijer RR . Item response theory: how Mokken scaling can be used in clinical practice . J Clin Nurs 2012 10 ;21 (19–20 ):2736 –46 . 10.1111/j.1365-2702.2011.03893.x 21883577 27 Sijtsma K , Emons WH , Bouwmeester S , Nyklicek I , Roorda LD . Nonparametric IRT analysis of Quality-of-Life Scales and its application to the World Health Organization Quality-of-Life Scale (WHOQOL-Bref) . Qual Life Res 2008 3 ;17 (2 ):275 –90 . 10.1007/s11136-007-9281-6 18246447 28 Sutterlin S , Paap MCS , Babic S , Kubler A , Vogele C . Rumination and age: some things get better . J Aging Res 2012 . 29 Beukers F , Houtzager BA , Paap MCS , Middelburg KJ , Hadders-Algra M , Bos AF , et al Parental psychological distress and anxiety after a successful IVF/ICSI procedure with and without preimplantation genetic screening: follow-up of a randomised controlled trial . Early Hum Dev 2012 9 ;88 (9 ):725 –30 . 22460061 30 Paap MCS , Kreukels BP , Cohen-Kettenis PT , Richter-Appelt H , de Cuypere G , Haraldsen IR . Assessing the utility of diagnostic criteria: a multisite study on gender identity disorder . J Sex Med 2011 1 ;8 (1 ):180 –90 . 10.1111/j.1743-6109.2010.02066.x 20946149 31 Paap MCS , Meijer RR , Cohen-Kettenis PT , Richter-Appelt H , de Cuypere G , Kreukels BP , et al Why the factorial structure of the SCL-90-R is unstable: comparing patient groups with different levels of psychological distress using Mokken Scale Analysis . Psychiatry Res 2012 12 30 ;200 (2–3 ):819 –26 . 10.1016/j.psychres.2012.03.012 22494703 32 Mokken RJ . A theory and procedure of scale analysis . The Hague : Mouton ; 1971 . 33 Sijtsma K , Molenaar IW . Introduction to Nonparametric Item Response Theory . Thousand Oaks : Sage Publications 2002 ;5 . 34 R: A language and environment for statistical computing [computer program ]. Vienna : R Foundation for Statistical Computing ; 2013 . 35 van der Ark LA . Mokken scale analysis in R . Journal of Statistical Software , 2007 ;20 (11 ):1 –19 . 36 Timmerman ME , Lorenzo-Seva U . Dimensionality assessment of ordered polytomous items with parallel analysis . Psychol Methods 2011 6 ;16 (2 ):209 –20 . 10.1037/a0023353 21500916 37 Ten Berge JMF , Sočan G . The greatest lower bound to the reliability of a test and the hypothesis of unidimensionality . Psychometrika 2004 ;69 (4 ):613 –25 . 38 Timmerman ME , Lorenzo-Seva U , Ceulemans E . The number of factors problem In: Irwing P , Booth T , Hughes DJ , editors. The Wiley-Blackwell Handbook of Psychometric Testing .Oxford : Wiley-Blackwell ; 2015 . 39 Lorenzo-Seva U , Ferrando PJ . FACTOR: a computer program to fit the exploratory factor analysis model . Behav Res Methods 2006 2 ;38 (1 ):88 –91 . 16817517 40 Lorenzo-Seva U . Promin: A Method for Oblique Factor Rotation . Multivariate Behavioral Research 1999 ;34 (3 ):347 –65 . 41 MacCallum R , Widaman KF , Zhang S , Hong S . Sample size in factor analysis . Psychological Methods 1999 ;4 (1 ):84 –99 . 42 Marsh H , Hau K , Balla J , Grayson D . Is more ever too much? The number of indicators per factor in confirmatory factor analysis . Multivariate Behavioral Research 1998 ;33 (2 ):181 –220 . 10.1207/s15327906mbr3302_1 26771883 43 Field A . Exploratory factor analysis Discovering statistics using SPSS . 3 ed SAGE Publications Inc. ; 2009 p. 627 –85 . 44 Wismeijer AA , Sijtsma K , van Assen MA , Vingerhoets AJ . A comparative study of the dimensionality of the self-concealment scale using principal components analysis and Mokken scale analysis . J Pers Assess 2008 7 ;90 (4 ):323 –34 . 10.1080/00223890802107875 18584441 45 Borsboom D . The attack of the psychometricians . Psychometrika 2006 9 ;71 (3 ):425 –40 . 19946599 46 Cattell R . The scree test for the number of factors . Multivariate Behavioral Research 1966 ;(1 ):245 –76 .26828106 47 Zwick W , Velicer W . Comparison of five rules for determining the number of components to retain . Psychological Bulletin 1986 ;99 :432 –42 . 48 Fabrigar L , Wegener D , MacCallum R , Strahan E . Evaluating the use of exploratory factor analysis in psychological research . Psychologicial Methods 1999 ;3 :272 –99 . 49 Straat JH , van der Ark LA , Sijtsma K . Minimum Sample Size Requirements for Mokken Scale Analysis . Educational and Psychological Measurement 2014 ;74 (5 ):809 –22 . 50 Hogarty KY , Hines CV , Kromrey JD , Ferron JM , Mumford KR . The quality of factor solutions in exploratory factor analysis: The influence of sample size, communality, and overdetermination . Educational and Psychological Measurement 2005 ;65 (202 ):226 . 51 Beaton DE , Bombardier C , Guillemin F , Ferraz MB . Guidelines for the process of cross-cultural adaptation of self-report measures . Spine (Phila Pa 1976 ; %2000 12 15 ;25 (24 ):3186 –91 .11124735 52 Bullinger M , Alonso J , Apolone G , Leplege A , Sullivan M , Wood-Dauphinee S , et al Translating health status questionnaires and evaluating their quality: the IQOLA Project approach. International Quality of Life Assessment . J Clin Epidemiol 1998 11 ;51 (11 ):913 –23 . 9817108 53 Wiesinger GF , Nuhr M , Quittan M , Ebenbichler G , Wolfl G , Fialka-Moser V . Cross-cultural adaptation of the Roland-Morris questionnaire for German-speaking patients with low back pain . Spine (Phila Pa 1976 ;1999 6 1 ;24 (11 ):1099 –103 .10361659 54 Harvey PW , Petkov JN , Misan G , Fuller J , Battersby MW , Cayetano TN , et al Self-management support and training for patients with chronic and complex conditions improves health-related behaviour and health outcomes . Aust Health Rev 2008 5 ;32 (2 ):330 –8 . 18447824

PMC005xxxxxx/PMC5001638.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 27564376PONE-D-15-4324110.1371/journal.pone.0161556Research ArticleResearch and Analysis MethodsDatabase and Informatics MethodsResearch and Analysis MethodsImaging TechniquesPhysical SciencesMathematicsAlgebraLinear AlgebraEigenvaluesMedicine and Health SciencesOphthalmologyRetinal DisordersRetinopathyDiabetic RetinopathyPhysical SciencesMathematicsApplied MathematicsAlgorithmsResearch and Analysis MethodsSimulation and ModelingAlgorithmsEngineering and TechnologySignal ProcessingSignal FilteringMatched FiltersPhysical SciencesMathematicsProbability TheoryRandom VariablesCovarianceBiology and Life SciencesAnatomyCardiovascular AnatomyBlood VesselsRetinal VesselsMedicine and Health SciencesAnatomyCardiovascular AnatomyBlood VesselsRetinal VesselsBiology and Life SciencesAnatomyOcular SystemOcular AnatomyRetinal VesselsMedicine and Health SciencesAnatomyOcular SystemOcular AnatomyRetinal VesselsRetinal Microaneurysms Detection Using Gradient Vector Analysis and Class Imbalance Classification Retinal Microaneurysms DetectionDai Baisheng 1Wu Xiangqian 1*Bu Wei 2 1 School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China 2 Department of New Media Technologies and Arts, Harbin Institute of Technology, Harbin, China Woloschak Gayle E. Editor Northwestern University Feinberg School of Medicine, UNITED STATES Competing Interests: The authors have declared that no competing interests exist. Conceived and designed the experiments: BD XW. Performed the experiments: BD. Analyzed the data: BD WB. Wrote the paper: BD XW WB. * E-mail: xqwu@hit.edu.cn2016 26 8 2016 11 8 e01615561 10 2015 9 7 2016 © 2016 Dai et al2016Dai et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Retinal microaneurysms (MAs) are the earliest clinically observable lesions of diabetic retinopathy. Reliable automated MAs detection is thus critical for early diagnosis of diabetic retinopathy. This paper proposes a novel method for the automated MAs detection in color fundus images based on gradient vector analysis and class imbalance classification, which is composed of two stages, i.e. candidate MAs extraction and classification. In the first stage, a candidate MAs extraction algorithm is devised by analyzing the gradient field of the image, in which a multi-scale log condition number map is computed based on the gradient vectors for vessel removal, and then the candidate MAs are localized according to the second order directional derivatives computed in different directions. Due to the complexity of fundus image, besides a small number of true MAs, there are also a large amount of non-MAs in the extracted candidates. Classifying the true MAs and the non-MAs is an extremely class imbalanced classification problem. Therefore, in the second stage, several types of features including geometry, contrast, intensity, edge, texture, region descriptors and other features are extracted from the candidate MAs and a class imbalance classifier, i.e., RUSBoost, is trained for the MAs classification. With the Retinopathy Online Challenge (ROC) criterion, the proposed method achieves an average sensitivity of 0.433 at 1/8, 1/4, 1/2, 1, 2, 4 and 8 false positives per image on the ROC database, which is comparable with the state-of-the-art approaches, and 0.321 on the DiaRetDB1 V2.1 database, which outperforms the state-of-the-art approaches. http://dx.doi.org/10.13039/501100001809National Natural Science Foundation of China61472102Bu Wei This work was supported by the National Natural Science Foundation of China (http://www.nsfc.gov.cn/) under Grants 61472102 (Received by WB). Data AvailabilityAll relevant data are available from the Retinopathy Online Challenge and the DiaRetDB1 V2.1 database (doi:10.1109/TMI.2009.2033909, doi:10.1155/2013/368514).Data Availability All relevant data are available from the Retinopathy Online Challenge and the DiaRetDB1 V2.1 database (doi:10.1109/TMI.2009.2033909, doi:10.1155/2013/368514). ==== Body Introduction Diabetic retinopathy (DR) is the commonest complication of diabetes and one of the major causes of blindness. The early diagnosis and treatment of DR are very important to prevent vision impairment. The microaneurysms (MAs) on retina are the small saccular bulges in the walls of retinal capillary vessels [1] and generally appear near to the macula [2]. MAs are the first sign of DR [3]. According to the Early Treatment Diabetic Retinopathy Study (ETDRS) [4], the presence of only even 1 or 2 MAs shows the symptom of the mild non-proliferative diabetic retinopathy (i.e., ETDRS level 20). The more MAs, the higher risk of the progression of retinopathy [5, 6]. However, screening of MAs is usually performed manually by ophthalmologist through visual inspection of the color fundus image [5–7], which is a time-consuming, repetitive, tiring, subjective and error-prone process. Therefore, it is necessary to investigate automated MAs detection in color fundus image. MAs always appear as dark red and small round spots in color fundus image. Fig 1 shows a color fundus image containing MAs and a corresponding enlarged part in green channel with some MAs indicated by yellow markers. As shown in Fig 1B, some of MAs are difficult to be found out and separated from the background noises (e.g., the subtle one indicated by the circle). In addition, some MAs may have an irregular shape (e.g., the pedunculated one indicated by the triangle), cluster together (e.g., the clustered ones indicated by the square) or close to vessels (e.g., the one indicated by the pentagon). Thus, automated MAs detection in the color fundus image is a challenging task. In general, there are two stages for automated MAs detection, i.e. candidate MAs extraction and classification. In the first stage, the candidate MAs are extracted, and in the second stage, the true MAs are identified from candidates by a classifier with a set of extracted features. 10.1371/journal.pone.0161556.g001Fig 1 An example of color fundus image containing MAs. (A) The color fundus image. (B) The corresponding enlarged part of (A) in green channel with indicated MAs (diamond: regular MA, circle: subtle MA, triangle: irregular MA, square: clustered MAs, pentagon: MA close to vessel). Reprinted from [8] under a CC BY license, with permission from Dr. Yalin Zheng, original copyright 2012. For candidate MAs extraction, most of the existing works could be roughly categorized into the matched filter based, the morphology based and the other approaches. In the matched filter based approaches, some special filters were designed to discriminate the MAs from other structures. Quellec et al. [9] and Zhang et al. [10] modeled the MAs with 2D Gaussian functions, and detected the MAs by using template matching [9] or multi-scale correlation filtering [10]. Giancardo et al. [11] used Radon transformation to extract the Gaussian-like circular MAs. Hatanaka et al. [12] proposed a double-ring filter to detect MAs by comparing the intensities between different circular regions. Matched filter based approaches [9–12] worked well when the shapes of MAs are similar with the shape of the filters, but failed to extract some irregular MAs (e.g., MAs with saccular, fusiform, or pedunculated shape [13]), clustered MAs or MAs close to vessels. In morphology based approaches, some morphology characteristics of MAs were used to extract the candidate MAs. Fleming et al. [14] and Ram et al. [15] detected MAs based on the morphological top-hat operation with different linear structural elements. Walter et al. [16] assumed that the diameters of MAs are always smaller than a threshold and proposed a morphological diameter closing operation to extract MAs. Rosas-Romero et al. [17] extracted MAs by using the bottom-hat and hit-or-miss transformations. Morphology based approaches can effectively extract the MAs whose shapes and sizes are similar with those of the structural elements. However, since the shapes and the sizes of different MAs vary largely in fundus images, it is very difficult to define a set of morphological features to characterize and detect all MAs. In the other approaches, such as in [18], a pixel classification followed the morphological operation was used to determine the candidate MAs. In [19], a moat operator was first applied to enhance the edge of candidate MAs, and a recursive region growing was then extracted the area of candidate MAs. It is noted that any true MA lost in this stage cannot be retrieved in the next stage. Hence, to improve the sensitivity of automated MAs detection, it is important to extract as many true MAs as possible in this stage. Due to the complexity of fundus image, besides the true MAs, the extracted candidate MAs also include a large amount of non-MAs. In candidate MAs classification stage, Fleming et al. [14] defined three types of features for each candidate, i.e., size, intensity and vesselness Boolean features, and then trained a kNN classifier to identify true MAs. A sensitivity of 0.54 at the level of 10 false positives per image (FPs/I) was reported on a private dataset. Niemeijer et al. [18] exploited shape, intensity and texture features for each candidate and also trained a kNN classifier to recognize the true MAs. A CPM score of 0.395 (competition performance metric, i.e., an average sensitivity at a set of particular false positives per image 1/8, 1/4, 1/2, 1, 2, 4 and 8 FPs/I) was achieved on the ROC database [3]. Giancardo et al. [20] trained a SVM classifier to identify true MAs with a set of features extracted from Radon space and obtained a CPM score of 0.375 on the ROC database. Hatanaka et al. [12] utilized shape, color, statistic and some filter response features to train an ANN for identifying true MAs from non-MAs and obtained a sensitivity of 0.68 at the level of 15 FPs/I on 25 images of ROC training dataset. Lazar et al. [21] proposed a set of intensity profile features and applied a naïve Bayes classifier in candidate MAs classification. A CPM score of 0.423 was achieved on the ROC database. In [22], Fegyver applied a set of features based on gradient directions and lengths and trained a naïve Bayes classifier to identify true MAs. The CPM score of this method is 0.422 on the ROC database. These traditional classifiers work well when the samples of different classes are balanced. However, in the extracted candidate MAs, the number of non-MAs is always more tremendous than the number of true ones. The ratio between the non-MAs and the true MAs is very high, e.g., this ratio reported in [18] was close to 50:1 (14591 non-MAs and 315 true MAs). That is, the true and the non-MAs are extremely imbalanced and those traditional classifiers may not work well for this extremely class imbalanced problem. Class-imbalanced classification should be introduced for MA classification. Hatanaka et al. [12] limited the maximum number of the candidate MAs in each image, and Lazar et al. [21] manually selected the negative samples (non-MAs) to construct a training set, in which the true MAs and the non-MAs classes are not too class-imbalanced. However, limiting the number of candidates may reduce the sensitivity of MAs detection, and the manually selection of samples from about ten thousand non-MAs is very time consuming and subjective, and lose much information of training samples. In [16], Walter et al. used a Bayesian risk minimization rule to classify the true MAs, where a misclassification cost parameter was introduced to alleviate the problem of class imbalance. But the specific cost information is rarely available, and the method obtained similar results with the kNN classifier as reported by the authors [16]. Séoud et al. [23] utilized a Random Forest classifier to overcome the problem of imbalanced training data. However, this classifier may not work well with highly imbalanced data as reported in [24]. Recently, Antal et al. [25] presented an ensemble approach to detect MAs, which selected five image preprocessing operations and five candidate extractors [10, 16, 26–28], to form 25 〈preprocessing operation, candidate extractor〉 pairs with 225 possible combinations. The final MAs were detected by the fusion of the MA candidates which were extracted by the individual pairs in the optimal ensemble. This ensemble strategy has outperformed most existing individual approaches on ROC database [3], in which almost all of the small dark spots in the images are MAs. Actually, in fundus images, besides MAs, there are other objects which are also shown as small dark spots, e.g. the small hemorrhages and scars left after PRP (Pan-Retinal Photocoagulation) treatment, etc. These small dark spots cannot be further discriminated by this approach since all candidate extractors used in [25] were designed for roughly extracting small dark spot objects. Therefore, if fundus images contains more complicated small dark spots, the performance of this approach will be deteriorated, as indicated by the results on DiaRetDB1 V2.1 database [29] reported in [25]. This paper proposes a novel automated method for MAs detection. The key idea is that, to preserve the MAs with different appearance, intensity and size as many as possible, we only reject the obvious non-MA objects, e.g., vessels and some background noises, with the gradient vector analysis, and all the remained small dark objects are taken as MA candidates which are further classified with a class imbalance classifier. There are four main contributions in this work. Firstly, a new vessel removal algorithm is proposed based on the multi-scale log condition number computed from gradient vectors, which can remove most of the vessels and preserve most of the true MAs. Secondly, a candidate MAs localization algorithm is presented based on the second order directional derivatives in different directions, which can accurately localize the MAs’ centers. Thirdly, a new set of features are extracted for candidate MA classification, which can effectively discriminate the true MAs and the non-MAs. In addition, class-imbalance classification is introduced and analyzed in candidate MAs classification, which can effectively identify the true MAs from a large number of non-MAs. The remainder of this paper is organized as follows. The proposed candidate MAs extraction and classification are described in detail in Section Methods. The two public available databases and the evaluation metrics used in this work are introduced in Section Databases and Evaluation Metrics. The experimental results are provided and analyzed in Section Results and Analysis. Finally, this paper is concluded in Section Conclusions. Methods As shown in Fig 2, the proposed MAs detection consists of two main steps, i.e., candidate MAs extraction and classification. In the candidate MAs extraction step, the main interferences, i.e., retinal vessels, are first suppressed, and then the candidate MAs are localized and segmented. In the candidate MAs classification step, a set of discriminative features are extracted and a class-imbalanced classifier is trained and used to identify the true MAs from amount of non-MAs. 10.1371/journal.pone.0161556.g002Fig 2 Schema of the proposed method. The solid line path summarizes the training process of MAs detection, and the dashed line path summarizes the test process. Candidate MAs Extraction The proposed candidate MAs extraction includes vessel removal, candidate MAs localization and segmentation, and is described in detail as follows. As suggested in [18], the green channel of the color fundus image, where MAs have the highest contrast with the background, is used as the input image in this work. Before extracting candidate MAs, to smooth the image noises while preserving the boundary of MAs, an edge-preserving smoothing method [30] is first applied, which have been proved to be effective and can avoid introducing artifacts (e.g. ringing) that can deteriorate the performance of MAs detection. Next, a shade-correction method [26], which has been successfully employed in fluorescein image, is then used to reduce the uneven illumination of the smoothed input image. The final preprocessed image is denoted as Ip. Vessel Removal Because MAs are situated on capillaries and capillaries are not visible in color fundus images, they generally appear disconnected from the retinal vessel network [16]. Considering the vessels may affect MAs detection, we should remove vessels before extracting MA candidates. However, many existing vessel removal strategies may mistakenly remove some true MAs [18]. In this paper, we intend to remove vessels while preserving MAs with the gradient information. By observing the fundus images, we can find that retinal vessels always appear as piecewise linear structures, while the MAs generally appear as small round spots. Fig 3 shows the examples of these two structures and the distribution of their gradient vectors, where Fig 3A and 3C illustrate a linear structure and a round spot with their gradient vectors, and Fig 3B and 3D plot the distributions of the gradient vectors in Fig 3A and 3C. Let S denote a local support region, (xi, yi)(i = 1, …, N) denote all points in S, and (Δxi,Δyi) denote the gradient vector of the point (xi, yi). If S is a part of a vessel segment, which always can be regarded as a linear structure, the intensities always gradually increase from the centerline to the boundary along the perpendicular directions of this vessel (see Fig 3A), and the gradient vector directions (GVD) of most points on this vessel segment are perpendicular to the vessel segment (see Fig 3A). That is, the perpendicular direction of the vessel segment is the dominant principle direction of the distribution of (Δxi,Δyi) (see Fig 3B). While, if S located at the center of a true MA, which is always a blob-like structure, the intensities gradually increase from the center point to the boundary along the radial directions of this MA (see Fig 3C), and the GVD of the points in this MA are always similar with their corresponding radial directions. Since the radial directions of a blob can be arbitrary, no principle direction of the distribution of (Δxi,Δyi) is dominant (see Fig 3D). Therefore, we can determine if S is part of the retinal vessel or not by analyzing the principle direction of the distribution of (Δxi,Δyi). In this work, the covariance matrix of (Δxi,Δyi) are constructed and then the corresponding eigenvalues are computed to analyze the principle directions. For the vessel, there should be one dominant eigenvalue of the covariance matrix. While for the MAs, the two eigenvalues should be approximately equal. 10.1371/journal.pone.0161556.g003Fig 3 Different distributions of gradient vectors of the vessel-like and the MA-like objects. (A) The gradient field of a vessel-like object. (B) The distribution of gradient vectors in (A). (C) The gradient field of a MA-like object. (D) The distribution of gradient vectors in (C). More specifically, for each pixel (x, y) in the preprocessed image image Ip, a circular support region with radius r centered at (x, y) is denoted as Sr, and the covariance matrix of the gradient vectors in Sr is denoted as C(x, y, r). The eigenvalues of C(x, y, r) denotes as λ1 and λ2, and λ1 ≥ λ2. The above-mentioned eigenvalue relations now can be reflected by the ratio λ1/λ2, i.e., the condition number of C(x, y, r), which is denoted as κ(C(x, y, r)). The condition number is closer to 1, the object is more circular, while the higher the value, the more elongated the object. This number thus gives a high discriminability between vessel-like and MA-like objects. To deal with the size variation of both vessels and MAs in fundus image, we compute the condition number in a multi-scale manner by changing the radius of support region. A multi-scale log condition number K at (x, y) is defined by K(x,y)=ln∏r=rminrmaxκ(C(x,y,r))=∑r=rminrmaxln(κ(C(x,y,r))),(1) where the logarithm function is used to prevent overflow. For our experimental data, the radius of most vessels and MAs is both generally varied from 1 to 6 pixels in the image with the smallest image width (768 pixels). We thus choose rmin = 2ρ, rmax = 7ρ, where ρ = (image width)/768. Fig 4 shows the process of K map computation of the preprocessed image (Fig 4A). Fig 4B and 4C are the log condition number maps with respect to the support regions with different radius, and Fig 4D shows the final K map. As can be seen, most of the vessels, including the small and the large ones, have higher value of K. Fig 4E is a patch of the preprocessed image containing 5 true MAs marked with ‘□’, and Fig 4F is the corresponding patch of K map, from which we can see that all five true MAs of different size have lower responses than most of vessels. According to this figure, in K map, the vessel structures have been enhanced, while the MAs-like objects have been suppressed. We will separate vessels from MA-like objects according to the K map. 10.1371/journal.pone.0161556.g004Fig 4 The process of K map computation. (A) The preprocessed input image. (B)-(C) The log condition number maps with different support regions. (D) The final K map. (E) An image patch of preprocessed image contains 5 true MAs marked with ‘□’. (F) The corresponding patch of K map. A similar principle for distinguishing the vessels and MAs was presented in [31], where the production and mean of the eigenvalues of Hessian matrix was exploited. Compared with the covariance matrix of a region centered at a pixel used in this work, the Hessian matrix of a pixel is more sensitive to noises. Unlike the proposed covariance matrix based method, the Hessian matrix based approach in [31] cannot use the information of a pixel’s surrounding pixels, which are very important to judge the pixel as a MA or vessel pixel. The estimation of Hessian with multi-scale Gaussian smoothing in [31] also easily impair the extraction of faint MAs and MAs near vessels. Additionally, the production and the mean of eigenvalues may not effectively distinguish the faint MAs (low contrast and small variety of intensity) with two small eigenvalues and the vessels with one large eigenvalue and one small eigenvalue. From Fig 4E and 4F, since the MA marked with the yellow ‘□’ has an irregular shape, its response is slightly higher than the response of other MAs. And some vessel segments, e.g., the one indicated with the cyan arrow, also have lower K responses. Therefore it is difficult, if not impossible, to obtain the vessels map without MAs by directly applying threshold segmentation on K map. In this work, we remove vessels based on the morphological grayscale reconstruction algorithm [32] with the K map. Fig 5 shows the process of vessel removal. An empirical threshold is firstly applied to K map to get a binary image, denoted as Ibw, which includes the most vessel-like objects. The threshold value is set to 14.5 in this work according to the preliminary experiments on the training set. Then Ibw serves as a marker to reconstruct vessel map from the complement of the preprocessed image Ip, denoted as Iv (Fig 5A). Since the marker Ibw only contains the rough vessel structures, only the vessel map can be morphologically reconstructed. Since MAs are not connected with vessels and not appeared in the marker Ibw, MAs cannot be reconstructed. Thus, by algebraic subtracting Iv from Ip, most vessel structures are removed efficiently, while the MAs-like objects with a variety of appearance, intensity and size are preserved, as shown in Fig 5B. Fig 5C is an enlarged part of the marked region in Fig 5B, which is also the corresponding vessel removed result of Fig 4E with the K map in Fig 4F. As shown in this figure, the vessel segments with low value of K are also removed, while the MAs with irregular shape are preserved successfully. Fig 6 also shows some results of vessel removal, in which the MAs, including the subtle ones (Fig 6A) and the one near vessel (Fig 6C), are successfully preserved. 10.1371/journal.pone.0161556.g005Fig 5 The process of vessel removal. (A) The image morphologically reconstructed by using a binary K map. (B) The vessel removed image by subtracting (A) from Fig 4A. (C) The enlarged part of marked region in (B). 10.1371/journal.pone.0161556.g006Fig 6 Some results of vessel removal in special cases. (A) The image patch containing subtle MAs with different size. (B) The corresponding result after vessel removed. (C) The image patch containing a MA near vessel. (D) The corresponding result after vessel removed. In [17], the bottom-hat and hit-or-miss transformations were used to identify the MAs and vessels by considering the sizes of different structures in the fundus image. However, due to the size variety of both MAs and vessels, this approach may not work effectively. To extract candidate MAs, Séoud et al. [23] used the morphological gray reconstruction based on only the contrast information without considering the shape criterion, which may mistakenly extract many false MAs and miss some ture MAs. Candidate MAs Localization Besides the MAs, there also exists some tiny vessel segments, many dark background noises or other dark round objects in the vessel removed image. We need to localize the candidate MAs and, at the same time, suppress these non-MAs as many as possible. Considering that the intensities of some non-MAs may be similar to the intensities of some subtle MAs, it is not feasible to localize candidates only based on their intensities. In general, MAs exhibit a Gaussian like intensity distribution in all directions [9, 10, 21]. According to this prior, Pereira et al. [33] used the gradient patterns and Gaussian fitting parameter in different directions to exclude the false MAs. In this work, we compute the second derivatives of grayscale profiles in different directions for MA localization, which can preserve the true MAs and exclude the false ones as many as possible, and at the same time, obtain more accurate MAs’ positions. Since MAs centers have the minimum intensity along 1-D grayscale profile in different directions, the center of MAs will have the positive local maximum of second derivatives of the grayscale profiles in all directions. While for the other positions of MAs or the positions of some non-MAs, the second derivatives in some directions will decrease or even be close to or less than zero. Fig 7 shows the distribution of second derivatives at different positions with different direction varied from 0° to 360° (all negative second derivatives are set to zero), where Fig 7A and 7B shows the different positions of the same MA, and Fig 7E and 7F are the polar coordinate plots of the direction versus the value of second derivative in the direction at these positions; Fig 7C and 7D shows the positions centered at one subtle MA and one non-MAs with the similar intensity as the subtle MA, and Fig 7G and 7H are the corresponding polar coordinate plots. From this figure, we can see that only the centers of MAs have high positive value of second derivatives in all directions, while other positions have low value of second derivatives close or equal to zero in some directions. Based on this observation, we localize the centers of the candidate MAs in vessel removed image with the second derivatives in multiple directions, i.e., the second directional derivatives [34], which can be computed by using gradient vectors. 10.1371/journal.pone.0161556.g007Fig 7 The illustration of Iθ′′˜ at different position with different direction θ. (A)-(B) The different positions of the same MA. (C)-(D) The positions centered at one subtle MA and one non-MA. (E)-(H) are the polar coordinate plots of the direction θ versus the value of Iθ′′˜ in the direction θ at these positions. Given a discrete image I, for each point (x, y), a continuous surface is first fitted using a facet model [34, 35] over the intensity values in a local window with a size of 7 × 7 centered at that point. The partial derivatives of the discrete intensity surface at (x, y) are then approximated by the corresponding partial derivatives of the continuous surface at that point [35]. Let’s consider a direction θ and the corresponding unit vector u = [cosθ, sinθ]T. The gradient vector of I here is denoted by ∇I=[∂I∂x,∂I∂y]T, and the directional derivative of I at the point (x, y) in the direction θ, denoted by Iθ′(x,y), is given by Iθ′(x,y)=∂I(x,y)∂xcosθ+∂I(x,y)∂ysinθ=uT∇I(x,y).(2) The second directional derivative of I at the point (x, y) in the direction θ, denoted by Iθ′′(x,y), is then computed as Iθ′′=∂Iθ′∂xcosθ+∂Iθ′∂ysinθ=∂2I∂x2cos2θ+2∂2I∂x∂ycosθsinθ+∂2I∂y2sin2θ=uT∇Iθ′=uT∇(uT∇I).(3) For exploiting the intensity distribution of MAs efficiently, we analyzed the intensity distribution by computing the directional derivatives from 0° to 360°. Considering that a negative Iθ′ representing the intensity value along the direction θ is monotonically decreasing, which is not consistent with the intensity distribution along the MA center to the boundary in that direction, we set the negative Iθ′ value to zero in the computation of second order directional derivatives in direction θ, and define a modification of Iθ′′, denoted as Iθ′′˜, as follows: Iθ′′˜=∂⌊Iθ′⌋∂xcosθ+∂⌊Iθ′⌋∂ysinθ=uT∇⌊Iθ′⌋=uT∇(⌊uT∇I⌋),(4) where ⌊⋅⌋ denotes that the enclosed quantity is equal to itself when its value is positive, and zero otherwise. These multiple second directional derivatives in different direction θ are then integrated by the following equation: P=∏allθ⌊Iθ′′˜⌋=∏allθ⌊uT∇⌊Iθ′⌋⌋=∏allθ⌊uT∇(⌊uT∇I⌋)⌋,(5) Obviously, the center of MA-like objects can get much higher P values than other positions. In this work, we considered 36 directions from 0° to 360° with 10° step for P map computation. Finally, we find all local maxima on the P map, and consider those whose P values are greater than a threshold as the final locations of candidates. The threshold here is empirically chose as the 0.1 times of the maximal P value. Fig 8 shows the results of candidate MAs localization of Fig 4A, in which Fig 8A is the P map, and Fig 8B is the final location of candidate MAs, indicated with the green ‘×’. As shown in this figure, the proposed technique can localize almost all of the true MAs (marked with ‘□’) annotated by medical experts. It should be noted that the MAs clustered together have also been localized separately, as shown in the enlarged patch of Fig 8, which often be treated as one candidate by other algorithms. 10.1371/journal.pone.0161556.g008Fig 8 The result of candidate MAs localization. (A) The P map. (B) The final location of candidates MAs, where all candidates indicated with ‘×’ and the true MAs provided by medical expert marked with ‘□’. Candidate MAs Segmentation After localizing all the MA candidates, the whole regions of these candidates should be segmented for following candidate MAs classification. To be robust to the variability of intensity across the fundus image, we adopt a localized-based level set model [36] to segment the whole regions of these candidates, in which a localized based Chan-Vese energy [36] is applied to drive the evolution of the level set. The localized based Chan-Vese energy is formed by replacing global means (of interior and exterior regions) in the original Chan-Vese energy with the means of the local regions of each active point. To improve the efficiency of the segmentation, we compute and update the values of level set in abovementioned model based on the sparse field technique [37], with which an efficient representation of level set can be maintained. The sparse field technique uses the linked-lists of the active points of the zero level set and their neighbor points to efficiently represent the level set, in which only the varying active points and their neighbors are updated. Fig 9 shows some results of segmentation, in which candidates with different size and various local contrast are successfully segmented. 10.1371/journal.pone.0161556.g009Fig 9 Segmentation results of candidate MAs. (A), (C) and (E) The image patches containing MAs. (B), (D) and (F) The segmentation results. Candidate MAs Classification Feature Extraction To distinguish the true MAs and the non-MAs, this work extracts seven types of features for each candidate MA, i.e. geometric, contrast, intensity, edge, texture, region descriptors and other features. Excepting the region descriptors, the remaining extracted features (called common features) are commonly used to recognize the true MAs from the non-MAs based on the round shape and the color prior of true MAs, and most of them are adjusted from [10, 18, 22]. Region descriptors are introduced to exploit the local information of the candidate MA and its surrounding area for MAs classification. Geometric features: The ratio r1 between the minor and the major axis length of the candidate region Ω. The ratio r2 between the diameter of a circle with the same area as Ω and its major axis length. The area a, the circularity c, the eccentricity e and the compactness v of Ω [18]. Contrast features: The intensity difference ξ between the maximum intensity value of the inside pixels and the minimum intensity value of the outside pixels of Ω in the preprocessed image Ip and each channel of the RGB, LUV, and HSI color spaces of the original color fundus image Io. Notice that the outer region of Ω here is the region obtained by removing Ω from its morphological dilated version. Intensity features: The total intensity Σ of Ω, the normalized intensity ni and the normalized mean intensity nm in Ip and the green channel of Io [18]. The mean intensities μin, μout and their corresponding standard deviations σin, σout of the inside pixels and the outside pixels of Ω in each channel of the RGB, LUV, HSI color spaces of Io. Edge features: The mean value μe of the gradient magnitude of pixels on the boundary of Ω in Ip. Texture features: The mean value μg (μl) and the standard deviation σg (σl) of Ω in the Gaussian (LoG) filter responses of Ip with σ = 1, 2, 4, and 8 [10]. Region descriptors: The region descriptors, i.e. HOG [38], SURF [39] and GIST [40] descriptors, of the local image patch centered at Ω in Ip. Since the radius of the manually labeled MA mask is commonly between 5∼10 pixels, to include the information of both the candidate MA region and its surrounding region in the region descriptors, the patch size is set to 31 × 31 pixels (about a radius of 15 pixels). The HOG descriptor computes locally normalized histograms of gradient orientation in 5 × 5 grids for the local patch, in which each cell have 31 features. The SURF descriptor computes 8 features about Haar wavelet responses for each cell of the 4 × 4 grids of the patch. The GIST descriptor utilizes the Gabor filter responses in 3 scales and 8 orientations in 3 × 3 grids to provide a rough representation of the patch. The total amount of HOG, SURF and GIST features are 775, 128 and 216 respectively. Other features: The mean value μξ and the standard deviation σξ of the angle differences between the gradient vector and the unit vector along the different sampling directions within Ω computed in Ip [22]. The mean values μκ, μci and μdiv and the standard deviation σκ, σci and σdiv of the inside pixels and the outside pixels of Ω in the condition number map, the convergence index map [41] and the divergence map [42] computed in Ip. The mean value μii and the standard deviation σii of Ω in the isolated index map, which is the ratio of the mean intensity to the standard deviation in a ring region with width of 3 pixels outside a circular support region with radius of 7 computed in Ip. The mean value μwi of candidate pixels in the product image of condition number, convergence index, isolated index and divergence maps. In summary, the total feature set contains 1247 features. Since those region descriptors are seldom applied in candidate MAs classification, we will investigate their contributions for identifying MAs in Section Results and Analysis. MAs Recognition After extracting features for candidates, the true MAs should be identified in these candidates. However, as mentioned previously, in the extracted candidate MAs, the non-MAs are much more than the true MAs. In this work, since we intend to preserve more MAs in the extraction stage, the ratio is much higher, which can be close to 500:1, as shown in Section Introduction. To address this issue, a simple exclusion criterion is firstly used to remove some obvious false positives according to the discrimination table in [10] and our preliminary experiments, such as the candidates whose area a are out of the range from 2 to 150 or whose ratio r1 are less than 0.3. And then a class-imbalance classifier is trained to recognize the true MAs. The RUSBoost learning algorithm [43], which embedded the technique of random undersampling into the AdaBoost.M2 algorithm [44], has been previously shown to be very effective at alleviating the problem of class imbalance. Hence a RUSBoost classifier is trained to recognize the true MAs from a large number of non-MAs in this work. Given the minority training set R and the majority training set S, where |R| ≪ |S|, the RUSBoost randomly undersamples a subset S′ from S in each iteration of AdaBoost.M2, and construct the temporary class-balanced training set R ∪ S′ to train weak learners, where |S′| < |S|, and usually, |S′| = |R|. The final strong classifier H(x) is a weighted combination of T weak classifier ht (t = 1, 2, …, T), which are trained with class-balanced subset R∪St′ instead of R ∪ S in round t. Databases and Evaluation Metrics The proposed method is evaluated on two public database, i.e., the Retinopathy Online Challenge (ROC) competition [3] and DiaRetDB1 V2.1 database [29]. The ROC database The ROC database consists of 100 images with different resolutions [3], which are randomly split into a training and a test set, each containing 50 images. ROC only provide the ground truth of training set, which contains the location and the radius of each manually labeled MA. The proposed method is trained on the training set, and the ROC organizer evaluates it on the test set. The DiaRetDB1 V2.1 database The DiaRetDB1 V2.1 (denoted as DRDB) database contains 89 color fundus images with the fixed 1500 × 1152 resolution [29]. Among these fundus images, 28 images are given for training and the remaining 61 images are for testing. For each image in both the training and the test set, four ground truths annotated by different experts are provided, which include the location, the radius, and the confidence level of each marked MA. Evaluation Metrics In this work, a finding is defined as a “hit” of a manually labeled MA if the MA is the closest one to the finding and the distance between the center of the finding and the center of the MA is smaller than the provided radius. A finding is defined as a “hit-miss” if the distance between the center of the finding and the center of any manually labeled MA is larger than the provided radius. One marked MA may have multiple “hits”, but we only count a single “hit” of each MA as a true positive (TP), and count the other “hit” of the same MA as a false positive (FP), as defined in the ROC competition [3]. All “hit-miss” are also counted as the false positives. The sensitivity is then computed as #TPs/#Ttrue, where #TPs is the number of the true positives, and #Ttrue is the number of the manually labeled MAs. An average number of false positives per image (FPs/I) is also used to analyze the sensitivity in the evaluation. For fairly evaluating different methods, the ROC competition organizer do not provide the ground truth of the ROC test dataset to avoid training parameters on this dataset. Therefore, we cannot compare our output results with the ground truth images to get the sensitivity and specificity on this dataset. The detected MAs of different methods are submitted to the ROC organizer, and the organizer computes and returns the sensitivities at some levels of FPs/I and the competition performance metric (CPM, an average sensitivity at a set of particular false positives per image 1/8, 1/4, 1/2, 1, 2, 4 and 8 FPs/I) [3], which are used as the evaluation of different methods on the ROC test dataset. We thus evaluate the performance of the proposed method with these metrics on both ROC and DRDB database. Besides these metrics, we also evaluate our method with the free-response receiver operating characteristic (FROC) curve [3] and the partial area under the curve (AUC) of the FROC curve [25]. The FROC curve plots the sensitivity against the number of FPs/I. The partial AUC is the partial area under the FROC curve between 1/8 and 8 FPs/I. Results and Analysis Evaluation of Candidate MAs Extraction Table 1 compares the sensitivities of the proposed candidate extractor and the state-of-the-art candidate extractor algorithms [10, 16, 26–28] with the same level of FPs/I on the ROC training set, where our results are listed in the columns titled “Our Sen.”. The results of state-of-the-art approaches are reported in [25] and [45]. As Table 1 demonstrated, the proposed method notably improves the sensitivities of candidate MAs extraction. Although sensitivity is slightly lower than Lazar et al. [28] at the level of 73.94 FPs/I, our method achieves much higher sensitivity at the level of 569.39 FPs/I. Therefore, the proposed candidate extractor can characterize MAs better than other approaches. 10.1371/journal.pone.0161556.t001Table 1 Comparison of different candidate extractors on ROC training set. Method FPs/I Sen. Ours Sen. Spencer et al. [26] 20.30 0.12 0.29 Lazar et al. [28] 73.94 0.48 0.47 Walter et al. [16] 154.42 0.36 0.55 Zhang et al. [10] 328.30 0.33 0.62 Abdelazeem [27] 505.85 0.28 0.68 Lazar et al. [28] 569.39 0.598 0.691 After filtering out some obvious non-MAs from the candidate MAs, the remaining candidates form the final candidates set for the following classification. When a manually labeled MA has multiple “hits”, we cannot tell apart which one is true. To avoid losing the positive samples in the candidate MAs classification stage, we take all of the “hits” as positive ones (called positive candidates) and all of the “hit-miss” as negative ones (called negative candidates) when training the classifier. Table 2 lists the number of positive candidates (#PCs), the number of negative candidates (#NCs), and the ratio between #NCs and #PCs (Ratio; referred to the level of imbalance) of the final candidates sets extracted from the training set of different databases. In particular, we list all results of DRDB database with different ground truths in this table. Since the experts have different clinical experiences, the ground truths annotated based on their experiences in DRDB database vary largely, which result in the large variance of the statistic results in Table 2. Especially, for the ground truth annotated by the Expert 2, the marked areas are much larger than those marked by other experts, therefore much more candidates fall into these areas. It is the reason that the #PCs with Expert 2 are much larger and the Ratio is much smaller than those with other experts. According to this Table, the ratios between the negative candidates and the positive candidates are very high, the candidate MAs classification is thus an extremely class imbalanced problem in this work. 10.1371/journal.pone.0161556.t002Table 2 The statistical results of final candidate sets extracted from training set of different databases. Training set #Images #PCs #NCs Ratio ROC 50 284 123042 433:1 DRDB w/ Exp. 1 28 317 81734 258:1 DRDB w/ Exp. 2 28 4424 77627 18:1 DRDB w/ Exp. 3 28 162 81889 505:1 DRDB w/ Exp. 4 28 314 81737 260:1 Evaluation of Class Imbalanced Candidates Classification In our experiments, we use the decision trees as the weak learners in the RUSBoost model. To get the optimal number of weak learners T used in RUSBoost, we randomly split the ROC training set into two subsets, and perform 8 times two-folds cross validation for different T. Fig 10 shows the performance of each round of twofold cross validation on the ROC training set with the different number T in the RUSBoost. The classifier achieves a relatively stable sensitivity of 0.24 ± 0.01 when T is about 500 to 3000, and the overfitting occurs when T > 3000. We set T = 500 trees in this work to test the proposed method on both ROC and DRDB database. 10.1371/journal.pone.0161556.g010Fig 10 The sensitivities of cross validation on ROC training set with different T in RUSBoost classifier. To evaluate the performance of class imbalance classifier in MAs detection, we choose three ensemble classifiers, i.e., RUSBoost [43], EasyEnsemble [24] and AdaBoost [44], where the first two classifiers are especially designed for class-imbalance learning. We also test the kNN classifier, which has been successfully used for MAs classification in [14, 18, 46]. The performance of a Random Forest classifier is also tested, which has been used in [23] to recognize true MAs from class-imbalanced data. The FROC curves of these classifiers on ROC training set are plotted in Fig 11. Since the samples are extremely imbalanced, it is not surprised that the class imbalance classifiers notably outperform the other classifiers in MAs classification. Although the Random Forest has slightly improved the performance of kNN and AdaBoost for class-imbalanced problem, the RUSBoost, which integrated a under-sampling and boosting techniques, obtained the best result and improved the classification performance significantly. 10.1371/journal.pone.0161556.g011Fig 11 FROC curves produced by different classifiers. Evaluation of the Overall MAs Detection We next test the overall performance of MAs detection with the proposed method on both ROC and DRDB database. Table 3 lists the sensitivities at the predefined FPs/I, the ranked CPM evaluated by ROC organizer, and the partial AUC of 15 participating teams of ROC. The proposed method was ranked in the second place among these approaches with the score of 0.433, which is very close to the score of 0.434 of the first placed ensemble approach DRSCREEN [25]. And the proposed method obtains the highest partial AUC (0.553) among all of these approaches. The FROC of these approaches are plotted in Fig 12, from which, we can see that the proposed method performs very well especially after the level of 2 FPs/I. According to Table 3 and Fig 12, the proposed method can get a comparable performance with the DRSCREEN approach. 10.1371/journal.pone.0161556.t003Table 3 Quantitative results of the ROC competition for each participating team. Team Name 1/8 1/4 1/2 1 2 4 8 CPM AUC DRSCREEN [25] 0.173 0.275 0.380 0.444 0.526 0.599 0.643 0.434 0.551 Our method 0.219 0.257 0.338 0.429 0.528 0.598 0.662 0.433 0.553 Lazar [21] 0.251 0.312 0.350 0.417 0.472 0.542 0.615 0.423 0.510 Fegyver [22] 0.248 0.309 0.341 0.417 0.487 0.554 0.601 0.422 0.514 Niemeijer [18] 0.243 0.297 0.336 0.397 0.454 0.498 0.542 0.395 0.469 LaTIM [9] 0.166 0.230 0.318 0.385 0.434 0.534 0.598 0.381 0.489 ISMV [20] 0.217 0.270 0.366 0.407 0.440 0.459 0.468 0.375 0.435 OKmedical II [47] 0.175 0.242 0.297 0.370 0.437 0.493 0.569 0.369 0.465 Adal et al. [46] 0.204 0.255 0.297 0.364 0.417 0.478 0.532 0.364 0.446 OKmedical [10] 0.198 0.265 0.315 0.356 0.394 0.466 0.501 0.357 0.430 GIB Valladolid [48] 0.190 0.216 0.254 0.300 0.364 0.411 0.519 0.322 0.399 Fujita Lab [49] 0.181 0.224 0.259 0.289 0.347 0.402 0.466 0.310 0.378 IRIA Group [15] 0.041 0.160 0.192 0.242 0.321 0.397 0.493 0.264 0.368 Pereira et al. [33] 0.053 0.083 0.135 0.187 0.276 0.407 0.540 0.240 0.366 Waikato [50] 0.055 0.111 0.184 0.213 0.251 0.300 0.329 0.206 0.273 10.1371/journal.pone.0161556.g012Fig 12 The FROC curves for each of different approaches on the test set of the ROC competition. Recently, Soares et al. [23] reported a sensitivity of 0.47 at 37.8 FPs/I on the training set of ROC database, while our method achieved a sensitivity of 0.49 at the same FPs/I on the same set. DRDB database provides four ground truths annotated by different experts, respectively denoted as GT1, GT2, GT3 and GT4. Since there are disagreements among four experts’ annotations, we take a consensus of 75% agreement as the fusion ground truth, denoted as FGT. The proposed method is evaluated by using these ground truths. For comparison, a fused ground truth used by the DRSCREEN approach [25] (denoted as DFGT), provided by the authors, is also employed for evaluation. The FROCs of the proposed method with different ground truths and the FROC of DRSCREEN with DFGT, reproduced from the original work in [25] with the kind support from the authors, are plotted in Fig 13 and the sensitivities at the predefined FPs/I, the CPM scores and the partial AUC of the proposed method with different ground truths, and the results of DRSCREEN [25] with DFGT are listed in Table 4. According to this figure and table, the proposed method outperforms the DRSCREEN approach at all predefined FPs/I with all ground truths. Particularly, the CPM and partial AUC of the proposed method (0.180 and 0.270) are much higher than the those of DRSCREEN (0.070 and 0.130) with DFGT. The possible reason is that the small dark objects in DRDB database are more complicated, which contains not only MAs but also some small hemorrhages and scars left after PRP treatment. However, DRSCREEN only selects the optimal combination of several different preprocessing operations and several different candidate extractors, which are designed to extract small dark objects from fundus images, without further discriminating them. Therefore, DRSCREEN cannot discriminate these different small dark objects. While in this work, after extracting candidates, we have trained a classifier with a set of features for classifying these different small dark objects. Therefore, the proposed method can outperform the DRSCREEN approach [25] on DRDB database. 10.1371/journal.pone.0161556.g013Fig 13 FROC curves of the proposed method and the DRSCREEN approach on the DRDB database. The FROC curve reproduced from the original work of DRSCREEN [25] on the same database. 10.1371/journal.pone.0161556.t004Table 4 The results of the proposed method and the DRSCREEN approach on the DRDB database. Method 1/8 1/4 1/2 1 2 4 8 CPM AUC Ours w/ DFGT 0.043 0.071 0.092 0.155 0.198 0.301 0.398 0.180 0.270 DRSCREEN 0.001 0.003 0.009 0.020 0.059 0.140 0.257 0.070 0.130 Ours w/ GT1 0.042 0.061 0.100 0.191 0.296 0.390 0.504 0.226 0.352 Ours w/ GT2 0.018 0.053 0.094 0.137 0.213 0.274 0.331 0.160 0.244 Ours w/ GT3 0.075 0.116 0.209 0.288 0.408 0.526 0.671 0.328 0.482 Ours w/ GT4 0.031 0.066 0.119 0.179 0.268 0.372 0.490 0.218 0.336 Ours w/ FGT 0.035 0.058 0.112 0.254 0.427 0.607 0.755 0.321 0.527 Please note that the results of the proposed method evaluated with GT2 are much worse than those of the other ground truths (See Table 4). The reason is that the regions annotated by Expert 2 are too large and hence the label information of samples are too ambiguous, which result in much noise samples in classification and degenerate the performance. Zhang et al. [10] reported an average sensitivities of 0.713 at three levels of FPs/I (i.e. 1/2, 1, 2) on the DRDB database with the FGT. It is much higher than the value of 0.264 of our method at those levels, but that approach only used 11 images randomly selected from the database for both training and testing purpose. Noted that our method was trained with the training set of 28 images and tested on the whole test set of 61 images. Recently, Adal et al. [46] achieved a sensitivity of 0.6462 at 10 FPs/I on the DRDB database with the FGT, while our method have achieved a higher sensitivity of 0.7753 at the same FPs/I. Rosas-Romero et al. [17] reported a sensitivity of 0.9232 at the specificity of 0.9387 on the DRDB database, which is tested on the both training and test set with a selected subset of 88 images. Noted that there is no description of the selection of the ground truth and the training set in [17]. We here trained our classifier with the training set of 28 images, and tested it on the subset of 88 images as used in [17] with the FGT. We achieved a sensitivity of 0.9337 at the same specificity. These improvements also demonstrates the effectiveness of our method in MAs detection. Evaluation of Region Descriptors in Classification To investigate the contribution of region descriptors for candidates classification, we test the performances of different combinations of region descriptors and the common features. Table 5 shows the classification results on ROC training set with different feature combinations, i.e., without (w/o), only, or with (w/) the HOG (H), SURF (S) and GIST (G) features. According to this table, the common features achieve a CPM score of 0.235, while only the region descriptors can also achieve a score of 0.206. Although the score of the region descriptors is inferior to that of the common features, the region descriptors can achieve higher sensitivities than the common ones at the low FPs/I of 1/8 and 1/4, and the feature sets combined common features and region descriptors have improved the overall performance of candidate classification. More specially, the combination included common features and all three region descriptors have achieved the highest CPM score of 0.264. It is also noteworthy that the improvement of the combinations with HOG features are more significant than those without HOG, such as the score of ‘w/ H’, ‘w/ H+S’ and ‘w/H+G’ is 0.261, 0.263 and 0.262 respectively. The possible reason is that the HOG features are extracted in dense overlapping grids, such that they may provide more supplementary information than other descriptors for common features in MAs classification. 10.1371/journal.pone.0161556.t005Table 5 Classification results of different feature combinations on ROC training set. Features 1/8 1/4 1/2 1 2 4 8 CPM w/o H+S+G 0.044 0.079 0.180 0.252 0.298 0.365 0.428 0.235 only H+S+G 0.078 0.107 0.139 0.226 0.240 0.296 0.356 0.206 w/ H 0.079 0.111 0.184 0.255 0.349 0.404 0.445 0.261 w/ S 0.043 0.076 0.174 0.245 0.338 0.399 0.438 0.245 w/ G 0.064 0.092 0.164 0.243 0.346 0.392 0.434 0.248 w/ H+S 0.093 0.121 0.181 0.259 0.349 0.397 0.440 0.263 w/ H+G 0.087 0.110 0.170 0.264 0.353 0.405 0.444 0.262 w/ S+G 0.087 0.088 0.154 0.278 0.345 0.398 0.435 0.255 w/ H+S+G 0.088 0.122 0.189 0.287 0.343 0.370 0.450 0.264 Analysis There are some errors of MAs detection with the proposed method in the experiments. Some true MAs with low contrast are not detected, as shown in Fig 14A and 14B. It is mainly because that the images in the databases are provided with a compressed format, by which these subtle MAs are heavily blurred and cannot be found out even by the naked eyes. 10.1371/journal.pone.0161556.g014Fig 14 Examples of error detection. (A)-(B) The true MAs with low contrast. (C) The vessel crossing with high contrast. (D)-(E) The non-MAs have very similar appearance as true MAs. Some non-MAs were mistakenly detected as MAs, as shown in Fig 14C–14E. Fig 14C is a vessel crossing remained after vessel removal which is misclassified as MAs in classification stage. This is due to that these non-MAs have similar distribution of gradient vectors with the true MAs and have been extracted as candidates at the candidate MAs extraction stage, and the extracted features cannot discriminate these false positives at the classification stage. Fig 14D and 14E lists some isolated spots which are mistakenly detected as MAs. These non-MAs appeared almost identical to the true MAs, which are challenging to be distinguished even by ophthalmologists. Conclusions This paper proposed a novel automated method for MAs detection in color fundus images, which contains two stages, i.e. candidate MAs extraction and classification. In the first stage, the vessels can be effectively removed and the candidate MAs can be accurately localized by analyzing the gradient vectors of the images, which demonstrates that the gradient vectors can reflect the characteristic of different objects on fundus images. Most of the true MAs can be effectively extracted in this stage. To classify the non-MAs and the true MAs, seven types of features, i.e. geometry, contrast, intensive, edge, texture, region descriptors and other features, are extracted in the second stage. These features, especially the region descriptors, can well characterize the true MAs and have high discriminability to the true and the non-MAs. Since the non-MAs and the true MAs are extremely class imbalanced in the extracted candidates, the class imbalanced classifiers, e.g. RUSBoost, can effective classify these candidates. The proposed method performs well on different databases. The authors are most grateful for the ROC competition and the DiaRetDB1 project. We would highly appreciate the kind supports of Dr. Meindert Niemeijer and Dr. Li Tang for competition results evaluation. Special thanks to Dr. Bálint Antal for providing the detailed evaluation information of their microaneurysms detection approach. ==== Refs References 1 Jelinek HF , Cree MJ . Automated image detection of retinal pathology . CRC Press ; 2010 . 2 Orient JM . Sapira’s art and science of bedside diagnosis . Lippincott Williams & Wilkins ; 2012 . 3 Niemeijer M , van Ginneken B , Cree M , Mizutani A , Quellec G , Sánchez C , et al Retinopathy online challenge: automatic detection of microaneurysms in digital color fundus photographs . IEEE Trans Med Imag . 2010 1 ;29 (1 ):185 –195 . 10.1109/TMI.2009.2033909 4 Browning DJ . Diabetic Retinopathy: Evidence-Based Management . Springer Science & Business Media ; 2010 . 5 Bernardes R , Nunes S , Pereira I , Torrent T , Rosa A , Coelho D , et al Computer-Assisted Microaneurysm Turnover in the Early Stages of Diabetic Retinopathy . Ophthalmologica . 2009 ;223 :284 –291 . 10.1159/000213638 19372722 6 Sjølie AK , Klein R , Porta M , Orchard T , Fuller J , Parving HH , et al Retinal microaneurysm count predicts progression and regression of diabetic retinopathy. Post-hoc results from the DIRECT Programme . Diabetic Med . 2011 ;28 (3 ):345 –351 . 21309844 7 Torrent-Solans T , Duarte L , Monteiro R , Almeida E , Bernardes R , Cunha-Vaz J . Red dots counting on digitalized fundus images of mild nonproliferative retinopathy in Diabetes type 2 . Invest Ophthalmol Vis Sci . 2004 ;45 (13 ):2985 .15326111 8 Zheng Y , Hijazi MHA , Coenen F . Automated “disease/no disease” grading of age-related macular degeneration by an image mining approach . Invest Ophthalmol Vis Sci . 2012 ;53 (13 ):8310 –8318 . 10.1167/iovs.12-9576 23150624 9 Quellec G , Lamard M , Josselin PM , Cazuguel G , Cochener B , Roux C . Optimal wavelet transform for the detection of microaneurysms in retina photographs . IEEE Trans Med Imag . 2008 9 ;27 (9 ):1230 –1241 . 10.1109/TMI.2008.920619 10 Zhang B , Wu X , You J , Li Q , Karray F . Detection of microaneurysms using multi-scale correlation coefficients . Pattern Recognit . 2010 ;43 (6 ):2237 –2248 . 10.1016/j.patcog.2009.12.017 11 Giancardo L, Mériaudeau F, Karnowski TP, Tobin KW, Li Y, Chaum E. Microaneurysms detection with the radon cliff operator in retinal fundus images. In: Proc. SPIE. vol. 7623; 2010. p. 76230U–76230U–8. 12 Hatanaka Y , Inoue T , Okumura S , Muramatsu C , Fujita H . Automated microaneurysm detection method based on double-ring filter and feature analysis in retinal fundus images . In: Proc. CBMS. IEEE ; 2012 p. 1 –4 . 13 Dubow M , Pinhas A , Shah N , Cooper RF , Gan A , Gentile RC , et al Classification of human retinal microaneurysms using adaptive optics scanning light ophthalmoscope fluorescein angiography . Investigative Ophthalmology & Visual Science . 2014 ;55 (3 ):1299 –309 . 10.1167/iovs.13-13122 24425852 14 Fleming AD , Philip S , Goatman KA , Olson JA , Sharp PF . Automated microaneurysm detection using local contrast normalization and local vessel detection . IEEE Trans Med Imag . 2006 9 ;25 (9 ):1223 –1232 . 10.1109/TMI.2006.879953 15 Ram K , Joshi G , Sivaswamy J . A successive clutter-rejection-based approach for early detection of diabetic retinopathy . IEEE Trans Biomed Eng . 2011 3 ;58 (3 ):664 –673 . 10.1109/TBME.2010.2096223 21134810 16 Walter T , Massin P , Erginay A , Ordonez R , Jeulin C , Klein J . Automatic detection of microaneurysms in color fundus images . Med Image Anal . 2007 ;11 (6 ):555 –566 . 10.1016/j.media.2007.05.001 17950655 17 Rosas-Romero R , Martínez-Carballido J , Hernández-Capistrán J , Uribe-Valencia LJ . A method to assist in the diagnosis of early diabetic retinopathy: Image processing applied to detection of microaneurysms in fundus images . Comput Med Imaging Graph . 2015 ;44 :41 –53 . 10.1016/j.compmedimag.2015.07.001 26245720 18 Niemeijer M , van Ginneken B , Staal J , Suttorp-Schulten M , Abràmoff M . Automatic detection of red lesions in digital color fundus photographs . IEEE Trans Med Imag . 2005 5 ;24 (5 ):584 –592 . 10.1109/TMI.2005.843738 19 Sinthanayothin C , Boyce JF , Williamson TH , Cook HL , Mensah E , Lal S , et al Automated detection of diabetic retinopathy on digital fundus images . Diabetic Med . 2002 ;19 (2 ):105 –112 . 10.1046/j.1464-5491.2002.00613.x 11874425 20 Giancardo L , Meriaudeau F , Karnowski TP , Li Y , Tobin KW , Chaum E . Microaneurysm detection with radon transform-based classification on retina images . In: Proc. EMBC. IEEE ; 2011 p. 5939 –5942 . 21 Lazar I , Hajdu A . Retinal microaneurysm detection through local rotating cross-section profile analysis . IEEE Trans Med Imag . 2013 2 ;32 (2 ):400 –407 . 10.1109/TMI.2012.2228665 22 Fegyver Z . A Gradient Based Microaneurysm Detector . Budapest University of Technology and Economics ; 2012 Available from: http://roc.healthcare.uiowa.edu/results/documentation/zfegyer.pdf 23 Séoud L , Faucon T , Hurtut T , Chelbi J , Cheriet F , Pierre Langlois JM . Automatic detection of microaneurysms and haemorrhages in fundus images using dynamic shape features . In: ISBI. IEEE ; 2014 p. 101 –104 . 24 Liu X , Wu J , Zhou Z . Exploratory Undersampling for Class-Imbalance Learning . IEEE Trans Syst Man Cy B . 2009 4 ;39 (2 ):539 –550 . 10.1109/TSMCB.2008.2007853 25 Antal B , Hajdu A . An ensemble-based system for microaneurysm detection and diabetic retinopathy grading . IEEE Trans Biomed Eng . 2012 6 ;59 (6 ):1720 –1726 . 10.1109/TBME.2012.2193126 22481810 26 Spencer T , Olson J , McHardy K , Sharp P , Forrester J . An image-processing strategy for the segmentation and quantification of microaneurysms in fluorescein angiograms of the ocular fundus . Comput Biomed Res . 1996 ;29 (4 ):284 –302 . 10.1006/cbmr.1996.0021 8812075 27 Abdelazeem S . Microaneurysm detection using vessels removal and circular hough transform . In: Proc. NRSC. IEEE ; 2002 p. 421 –426 . 28 Lazar I , Hajdu A . Microaneurysm detection in retinal images using a rotating cross-section based model . In: Proc. ISBI. IEEE ; 2011 p. 1405 –1409 . 29 Kauppi T , Kämäräinen JK , Lensu L , Kalesnykiene V , Sorri I , Uusitalo H , et al Constructing benchmark databases and protocols for medical image analysis: Diabetic retinopathy . Comput Math Methods Med . 2013 ;2013 :1 –15 . 10.1155/2013/368514 30 Farbman Z , Fattal R , Lischinski D , Szeliski R; ACM . Edge-preserving decompositions for multi-scale tone and detail manipulation . ACM Trans Graph . 2008 8 ;27 (3 ):67 :1 –67 :10. 10.1145/1360612.1360666 31 Soares I, Castelo-Branco M, Pinheiro A. Microaneurysms detection using a novel neighborhood analysis. In: Proceedings of OMIA. Iowa Research Online; 2014. p. 65–72. 32 Vincent L . Morphological grayscale reconstruction in image analysis: applications and efficient algorithms . IEEE Trans Image Process . 1993 4 ;2 (2 ):176 –201 . 10.1109/83.217222 18296207 33 Pereira C , Veiga D , Mahdjoub J , Guessoum Z , Gonçalves L , Ferreira M , et al Using a multi-agent system approach for microaneurysm detection in fundus images . Artif Intell Med . 2014 ;60 (3 ):179 –188 . 10.1016/j.artmed.2013.12.005 24423809 34 Haralick R . Digital step edges from zero crossing of second directional derivatives . IEEE Trans Pattern Anal Mach Intell . 1984 1 ;6 (1 ):58 –68 . 10.1109/TPAMI.1984.4767475 21869165 35 Tico M , Onnia V , Kuosmanen P . Fingerprint Image Enhancement Based on Second Directional Derivative of the Digital Image . EURASIP J Adv Sig Pr . 2002 ;2002 (10 ):327301 . 36 Lankton S , Tannenbaum A . Localizing region-based active contours . IEEE Trans Image Process . 2008 11 ;17 (11 ):2029 –2039 . 10.1109/TIP.2008.2004611 18854247 37 Whitaker RT . A level-set approach to 3D reconstruction from range data . Int J Comput Vis . 1998 ;29 (3 ):203 –231 . 10.1023/A:1008036829907 38 Felzenszwalb PF , Girshick RB , McAllester D , Ramanan D . Object Detection with Discriminatively Trained Part-Based Models . IEEE Trans Pattern Anal . 2010 9 ;32 (9 ):1627 –1645 . 10.1109/TPAMI.2009.167 39 Bay H , Tuytelaars T , Van Gool L . SURF: Speeded up robust features In: Proc. ECCV . vol. 3951 Springer ; 2006 p. 404 –417 . 40 Oliva A , Torralba A . Modeling the shape of the scene: A holistic representation of the spatial envelope . Int J Comput Vis . 2001 ;42 (3 ):145 –175 . 10.1023/A:1011139631724 41 Kobatake H , Hashimoto S . Convergence index filter for vector fields . IEEE Trans Image Process . 1999 8 ;8 (8 ):1029 –1038 . 10.1109/83.777084 18267518 42 Korn G , Korn T . Mathematical handbook for scientists and engineers: definitions, theorems, and formulas for reference and review . Courier Dover Publications ; 2000 . 43 Seiffert C , Khoshgoftaar T , Van Hulse J , Napolitano A . RUSBoost: A hybrid approach to alleviating class imbalance . IEEE Trans Syst Man Cy A . 2010 1 ;40 (1 ):185 –197 . 10.1109/TSMCA.2009.2029559 44 Freund Y , Schapire R . Experiments with a new boosting algorithm . In: Proc. ICML . vol. 96 ; 1996 p. 148 –156 . 45 Antal B , Hajdu A . Improving microaneurysm detection using an optimally selected subset of candidate extractors and preprocessing methods . Pattern Recognit . 2012 ;45 (1 ):264 –270 . 10.1016/j.patcog.2011.06.010 46 Adal K , Sidibé D , Ali S , Chaum M , Karnowski T , Mériaudeau F . Automated Detection of Microaneurysms Using Scale-Adapted Blob Analysis and Semi-Supervised Learning . Comput Meth Prog Bio . 2014 ;114 (1 ):1 –10 . 10.1016/j.cmpb.2013.12.009 47 Zhang B , Karray F , Li Q , Zhang L . Sparse Representation Classifier for microaneurysm detection and retinal blood vessel extraction . Inform Sciences . 2012 ;200 :78 –90 . 10.1016/j.ins.2012.03.003 48 Sánchez CI , Hornero R , Mayo A , García M . Mixture model-based clustering and logistic regression for automatic detection of microaneurysms in retinal images . In: Proc. SPIE . vol. 7260 ; 2009 p. 72601M–72601M –8 . 49 Mizutani A , Muramatsu C , Hatanaka Y , Suemori S , Hara T , Fujita H . Automated microaneurysm detection method based on double ring filter in retinal fundus images . In: Proc. SPIE . vol. 7260 ; 2009 p. 72601N–72601N –8 . 50 Cree M . The Waikato Microaneurysm Detector . University of Waikato ; 2008 Available from: http://roc.healthcare.uiowa.edu/results/documentation/waikato.pdf

PMC005xxxxxx/PMC5001639.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 27564546PONE-D-15-4333710.1371/journal.pone.0161621Research ArticleBiology and Life SciencesOrganismsAnimalsVertebratesAmniotesMammalsBovinesCattleBiology and Life SciencesAgricultureLivestockCattleBiology and Life SciencesOrganismsAnimalsVertebratesAmniotesMammalsRuminantsCattleBiology and Life SciencesDevelopmental BiologyFibrosisPhysical SciencesMathematicsProbability TheoryRandom VariablesCovarianceMedicine and Health SciencesParasitic DiseasesResearch and Analysis MethodsImmunologic TechniquesImmunoassaysEnzyme-Linked ImmunoassaysBiology and Life SciencesVeterinary ScienceVeterinary MedicineVeterinary DiagnosticsBiology and Life SciencesAgricultureAnimal ProductsMeatBiology and Life SciencesNutritionDietFoodMeatMedicine and Health SciencesNutritionDietFoodMeatBiology and Life SciencesAnatomyLiverBiliary SystemGallbladderMedicine and Health SciencesAnatomyLiverBiliary SystemGallbladderEvaluation of the Performance of Five Diagnostic Tests for Fasciola hepatica Infection in Naturally Infected Cattle Using a Bayesian No Gold Standard Approach Fasciola hepatica Diagnostic Test Evaluation in Cattle in the UKhttp://orcid.org/0000-0002-4591-8267Mazeri Stella 1¤*Sargison Neil 2Kelly Robert F. 3Bronsvoort Barend M. deC. 1Handel Ian 2 1 The Roslin Institute, University of Edinburgh, Edinburgh, United Kingdom 2 Royal (Dick) School of Veterinary Studies, University of Edinburgh, Edinburgh, United Kingdom 3 Farm Animal Clinical Sciences, School of Veterinary Medicine, University of Glasgow, Glasgow, United Kingdom Yu Xue-jie Editor University of Texas Medical Branch, UNITED STATES Competing Interests: The authors have the following interests: This study was funded by Scotbeef Limited and used samples from Scotbeef. There are no patents, products in development or marketed products based on the results presented in this paper to declare. This does not alter the authors’ adherence to all the PLOS ONE policies on sharing data and materials. Conceived and designed the experiments: SM NS RK BB IH. Performed the experiments: SM NS RK BB. Analyzed the data: SM IH. Contributed reagents/materials/analysis tools: SM NS BB IH. Wrote the paper: SM NS BB IH. ¤ Current address: The Roslin Institute, The University of Edinburgh, Easter Bush, Midlothian, EH25 9RG, Scotland, United Kingdom * E-mail: stellamazeri@gmail.com2016 26 8 2016 11 8 e01616211 10 2015 9 8 2016 © 2016 Mazeri et al2016Mazeri et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.The clinical and economic importance of fasciolosis has been recognised for centuries, yet diagnostic tests available for cattle are far from perfect. Test evaluation has mainly been carried out using gold standard approaches or under experimental settings, the limitations of which are well known. In this study, a Bayesian no gold standard approach was used to estimate the diagnostic sensitivity and specificity of five tests for fasciolosis in cattle. These included detailed liver necropsy including gall bladder egg count, faecal egg counting, a commercially available copro-antigen ELISA, an in-house serum excretory/secretory antibody ELISA and routine abattoir liver inspection. In total 619 cattle slaughtered at one of Scotland’s biggest abattoirs were sampled, during three sampling periods spanning summer 2013, winter 2014 and autumn 2014. Test sensitivities and specificities were estimated using an extension of the Hui Walter no gold standard model, where estimates were allowed to vary between seasons if tests were a priori believed to perform differently for any reason. The results of this analysis provide novel information on the performance of these tests in a naturally infected cattle population and at different times of the year where different levels of acute or chronic infection are expected. Accurate estimates of sensitivity and specificity will allow for routine abattoir liver inspection to be used as a tool for monitoring the epidemiology of F. hepatica as well as evaluating herd health planning. Furthermore, the results provide evidence to suggest that the copro-antigen ELISA does not cross-react with Calicophoron daubneyi rumen fluke parasites, while the serum antibody ELISA does. Scotbeef LimitedScotbeef Limited, http://www.scotbeef.com/. BB thanks the BBSRC for their support through the Institute Strategic Programme (BB/J004235/1). The funders have read and agreed to publish this material, but had no role in study design, data collection and analysis, or preparation of the manuscript. Data AvailabilityAll relevant data for the main model shown in the paper are within the paper and its Supporting Information files. Data on location or identification of the farms the animals originated from are confidential and therefore cannot be shared. Nevertheless they are not critical for the main analysis.Data Availability All relevant data for the main model shown in the paper are within the paper and its Supporting Information files. Data on location or identification of the farms the animals originated from are confidential and therefore cannot be shared. Nevertheless they are not critical for the main analysis. ==== Body Introduction Fasciolosis, first reported in 1379, has been recognised as a clinically and economically important disease for centuries [1]. The infection caused by trematode parasites of the genus Fasciola can infect many mammals including sheep, cattle, goats, deer and humans [2]. In cattle, fasciolosis mainly manifests in its chronic form, which can lead to weight loss, anaemia and hypoproteinaemia. Clinical signs are often mild and may present as loss of productivity, while in severe cases sub-mandibular oedema may be seen. Unlike sheep, cattle liver pathology includes bile duct calcification and gallbladder enlargement [2, 3]. Globally, the infection is estimated to cost the livestock industry €2.5 billion per year [4], while losses due to liver fluke have been estimated to range between €1100-2000 million per year in the European Union [5]. In the UK and other temperate regions, F. hepatica is the most common aetiological agent of fasciolosis [2]. F. hepatica has a complicated multi-host, highly climate dependent life cycle which takes typically between 18 and 30 weeks to be completed. The mud snail, Galba truncatula is the most common intermediate host of F. hepatica in Europe [3, 6]. Temperature and moisture levels play an important role in the parasite’s life cycle and it is generally accepted that average daily temperatures of more than 10°C and high moisture levels are required for both the egg development and the reproduction of the parasite within the snail [7]. This results in seasonal increases of the incidence of infection, which vary between years depending heavily on climatic conditions. The incidence of fasciolosis in the UK has been reported to have increased during the last decade and more importantly its distribution has changed. In the past, fasciolosis was most commonly seen in the wetter western regions of the country, while it is now evident that the disease has become endemic in the previously drier eastern regions [8, 9]. Reasons for the changing epidemiology of F. hepatica are thought to include climate change, increasing animal movements and development of triclabendazole resistance [10]. Unpredictable weather conditions and resistance to anthelmintic treatment make control strategies less straightforward to plan. This increases the need for appropriate use of diagnostic tests, which along with improved knowledge and consideration of their limitations, can enhance implementation of more effective management strategies. The development of tests for the correct diagnosis of the infection has been going on for years, yet no test developed so far has been shown to have adequately high sensitivity and specificity in the field setting. Research on performance of available diagnostic tests in cattle and especially the copro-antigen ELISA is far from complete. The faecal egg count test, is commonly used in practice but can only detect patent infections. The serum antibody ELISA has the limitation of providing information on exposure rather than current infection but can detect exposure even at pre-patent stages of infection [11]. On the other hand the copro-antigen ELISA, which detects F. hepatica excretory-secretory antigens in faeces, is reported to detect early stages of infection without the limitation of giving positive results due to past exposure [12, 13]. This test has been evaluated by different research groups with varying results in sheep, but little has been reported on its performance in cattle [14]. Furthermore, inspection of livers of cattle slaughtered in abattoirs across Europe for signs of liver fluke is mandatory according to Regulation (EC) No 854/2004. In a previous study in Switzerland, Rapsch et al. (2006) [15] estimated the sensitivity of abattoir liver inspection to be 63.2%. Such estimates are expected to vary between countries, hence it is important to be able to obtain estimates specific to each country. Lastly, detailed liver necropsy techniques including gall bladder egg counts are available for research purposes, but impractical and expensive for routine use. These are expected to be extremely sensitive, even though there is still a window of error in case of very early stage infections. Moreover they can provide information on the severity of infection according to the degree of damage, as well as the fluke burden. In this study we have used the above diagnostic tests on samples taken from Scotbeef, one of Scotland’s largest red meat abattoirs, receiving animals from all around Scotland, northern England and Northern Ireland in an attempt to improve our knowledge on the performance of these diagnostic tests in the UK setting. More precisely this analysis aims to estimate: i) the performance of meat inspection as a tool for diagnosis of F. hepatica infection; and ii) the performance of liver necropsy, serum antibody ELISA, the copro-antigen ELISA and faecal egg count diagnostic tests. Materials and Methods Abattoir Based Sampling Samples were collected from Scotbeef Limited, Scotland’s largest red meat abattoir during three sampling periods. Sampling period A (June-July 2013) will be referred to as “summer 2013”. Sampling period B (January-beginning of March 2014) will be referred to as “winter 2014”. Lastly, sampling period C, which took place between the end of August 2014 and October 2014 will be referred to as “autumn 2014”. Each period consisted of six sampling days, one per week and 32-36 animals were sampled each time. The day and number of animals sampled each day were constrained by logistics. We used systematic sampling, collecting samples from one cattle in every 10 slaughtered to allow time for processing and represent animals slaughtered during the whole day. Animals to be sampled were clearly labeled at the time of bleeding and labels were maintained at all sampling stages to ensure that the correct samples were taken. Samples included blood, faecal samples as well as whole livers and gall bladders from each animal. Whole livers and gall bladders were stored at 4°C and were analysed within 72 and 96 hours respectively. Blood samples were stored at 4°C for 24 hours before sera were obtained and stored at -20°C. 2g of faeces were stored at -20°C, while the rest was stored at 4°C for egg counting which took place within a week post sampling. Diagnostics Tests Necropsy a) Liver dissection Livers were laid out on a tray and incisions parallel to and approximately 1 cm apart from the meat inspector’s incisions were made. Grades from 0 to 3 (no, mild, moderate, severe) were given in terms of signs of fibrosis; 0—no signs of fibrosis, 1—mild focal fibrosis, 2—severe local fibrosis or mild generalised fibrosis, 3—severe local fibrosis with calcified biled ducts or severe generalised fibrosis. Fibrosis scores were assigned before slicing the liver further in order to mimic what a meat inspector would be able to see on the offal line in the abattoir. The liver was cut into 1-2 cm slices thick and each slice was squeezed in order to collect flukes present. The slices were then placed in a bucket containing hot water for approximately 30 minutes. Water contents were then poured through 200μm sieves and inspected to retrieve flukes. Each slice was squeezed so that fluke exited the bile ducts, rinsed with water flowing in the bucket and discarded. Water remaining in the bucket was poured through 200μm sieves and inspected to retrieve remaining flukes. Flukes were then counted and stored in formalin. The total number of flukes was based on the number of whole flukes plus the number of anterior or posterior fluke parts depending on which one was greater [16, 17]. b) Gall bladder egg count Gall bladder contents were sieved through a series of 250 and 150μm sieves and collected in a measuring flask. The content was allowed to sediment for 3 minutes, excess liquid was removed and the remaining liquid was agitated and poured into a narrow bottomed glass. Water was added to the flask and poured into the glass to ensure no eggs remained in the flask. This process was repeated and liquid was poured in a 15ml falcon tube and allowed to sediment for 3 minutes. The sediment was collected in a petri dish, one drop of 0.5% methylene blue was added and all the eggs on the plate were counted using a stereoscopic dissecting microscope [13]. An animal was classified as positive for liver necropsy when 1 or more parasites were found in the liver and/or 1 or more eggs were found in the gall bladder. Faecal egg count (FEC) The faecal sample was mixed using a spatula and 5g were weighed out in a measuring cylinder. Water was added up to the 40ml mark and contents were mixed using a stirring rod. Contents were sieved through a coffee strainer and collected in a 250ml beaker for removal of coarse faecal material. The contents were then sieved through a 150μm sieve, collected into a narrow bottomed glass and allowed to sediment for 3 minutes. Excess liquid was syringed off and sediment was transferred into a 15ml falcon tube and allowed to sediment for 3 minutes. Excess liquid was syringed off and the sediment was transferred onto a petri dish. One drop of 0.5% methylene blue was added and all the eggs on the plate were counted using a stereoscopic dissecting microscope [18]. A sample was classified as positive when 1 or more eggs were found in the sample. Copro antigen ELISA (cELISA) Faecal samples were tested for the presence of excretory-secretory antigens using the commercially available Fasciola hepatica antigen ELISA kit (Bio-X Diagnostics, Belgium). The test was performed following the manufacturer’s instructions [12] and results were expressed as the sample optical density (OD) as a percentage of the mean positive control OD. Percentpositive=SampleODMeanpositivecontrolOD*100 Samples were classified as positive or negative according to the cut-offs provided by the manufacturer for each batch. Serum antibody ELISA (sELISA) Serum samples were analysed using the excretory/secretory (ES) antibody ELISA developed by the Liverpool School of Tropical Medicine [11]. The procedure described by Salimi-Bejestani et al (2005) [11] was performed with the following modifications: 1:8000 monoclonal mouse anti-bovine IgG conjugate (AbD Serotec, Bio-Rad Laboratories Inc, Hertfordshire, UK) was used. A new positive control was used so the equation used for calculating the results was slightly varied to obtain comparable results to previous controls. The percent positive (PP) value was obtained by the quotient of the mean sample OD (based on two duplicates) divided by the mean positive control OD (four duplicates), which was then multiplied by 111 instead of 100 to account for the new positive control as suggested by the test developers at Liverpool (Prof. D. Williams, 2014, pers.comm., 1 Dec). Percentpositive=MeantestsampleODMeanpositivecontrolOD*111 Samples were classified as positive if they had a PP greater or equal to 10. Liver inspection by the Meat Hygiene Service (MHS) The final test included in this analysis is liver inspection carried out at the abattoir by the Meat Hygiene Service. According to the manual for official controls, liver inspection requirements include visual inspection, palpation and incision of the gastric surface of the liver [19]. Livers with signs of liver fluke related pathology then have to be condemned. At Scotbeef, since 2012, MHS decision regarding liver condemnation is recorded as ‘Active’, ‘Historic’ or ‘No fluke’. ‘Active’ is roughly defined as livers in which parasites were seen, while ‘Historic’ describes livers with liver fluke related pathology but no signs of current infection. Both ‘Active’ and ‘Historic’ livers have to be condemned. This is unlike most UK abattoirs and for the purposes of this analysis we will be using the standardised classification, considering ‘Active’ and ‘Historic’ livers as positive and ‘No fluke’ livers as negative. Forestomach inspection for presence of rumen fluke During the second and third sampling seasons of the study (autumn and winter 2014) forestomachs of sampled animals were inspected for the presence of rumen fluke parasites. This procedure was carried out as part of a separate study [20], but results will be used here to assess the copro-antigen and the serum-antibody ELISAs for cross-reactivity with rumen fluke. Statistical Analysis A. The No Gold Standard (NGS) estimation of diagnostic test performance NGS, introduced by Hui & Walter [21], is a latent class approach to the evaluation of diagnostic tests when a “gold standard” is not available. The Bayesian version incorporates prior knowledge by specifying prior distributions for test properties and prevalence. If no prior information is available, vague, uniform priors are set. Probabilities of all the possible combinations of test outcomes conditional on the unknown disease status are specified using the sensitivity (Se) and specificity (Sp) of each test and the prevalence (p) of each sub-population, in this case periods “summer 2013”, “winter 2014” and “autumn 2014” [15, 22]. Animals can be positive or negative for each of the five tests included in this analysis so there are 25 (i.e. 32) possible combinations of test results. Hence, for each sub-population the counts of animals (Oi) of each combination of test results, in this case 32 (S) combinations for the five tests (T), follow a multinomial distribution [23, 24]: Oi|Sej,Spj,pi∼Multinomial(Pri,ni)fori=1,2,…,Sandj=1,2,…,T where Pri is the probability of observing the ith combination of test results. Examples of how to specify two such probabilities are shown below: Probability of obtaining a positive result in all five tests Pr(T1+, T2+, T3+, T4+, T5+) = Se1Se2Se3Se4Se5pi + (1 − Sp1)(1 − Sp2)(1 − Sp3)(1 − Sp4)(1 − Sp5)(1 − pi) Probability of obtaining a positive result in the first four tests and a negative result in the fifth test Pr(T1+, T2+, T3+, T4+, T5−) = Se1Se2Se3Se4(1 − Se5)pi + (1 − Sp1)(1 − Sp2)(1 − Sp3)(1 − Sp4)Sp5(1 − pi) The ratio of acute versus chronic infection is expected to be different, according to the known lifecycle of the parasite, between the three different times of the year which may affect the sensitivities and/or specificities of certain tests. Therefore, different estimates for the sensitivities of FEC, the copro-antigen and the serum antibody ELISA tests were obtained for each season as well as the specificity of the serum antibody ELISA. This was done for two reasons. Firstly, as shown by Toft et al (2005) [23], if estimates vary between sub-populations the combined estimate will be biased towards the estimate supported by most data i.e the one from the sub-population with the highest prevalence. Secondly, this can provide information on which tests are more appropriate at different times of the year. Model Assumptions Tests are conditionally independent. In other words, the misclassification errors of each test are unrelated conditional on the true disease status of the animal. For example, the probability of a truly diseased animal testing positive in test 2 (sensitivity), is not altered by the result of test 1 [25, 26]. There are various models for accounting for conditional dependence. In this case we have used the model suggested by Vacek (1985) as described below [23, 25]. Ten models including covariance terms (γSe and γSp) for one combination of two tests at a time were specified in order to inspect the effect of adjusting for covariance for each test combination on the sensitivity and specificity estimates of all tests. For example: Probability of obtaining a positive result in all five tests accounting for covariance between tests 1 and 2 Pr(T1+, T2+, T3+, T4+, T5+) = (Se1Se2+γSe)Se3Se4Se5pi + ((1 − Sp1)(1 − Sp2)+γSp)(1 − Sp3)(1 − Sp4)(1 − Sp5)(1 − pi) Probability of obtaining a negative result in the first test and a positive result in all other tests accounting for covariance between tests 1 and 2 Pr(T1−, T2+, T3+, T4+, T5+) = ((1 − Se1)Se2 − γSe)Se3Se4Se5pi + (Sp1(1 − Sp2) − γSp)(1 − Sp3)(1 − Sp4)(1 − Sp5)(1 − pi) Probability of obtaining a negative result in the first two tests and a positive result in all other tests accounting for covariance between tests 1 and 2 Pr(T1+, T2+, T3+, T4+, T5+) = ((1 − Se1)(1 − Se2) + γSe)Se3Se4Se5pi + (Sp1Sp2+γSp)(1 − Sp3)(1 − Sp4)(1 − Sp5)(1 − pi) Test sensitivities and specificities are constant between populations. Prevalences vary between populations The original Hui & Walter model contained two tests and two populations. Assumptions 3 and 4 were there to ensure that there are enough degrees of freedom to ensure the model’s identifiability. As liver fluke infection levels vary throughout the year and between years, we were able to assume that the prevalence will vary between the three sampling seasons. Additionally, according to Toft et al (2005) when three or more tests are compared one population is enough [23] to have sufficient degrees of freedom. As this model is an adaptation of the original model, with three sub populations and five tests, we ensure that we have enough degrees of freedom to be able to allow the stated sensitivities and specificities to vary between sub-populations and to include covariance terms for one combination of tests at a time. MCMC diagnostics Markov chain Monte Carlo (MCMC) chain convergence was assessed by visual inspection of the three sample chains using trace and Gelman-Rubin diagnostic plots for each variable in the model [27]. A correlation matrix of each chain was plotted to check for high correlation between variables. Priors As a Bayesian framework is used in this analysis, prior distributions were specified for the prevalence of each sub-population, sensitivities and specificities of each test. Vague, uniform priors with an interval between 0 and 1 were used for the prevalence of each sub-population. p∼dbeta(1,1) Similarly for the sensitivities and most of the specificities of evaluated tests, a wide distribution with an interval between 0 and 1 was used to reflect the fact that there is scarce knowledge on the performance of most of these tests in a real life scenario. Se∼dbeta(2,1)Sp∼dbeta(2,1) Liver necropsy was the only test where the prior distribution given for the specificity was highly informative. As mentioned before an animal was classified as positive for liver necropsy when either at least one fluke was found in the liver and/or when at least one egg was seen in the bile sample. It is therefore very unlikely that an animal can be wrongly classified as positive as liver flukes are easily identifiable and no other eggs similar to Fasiola hepatica eggs are expected to be seen in the bile. In a previous study by Rapsch et al (2006) [15] a similar test was assigned a specificity of 1 for the reasons explained. In order to account for the possibility of egg sequestration in the gall bladder for up to three weeks post treatment [28] we chose the following prior distribution instead. Splivernecropsy∼dbeta(9,1) The analysis was repeated using priors dbeta(1,1) for the Se and Sp of all tests to assess the effect of priors. Priors for the covariance variables, γSe and γSp, were uniform distributions using the following maximum and minimum limits [22, 29]. (Se1-1)(1-Se2)≤γSe≤min(Se1,Se2)-Se1Se2(Sp1-1)(1-Sp2)≤γSp≤min(Sp1,Sp2)-Sp1Sp2 Model implementation The model was implemented in JAGS [30], a software which uses MCMC simulations to construct posterior distributions for the analysis of Bayesian hierarchical models. JAGS was run within R (Version 3.0.3) [31] using the rjags package [32]. The first 20,000 iterations were discarded as burn-in and the following 20,000 iterations were used to construct the posterior distributions. The model specification is included in “S1 R script”. R Package coda [27] was used to carry out MCMC diagnostics and package corrplot [33] was used to visualize the correlation matrix between variables. The results were plotted using ggplot2 [34]. A map showing the distribution of sampled animals was plotted using ggmap [35] and the map tiles were sourced from Stamen Design (using data by OpenStreetMap), which are freely available under CC BY 3.0 license. Positive and Negative Predictive Values Sensitivity and specificity estimates report diagnostic test validity however positive (PPV) and negative predictive values (NPV) are the appropriate measure for interpreting tests in a specific population. They are the probability that a test positive or negative animal is truly positive or negative respectively. This is more easily interpreted by both farmers and vets, but its value depends on the true prevalence of the disease in the population [36]. Based on the Bayes formula [37], presented below, one can estimate the predictive values using estimates for sensitivity (Se), specificitiy (Sp) and the true population prevalence (p) [38]. PPV=Se*p(Se*p)+(1-Sp)*(1-p)NPV=Sp*(1-p)(Sp*(1-p))+(1-Se)*p PPVs and NPVs of the MHS liver inspection and FECs were calculated using the Se and Sp estimated by the NGS model over a range of possible prevalences to demonstrate this. Results Descriptive statistics In total, 619 cattle were sampled, 207 during summer 2013, 204 during winter 2014 and 208 during autumn 2014. Cattle age ranged from 369 to 1121 days old (Fig 1) and cattle of a variety of breeds were sampled as shown in Fig 2. As Fig 3 shows, cattle sampled came from Scotland, northern England and Northern Ireland i.e the geographical distribution of the general population of cattle slaughtered at the abattoir was well represented. Samples from every cattle were tested with the five tests mentioned. 10.1371/journal.pone.0161621.g001Fig 1 Distribution of cattle age per period. The age of cattle sampled ranged from 369 to 1121, with a mean of 720 days old. 10.1371/journal.pone.0161621.g002Fig 2 Distribution of cattle breed per period. Cattle sampled were of a range of different breeds found in the UK. 175 cattle were Aberdeen Angus cross, 118 were Limousin cross, 73 were Charolais cross, 48 were Aberdeen Angus, 37 were Simmental, 33 Holstein Friesian, 26 Limousin, 18 British Blue cross, 12 Charolais and 85 were of other less common breeds. 10.1371/journal.pone.0161621.g003Fig 3 Geographical distribution of cattle sampled. Samples used in this study were taken from Scotbeef, one of Scotland’s largest red meat abattoirs, receiving animals from all around Scotland, northern England and Northern Ireland. Figure shows the distribution of cattle sampled i.e the geographical distribution of the general population of cattle slaughtered at the abattoir was well represented. The map was plotted using R package ggmap [35] using tiles sourced from Stamen Design (using data by OpenStreetMap). Diagnostic test results Table 1 shows the binary results of each test per sampling period. Fig 4 shows the distribution of parasite burden per fibrosis score as recorded during liver necropsy. Among livers where flukes were found, parasite burden ranged from 1 to 86 parasites, with a mean of 8.5 and a median of 4. As previously described a fibrosis score was assigned based on a presentation of the liver mimicking the one presented to the MHS. The colour of the points shows the decision taken by the MHS during liver inspection at the abattoir. Higher fibrosis scores appear to have higher parasite burden, but it is also important to note that livers with no signs of fibrosis, that were also not rejected at the abattoir were found to have parasites. Furthermore, many livers which were classified as “Historic” by the MHS (green) were found to have parasites. Lastly, livers classified as “Active” by the MHS (red) appear to be spread evenly among fibrosis scores 1 to 3, while there were a few livers with a fibrosis score 0 which were classified as “Active”. This might mean that what was presented at the liver necropsy was not always the same as what was seen by the MHS. 10.1371/journal.pone.0161621.t001Table 1 Proportions of test positives for each test and number of animals sampled. summer 2013 winter 2014 autumn 2014 Overall Number sampled 207 204 208 619 MHS inspection 0.32 0.29 0.25 0.29 Necropsy 0.39 0.33 0.23 0.32 cELISA 0.29 0.25 0.18 0.24 FEC 0.31 0.25 0.13 0.23 sELISA 0.35 0.36 0.37 0.36 10.1371/journal.pone.0161621.g004Fig 4 Distribution of parasite counts by fibrosis score and MHS classification. Figure shows the distribution of parasite burden per fibrosis score as recorded during liver necropsy. Among livers where flukes were found, parasite burden ranged from 1 to 86 parasites, with a mean of 8.5 and a median of 4. A fibrosis score was assigned based on a presentation of the liver mimicking the one presented to the MHS. The colour of the points shows the decision taken by the MHS during liver inspection at the abattoir. Estimates of diagnostic test sensitivity and specificity Fig 5 is a plot of mean estimates and 95% Bayesian Credible Interval (BCI) for each model parameter. The precise mean estimates and 95% BCIs for each variable are shown in Table 2. F. hepatica infection prevalence during summer 2013, winter 2014 and autumn 2014 sampling periods was estimated to be 0.38, 0.31 and 0.23 respectively. Liver necropsy was, as expected, a near perfect test with a sensitivity estimate of 0.99 and a specificity of 0.98. Liver inspection by the abattoir Meat Hygiene Service had a sensitivity estimate of 0.68 and a specificity of 0.88. The sensitivity estimates of the copro-antigen ELISA were allowed to vary between seasons, but were estimated as 0.77 for all three sampling seasons. cELISA was estimated to have a very high specificity of 0.99. The Faecal Egg Count sensitivity values varied greatly between sampling seasons and were estimated as 0.81, 0.77 and 0.58 respectively. The test was shown to be highly specific, 0.99. Lastly, both the sensitivity and the specificity of the serum antibody ELISA were allowed to vary between seasons. Sensitivity estimates varied between seasons with the mean sensitivity estimate being much higher during the winter sampling, 0.94, compared to 0.72 and 0.80 during the summer and autumn sampling periods respectively. Similarly the mean specificity estimate during the autumn sampling of 0.76 was comparatively lower than summer and winter estimates which were 0.87 and 0.89 respectively. The exact data used for this model can be found in “S1 Table”. 10.1371/journal.pone.0161621.g005Fig 5 Mean posterior estimates and 95% BCIs. Estimates of the prevalence (pi), sensitivity (Se) and specificity (Sp) for each period (summer 2013 (A), winter 2014 (B), autumn 2014 (C)). 10.1371/journal.pone.0161621.t002Table 2 Mean estimates and 95% BCIs of the prevalence and test sensitivity and specificity per period. Estimate (Season) Mean 2.5% BCI 97.5% BCI Estimate (Season) Mean 2.5% BCI 97.5% BCI Prevalences Summer 2013 (A) 0.38 0.31 0.45 Winter 2014 (B) 0.31 0.25 0.38 Autumn 2014 (C) 0.23 0.17 0.29 Sensitivities Specificities MHS inspection 0.68 0.61 0.75 MHS inspection 0.88 0.85 0.91 Necropsy 0.99 0.96 1 Necropsy 0.98 0.96 0.99 cELISA (A) 0.77 0.67 0.86 cELISA 0.99 0.98 1 cELISA (B) 0.77 0.67 0.87 cELISA (C) 0.77 0.64 0.88 FEC (A) 0.81 0.72 0.9 FEC 0.99 0.98 1 FEC (B) 0.77 0.66 0.86 FEC (C) 0.58 0.43 0.72 sELISA (A) 0.72 0.62 0.82 sELISA (A) 0.87 0.8 0.92 sELISA (B) 0.94 0.86 0.98 sELISA (B) 0.89 0.84 0.94 sELISA (C) 0.8 0.69 0.91 sELISA (C) 0.76 0.69 0.82 Model checking Supporting information contain figures to demonstrate the results of checking for conditional dependence, the effect of priors and correlation between model variables, respectively. As shown in “S1 Fig”, there are no major differences in estimates when accounting for covariance for the different combinations of tests and the model with no covariance terms. It was therefore justifiable to use a final model with no covariance terms. Furthermore, “S2 Fig”, show a comparison of prior and posterior distribution which reveals that results are mainly informed by the data. This is further supported by “S3 Fig”, which presents a comparison between the results presented in the paper and the results of the same model run using non-informative priors for the sensitivities and specificities of all tests, where results do not appear to be altered. Lastly, “S4 Fig” presents the cross correlation plots between the parameters included in the model showing that there is no obvious strong correlation between any combination of parameters. Predictive Values of diagnostic tests Fig 6 show the positive and negative predictive values of the MHS liver inspection and Feacal egg counts respectively, over a range of prevalences. Estimates for FEC sensitivity was allowed to vary over the 3 sampling seasons hence 3 plots are presented. Prevalence estimates of the 3 sampling periods are shown by dotted lines. It is important to note how predictive values change according to the population prevalence. Additionally, when the PPV values of the two tests are compared at low prevalence levels it is clear that PPV of FEC is higher and varies less than the PPV of MHS due to a much higher specificity estimate for FEC. 10.1371/journal.pone.0161621.g006Fig 6 Predictive values of a) FEC and b) MHS over a range of prevalences. Prevalence estimates for each sampling period are shown by dotted lines. Discussion The main aims of this study was the evaluation of the performance of tests available for the diagnosis of F. hepatica. The no gold standard approach introduced by Hui & Walter [21] was used within a Bayesian framework in order to compare the binary results of the five diagnostic tests. Estimates of sensitivity and specificity for liver necropsy were 0.99 (95% BCI 0.96-1.00) and 0.98 (95% BCI 0.96-0.99) respectively. Liver necropsy is not readily used for disease diagnosis by veterinarians as it is a very time consuming procedure and it can only be carried out post mortem. Its role in this study was to provide a measure of infection and fibrosis levels to better describe the sample. Additionally, as a test previously used as a gold standard in assessments of F. hepatica diagnostic tests [39] it was expected to provide near perfect results and therefore be highly informative. A gold standard analysis was not chosen due to the possibility of gall bladder egg sequestration in animals where infections has been successfully treated causing false positive results and very early infections being difficult to detect due to the small size of flukes causing false negative results. The results of this study show that liver necropsy has a near perfect sensitivity and a very high specificity and must have contributed greatly in the evaluation of the rest of the tests by our model. Liver inspection is routinely carried out at the abattoir according to Regulation (EC) No 854/2004. The only previously reported estimate of its sensitivity in a European setting identified by the author was from a study in Switzerland by Rapsch et al. (2006) [15] which was 63.2%. In the current study the sensitivity estimate of liver inspection appeared to be lower than all other diagnostic tests, except that of FEC during Autumn 2014. Similarly, specificity appeared to be similar to the serum antibody ELISA, but lower than all other tests. More precisely the sensitivity was estimated to be 0.68 (95% BCI 0.61-0.75) and the specificity 0.88 (95% BCI 0.85-0.91). Estimates for meat inspection are expected to vary between countries and potentially between abattoirs. It is therefore relevant to report estimates for liver inspection from one of the biggest abattoirs in Scotland as this can provide a way to more accurately estimate the prevalence of F. hepatica infection in the UK accounting for imperfectness of this technique. Additionally, liver inspection can provide a useful and practical tool for evaluation of the effectiveness of health planning programmes used on farms. In this setting it is possibly more intuitive to use positive and negative predictive values, which can readily be estimated based on population prevalence as shown in the results section. Mezo et al. (2004) presented a new copro-antigen ELISA which was reported to have a sensitivity of 100% in detecting cattle with fluke burden of two or more parasites and be highly specific with no cross reactivity with parasites including Moniezia, Dicrocoelium, Echinococcus and Paramphistomum cervi [12, 14]. This ELISA is commercially available by Bio-X Diagnostics in Belgium. The protocol used in the commercial test is a considerable modification of the original, and its performance in the field setting has been poorly assessed, especially in cattle. In this study sensitivity estimates of the copro-antigen ELISA were allowed to vary between seasons, but were in fact very similar. They were estimated to be 0.77 (95% BCI 0.67-0.86), 0.77 (95% BCI 0.67-0.87), 0.77 (95% BCI 0.64-0.88) during summer 2013, winter 2014 and autumn 2014 sampling periods respectively. These estimates were considerably lower compared to Charlier et al. (2008) [39] who reported a sensitivity of 94%. This might be because liver necropsy without detection of eggs in the gall bladder was used as the gold standard, potentially missing a proportion of infected animals and therefore overestimating the sensitivity. Additionally, a lower cut-off than the one recommended in the protocol was used which might increase the sensitivity. Our estimate was similar to that of Palmer et al (2014) [40] who estimated the sensitivity to be 0.80 when the cut-off recommended by the manufacturer was used. When using a lower cut-off Palmer et al (2014) estimated the sensitivity to be 87%. The specificitiy of copro-antigen ELISA was estimated to be 0.99 (95% BCI 0.98-1.00). This is comparable to Palmer et al (2014) who estimated the specificity to be 1 using the manufacturer’s cut off and >99% using their own cut off [40]. On the contrary, Charlier et al (2008) estimated the specificity to be 93%. This might be a result of their cut-off adaptation. As the cut off adjustment used by Palmer et al. (manufacturer’s cut off multiplied by 0.67) provided greater improvement in the test performance, the model was rerun using the modified cut-off for the cELISA. Sensitivity was estimated as 0.80 (95% BCI 0.71-0.89), 0.85 (95% BCI 0.75-0.93), 0.87 (95% BCI 0.76-0.95) during summer 2013, winter 2014 and autumn 2014 sampling periods respectively. The specificity remained 0.99 (95% BCI 0.98-1.00) confirming that this cut-off modification can improve test sensitivity without compromising specificity. Estimates regarding the other four tests were not altered (results not shown). Gordon et al. (2013) [41] identified rumen fluke from a range of cattle and sheep samples across the UK to be Calicophoron daubneyi instead of P. cervi which was previously thought to be the species found in the UK. Even though lack of cross-reactivity with P. cervi has already been reported [42], this emphasises that it is important to also check for cross-reactivity of cELISA with C. daubneyi. In our study, during the second and third sampling seasons 53 cattle with negative liver necropsy results were found to have at least one fluke in the rumen. None of those samples had a positive copro-antigen result (using both manufacturer’s and adjusted cut-off). Rumen flukes collected have not been speciated, but based on the findings of Gordon et al. (2013) it is reasonable to assume that a great proportion of those were C. daubneyi. This suggests that cELISA does not cross react with this parasite in cattle, which agrees with the results of a similar comparison with cELISA in sheep [41]. This is becoming increasingly important in the UK as levels of rumen fluke infection appear to be rising and will further complicate fasciolosis control. Diagnosis of F. hepatica infection by detection of eggs in faecal samples has been around for decades and various protocols exist. The main drawbacks, of this otherwise easy to learn method, are that by definition it can only diagnose patent infections and that it is time consuming and therefore costly or undercharged. It is generally accepted that the specificity of faecal egg counting is almost perfect. In the UK this might be compromised by the increasing levels of rumen fluke infection as the eggs are of similar shape [41], even though the trained eye should be able to discriminate between the two kinds of eggs as they are of different colour. As vets and technicians become more aware of the increasing chance of finding rumen fluke eggs in faeces this problem is expected to be reduced. On the other hand, the sensitivity of the test has been reported to vary from well below 50% to moderate values and depends on various factors mainly based on the protocol used, for example volume of faeces [15, 39] and levels of infection in the population [43]. In the current context FEC sensitivity was estimated to be 0.81 (95% BCI 0.72-0.90), 0.77 (95% BCI 0.66-0.86) and 0.58 (95% BCI 0.43-0.72) during summer 2013, winter 2014 and autumn 2014 respectively. As expected the specificity was close to perfect and comparable to the copro-antigen ELISA (0.99, 95%, BCI 0.98-1.00). The sensitivity of FEC was shown to be comparable to cELISA during the first two sampling seasons, while it dropped significantly during autumn 2014. This shows that FEC still remains a very useful test during periods where infection is expected to be mainly chronic, and even superior to antibody ELISA tests as it has a higher specificity. As shown here it is important to remember that when recent infections are expected, for instance at the start of a new liver fluke season, this test performs a lot worse than other tests due to its inability to detect pre-patent infections. The last test evaluated in this study was the excretory/secretory antibody ELISA developed by the Liverpool School of Tropical Medicine [11]. This is the only test included that is developed to also detect past exposure to the parasite. Therefore, both the sensitivity and the specificity were allowed to vary between seasons. Sensitivity appeared to be much higher during the winter sampling, 0.94 (95% BCI 0.86-0.98) when compared to 0.72 (95% BCI 0.62-0.82) and 0.80 (95% BCI 0.69-0.91) during the summer and autumn sampling periods respectively. It was particularly interesting to see whether the false positive rates differed as well. Indeed, specificity during the autumn sampling was estimated to be 0.76 (95% BCI 0.69-0.82), which was comparatively lower than summer and winter estimates of 0.87 (95% BCI 0.80-0.92) and 0.89 (95% BCI 0.84-0.94) respectively. Serum antibody ELISA tests for the diagnosis of F. hepatica have been around for decades and have various reported sensitivities and specificities ranging from 91.7% to 100% and 94.6% to 100% respectively [44]. The ELISA used in this study is not commercially available and was first presented by Salimi-Bejestani et al. in 2005 with a sensitivity of 98% and a specificity of 96%. For their test evaluation they used FEC positive cattle, while their negative samples came from zero-grazed cattle of no known previous exposure to the parasite. Our sensitivity estimates are much lower and this is thought to be because the test was evaluated using an abattoir random sample of a range of levels of infection, including ones not detectable by FEC. Similarly, our specificity estimates are lower than previously reported. This is believed to be a result of the inability of antibody ELISAs to distinguish between current and previous exposure as it is highly possible that our sample included animals that were previously infected with the parasite, but who have received treatment, unlike the sample used in the initial evaluation. It is therefore possible that our estimates reflect a more realistic evaluation of this test in the field. Another issue with serum antibody ELISA tests in general is cross-reactivity with other trematodes [11]. The current ELISA showed no cross-reaction with D. viviparus, N. helvetianus and O. ostertagi, while cross-reaction with rumen flukes has not been reported [11]. Out of the 53 cattle with rumen flukes identified in the rumen and a negative liver necropsy result, 18 had positive sELISA results. While we cannot know whether those were animals with previous exposure to F. hepatica this may be an indication of cross-reactivity which can be further supported by Ibarra et al (1998) who reported cross-reactivity of an ES antigen ELISA first described by Arriaga de Morilla et al in 1989 with Paramphistomum spp. [45]. In our study most animals had a low fluke burden with a mean and median of 8.5 and 4 respectively. Firbrosis scores appeared to be related to burden, but it is important to note that our results explain the limitations of liver inspection by the MHS reflected in its imperfect sensitivity and specificity estimates. Presence of parasites in the liver did not always correspond to obvious fibrosis signs at inspection. Additionally, it is unclear whether “active” or “historic” is a useful classification as many of the livers classified as historic were found to harbour at least one fluke. Even though knowing whether the infection is absent or present in an animal is highly important, one could argue that the level of infection present could also be important in the control of fasciolosis especially in cattle. As fasciolosis is a chronic disease in cattle causing mostly sub-clinical disease, it might be meaningful to farmers to know what the intensity of infection is and how that translates to production losses. This information might therefore be used to decide what treatment strategy if any they might decide to use [43]. Such an investigation was beyond the scope of this paper, but it is one definitely worth pursuing to investigate the use of available diagnostic tests in quantifying infection or level of production loss attributed to the infection for a more cost effective control of F. hepatica infection in cattle. The present study has several strengths and limitations. We have used systematically chosen samples from naturally infected animals slaughtered at one of Scotland’s biggest abattoirs, therefore obtaining a sample more representative of the field situation than if experimentally infected animals where used [46]. Whilst we were not able to use simple random sampling due to logistics, we believe that this sampling method enabled us to represent animals arriving at the abattoir during the whole day. Five different tests were used in order to enable us to run a no gold standard analysis, avoiding the limitations of using an imperfect test as a gold standard. This approach certainly does not come without biases. In order to determine whether our proposed model could reclaim tests parameters using the sample size available, test results were simulated for three sub-populations of animals, representing the three sampling periods, under a range of plausible diagnostic test sensitivities and specificities. The model was run using this data and was able to recover pre-determined estimates of diagnostic test sensitivities/specificities and prevalence with reasonable precision for each sampling period. Furthermore, we checked for conditional dependence between tests and carried out appropriate MCMC diagnostics. Moreover, this is the first study to provide information on the appropriateness of available diagnostic tests during three different seasons, even though a first attempt at this was carried out by Charlier et al 2008 [39] using two sampling seasons, a much smaller dataset and a gold standard analysis. Limitations of this study include the fact that we have not been able to account for differences in meat inspection results depending on which meat inspector carried out the inspection, as well as the fact that seasonal differences were described only during one year. If results of tests are dependent on the liver fluke life cycle, which in turn is heavily dependent on climatic factors, the appropriateness of diagnostic tests in each season might need to be tested over more years to confirm the differences or similarities described here. Additionally, we have assessed the assumption of conditional independence using pairwise dependency models. It is important to bear in mind that it is possible that more complicated dependencies might exist, which we were unable to account for. Nevertheless, due to the fact that the tests compared are looking for five different signals; flukes, fluke damage, eggs, faecal antigens and serum antibodies, it is unlikely that there are biologically likely common proxies of disease that might result in important covariance structures. This is supported by the absence of any considerable change in estimates using the 10 possible covariance pairs. Overall, our study has provided a valuable insight in the performance of tests available for the diagnosis of F. hepatica infections in a population of cattle believed to be representative of the field situation. Knowing its limitations and being able to adjust for them, abattoir liver inspection, can be a valuable tool in monitoring and understanding the changing epidemiology of F. hepatica as well as evaluating farm health plans. Faecal egg counting has been shown to still be a valuable tool in the diagnosis of current F. hepatica infections, but one has to bear in mind that it is a weak test during periods where recent infections are expected. The copro-antigen ELISA is a comparable test that can be used throughout the year, with evidence to suggest that there is no cross-reaction with the increasingly prevalent rumen fluke parasite. This study also provided further evaluation of an in house ES antigen ELISA showing that while being a valuable test, its sensitivity and specificity estimates are lower in the field setting that previously reported. Liver fluke control is becoming increasing challenging in the UK, hence the qualitative and quantitative evaluation of available diagnostic tests, as well as development of better ones is an area where ongoing investigation is required. Supporting Information S1 R Script Model specification. This script contains the code used for comparison of 5 diagnostic tests during 3 sampling periods. This model is an adaptation of the Hui & Walter [21] approach for the evaluation of diagnostic tests when a “gold standard” is not available. (R) Click here for additional data file. S1 Table Data used in Bayesian no gold standard model. Table shows the data used in the Bayesian no gold standard model. For each period there were 32 possible combinations of test results and the number of animals for each combination is shown here. A negative test result is shown by 0 and a positive test result is shown by 1. (PDF) Click here for additional data file. S1 Fig Conditional Dependence. Figure shows the mean estimates of sensitivity and specificity of each test as estimated by the 10 different models accounting for covariance of one combination of two tests at a time. For example S1S2 is the model including covariance terms for tests 1 and 2 i.e. MHS liver inspection and liver necropsy and so on. The last estimate (NoCov) as well as the horizontal line on each plot shows the mean as estimated by the model with no covariance terms. Plots such as Se4 containing 3 lines show Se or Sp estimates that were allowed to vary between season. Based on this figure we concluded that even though estimates vary slightly above or below the lines, there are no major differences in estimates when accounting for covariance for the different combinations of tests and the model with no covariance terms. It was therefore justifiable to use a final model with no covariance terms. (PDF) Click here for additional data file. S2 Fig Effect of priors. A comparison between prior and posterior distributions of model parameters is shown in these two figures. The top figure shows the mean and 95% Bayesian credibility intervals of each model parameter. Bayesian credibility intervals of posterior distributions are much narrower than the priors showing that results are heavily informed by the data. As described in the methodology the only informative prior was the one for the specificity of the liver necropsy, Sp2. This figure shows that even though the prior distribution is more informative the result is also informed by the data. Similarly the bottom figure shows the density plots of prior and posterior distributions and how prior distributions (except Sp2) are vague and posterior distributions are highly data driven being much narrower than the prior distributions. (PDF) Click here for additional data file. S3 Fig Comparison of results of original model and model using non-informative priors. Figure shows the results of the original model and of a model using non-informative priors for comparison. The analysis was repeated using priors dbeta(1,1) for the Se and Sp of all tests to assess the effect of priors. There is no obvious alterations of results. (PDF) Click here for additional data file. S4 Fig Correlation between model parameters. Figure shows the cross correlations between the parameters included in the model in each of the 3 MCMC chains. There is no obvious strong correlation between any combination of parameters. (PDF) Click here for additional data file. We acknowledge Scotbeef Limited who fully funded this project and allowed us to carry out the sampling and collect material. We acknowledge our colleagues who helped with sampling as well as the slaughterhouse workers and the Meat Hygiene Service for making this work possible. Lastly, we acknowledge the Veterinary Parasitology Research Group of Institute of Infection and Global Health of the University of Liverpool for providing serum-antibody ELISA training and Danielle Gordon for providing Faecal Egg Count and copro-antigen ELISA training. ==== Refs References 1 Rojo-Vázquez F , Meana A , Valcárcel F , Martínez-Valladares M . Update on trematode infections in sheep . Veterinary Parasitology . 2012 ;189 (1 ):15 –38 . 10.1016/j.vetpar.2012.03.029 22521973 2 Andrews SJ . The life cycle of Fasciola hepatica In: Dalton JP , editor. Fasciolosis . Oxon : CABI publishing ; 1999 p. 1 –30 . 3 Taylor MA , Coop LR , Wall LR . Veterinary Parasitology . 3rd ed Oxford : Blackwell Publishing ; 2007 . 4 European Commision, Directorate-General for Research and Innovation. A decade of EU-funded Animal Health. Luxembourg; 2012. 5 Novobilský A , Sollenberg S , Höglund J . Distribution of Fasciola hepatica in Swedish dairy cattle and associations with pasture management factors . Geospatial health . 2015 ;9 (2 ):293 –300 . 10.4081/gh.2015.351 25826310 6 Graczyk TK , Bernard F . Development of Fasciola hepatica in the Intermediate Host In: Dalton JP , editor. Fasciolosis . Oxon : CABI publishing ; 1999 p. 31 –46 . 7 Ollerenshaw CB , Rowlands WT . A method of forecasting the incidence of fasciolosis in Anglesey . The Veterinary record . 1959 ;71 (29 ):591 –598 . 8 Mitchell G . Update on fasciolosis in cattle and sheep . In Practice . 2002 ;24 (7 ):378 –385 . 10.1136/inpract.24.7.378 9 Sargison ND , Scott PR . Diagnosis and economic consequences of triclabendazole resistance in Fasciola hepatica in a sheep flock in south-east Scotland . The Veterinary record . 2011 ;168 (6 ):159 10.1136/vr.c5332 21493511 10 Gordon DK , Zadoks RN , Stevenson H , Sargison ND , Skuce PJ . On farm evaluation of the coproantigen ELISA and coproantigen reduction test in Scottish sheep naturally infected with Fasciola hepatica . Veterinary parasitology . 2012 ;187 (3-4 ):436 –44 . 10.1016/j.vetpar.2012.02.009 22421492 11 Salimi-Bejestani MR , McGarry JW , Felstead S , Ortiz P , Akca A , Williams DJL . Development of an antibody-detection ELISA for Fasciola hepatica and its evaluation against a commercially available test . Research in veterinary science . 2005 ;78 (2 ):177 –81 . 10.1016/j.rvsc.2004.08.005 15563926 12 Mezo M , González-Warleta M , Carro C , Ubeira FM , Gonza M . An ultrasensitive capture ELISA for detection of Fasciola hepatica coproantigens in sheep and cattle using a new monoclonal antibody (MM3) . Journal of Parasitology . 2004 ;90 (4 ):845 –852 . 10.1645/GE-192R 15357080 13 Flanagan AM , Edgar HWJ , Forster F , Gordon A , Hanna REB , McCoy M , et al Standardisation of a coproantigen reduction test (CRT) protocol for the diagnosis of resistance to triclabendazole in Fasciola hepatica . Veterinary parasitology . 2011 ;176 (1 ):34 –42 . 10.1016/j.vetpar.2010.10.037 21093156 14 Brockwell YM , Spithill TW , Anderson GR , Grillo V , Sangster NC . Comparative kinetics of serological and coproantigen ELISA and faecal egg count in cattle experimentally infected with Fasciola hepatica and following treatment with triclabendazole . Veterinary parasitology . 2013 ;196 (3-4 ):417 –426 . 10.1016/j.vetpar.2013.04.012 23643623 15 Rapsch C , Schweizer G , Grimm F , Kohler L , Bauer C , Deplazes P , et al Estimating the true prevalence of Fasciola hepatica in cattle slaughtered in Switzerland in the absence of an absolute diagnostic test . International journal for parasitology . 2006 ;36 (10-11 ):1153 –8 . 10.1016/j.ijpara.2006.06.001 16843470 16 Clery D , Torgerson P , Mulcahy G . Immune responses of chronically infected adult cattle to Fasciola hepatica . Veterinary parasitology . 1996 ;62 (1-2 ):71 –82 . 10.1016/0304-4017(95)00858-6 8638395 17 De Bont J , Claerebout E , Riveau G , Schacht AM , Smets K , Conder G , et al Failure of a recombinant Schistosoma bovis-derived glutathione S-transferase to protect cattle against experimental Fasciola hepatica infection . Veterinary Parasitology . 2003 ;113 (2 ):135 –44 . 10.1016/S0304-4017(02)00450-8 12695038 18 Happich FA , Boray JC . Quantitative diagnosis of chronic fasciolosis 1. Comparative Studies on Quantitative Faecal Examinations for Chronic . Australian Veterinary Journal . 1969 ;45 (7 ):326 –28 . 10.1111/j.1751-0813.1969.tb05009.x 5817298 19 Food Standards Agency. Post-Mortem, Health and Identification Marking. In: Manual for official controls; 2015. 20 Sargison N , Francis E , Davison C , deC Bronsvoort BM , Handel I , Mazeri S . Observations on the biology, epidemiology and economic relevance of rumen flukes (Paramphistomidae) in cattle kept in a temperate environment . Veterinary Parasitology . 2016 ;219 :7 –16 . 10.1016/j.vetpar.2016.01.010 26921033 21 Hui S , Walter S . Estimating the error rates of diagnostic tests . Biometrics . 1980 ;36 (1 ):167 –71 . 10.2307/2530508 7370371 22 Branscum AJ , Gardner IA , Johnson WO . Estimation of diagnostic test sensitivity and specificity through Bayesian modeling . Preventive veterinary medicine . 2005 ;68 (2-4 ):145 –63 . 10.1016/j.prevetmed.2004.12.005 15820113 23 Toft N , Jørgensen E , Højsgaard S . Diagnosing diagnostic tests: Evaluating the assumptions underlying the estimation of sensitivity and specificity in the absence of a gold standard . Preventive Veterinary Medicine . 2005 ;68 (1 ):19 –33 . 10.1016/j.prevetmed.2005.01.006 15795013 24 deC Bronsvoort BM , Koterwas B , Land F , Handel IG , Tucker J , Morgan KL , et al Comparison of a flow assay for brucellosis antibodies with the reference cELISA test in West African Bos indicus . PloS ONE . 2009 ;4 (4 ):e5221 10.1371/journal.pone.0005221 19381332 25 Vacek PM . The effect of conditional dependence on the evaluation of diagnostic tests . Biometrics . 1985 ;41 (4 ):959 –968 . 10.2307/2530967 3830260 26 Gardner IA , Stryhn H , Lind P , Collins MT . Conditional dependence between tests affects the diagnosis and surveillance of animal diseases . Preventive veterinary medicine . 2000 ;45 (1-2 ):107 –22 . 10.1016/S0167-5877(00)00119-7 10802336 27 Martyn P , Best N , Cowles K , Vines K . CODA: Convergence Diagnosis and Output Analysis for MCMC . R News . 2006 ;6 (1 ):7 –11 . 28 Chowaniec W , Darski J . Investigations on excretion time of liver fluke eggs after killing the parasite . Bulletin of the Veterinary Institute in Pulawy . 1970 ;14 (3-4 ):108 –110 . 29 Dendukuri N , Joseph L . Bayesian approaches to modeling the conditional dependence between multiple diagnostic tests . Biometrics . 2001 ;57 (1 ):158 –167 . 10.1111/j.0006-341X.2001.00158.x 11252592 30 Martyn P. JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. In: Proceedings of the 3rd International Workshop on Distributed Statistical Computing; 2003. 31 R Core Development team. R: A Language and Environment for Statistical Computing; 2014. Available from: http://CRAN.R-project.org/package=rjags 32 Martyn P. rjags: Bayesian graphical models using MCMC; 2014. 33 Wei T. corrplot: Visualization of a correlation matrix; 2013. Available from: http://CRAN.R-project.org/package=corrplot 34 Wickam H. ggplot2: elegant graphics for data analysis. Springer New York; 2009. Available from: http://had.co.nz/ggplot2/book 35 Kahle D , Wickham H . ggmap: Spatial Visualization with ggplot2 . The R Journal . 2013 ;5 (1 ):144 –161 . 36 Akobeng AK . Understanding diagnostic tests 1: sensitivity, specificity and predictive values . Acta paediatrica . 2007 ;96 (3 ):338 –341 . 10.1111/j.1651-2227.2006.00180.x 17407452 37 Price P , Bayes T . An Essay towards solving a problem in the doctrine of chances. by the Late Rev. Mr. Bayes, F. R. S. Communicated by Mr. Price, in a Letter to John Canton, A. M. F. R. S . Philosophical Transactions of the Royal Society . 1763 ;53 :370 –418 . 10.1098/rstl.1763.0053 38 Linn S . A New Conceptual Approach to Teaching the Interpretation of Clinical Tests . Journal of Statistics Education . 2004 ;12 (3 ):1 –9 . 39 Charlier J , De Meulemeester L , Claerebout E , Williams DJL , Vercruysse J . Qualitative and quantitative evaluation of coprological and serological techniques for the diagnosis of fasciolosis in cattle . Veterinary parasitology . 2008 ;153 (1-2 ):44 –51 . 10.1016/j.vetpar.2008.01.035 18329811 40 Palmer D , Lyon J , Palmer M , Forshaw D . Evaluation of a copro-antigen ELISA to detect Fasciola hepatica infection in sheep, cattle and horses . Australian Veterinary Journal . 2014 ;92 (9 ):357 –361 . 10.1111/avj.12224 25156056 41 Gordon DK , Roberts LCP , Lean N , Zadoks RN , Sargison ND , Skuce PJ . Identification of the rumen fluke, Calicophoron daubneyi, in GB livestock: possible implications for liver fluke diagnosis . Veterinary parasitology . 2013 ;195 (1-2 ):65 –71 . 10.1016/j.vetpar.2013.01.014 23411375 42 Kajugu PE , Hanna REB , Edgar HW , Forster FI , Malone FE , Brennan GP , et al Specificity of a coproantigen ELISA test for fasciolosis: lack of cross-reactivity with Paramphistomum cervi and Taenia hydatigena . Veterinary Record . 2012 ;171 (20 ):502 10.1136/vr.101041 23077134 43 Charlier J , Vercruysse J , Morgan E , van Dijk J , Williams DJL . Recent advances in the diagnosis, impact on production and prediction of Fasciola hepatica in cattle . Parasitology . 2014 ;141 (3 ):326 –35 . 10.1017/S0031182013001662 24229764 44 Alvarez Rojas CA , Jex AR , Gasser RB . Chapter Two Techniques for the Diagnosis of Fasciola Infections in Animals: Room for Improvement . Advances in Parasitology . 2014 ;85 :65 –107 . 10.1016/B978-0-12-800182-0.00002-7 24928180 45 Arriaga de Morilla C , Paniagua R , Ruiz-Navarrete A , Bautista C , Morilla A . Comparison of dot enzyme-linked immunosorbent assay (Dot-ELISA), passive haemagglutination test (PHT) and thin layer immunoassay (TIA) in the diagnosis of natural or experimental Fasciola hepatica infections in sheep . Veterinary parasitology . 1989 ;30 (3 ):197 –203 . 10.1016/0304-4017(89)90015-0 2705286 46 Salimi-Bejestani MR , Cripps P , Williams DJL . Evaluation of an ELISA to assess the intensity of Fasciola hepatica infection in cattle . The Veterinary record . 2008 ;162 (4 ):109 –111 . 10.1136/vr.162.4.109 18223266

PMC005xxxxxx/PMC5001640.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756406110.1371/journal.pone.0161954PONE-D-16-18397Research ArticleBiology and Life SciencesAnatomyHeadEarsInner EarCochleaMedicine and Health SciencesAnatomyHeadEarsInner EarCochleaBiology and Life SciencesGeneticsGene ExpressionBiology and Life SciencesAnatomyRenal SystemKidneysMedicine and Health SciencesAnatomyRenal SystemKidneysMedicine and Health SciencesPharmacologyDrugsCannabinoidsResearch and Analysis MethodsImmunologic TechniquesImmunoassaysImmunofluorescenceBiology and Life SciencesImmunologyImmune System ProteinsImmune ReceptorsMedicine and Health SciencesImmunologyImmune System ProteinsImmune ReceptorsBiology and Life SciencesBiochemistryProteinsImmune System ProteinsImmune ReceptorsBiology and Life SciencesCell BiologySignal TransductionImmune ReceptorsResearch and Analysis MethodsHistochemistry and Cytochemistry TechniquesImmunohistochemistry TechniquesResearch and Analysis MethodsImmunologic TechniquesImmunohistochemistry TechniquesBiology and Life SciencesPhysiologyElectrophysiologyMembrane PotentialReceptor PotentialsMedicine and Health SciencesPhysiologyElectrophysiologyMembrane PotentialReceptor PotentialsSpontaneous Cannabinoid Receptor 2 (CB2) Expression in the Cochlea of Adult Albino Rat and Its Up-Regulation after Cisplatin Treatment CB2 Receptor Expression in the Cochlea of Adult Albino Rathttp://orcid.org/0000-0002-9506-8483Martín-Saldaña Sergio 1*Trinidad Almudena 1Ramil Elvira 2Sánchez-López Antonio J. 3Coronado Maria José 4Martínez-Martínez Esther 5García José Miguel 5García-Berrocal José Ramón 1Ramírez-Camacho Rafael 11 Ear Research Group, Universitary Hospital Puerta de Hierro Majadahonda- Health Research Institute Puerta de Hierro, Madrid, Spain2 Sequencing and Molecular Biology Unit, Universitary Hospital Puerta de Hierro Majadahonda, Health Research Institute Puerta de Hierro, Madrid, Spain3 Neuroimmunology Unit, Universitary Hospital Puerta de Hierro Majadahonda- Health Research Institute Puerta de Hierro, Madrid, Spain4 Confocal Microscopy Unit, Universitary Hospital Puerta de Hierro Majadahonda, Health Research Institute Puerta de Hierro, Madrid, Spain5 Department of Medical Oncology, Universitary Hospital Puerta de Hierro Majadahonda- Health Research Institute Puerta de Hierro, Madrid, SpainDehghani Faramarz EditorMartin Luther University, GERMANYCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: RRC JMG AT SMS AJSL JRGB. Formal analysis: SMS. Funding acquisition: RRC. Investigation: SMS AT ER MJC EMM. Methodology: SMS AT RRC. Project administration: RRC SMS. Resources: RRC MJC ER. Supervision: RRC AT. Validation: SMS MJC. Visualization: SMS AT RRC. Writing – original draft: SMS AT. Writing – review & editing: SMS AT RRC JRGB. * E-mail: smartinsaldana@gmail.com26 8 2016 2016 11 8 e016195411 5 2016 15 8 2016 © 2016 Martín-Saldaña et al2016Martín-Saldaña et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.We provide evidence for the presence of cannabinoid CB2 receptors in some cellular types of the cochlea of the adult albino rat. Cannabinoids and their receptors are increasingly being studied because of their high potential for clinical use. As a hyperspecialized portion of the peripheral nervous system, study of the expression and function of cannabinoid receptors in the hearing organ is of high interest. Stria vascularis and inner hair cells express CB2 receptor, as well as neurites and cell bodies of the spiral ganglion. Cellular types such as supporting cells and outer hair cells, in which the expression of other types of functional receptors has been reported, do not significantly express CB2 receptors in this study. An up-regulation of CB2 gene expression was detected after an ototoxic event such as cisplatin treatment, probably due to pro-inflammatory events triggered by the drug. That fact suggests promising potential of CB2 receptor as a therapeutic target for new treatments to palliate cisplatin-induced hearing loss and other ototoxic events which triggers inflammatory pathways. Instituto de Salud Carlos III (ES)PI 11/00742Ramírez-Camacho Rafael Plan Nacional I+D+I 2008-2011Martin-Saldaña Sergio Fondo Europeo de Desarrollo RegionalRamírez-Camacho Rafael http://dx.doi.org/10.13039/501100004593Universidad Autónoma de MadridRamírez-Camacho Rafael Fundación Universidad Autónoma de MadridMartin-Saldaña Sergio This work was supported by Instituto de Salud Carlos III PI 11/00742 (to RRC) http://www.isciii.es/, Plan Nacionall I+D+I 2008-2011 (to RRC) http://www.idi.mineco.gob.es/portal/site/MICINN/menuitem.7eeac5cd345b4f34f09dfd1001432ea0/?vgnextoid=fe5aec1eb658c310VgnVCM1000001d04140aRCRD, Fondo Europeo de Desarrollo Regional as a cofunder (to RRC) http://www.red.es/redes/%C2%BFquienes-somos/feder, Universidad Autónoma de Madrid (to RRC) http://www.uam.es/, Fundación Universidad Autónoma de Madrid SMS http://fuam.es/. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction There is an increasing body of research on molecular signalling in the cochlea with the expectation that a deeper understanding of cell damage and regeneration pathways could uncover potential targets for the treatment and prevention of auditory damage. In the context of searching for potential therapeutic targets against inner ear disorders, the study of cannabinoid receptors expression in the auditory organ is of great interest. Cannabinoid receptors modulate a large number of normal brain and bodily functions and have been implicated as potential drug targets in a wide variety of diseases from cancer [1] to neurodegenerative disorders [2]. The endocannabinoid system is increasingly being studied because of the high potential for clinical use of cannabinoids. The endocannabinoid system is a cell communication system which consists of two receptors, CB1 and CB2, their endogenous ligands (anandamide and 2-AG), and the enzymes that produce and metabolize these ligands [3]. Cannabinoids exert a wide spectrum of therapeutic effects through CB1 and CB2 receptors [4]. Cannabinoid receptors are of a class of cell membrane receptors under the G protein-coupled receptor superfamily that contain seven transmembrane spanning domains [3]. Also, the existence of other cannabinoid receptors is suspected due to the actions of compounds that produce cannabinoid-like effects without activating either CB1 or CB2 [5]. CB receptors signaling pathways include regulation of adenylyl cyclase, MAP kinase, intracellular Ca 2+, and ion channels [6]. The first cannabinoid receptor (CB1) was characterized in rat brain in 1988 [7] and is mainly expressed in lung [8], liver [9] and kidneys [10]. CB2 was isolated in the rat spleen in 1993 [11]. CB2 receptor has been found to be 10–100 times more abundant than CB1 in immune cells such as macrophages, B cells, “natural killer” (NK) cells, monocytes, neutrophills and T cells [12]. It has also been described in non-immunological cells such as gastrointestinal system cells [13]; lung endotelium [14]; osteocytes and osteoclasts of bone [15] mouse spermatogenesis [16] and other aspects of reproduction [17]; in trabecular meshwork cells of the eye [18]; in adipocytes [19]; and in cells from cirrhotic liver (but not in the healthy organ) [20]. Cisplatin (CDDP) was the first platinum-based drug to be used and nowadays is widely employed in the treatment for some neoplastic entities such as head and neck, ovary, bladder, lung or brain. However, it presents severe side effects such as nephrotoxicity, bone marrow toxicity, gastrointestinal toxicity, liver toxicity, peripheral nervous system toxicity and ototoxicity [21]. Most of these related effects can be treated. However, CDDP- induced ototoxicity, that implies several cochlear damage which evolves in hearing impairment, tinnitus and dizzines, is one of the main reasons to chemotherapy discontinue.Inflammation has been recently shown to play a role in auditory cells apoptosis induced by CDDP, and is the main mechanism involved in immune-mediated hearing loss [22]. The inflammatory mediator inducible nitric oxide synthase (iNOS), is activated through nuclear factor-kB (NF-kB) and is related with cochlear wound by increasing cytokine expression and apoptosis in the cochlea [23]. Therefore, CDDP induce an over expression of COX-2, iNOS and TNF-α, that are regulated in cell nucleus by STAT1. STAT1 has an important role in inflammation and apoptosis in the cochlea like Kaur et al. demonstrated when used short interfering RNA against STAT1 to ameliorate CDDP-induced ototoxicity reducing the inflammation [24]. There is a report on the expression of CB2 receptors in the auditory HEI-OC1 (House Ear Institute-Organ of Corti) cell line [25]. In this study, researchers were able to demonstrate that JWH-015 (a synthetic CB2 agonist) could inhibit CDDP-induced apoptosis in HEI-OC1 cells. Up to our knowledge, CB2 receptor has been studied in vivo in relation to inflammatory processes in the outer ear [26] and the middle ear [27], but there are no reports on CB2 receptor expression or function in the inner ear. The aim of this study was to explore the expression of CB2 receptors in the mature inner ear of the Wistar rat by means of immunohistochemical staining in post-mortem tissue sections of the cochlea. CB2 gene expression was also measured by RT-qPCR in healthy and CDDP treated animals, using a well-established murine model of CDDP-induced ototoxicity, in order to study if an ototoxic event triggers an up- or down-transcriptional modulation of this receptor in the cochlea. These results may have important preliminary implications for the therapy of CDDP-induced hearing loss and other inflammatory inner ear diseases using CB2 receptor as a therapeutic target to ameliorate ototoxicity events. Material and Methods Animals Thirty three female Wistar rats (14–16 weeks old) weighting between 200 and 280 g were used in the present study. Animals were breaded and handled at the animal facility of the Health Research Institute Puerta de Hierro in controlled temperature rooms, with light–dark cycles, and with free access to food and water. Rats with signs of present of past middle ear infection were discarded. The animals were handled according to Spain Royal Law 53/2013 and the European Directive 2010/63/EU. The study was approved by the Clinical Research and Ethics Committee of Hospital Universitario Puerta de Hierro and the Autonomic Community of Madrid (PROEX 022/16). Auditory steady-state responses (ASSR) The animals were randomly assigned in two different groups: one group (n = 20; 10 for histology and 10 for RT-PCR analysis) was administered with phosphate buffered saline (PBS) as a control, and the other group (n = 13; 3 for histology and 10 for RT-PCR) received 10 mg/kg of CDDP (Accord Healthcare, Barcelona, Spain) dose. After 72 hours animals were euthanized by CO inhalation and cochleae, spleen and kidneys of each animal were extracted. Animals were anesthetized with intraperitoneal ketamine (100 mg/kg) and diazepam (0.1 mg/kg) before the procedure. Subcutaneous electrodes were placed over the vertex (active) and in the pinna of each ear (reference). An insert earphone (Etymotic ER-2) was placed directly into the external auditory canal. Ground electrodes were placed over the neck muscles. ASSR were recorded using an evoked potential averaging system (Intelligent Hearing System Smart-EP, FL, USA) in an electrically shielded, double-walled, sound-treated booth in response to 100 ms clicks or tone bursts, at 8, 12, 16, 20, 24, 28 and 32 kHz with 10 ms plateau and 1 ms rise/fall time. Intensity was expressed in decibels sound pressure level (dB SPL) peak equivalent. Intensity series were recorded, and an ASSR threshold was defined by the lowest intensity able to induce a replicable visual detectable response. ASSR was measured before treatments with PBS or CDDP, and 72 hours after the administration. Tissue extraction, fixation and decalcification Thirteen animals were euthanized by CO inhalation and subsequently decapitated. Temporal bones were removed and structures of middle ear dissected for isolation of the cochlear portion. Cochleae were fixed by intralabyrinthine perfusion with 4% paraformaldehyde (PFA) in phosphate buffer (pH = 7) and then immersed for 24 hours in 4% PFA solution. Afterwards, cochleae were decalcified in 0.12 M Ethylenediaminetetracetic acid dipotassium salt dihydrate (Sigma-Aldrich, St.Louis, USA) for 7–10 days. Cochleae were finally maintained in 1% PFA in 0.1 M PBS. Spleen and kidneys of each animal were extracted and submitted to the same protocol (except the decalcification step). Histological processing In order to obtain complementary images from the cellular structures of the organ of Corti, the two cochleae from one animal were each randomly processed by one of the following techniques: Inner ear surface preparations Cochleae were dissected following Liberman’s technique. In short, cochlea was bisected along a mid-modiolar plane and the half-turns cut apart, ensuring that the spiral ligament and tectorial membrane were pulled off and the resulting pieces of tissue were mounted on a microscope slide for examination in a confocal laser scanning microscope. Mid-modiolar sections The cochleae were processed for paraffin embedding. Careful placement of the cochlear portion was performed to obtain transverse serial sections of 6 μm (Leica microtome RM 2235), which were orthogonal and complementary to the surface preparations. Immunolabeling CB2 receptors were immunostained with the polyclonal anti-cannabinoid receptor II antibody (ab45942, Abcam, USA; dilution 1:50), using Alexa 488-conjugated anti-rabbit raised in goat (Molecular Probes–Invitrogen; dilution 1:400) as the secondary antibody. Antibodies were diluted in PBS supplemented with 1% bovine serum albumin (BSA) and 0.04% triton. All preparations were contrasted with TOPRO-3 iodide (Life Technologies, USA; dilution 1:500). Specificity was confirmed in separated experiments with additional negative controls, including tissue sections incubated with primary antibody pre-adsorbed to an excess of control peptide (ab45941, Abcam, USA; dilution 1:25). Cochleae assigned to surface preparations were immunolabeled once bisected into two halves and before final dissection, by means of a “free floating” technique. Transversal sections were immunolabeled once mounted on the slide and conveniently dried. As a positive control of the primary antibody used, spleen and kidneys [28] of each animal were submitted to the same protocol. Microscopic analysis Samples were visualized by means of a confocal laser scanning microscope Nikon Eclipse C1 coupled to a Nikon 90i microscope with a camera DXM1200F (Nikon, Haarlem, Netherlands). For validation of the specificity of the CB2 immunolabelling, immunofluorescence patterns were compared to negative controls in which primary antibody was pre-adsorbed with blocking peptide, and kidney and spleen were used as a positive control. Measurement of immunofluorescence intensity For relative quantification of CB2 fluorescence intensity, standardized settings for image acquisition and processing were used. To obtain values for CB2 immunofluorescence intensity in cell types of the cochlear duct, morphological boundaries of each cell type were determined on phase contrast images. The cell type-specific outlines were plotted with the corresponding grey-scaled CB2 immunofluorescence images. Single measurements of fluorescence intensity were performed using ImageJ software (version 1.40 g; National Institutes of Health, US). CB2 fluorescence intensity values for each cell type of the cochlear duct were averaged from ten independent measurements of immunolabelled cochlear mounts and sections from ten specimens. The fluorescence intensity of CB2 labelling was plotted using arbitrary units ranging from 0 to 2500. Following the same protocol used to determine relative fluorescence intensity in cell types of the cochlea, effect of CDDP treatment in the immunofluorescence was compared in stria vascularis and IHC of three animals treated with CDDP and three animals treated with PBS. Measurement of CB2 gene expression by RT-qPCR in healthy and CDDP treated animals Total RNA purification Dissected cochleae obtained from CDDP-treated and control Wistar rats were embebed in RNA later overnight and kept at -80°C until purification. Tissues were resuspended in Trizol and homogenized by the MagNA Lyser System (Roche). Total RNA was isolated by RNeasy kit (QIAGEN) including on-column DNAse Digestion, following protocol previously described [29]. RNA concentration was determined by spectrophotometry and each sample was reverse-transcripted to cDNA by using the First-Strand cDNA Synthesis Protocol (Agilent Technologies). Relative Quantification of CB2 and CB1 gene expression qPCR was performed in a LC480 (Roche) in order to quantify CB2 and CB1 levels of gene expression in relation to the level of the reference TATA-binding protein (TBP), in the cochlear tissue using Real Time Ready Single Assay (TBP Transcript ID ENSRNOT0000002038, amplicon length of 75 bps from 1029 to 110; CB2 Transcript ID ENSRNOT00000012342, amplicon length of 65 bps from 344 to 408; CB1 Transcipt ID ENSRNOT00000010850, amplicon length of 66 bps from 3244 to 3309; Roche). CB1 gene expression was used as a control (see S1 Fig). Statistical analysis Statistical analysis was performed with GraphPad Prism6 (San Diego, USA). One-way ANOVA was used to analyze for statistical significance of the results. Tukey test was used to identify significant differences between the paired treatments. p<0.05 was considered statistically significant. Results Ototoxic effect of CDDP-treatment Post-treatment ASSR recordings were found to be higher than pre-treatment ASSR recordings in the CDDP-treated animals, as it was previously described [30]. PBS-treated animals did not suffer hearing loss (Fig 1). 10.1371/journal.pone.0161954.g001Fig 1 Hearing thresholds before and after intraperitoneal administration of 10 mg/kg dose of CDDP or PBS. CDDP-induced hearing loss was statistically significant for all frequencies with respect to pre-treatment hearing. The diagrams include the mean, the standard deviation (n = 33), and the ANOVA results (difference statistically significant respect tectorial membrane *p<0.05). CB2 receptor detection in the cochlea by immunohistochemistry CB2 labeling was detected in the organ of Corti, specifically in inner hair cells (IHC) (Fig 2). 10.1371/journal.pone.0161954.g002Fig 2 Detailed view of the sensory hair cells of the organ of Corti. Detailed view of the 3 rows of OHC (arrows) and the single row of IHC (arrowhead). (A) Mid-modiolar section. (B) Cochlear whole mount. Immunofluorescence labeling of CB2 receptors is shown in green. To-Pro (blue) stains nuclei of cells. Intense fluorescence is observed in IHC (arrowhead), but not in OHC. Labeling of nerve fibers can be seen at the bottom of the figure (FN). IHC: inner hair cell. OHC: outer hair cell. (Scale bar = 30μm). In the stria vascularis, the CB2 labeling pattern was mostly coincident with an intermediate region within the stria vascularis, although some of the clearly immunolabeled cells occupied a luminal position and also the basal region in some areas. This is compatible with the distribution of intermediate cells as well as at least some of the marginal and basal cells (Fig 3). 10.1371/journal.pone.0161954.g003Fig 3 Image showing intense CB2 immunolabeling in cellular types of the stria vascularis (SV). Intermediate cells are indicated by long arrow. Marginal cells (short arrows); basal cells (arrowheads) might show labeling at some points. Fibrocytes of the spiral ligament (Fc) do not show expression of CB2 receptors. (Scale bar = 30μm) A strong CB2 labeling was also observed in the soma and neurites of the spiral ganglion neurons (Fig 4). 10.1371/journal.pone.0161954.g004Fig 4 Close view of CB2 immunolabeling of nerve fibers and spiral ganglion. Neurites are intensely labeled (short arrows), as well as the soma (long arrow) of the spiral ganglion (SG) neurons (Scale bar = 30μm). As a control, the same protocol was assayed in the kidney and spleen. A strong marking was observed in the spleen related to the B cells (Fig 5A). A strong labeling of CB2 receptor in the epithelial cells of renal tubules was observed in the kidney. A slightly marking in the glomerule was also detected and the pattern of staining was suggestive of podocyte labeling, as previously described by Barutta et al. (Fig 5B). Tissue sections incubated with CB2 primary antibody pre-adsorbed with control peptide did not show immunofluorescence, demonstrated the specificity of primary antibody used against the CB2 receptor (S1 Fig). 10.1371/journal.pone.0161954.g005Fig 5 Detailed view of immunolabeling of CB2 in the spleen and the kidney. Detailed view of CB2 immunolabeling in the spleen (A), used as a positive control of the primary antibody (Scale bar = 100μm). Detailed view of CB2 immunolabeling of epithelial cells of the renal tubules (RT) and the glomerule (G) of the kidney (B). (Scale bar = 30μm). Relative quantification of CB2 immunofluorescence labeling intensity in the IHC and OHC, the stria vascularis, modiolus, Reissner and tectorial membranes and the supporting cells of the rat inner ear is presented in Fig 6. OHC as well as supporting cells were almost completely devoid of CB2 staining. Relative immunofluorescence of OHC was remarkably lower when compared to labeling of IHC and was not statistically significant (Fig 6). The same protocol was assayed to compare relative fluorescence in in the stria vascularis (Fig 7A and 7B) and IHC (Fig 7C and 7D) of healthy and CDDP-treated animals. CB2 immunofluoresce was higher in the stria vascularis of CDDP-treated animals in a statistically significant way (Fig 7E). No difference between IHC of healthy and CDDP-treated animals was achieved. 10.1371/journal.pone.0161954.g006Fig 6 Relative quantification of CB2 immunofluorescence labeling intensity in different structures of the cochlea by ImageJ. Support cells: supporting cells (including Deiters’, Hensen’s, Claudius and pillar cells). Tectorial m: tectorial membrane (acellular). Reissner m.: Reissner membrane. OHC: outer hair cells. IHC: inner hair cells. Relative immunofluorescence in the IHC and in the stria vascularis was significantly than in the other cell types. The diagrams include the mean, the standard deviation (n = 10), and the ANOVA results (difference statistically significant respect tectorial membrane *p<0.05). 10.1371/journal.pone.0161954.g007Fig 7 Relative quantification of CB2 immunofluorescence labeling intensity in healthy and CDDP-treated animals by ImageJ. Detailed view of the stria vascularis (SV) and the IHC (arrowhead) of healthy (A and C) and CDDP-treated animals (B and D) (Scale bar = 100μm). Relative immunofluorescence in the IHC and in the stria vascularis of healthy and CDDP-treated animals (E). CB2 relative immunofluorescence was significantly higher only in the stria vascularis of animals treated with CDDP if compared with the control group. The diagram include the mean, the standard deviation (n = 6; 3 control and 3 CDDP-treated), and the ANOVA results (difference statistically significant in each cellular structure between healthy and CDDP-treated animals *p<0.05). CB2 gene expression in the cochlea and CDDP induced an up-regulation in their expression Gene expression of CB2 receptors was studied in the cochlea of healthy animals with respect to TBP reference gene. The relative levels of expression showed up-transcriptional regulation of CB2 under CDDP treatment in a statistically significant way. Similar results were observed in kidney of animals untreated, an under CDDP treatment (Fig 8). 10.1371/journal.pone.0161954.g008Fig 8 Measure of relative CB2 gene expression by RT-qPCR. Measure of CB2 gene expression in the cochlea and kidney of healthy (control) and CDDP treated animals, respect TBP reference. CB2 gene expression was significantly higher in animals treated with CDDP than those in control group. The diagrams include the mean, the standard deviation (n = 10), and the ANOVA results (difference statistically significant respect control *p<0.05). Level of CB1 gene expression in both organs was also measured as a control. CB1 receptors expression was detected in the cochlea and in the kidney. A slightly down-transcriptional regulation in the CDDP treated animals was observed in the cochlea, but there was no statistically significant difference between control and CDDP-treated groups (see S2 Fig). Discussion The expression of cannabinoid receptors in a specific tissue opens the door to potential treatments with agonists and antagonists that could trigger certain cell signaling pathways. Expression of CB2 has also been described in astrocytes and some neuron subpopulations. Role of CB2 receptor in the central nervous system is related to neuroinflammation and neurodegeneration [31]. In the peripheral nervous system, CB2 agonists have analgesic effects by acting at dorsal root ganglia (as well as at spinal cord) [32]. Other authors have reported CB2 expression on human skin nerve fibers [33]. Baek et al. previously described expression of CB2 protein receptor in the vestibular and cochlear nuclei of the brain in a rat model, suggesting that it could have an important role in the control of balance and hearing function [34]. As part of the peripheral nervous system [35], it is not surprising to find such a remarkable immunolabelling of CB2 receptor in organ of Corti cells, although its functional role has still to be ascertained. Immunolabeling to CB2 receptor resulted to be higher in the stria vascular of animals which received the chemotherapeutic, but no difference could be appreciate in the IHC of healthy and CDDP-treated animals. Due to these limitations of immunohistochemistry to appreciate differences in the transcriptional up-regulation of CB2 between healthy and CDDP-treated animals, which suffered big structural tissue degradation, RT-qPCR was assayed to determine the regulation of CB2 gene expression. In 2005 meeting of the Association for Research in Otolaryngology (ARO), Fauser et al. reported preliminary studies of the expression of CB1 cannabinoid receptor in type I and type II spiral ganglion cells that was higher after intratympanic treatment with salicilate and glutamate in comparison with saline, with no specific stain in the organ of Corti or stria vascularis [36]. In this study high values of CB1 gene expression was observed and used as a control of cannabinoid gene expression (see S2 Fig), probably due to the presence of the spiral ganglion in the dissected cochleae. A slightly down-regulation of CB1 gene expression in the cochlea and in the kidney was observed when animals received CDDP treatment, being statistically significant in the kidney probably due to a primary down-regulation trying to ameliorate tissue damage promoted by CB1 in nephropathy models [37]. Recent studies go deeper in the role of CB1 receptor in hearing function. Zheng et al. use CBD and Δ-9-THC, and concluded that their results suggest that cannabinoids, such as Δ-9-THC and CBD, may actually aggravate tinnitus, probably because of the net effect of activation of CB1 receptors in the dorsal cochlear nucleus might be to increase the excitation of fusiform cells, thus exacerbating neuronal hyperactivity [38]. In a study published in Hearing Research, researchers showed in knockout CB1 mice, the role of the cannabinoid receptor in hearing. Animals without CB1 presents deficit in their audiograms for frequencies above 8 kHz, but they are still able to distinguished changes in frequencies as well as the wild type [39]. Up to our knowledge, this is the first in vivo report of the expression of CB2 receptors in the inner ear of mammals. Expression of CB2 gene had been previously described in the in vitro line HEI-OC1, by means of RT-qPCR [25]. The authors demonstrated that JWH-015 (a synthetic CB2 agonist) and HU210 (CB1 agonist), could inhibit CDDP-induced apoptosis in HEI-OC1 cells. Specifically, they found that JWH-015 inhibited CDDP-induced caspase-3 and caspase-8 activity; cytochrome c (proapoptotic molecule) release from the mitochondria; increased phosphorylation of ERK; blocked the increase of ROS produced by CDDP; and inhibited TNF-α production on HEI-OC1 cells. From this, they suggested that CB activation was important in CDDP-induced apoptosis of HEI-OC1 cells. HEI-OC1 cells express several molecular markers characteristic of organ of Corti sensory cells and could be considered as precursors. CB2 immunolabeling in our study was seen in inner hair cells, but not in outer hair cells. Previous studies suggested that hair cell types derived from different progenitor cells. Specifically, these studies suggest that IHC derive from the greater epithelial ridge and the OHC derive from the lesser epithelial ridge [40, 41]. This difference during ontogenesis has been related to the lateral process of inhibition that avoid cells from follow the same developmental pathway [42]. These differences in the development of IHC and OHC could explain the differential immunolabeling of CB2 receptor in sensory cells reported in our assays. Our study suggests the expression of CB2 receptor mostly in the intermediate region of the stria vascularis, although luminal and basal labeling could also be seen. We consider this as a limitation of our study and further studies would be needed to ascertain what cellular types are specifically expressing CB2 receptor. Potential labeling of residual erythrocytes within the capillary vessels of stria vascularis was contemplated, but we consider that it small size is negligible in relation to strial cells. The fact that CB2 immunolabeling is also located in the stria vascularis suggests that CB2 receptors might play a role in generation and/or maintenance of endocochlear potential. Potential interaction with structures identified in stria vascularis that are related to water and ion transport, antioxidant defenses and other homeostatic mechanisms merit further research. CDDP clinical use is limited due to the induced toxicity affecting nervous system, kidney function and hearing. CDDP induces apoptosis by binding to DNA, ROS accumulation, increased lipid peroxidation and Ca2+ influx and inflammation events [43]. In the present study CDDP treated animals presents an up-regulation in the expression of CB2 gene in cochleae and also in the kidney. Mukhopaday et al. used an agonist of CB2 receptor and CB2 knockout mice trying to understand the role of the endocannabinoid system in CDDP-induced kidney disease. They observed that treatment with CB2 agonist HU-308 ameliorates many CDDP-induced events like: ROS and inflammation markers; inflammatory cell infiltration and chemokine production; leading to a marker improvement of the renal function in CDDP-treated animals that also receive HU-308. An increase in inflammation, oxidative stress and cell death in knockout mice CB2 -/- when compared to wild type was also observed, suggesting the protective role of the CB2 receptor activation [44]. More recently, Mukhopaday et al. characterized the partial CB2 receptor agonist LEI-101 showing promising results. The partial agonist does not induce cannabimimetic in animals treated, and a considerable attenuation in CDDP-induced nephropathy, including inflammation and oxidative stress [45]. As shown in this study, CDDP treatment increased the gene expression levels of CB2 in the cochlea and in kidney (control) (Fig 8). This up-regulation of the inflammatory microenvironment induced by the treatment with CDDP promotes the overexpression of endocannabinoid system (increased CB2). CB2 overexpression may represent an adaptive response to control a deregulated balance in the ear maybe trying to restore normal conditions. A deeper understanding of the functional role of CB2 receptor in sensory and non-sensory elements of the cochlea may help develop therapies using CB2 as target for reducing damage to the hearing organ. That opens a big field of possibilities to the treatment of inner ear diseases which involves inflammation process, using drugs with non canabimimetic effects, because CB2 is the lack of psychoactive side effects after stimulation. Conclusions In conclusion, evidence for the presence of cannabinoid CB2 receptor by immunohystochemistry and by RT-qPCR was provided. An immunolabeling of CB2 antibodies in four structures of the adult rat cochlea was found. That was, stria vascularis, inner hair cells, auditory afferent nerves and cell bodies of the spiral ganglion. Up-regulation of CB2 gene expression in animals exposed to CDDP treatment was also detected, when compared with healthy animals. This fact was partially supported by the higher immunofluorescence observed in the stria vascularis of CDDP-treated animals if compared with the healthy ones. These results suggest a considerable promising potential of CB2 receptor as a target of new treatments against CDDP-induced ototoxicity, and probably against other inflammatory diseases in the inner ear. Further research is needed to determine the functionality of CB2 receptors in the organ of Corti and the potential therapeutic role of agonists and antagonists of these receptors. Supporting Information S1 Fig Detailed view of tissue sections pre adsorbed with CB2 blocking peptide. Detailed view of the cochlea pre adsoberd with CB2 blocking peptide. No immunofluorescence was observed in the stria vascularis (SV), the IHC or the spiral ganglion (SG) (Scale bar = 100μm). (TIF) Click here for additional data file. S2 Fig Measure of relative CB1 gene expression by RT-qPCR. Measure of relative CB1 gene expression in the cochlea and kidney of healthy (control) and CDDP treated animals, respect TBP reference. The diagrams include the mean, the standard deviation (n = 10), and the ANOVA results (difference statistically significant respect control *p<0.05). (TIF) Click here for additional data file. Authors acknowledge Carlos Vargas for his help in histology experiments, and PhD. Martin Santos, Maria Dolores Molina and Aitor Martín for their help in the animal facilities. ==== Refs References 1 Pérez-Gómez E , Andradas C , Blasco-Benito S , Caffarel MM , García-Taboada E , Villa-Morales M , et al Role of cannabinoid receptor CB2 in HER2 Pro-oncogenic signaling in breast cancer . J Natl Cancer Inst . 2015 4 8 10.1093/jnci/djv077 2 Fernández-Ruiz J , Moro MA , Martínez-Orgado J . Cannabinoids in Neurodegenerative Disorders and Stroke/Brain Trauma: From Preclinical Models to Clinical Applications . Neurotherapeutics . 2015 8 11 10.1007/s13311-015-0381-7 3 Pertwee R , Howlett A , Abood ME , Alexander S , Di Marzo V , Elphick M , et al International Union of Basic and Clinical Pharmacology. LXXIX. Cannabinoid receptors and their ligands: beyond CB1 and CB2 . Pharmacol Rev . 2010 12 10.1124/pr.110.003004 4 Mukherjee S , Adams M , Whiteaker K , Daza A , Kage K , Cassar S , et al Species comparison and pharmacological characterization of rat and human CB 2 cannabinoid receptors . Eur J Pharmacol . 2004 11 5 10.1016/j.ejphar.2004.09.058 5 McHugh D , Tanner C , Mechoulam R , Pertwee RG , Ross RA . Inhibition of human neutrophil chemotaxis by endogenous cannabinoids and phytocannabinoids: evidence for a site distinct from CB1 and CB2 . Mol Pharmacol . 2007 10 26 10.1124/mol.107.041863 6 Demuth DG , Molleman A . Cannabinoid signalling . Life Sci . 2005 8 18 10.1016/j.lfs.2005.05.055 7 Devane WA , Dysarz Fr , Johnson MR , Melvin LS , Howlett AC . Determination and characterization of a cannabinoid receptor in rat brain . Mol Pharmacol . 1988 11 1 2848184 8 Shire D , Carillon C , Kaghad M , Calandra B , Rinaldi-Carmona M , Le Fur G , et al An amino-terminal variant of the central cannabinoid receptor resulting from alternative splicing . J Biol Chem . 1995 2 24 10.1074/jbc.270.8.3726 9 Mallat A , Teixeira‐Clerc F , Deveaux V , Manin S , Lotersztajn S . The endocannabinoid system as a key mediator during liver diseases: new insights and therapeutic openings . Br J Pharmacol . 2011 8 10.1111/j.1476-5381.2011.01397.x 10 Deutsch DG , Goligorsky MS , Schmid PC , Krebsbach RJ , Schmid H , Das S , et al Production and physiological actions of anandamide in the vasculature of the rat kidney . J Clin Invest . 1997 9 15 10.1172/JCI119677 11 Munro S , Thomas KL , Abu-Shaar M . Molecular characterization of a peripheral receptor for cannabinoids . Nature . 1993 9 2 10.1038/365061a0 12 Berdyshev E . Cannabinoid receptors and the regulation of immune response . Chem Phys Lipids . 2000 11 10.1016/S0009-3084(00)00195-X 13 Storr M , Gaffal E , Saur D , Schusdziarra V , Allescher HD . Effect of cannabinoids on neural transmission in rat gastric fundus . Can J Physiol Pharmacol . 2002 1 31 10.1139/Y02-005 14 Zoratti C , Kipmen‐Korgun D , Osibow K , Malli R , Graier WF . Anandamide initiates Ca2+ signaling via CB2 receptor linked to phospholipase C in calf pulmonary endothelial cells . Br J Pharmacol . 2003 12 10.1038/sj.bjp.0705529 15 Ofek O , Karsak M , Leclerc N , Fogel M , Frenkel B , Wright K , et al Peripheral cannabinoid receptor, CB2, regulates bone mass . Proc Natl Acad Sci U S A . 2006 1 17 10.1073/pnas.0504187103 16 Grimaldi P , Orlando P , Di Siena S , Lolicato F , Petrosino S , Bisogno T , et al The endocannabinoid system and pivotal role of the CB2 receptor in mouse spermatogenesis . Proc Natl Acad Sci U S A . 2009 7 7 10.1073/pnas.0812789106 17 Maccarrone M . CB2 receptors in reproduction . Br J Pharmacol . 2008 1 10.1038/sj.bjp.0707444 18 He F , Song Z-H . Molecular and cellular changes induced by the activation of CB2 cannabinoid receptors in trabecular meshwork cells . Mol Vis . 2007 7 27 17679938 19 Roche R , Hoareau L , Bes-Houtmann S , Gonthier M-P , Laborde C , Baron J-F , et al Presence of the cannabinoid receptors, CB1 and CB2, in human omental and subcutaneous adipocytes . Histochem Cell Biol . 2006 8 10.1007/s00418-005-0127-4 20 Julien B , Grenard P , Teixeira-Clerc F , Van Nhieu JT , Li L , Karsak M , et al Antifibrogenic role of the cannabinoid receptor CB2 in the liver . Gastroenterol . 2005 3 10.1053/j.gastro.2004.12.050 21 Bokemeyer C , Berger C , Hartmann J , Kollmannsberger C , Schmoll H , Kuczyk M , et al Analysis of risk factors for cisplatin-induced ototoxicity in patients with testicular cancer . Br J Cancer . 1998 10 7 10.1016/S0022-5347(01)62317-6 22 Berrocal JRG , Ramírez-Camacho R . Immune response and immunopathology of the inner ear: an update . J Laryngol Otol . 2000 2 10.1258/0022215001905021 23 Watanabe K , Inai S , Jinnouchi K , Bada S , Hess A , Michel O , et al Nuclear-factor kappa B (NF-kappa B)-inducible nitric oxide synthase (iNOS/NOS II) pathway damages the stria vascularis in cisplatin-treated mice . Anticancer Res . 2001 11 12553036 24 Kaur T , Mukherjea D , Sheehan K , Jajoo S , Rybak L , Ramkumar V . Short interfering RNA against STAT1 attenuates cisplatin-induced ototoxicity in the rat by suppressing inflammation . Cell Death Dis . 2011 7 21 10.1038/cddis.2011.63 25 Jeong HJ , Kim SJ , Moon PD , Kim NH , Kim JS , Park RK , et al Antiapoptotic mechanism of cannabinoid receptor 2 agonist on cisplatin‐induced apoptosis in the HEI‐OC1 auditory cell line . J Neurosci Res . 2007 3 10.1002/jnr.21168 26 Mimura T , Oka S , Koshimoto H , Ueda Y , Watanabe Y , Sugiura T . Involvement of the endogenous cannabinoid 2 ligand 2-arachidonyl glycerol in allergic inflammation . Int Arch Allergy Immunol . 2012 5 31 10.1159/000336167 27 Oka S , Yanagimoto S , Ikeda S , Gokoh M , Kishimoto S , Waku K , et al Evidence for the involvement of the cannabinoid CB2 receptor and its endogenous ligand 2-arachidonoylglycerol in 12-O-tetradecanoylphorbol-13-acetate-induced acute inflammation in mouse ear . J Biol Chem . 2005 5 6 10.1074/jbc.M413260200 28 Barutta F , Piscitelli F , Pinach S , Bruno G , Gambino R , Rastaldi MP , et al Protective role of cannabinoid receptor type 2 in a mouse model of diabetic nephropathy . Diabetes . 2011 9 10.2337/db10-1809 29 Patil KV , Canlon B , Cederroth CR . High quality RNA extraction of the mammalian cochlea for qRT-PCR and transcriptome analyses . Hear Res . 2015 7 10.1016/j.heares.2015.03.008 30 Rybak LP , Whitworth CA , Mukherjea D , Ramkumar V . Mechanisms of cisplatin-induced ototoxicity and prevention . Hear Res . 2007 5 10.1016/j.heares.2006.09.015 31 Fernández-Ruiz J , Romero J , Velasco G , Tolón RM , Ramos JA , Guzmán M . Cannabinoid CB 2 receptor: a new target for controlling neural cell survival? Trends Pharmacol Sci . 2007 1 10.1016/j.tips.2006.11.001 32 Hsieh GC , Pai M , Chandran P , Hooker BA , Zhu CZ , Salyers AK , et al Central and peripheral sites of action for CB2 receptor mediated analgesic activity in chronic inflammatory and neuropathic pain models in rats . Br J Pharmacol . 2011 1 10.1111/j.1476-5381.2010.01046.x 33 Ständer S , Schmelz M , Metze D , Luger T , Rukwied R . Distribution of cannabinoid receptor 1 (CB1) and 2 (CB2) on sensory nerve fibers and adnexal structures in human skin . J Dermatol Sci . 2005 6 10.1016/j.jdermsci.2005.01.007 34 Baek J-H , Zheng Y , Darlington CL , Smith PF . Cannabinoid CB2 receptor expression in the rat brainstem cochlear and vestibular nuclei . Acta Otolaryngol . 2008 9 10.1080/00016480701796944 35 Osen KK , Furness DN , Hackney CM . The border between the central and the peripheral nervous system in the cat cochlear nerve: A light and scanning electron microscopical study . Hear Res . 2011 7 10.1016/j.heares.2011.03.012 36 Santi PA, editor. CB1 cannabinoid receptor is expressed in the mammalian cochlea. Abstracts of the Twenty-Eight Annual Midwinter Research Meeting of the Association for Research in Otolaryngology. 2005 Feb 19–24. New Orleans, Louisiana, USA. 37 Mukhopadhyay P , Pan H , Rajesh M , Bátkai S , Patel V , Harvey‐White J , et al CB1 cannabinoid receptors promote oxidative/nitrosative stress, inflammation and cell death in a murine nephropathy model . Br J Pharmacol . 2010 6 10.1111/j.1476-5381.2010.00769.x 38 Zheng Y , Reid P , Smith PF . Cannabinoid CB1 Receptor Agonists Do Not Decrease, but may Increase Acoustic Trauma-Induced Tinnitus in Rats . Front Neurol . 2015 3 18 10.3389/fneur.2015.00060 39 Toal KL , Radziwon KE , Holfoth DP , Xu-Friedman MA , Dent ML . Audiograms, gap detection thresholds, and frequency difference limens in cannabinoid receptor 1 knockout mice . Hear Res . 2016 2 10.1016/j.heares.2015.09.013 40 Lim DJ , Anniko M . Developmental morphology of the mouse inner ear: a scanning electron microscopic observation . Acta Otolaryngol Suppl . 1985 3877398 41 Lim DJ , Rueda J . Structural development of the cochlea In: Romand R , editor. Development of Auditory and Vestibular Systems 2 . El Sevier , Amsterdam 1992 pp. 33 –58 . 42 Zine A , Romand R . Development of the auditory receptors of the rat: a SEM study . Brain Res . 1996 5 20 10.1016/0006-8993(96)00147-3 43 Casares C , Ramírez-Camacho R , Trinidad A , Roldán A , Jorge E , García-Berrocal JR . Reactive oxygen species in apoptosis induced by cisplatin: review of physiopathological mechanisms in animal models . Eur Arch OtorhinoLaryngol . 2012 12 10.1007/s00405-012-2029-0 44 Mukhopadhyay P , Rajesh M , Pan H , Patel V , Mukhopadhyay B , Bátkai S , et al Cannabinoid-2 receptor limits inflammation, oxidative/nitrosative stress, and cell death in nephropathy . Free Rad Biol Med . 2010 2 1 10.1016/j.freeradbiomed.2009.11.022 45 Mukhopadhyay P , Baggelaar M , Erdelyi K , Cao Z , Cinar R , Fezza F , et al The novel, orally available and peripherally restricted selective cannabinoid CB2 receptor agonist LEI‐101 prevents cisplatin‐induced nephrotoxicity . Br J Pharmacol . 2016 2 10.1111/bph.13338

PMC005xxxxxx/PMC5001641.txt

==== Front PLoS CurrPLoS CurrPLoS CurrplosPLoS Currents2157-3999Public Library of Science San Francisco, USA 2761716810.1371/currents.outbreaks.007219ac3b9a2117418df7ab629686b6Research ArticleTaste and Safety: Is the Exceptional Cuisine Offered by High End Restaurants Paralleled by High Standards of Food Safety? Kanagarajah Sanch Field Epidemiology Service, National Infection Service, Public Health England, London, United Kingdom*Mook Piers Field Epidemiology Service, National Infection Service, Public Health England, London, United KingdomCrook Paul Field Epidemiology Service, National Infection Service, Public Health England, London, United KingdomAwofisayo-Okuyelu Adedoyin Gastrointestinal Infections, National Infection Service, Public Health England, London, United Kingdom; NIHR Health Protection Research Unit in Gastrointestinal Infections, United KingdomMcCarthy Noel Field Epidemiology Service, National Infection Service, Public Health England, London, United Kingdom; NIHR Health Protection Research Unit in Gastrointestinal Infections, United Kingdom; Warwick Medical School, Division of Health Sciences, University of Warwick, Coventry, United Kingdom2 8 2016 8 ecurrents.outbreaks.007219ac3b9a2117418df7ab629686b6© 2016 Kanagarajah, Mook, Crook, Awofisayo-Okuyelu, McCarthy, et al2016Kanagarajah, Mook, Crook, Awofisayo-Okuyelu, McCarthy, et alThis is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited.Introduction: Restaurant guides such as the Good Food Guide Top 50 create a hierarchy focussing on taste and sophistication. Safety is not explicitly included. We used restaurant associated outbreaks to assess evidence for safety. Methods: All foodborne disease outbreaks in England reported to the national database from 2000 to 2014 were used to compare the Top 50 restaurants (2015) to other registered food businesses using the Public Health England (PHE) outbreak database. Health Protection Teams were also contacted to identify any outbreaks not reported to the national database. Among Good Food Guide Top 50 restaurants, regression analysis estimated the association between outbreak occurrence and position on the list. Results: Four outbreaks were reported to the PHE national outbreak database among the Top 50 giving a rate 39 times higher (95% CI 14.5–103.2) than other registered food businesses. Eight outbreaks among the 44 English restaurants in the Top 50 were identified by direct contact with local Health Protection Teams. For every ten places higher ranked, Top 50 restaurants were 66% more likely to have an outbreak (Odds Ratio 1.66, 95% CI 0.89–3.13). Discussion: Top 50 restaurants were substantially more likely to have had reported outbreaks from 2000-2014 than other food premises, and there was a trend for higher rating position to be associated with higher probability of reported outbreaks. Our findings, that eating at some of these restaurants may pose an increased risk to health compared to other dining out, raises the question of whether food guides should consider aspects of food safety alongside the clearly important complementary focus on taste and other aspects of the dining experience. campylobactercookingfood poisoninggastroenteritisgourmetnorovirusnovel cookingOutbreakThe research was funded by the National Institute for Health Research Health Protection Research Unit (NIHR HPRU) in Gastrointestinal Infections at University of Liverpool in partnership with Public Health England (PHE), in collaboration with University of East Anglia, University of Oxford and the Institute of Food Research. Noel McCarthy and Adedoyin Awofisayo-Okuyelu are based at the University of Oxford. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, the Department of Health or Public Health England. ==== Body INTRODUCTION Restaurant guides such as the UK Good Food Guide Top 50 create a hierarchy focusing on taste and sophistication but not explicitly including consideration of safety1. Since restaurants are a common setting for outbreaks of gastrointestinal illness in the United Kingdom2, which can affect large numbers of diners3, food safety may also be an important criterion for some diners. The Food Standards Agency (FSA), in cooperation with local authorities, offers information on compliance with basic food safety standards across the catering industry through their public facing Food Hygiene Rating Scheme http://www.food.gov.uk/business-industry/caterers/hygieneratings. As a way to assess whether there is evidence of different levels of risk of foodborne disease among top restaurants we have reviewed restaurant associated outbreaks in England including those restaurants in the UK Good Food Guide Top 50. METHOD Two complementary analyses were conducted. The first compared outbreaks in the Top 50 restaurants with outbreaks in all other restaurants in England using a national reporting database for foodborne disease outbreaks. The second tested whether position in the Top 50 was associated having an identified outbreak. FSA data on registered food premises in 2014 were used to estimate the total number of food premises in England and the Good Food Guide Top 50 (2015) to identify the number of these 50 that were in England. The Public Health England (PHE) electronic Foodborne and non-Foodborne Gastrointestinal Outbreak Surveillance System (eFOSS), which is a national reporting system collating reports of outbreaks from PHE Health Protection Teams (HPT) and laboratory notifications, was used to identify outbreaks linked to food outlets from 2000 to 2014, and which of these was associated with restaurants listed in the UK Good Food Guide Top. PHE was formed in 2013 and it took on the role of the Health Protection Agency (2004-2013) which had taken on the functions of the national Communicable Disease Surveillance Centre. The function of eFOSS remained the same during these organisational transitions. The risk of outbreaks in the Top 50 was compared with the risk of outbreaks in other restaurants in England and rate ratios were calculated. The 2015 edition of the UK Good Food Guide was also used to identify the position of individual English restaurants (excluding premises in Scotland and Wales) in the Top 50. HPTs in the areas where Top 50 restaurants in England were located were asked to identify outbreaks associated with these restaurants from 2000 to 2014 and to provide the following information on these outbreaks; the year of outbreak, organism identified, number of people affected, source of outbreak, agent, high risk food preparation processes, and the outbreak investigation report. The association between outbreak occurrence and position within the 2015 Top 50 was estimated using logistic regression analysis in this dataset. Possible risk factors identified in the investigation of the outbreaks were summarised. Data were analysed with Stata version 13 and OpenEpi version 3.30a. RESULTS FSA data identified 508,630 registered food premises in England in 2014. Forty four of the UK top 50 restaurants (88%) were located in England and included in this study, while the remaining restaurants were located in Wales (n=2) and Scotland (n=4). Among the 44 English restaurants identified in the Top 50, four had an outbreak reported to eFOSS (9.1%) compared to 1196 outbreaks reported from the other 508,586 registered food premises (0.23%) in England between 2000 and 2014; giving a Risk Ratio of 39 (95% CI 14.5-103.2). Eight outbreaks in seven restaurants among the 2015 Top 50 were identified from 2000 to 2014 following direct contact with HPTs across England (Table 1). Among the 44 English Top 50 restaurants; a position higher on the ranking list was associated with a trend toward increased risk of an outbreak, with an odds ratio of 1.05 [0.99-1.12] for each place higher, equivalent to an odds ratio of 1.66 [0.89- 3.13] for being ten places higher in the list. Of the seven restaurants in the Top 50 list in 2015 identified as having one or more outbreaks during the study period, five were in the top 50 list in the year that the outbreaks occurred (Table 1). Outbreak E occurred in a restaurant that opened in 2014, the same year as the outbreak, and so was not eligible for the Top 50 list for that year but was subsequently included in 2015. Outbreak H occurred in a restaurant that did not make the top 50 list in 2010 when it experienced an outbreak but was in the list from 2011 to 2015. Table 1. Outbreaks in the 2015 Top 50 restaurants list reported in England from 2000-2014 Table 1_OBK0044This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Descriptions of the outbreaks Outbreaks B and C occurred in the same restaurant (“restaurant 2”) in 2009 and 2012 respectively (Table 1). Outbreak reports were available for four out of the eight outbreaks and these are described in more detail. Outbreak A was of Campylobacter associated in 2010. Sixteen cases were identified; 11 probable and five confirmed. Chicken liver parfait was determined to be the source but the organism was not identified via microbiological sampling from remaining food. The cooking method involved chicken livers soaked in milk, then pan fried with the temperature reaching 61 °C. The livers were then pureed in a heated blender with the temperature reaching 60 °C. The chicken livers were reportedly undercooked to avoid a grey appearance. Outbreak B was an outbreak of norovirus reported as associated with the consumption of oysters. There were 529 reports of illness in January and February 2009. Ten stool samples were positive for norovirus. An analytical study identified one food, the ‘Oyster, Passion Fruit Jelly, Lavender’ from the ‘Tasting Menu’ to have the strongest association with illness. The lengthy period of transmission at the restaurant may have involved a combination of person to person spread among staff, contamination of pre-prepared foods, and persistent infection in seafood3. Outbreak E was reported as an outbreak of norovirus associated with consumption of a plaice and cockle dish and subsequent contamination of other dishes in 2014. The first case was the chef followed by 39 staff members who reported being unwell during the first round of notifications. Seventy two cases were reported at the end of the outbreak with 15/20 stool samples testing positive for norovirus. Staff members who first became ill ate at a taster event. Although tweezers were used in part for food handling, environmental health officers identified substantial touching of food during preparation. They also noted that hand wash basins were also used for other purposes. An analytical study showed that cases were more likely to have eaten the plaice and cockles dish. Outbreak H was reported as an outbreak of norovirus infection associated with the consumption of oysters in 2010. Eleven diners and one kitchen staff member were identified with symptoms of diarrhoea and vomiting. Environmental and stool samples tested positive for norovirus4. An analytical study showed an association of illness with oyster consumption. DISCUSSION The Top 50 restaurants in England were more likely to have had reported outbreaks from 2000-2014 than other food premises, and there was a trend for higher rated position among the Top 50 to be associated with higher probability of reported outbreaks, although the latter findings were non-significant at the 5% level. Limitations of this study which should be considered when interpreting these findings. Detected and reported outbreaks were the measure of food safety accessible to us across both the Top 50 restaurants and other registered food premises in England. Analyses identifying a 39 fold higher rate of reported outbreaks among Top 50 premises compared to other food premises and of higher position being associated with increased odds of an identified outbreak could each have been prone to bias in either direction. For example, if food premises within or further up the list in the Top 50 had a higher profile, and higher expectation from diners not to get sick, reporting and detection might be more likely. This could create an apparent positive association between high ranking and the occurrence of outbreaks. In contrast, if more prestigious restaurants use internal measures to respond to incidents without informing public health authorities3 , 4 unless very large or persistent outbreaks occur this could create a bias in the opposite direction. It is not clear whether systematic biases affect reporting of locally identified outbreaks to the national reporting system eFOSS. Direct contact with the local health protection teams revealed that four outbreaks occurring in top 50 restaurants had not been reported to the eFOSS system. Analysis of the association of position within the Top 50 and identified outbreaks showed a point estimate of 1.66 for every ten places higher in the ranking order. This may be less prone to bias but the numbers in this analysis are small and the non-significant trend identified may be due to chance alone. Due to lack of available data, these findings are unadjusted for potential confounders such as the size of the restaurant, the years of operation, the number of cases visiting the restaurants, or the number of dishes served at each restaurant. We used the 2015 version of the UK Good Food Guide and outbreaks from 2000 to 2014 so that the associations described may be best seen as whether having had outbreaks affected future presence and position in the rankings rather than position predicting risk of outbreaks. All but two restaurants were also in the top 50 list in the year of the outbreak. Due to the limitations in this study, several hypotheses and issues of bias remain to be tested in further studies: whether high end restaurant clientele are more likely to report illness, if the number of meals served in a restaurant may increase the number of cases, and if cooking methods used by the various restaurants can be investigated to see if certain methods cause more illness in diners in these settings. A Food Standards Agency (FSA) Food Hygiene Rating Scheme is run by local authorities in England, Wales and Northern Ireland and applies to food businesses which are inspected by a food safety officer from the local authority. The score is calculated based on how hygienically the food is handled, the condition of the structure of the buildings and how the business manages and records what it does to make sure food is safe. The restaurants in the Top 50 were each awarded either 4 or 5 stars out of 5 on this FSA scheme during 2014. Despite scoring well on generic hygiene assessments there are a number of reasons as to why gourmet restaurants may have a higher rate of outbreaks. They may be more likely to serve foods with known risks such as raw oysters5; their complex dishes may require more handling and tasting by chefs; and novel and innovative cooking techniques may not robustly kill pathogens. There is some evidence for each of these factors from our study. Oysters or shell fish were identified as the most likely vehicle in four outbreaks, complex food processes were identified as a possible risk in two outbreaks3 , 4, and low cooking temperature of chicken liver parfait was identified in one outbreak. This range both overlaps and differs from findings of outbreaks in other settings as reported in the literature6 , 7 , 8 , 9. Similarities with wider studies on outbreaks includes the importance of lightly cooked chicken liver products prepared in restaurants that have been identified as an important source of gastroenteritis outbreaks in recent years 10. Differences include the lack of Salmonella compared to an England and Wales study that where these accounted for most outbreaks, particularly among Chinese, Indian, British and Italian cuisines from 1992 to 200911. The pathogen profile and risk processes also differ from those reported from domestic premises which were associated with Salmonella, inappropriate storage of food, and consumption of poultry, eggs, or sauces12. There is thus some evidence for differences in the types of food safety risks that predominate in high end restaurants compared to other catered settings. Overall the eight outbreaks identified at restaurants rated as in the top 50 in the country by the Good Food Guide affected more than 750 diners. Our findings, that eating at some of these restaurants may pose an increased risk to health compared to dining out at other establishments, raises the question of whether food guides should consider aspects of food safety alongside the clearly important complementary focus on taste and other aspects of the dining experience. Our study also gives some oversight into the range of foods that may contribute to the risk of outbreaks in these high end restaurants and may offer a steer to the more cautious diner’s selections in these settings. Competing Interest The authors have declared that no competing interests exist. Acknowledgments The authors would like to acknowledge the contribution of colleagues from the Public Health England Health Protection Teams for providing the data used for the analysis. Epidemgiological scientist at Public Health England ==== Refs References 1 Waitrose. 2015 Top 50 Restaurants 2015 [cited 2015 01/12/2014]. Available from: http://www.thegoodfoodguide.co.uk/news/2015-top-50-announced. 2 Gormley FJ, Little CL, Rawal N, Gillespie IA, Lebaigue S, Adak GK. A 17-year review of foodborne outbreaks: describing the continuing decline in England and Wales (1992-2008). Epidemiology and infection. 2011;139(5):688-99. 3 Smith AJ, McCarthy N, Saldana L, Ihekweazu C, McPhedran K, Adak GK, et al. A large foodborne outbreak of norovirus in diners at a restaurant in England between January and February 2009. Epidemiology and infection. 2012;140(9):1695-701 4 Baker K, Morris J, McCarthy N, Saldana L, Lowther J, Collinson A, et al. An outbreak of norovirus infection linked to oyster consumption at a UK restaurant, February 2010. Journal of public health. 2011;33(2):205-11. 5 Westrell T, Dusch V, Ethelberg S, Harris J, Hjertqvist M, Jourdan-da Silva N, et al. Norovirus outbreaks linked to oyster consumption in the United Kingdom, Norway, France, Sweden and Denmark, 2010. Euro surveillance. 2010;15(12). 6 Farmer S, Keenan A, Vivancos R. Food-borne Campylobacter outbreak in Liverpool associated with cross-contamination from chicken liver parfait: implications for investigation of similar outbreaks. Public health. 2012;126(8):657-9. 7 Doyle A, Barataud D, Gallay A, Thiolet JM, Le Guyaguer S, Kohli E, et al. Norovirus foodborne outbreaks associated with the consumption of oysters from the Etang de Thau, France, December 2002. Euro surveillance. 2004;9(3):24-6. 8 Gillespie IA, Adak GK, O'Brien SJ, Brett MM, Bolton FJ. General outbreaks of infectious intestinal disease associated with fish and shellfish, England and Wales, 1992-1999. Communicable disease and public health. 2001;4(2):117-23. 9 Haba JH. Incidence and control of Campylobacter in foods. Microbiologia. 1993;9 Spec No:57-65. 10 Little CL, Gormley FJ, Rawal N, Richardson JF. A recipe for disaster: outbreaks of campylobacteriosis associated with poultry liver pate in England and Wales. Epidemiology and infection. 2010;138(12):1691-4. 11 Gormley FJ, Rawal N, Little CL. Choose your menu wisely: cuisine-associated food-poisoning risks in restaurants in England and Wales. Epidemiology and infection. 2012;140(6):997-1007. 12 Ryan MJ, Wall PG, Gilbert RJ, Griffin M, Rowe B. Risk factors for outbreaks of infectious intestinal disease linked to domestic catering. Communicable disease report CDR review. 1996;6(13):R179-83.

PMC005xxxxxx/PMC5001642.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 27564393PONE-D-16-1880410.1371/journal.pone.0161788Research ArticlePhysical SciencesMathematicsStatistics (Mathematics)Statistical DataResearch and Analysis MethodsBioassays and Physiological AnalysisElectrophysiological TechniquesBrain ElectrophysiologyElectroencephalographyBiology and Life SciencesPhysiologyElectrophysiologyNeurophysiologyBrain ElectrophysiologyElectroencephalographyMedicine and Health SciencesPhysiologyElectrophysiologyNeurophysiologyBrain ElectrophysiologyElectroencephalographyBiology and Life SciencesNeuroscienceNeurophysiologyBrain ElectrophysiologyElectroencephalographyBiology and Life SciencesNeuroscienceBrain MappingElectroencephalographyMedicine and Health SciencesDiagnostic MedicineClinical NeurophysiologyElectroencephalographyResearch and Analysis MethodsImaging TechniquesNeuroimagingElectroencephalographyBiology and Life SciencesNeuroscienceNeuroimagingElectroencephalographyResearch and Analysis MethodsSimulation and ModelingResearch and Analysis MethodsMathematical and Statistical TechniquesStatistical MethodsForecastingPhysical SciencesMathematicsStatistics (Mathematics)Statistical MethodsForecastingBiology and Life SciencesAnatomyHeadEyesMedicine and Health SciencesAnatomyHeadEyesBiology and Life SciencesAnatomyOcular SystemEyesMedicine and Health SciencesAnatomyOcular SystemEyesBiology and Life SciencesNeuroscienceCognitive ScienceArtificial IntelligenceMachine LearningComputer and Information SciencesArtificial IntelligenceMachine LearningPhysical SciencesMathematicsDiscrete MathematicsCombinatoricsPermutationComputer and Information SciencesData VisualizationInfographicsGraphsAn Efficient Data Partitioning to Improve Classification Performance While Keeping Parameters Interpretable An Efficient Data Partitioning to Improve Classification PerformanceKorjus Kristjan 1*Hebart Martin N. 2Vicente Raul 1 1 Computational Neuroscience Lab, Institute of Computer Science, University of Tartu, Tartu, Estonia 2 Department of Systems Neuroscience, Universitätsklinikum Hamburg-Eppendorf, Hamburg, Germany Hsiao Chuhsing Kate Editor National Taiwan University, TAIWAN Competing Interests: The authors have declared that no competing interests exist. Conceptualization: KK RV MNH. Data curation: KK. Formal analysis: KK. Funding acquisition: RV. Investigation: KK RV. Methodology: KK RV MNH. Project administration: RV. Resources: KK MNH RV. Software: KK MNH. Supervision: RV MNH. Validation: KK. Visualization: KK. Writing – original draft: KK RV MNH. Writing – review & editing: KK RV MNH. * E-mail: korjus@gmail.com2016 26 8 2016 11 8 e01617889 5 2016 11 8 2016 © 2016 Korjus et al2016Korjus et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Supervised machine learning methods typically require splitting data into multiple chunks for training, validating, and finally testing classifiers. For finding the best parameters of a classifier, training and validation are usually carried out with cross-validation. This is followed by application of the classifier with optimized parameters to a separate test set for estimating the classifier’s generalization performance. With limited data, this separation of test data creates a difficult trade-off between having more statistical power in estimating generalization performance versus choosing better parameters and fitting a better model. We propose a novel approach that we term “Cross-validation and cross-testing” improving this trade-off by re-using test data without biasing classifier performance. The novel approach is validated using simulated data and electrophysiological recordings in humans and rodents. The results demonstrate that the approach has a higher probability of discovering significant results than the standard approach of cross-validation and testing, while maintaining the nominal alpha level. In contrast to nested cross-validation, which is maximally efficient in re-using data, the proposed approach additionally maintains the interpretability of individual parameters. Taken together, we suggest an addition to currently used machine learning approaches which may be particularly useful in cases where model weights do not require interpretation, but parameters do. http://dx.doi.org/10.13039/501100002301Eesti TeadusagentuurPUT438Vicente Raul RV thanks the financial support from the Estonian Research Council through the personal research grants program (PUT438 grant). Website of the funder: http://www.etag.ee/. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction The goal of supervised machine learning, in particular classification, is to find a model that accurately assigns data to separate predefined classes. To test the generality of a learned model, this model is typically applied to independent test data, and the accuracy of the prediction informs a researcher about the quality of the classifier [1]. Finding a classifier that performs optimally according to the researcher’s objective requires a set of assumptions and also a trade-off in model complexity: Too simple parameters lead to under-fitting, i.e. the model is not able to account for the complexity of the data. Too complex parameters at the same time lead to over-fitting, i.e. the model is too complex and fits to noise in the data. To test different assumptions and also optimize this so-called bias-variance trade-off [2], it is quite common to divide a data set into three parts: (1) a training set and (2) a validation set, which together are used iteratively for optimizing the parameters of the chosen classifier, and (3) a separate test set to validate the generalization performance of the final classifier. Machine learning methods in life sciences are used with different objectives: At one end of the spectrum, the goal is making predictions in real-world applications and building a maximally predictive model with interpretable weights and parameters that can be used in future applications. At the other end, the goal is to make an inference about the presence of information, where even the slightest discrimination performance indicates a statistical dependence between independent and dependent variables (e.g. classes and data). In the latter approach, the interpretability of weights and parameters and their reuse is not the focus of the research and performance is commonly evaluated using statistical analyses (e.g. [3]). This latter approach is quite common in the field of neuroimaging [4] and bioinformatics [5]. Data collection can be very expensive in biological and social sciences, and over time more data-efficient methods have emerged [6]. Cross-validation is a method that makes near-optimal use of the available data by repeatedly training and testing classifiers on different subsets of data, typically with a large training and a small validation set in each iteration [7]. For example, in 5-fold cross-validation 80 percent of the data are used for training, 20 percent for validation, and in the next iteration another 20 percent of data are chosen as a test set, etc. This process is repeated five times until all data have served as validation data once. Cross-validation is repeated with different parameter combinations, and once the best parameters have been found, the model is trained with the chosen parameters on all data that have previously been used for cross-validation and applied to the separate test set (see Fig 1). When the goal of a researcher is to build a model that generalizes well to unseen cases and that may be used in real life applications such as image recognition or text classification, this approach is probably the most generally used method for carrying out classification analyses, with five to ten fold cross-validation resulting in a good trade-off between over-fitting and sufficient training size of the classifier [8]. 10.1371/journal.pone.0161788.g001Fig 1 In the “Cross-validation and testing” approach, the data are divided into two separate sets (cross-validation set and test set) only once. First, different models are trained and validated with cross-validation and the best set of parameters is chosen. Prediction accuracy and statistical significance of the parameters are evaluated on the test set, after training on the cross-validation set. One difficulty of this approach is that the test set that is used to validate classification performance is limited in size. While cross-validation makes good use of training data, the estimation of the generalization performance of the classifier may still suffer by this limited size of the final test set. Increasing the size of the test set at the same time would come at the cost of diminishing classification performance. When data are scarce or expensive to acquire, this can become a large problem and may lead to a sub-optimal choice of classifiers and the associated parameters. One approach that has been used to overcome this difficulty is “Nested cross-validation” [9] (see Fig 2). Here, the test set is not kept completely separate, but cross-validation is extended to incorporate all available data (outer cross-validation). In that way, all data can serve as test set once, overcoming the aforementioned trade-off. In order to still be able to optimize parameters, for a particular cross-validation iteration the training set is again divided for nested cross-validation (inner cross-validation), and once the best parameters for this iteration have been found, they are used to train a model on the current training data, which is then applied to the current test set. This approach is most useful for a researcher who does not need to build a model that generalizes well to unseen data, but who would like to describe whether there is a meaningful statistical dependence between the class labels and the given dataset, in other words whether the dataset contains information about the labels. 10.1371/journal.pone.0161788.g002Fig 2 In the “Nested cross-validation” approach, first (outer) cross-validation is performed to estimate predictability of the data. In each iteration, data are divided into training and test sets. Before training, another (inner) cross-validation loop is used to optimize parameters. As model weights (fitted models) and parameters are different at every partition, it is not possible to report accuracy or statistical significance about a particular set of parameters or model weights. While a Nested cross-validation procedure makes more efficient use of data, it has some drawbacks: Due to the absence of a completely separate test set it is not possible to claim that a particular model, i.e. a particular set of weights and classifier parameters, could in the future be used to classify unseen data [2]. In addition, the chosen parameters and models may vary between cross-validation iterations, making it impossible to select one set of parameters or one model as the final choice. In other words, a separate model and a separate set of parameters are chosen in each iteration and choosing any one of them would mean returning to a simple cross-validation and testing approach which would annihilate the advantage gained by nested cross-validation. There are, however, cases in which the interpretation of parameters is desirable, even when the model is not used. For example, for certain applications it might be useful to report that the best parameter corresponds to linear models as opposed to quadratic ones, without the need to describe the specific model weights used by the linear models. In another example, when using neural network as the class of machine learning algorithms, the number of layers selected during the optimization, say 3, 4 or 5 layers, may be an important choice to be communicated to other researchers and that may lend some interpretation to the best combination of parameters and data. Here we describe a novel approach that improves the trade-off between training and test size for better generalization performance than “Cross-validation and testing”, while, in contrast to “Nested cross-validation”, maintains the interpretability of chosen parameters. In the case of the widely used “Cross-validation and testing” approach, a larger test set results in less data for choosing the best set of parameters and also less data for fitting the model. In contrast, with our newly proposed “Cross-validation and cross-testing” approach (Fig 3) a larger test set still means that we are left with less data for parameter selection, but it does not reduce the amount of data available for model fitting. This comes at the cost of losing the ability to generate a single predictive model that is used in the future for general application. However, for many researchers it is sufficient to (1) demonstrate that information is indeed present in the dataset [4] or (2) interpret the parameters of the classifier. The latter can also guide a number of important choices for future design of classifiers. The novel approach suggested in this work improves the trade-off by using data more efficiently for fitting the model but makes it still possible to choose, interpret and report a set of parameters that may be used in the future. In brief, it is a mixture of the approaches of “Cross-validation and testing” and “Nested cross-validation”, which allows reusing test data to improve the size of the training set, thereby improving classification performance. We term this method “Cross-validation and cross-testing”. 10.1371/journal.pone.0161788.g003Fig 3 For cross-validation and cross-testing, data are divided into two separate sets only once: a cross-validation set and a test set. Similar to typical cross-validation, a number of iterations are carried out to choose the best parameters for the final model on the test set. Once the best combination of parameters has been chosen, the prediction accuracy and statistical significance can be evaluated on the test set with a modified cross-validation such that for each fold the original cross-validation set is repeatedly added to the training data. Due to the similarity to cross-validation, we term this approach cross-testing. While making it impossible to pick one final model on additional unseen data, the parameters that have been chosen remain interpretable. The current article is structured as follows. In the methods section we will describe the data sets and parameters used in order to test the applicability of the proposed approach on simulated and real data. In the first part of the results section we will then outline the novel approach of “Cross-validation and cross-testing” in detail. In the second part of the results section, we will compare all approaches using numerical experiments described in the methods part and show that the novel approach has a superior classification performance than the common “Cross-validation and testing” approach. We also demonstrate that the novel pipeline is not biased by applying it to simulated data for which ground-truth is known. Methods In the present work we compare three different approaches for machine learning analysis methods: “Cross-validation and testing”, “Nested cross-validation”, and the newly proposed method “Cross-validation and cross-testing”. We predict that in terms of generalization performance, “Nested cross-validation” will outperform both other methods, at the cost of making the chosen parameters and weights uninterpretable. “Cross-validation and testing” is predicted to generate the worst performance, while allowing interpretation of both optimal parameters and the weights of the final model for generating future predictions on additional unseen data. Finally, we predict that “Cross-validation and cross-testing” will have prediction performance between the other two approaches, yielding no interpretable model weights, but fully interpretable model parameters. We will compare these different methods on four different data sets. Our approach is as follows. Each time, a subset of data is randomly sampled from a larger data set which serves as a reference data set. The three different methods are then evaluated using this reference data set. The mean classification accuracy is compared against chance, using a statistical cutoff of p < 0.05. This procedure is repeated 1,000 times and the proportion of significant results and the average accuracy are reported. Parameters To illustrate our theoretical predictions, we applied the three analysis approaches for optimization of two parameters. Firstly, for pre-processing the choice was between no preprocessing or performing a dimensionality reduction using principal component analysis (PCA) [10] with 70% of cumulative variance preserved. Secondly, three possible values (0.0001, 0.01 and 1) can be chosen for the penalty for misclassification (“c”) of the Support Vector Machine (SVM) algorithm [11, 12]. The combination of the two parameters give rise to a total of 6 possible parameter combinations. To speed the computational time hereafter we restrict ourselves to the before-mentioned 6 parameter combinations but many other parameters can be considered such as the type of normalization of the data set or the different types of classifier models. The k-fold parameter, the number of random partitions used in cross validations, is fixed to 5 in all the cases. We would also like to note that the term parameter has been used inconsistently in the literature, sometimes referring to the individual weights of a given model, and sometimes to the parameters that are used to optimize the learning algorithm. Here, we use the term parameter to refer to a variable that is used to optimize classification performance (which may incorporate choices not directly applied to a particular classifier, including the choice of pre-processing or the choice of classifier). For parameters related to the model itself, we use the term model weights. Please note that although our novel approach makes it possible to interpret the chosen parameters, the chosen parameters are not reported, for two reasons. First, they have been chosen quite arbitrarily for illustrative purposes only. Second, they have been repeatedly selected based on subsets of data that have been repeatedly sampled. For those reasons, we believe that interpretation of parameters in this context is not very meaningful. Size of the test set The “Cross-validation and testing” and “Cross-validation and cross-testing” approaches both have an important hyper-parameter, size of the test set, that must be determined before proceeding to any further analysis. With each of our 4 data sets, the trade-off between test and training data was varied in two different ways. First, we varied the size of the whole data set while fixing the test set proportion to 50%. Second, we kept the total amount of data fixed and examined how the test set size affected the results. The size of the full data set was fixed between two extremes (always and never finding a statistically significant result) in order to illustrate the differences in the relative sizes of the test set. The size of the data set was fixed to 50 for EEG data and to 100 in all other cases. The following sizes for the test set were used: 10%, 30%, 50%, 70% and 90%. Datasets We used four datasets, three of which contain information and are classifiable. The three classifiable datasets consist of a simulated data set and and two datasets from neural recordings. The neural datasets contain two types of electrophysiological data: electroencephalogram (EEG) and spiking data. The fourth dataset consists of randomly generated data with no classifiable information. A brief description of the datasets is given below. More information about the collection of the datasets can be found in the respective references. All datasets can be found in the Supplementary Information, S1 Public Repository. Simulated data We generated 4,000 data samples of two classes each (A and B) with 6 dimensions (features) sampled from a uniform distribution between 0 and 1. For class A, 2,000 data points were left unchanged, but we added a scalar 0.8 to the first two dimensions of the remaining 2,000 data points. For class B, we added 0.8 to the first dimension of the first 2,000 data points and 0.8 to the second dimension of the remaining 2,000 data points. These modifications make the classes linearly inseparable. The final generated dataset has 6 features with 8,000 data points from both of the two classes. EEG data The resting state EEG data analyzed here were collected prior to 12 different cognitive EEG experiments conducted at the University of Tartu. [13] The measurements took place in a dimly lit and quiet room. Participants sat in a comfortable office chair 1 or 1.15 m away from a computer screen. They were instructed to relax and avoid excessive body and eye movements. There were two conditions: eyes open and eyes closed. During the eyes open condition subjects were also required to fixate on a black cross in the middle of a gray screen. From the EEG time series the power spectral density values were computed using the Fourier transform with Hamming tapered 2 s windows. The two labels of the data are “eyes open” and “eyes closed”. The features consist of 91 power spectral density values from 5 occipital electrodes (‘PO3’ ‘PO4’, ‘O1’, ‘Oz’, ‘O2’). The final “EEG dataset” consists of 455 features with 289 samples from two classes. Spike train data The neural spiking dataset contains multiple single unit recordings from different rat hippocampal and entorhinal regions while the animals were performing multiple behavioral tasks. In particular, we used a single session of a single rat from the CA1 region in hippocampus. The dataset (“hc3”) is accessible from the neural data repository crcns.org [14]. The data used includes spikes of 61 neurons, and we sampled 2,000 data points while the rat was engaged in active spatial navigation in a square-shaped arena. We divided the square into two areas to assign two types of labels to each neural recording data point according depending on whether at that moment the rat was at the “upper part” or “lower part” of the arena. In total, the “spike train dataset” has 61 features with 2,000 data points from two classes. Random data We generated 10,000 data samples with a random assignation to two classes. Each data point has 20 dimensions sampled from a uniform distribution between 0 and 1. Thus, the “random dataset” has 20 features with 10,000 data points from two classes. Statistical significance In addition to reporting general classification accuracy for each of the datasets, we also report the proportion of results that were found to indicate accuracies significantly above chance. In the context of classification, statistical significance indicates whether a statistical dependence between labels and data can be assumed. A statistically significant results would reject the null hypothesis that the association of data and labels comes about by chance, and the p-value would indicate the probability of false rejection of the null hypothesis. Statistical significance was evaluated with a permutation test (number of permutations was fixed to 1,000 and statistical significance level to 0.05) [15]. It is worth noting that our datasets are sometimes very small and we have to repeat the permutation test many times. For that reason, it becomes important to correct the p-value for the bias introduced by the discrete nature of classification results as described in [16]. If in one permutation test the resulting p-value was in between two possible p-values around the significance level a probabilistic approach was used to determine if the given run was significant or not. Otherwise the nominal alpha level could never be approximated. The implementation of it can be found from the code in S1 Public Repository. The effect of the correction was small and did not change the results. Results In this section, we will describe the novel approach that we term “Cross-validation and cross-testing”. In particular, we will explain the algorithm and compare it with two other common approaches known as “Cross-validation and testing”, and “Nested cross-validation”. We note that the proposed approach is a natural interpolation between the two classical approaches. In the second part of this section, we will apply the proposed pipeline to simulated data as well as biological datasets. Cross-validation and cross-testing When a researcher is interested in publishing their model parameters, then typically the efficient and popular approach called “Cross-validation and testing” is used (Fig 1). In the cross-validation set, the best parameters are chosen—usually according to highest cross-validation accuracy—and the test set is used for out-of-sample accuracy estimation. An even more data-efficient approach for data analysis is called “Nested cross-validation” (Fig 2). The approach is similar to cross-validation and testing, but the test set becomes part of an outer cross-validation loop, while parameters are optimized in inner cross-validation iterations, using only the training data of the current outer cross-validation iteration. The whole data set can be used for estimating the final accuracy and therefore has a maximum statistical power for significance analysis. However, this approach does not make it possible to publish parameters which might be sometimes desirable as discussed in the Introduction section. Our proposed pipeline can be seen as a natural extension between the two extremes described before. See Table 1 for a comparison of the three approaches in terms of data efficiency and parameter and model weights interpretability. 10.1371/journal.pone.0161788.t001Table 1 Comparison of the approaches. Approach Data efficiency Possible to interpret parameters Possible to interpret fitted model Cross-validation and testing Low Yes Yes Proposed: Cross-validation and cross-testing Intermediate Yes No Nested cross-validation High No No Comparison of the newly proposed “Cross-validation and cross-testing”, classical “Cross-validation and testing” and “Nested cross-validation” with respect to data efficiency, and parameter and model interpretability. “Cross-validation and cross-testing” starts by carrying out the common “Cross-validation and testing” approach: The best parameters are chosen with cross-validation as described previously. Once the parameters are fixed, the remaining data are used for testing the classifier. The novelty of the approach is introduced by how prediction accuracy and statistical significance are evaluated (see Fig 3). Rather than keeping the test set entirely separate, a modified training set is iteratively introduced, where in each iteration the original training data plus part of the originally separate test data are used for training a classifier, and the remaining test data for testing the classification performance. We term this iterative procedure “cross-testing”, because this term more accurately describes the process that is repeated than “cross-validation”. Importantly, this approach maintains independence between training and test data, but makes more efficient use of test data by augmenting training data for more accurate predictions. In the next iteration, a different part of test data is added to the training data, and this process is repeated until each part of the test data has once been added to training data. The mean prediction accuracy across these different cross-validation iterations is then averaged. We refer to this novel approach as “Cross-validation and cross-testing”. This proposed approach is expected to provide more accurate results than classical “Cross-validation and testing” by construction because it simply uses more data for model fitting, while the system for choosing the best set of parameters remains the same. Simulated data To compare the three approaches we first analyzed the simulated data set containing two classes that are linearly inseparable. We start by varying the data set size while fixing the test set size at 50%. It can be seen from Fig 4 that larger data set size increases the average accuracy and the proportion of significant results. We can also see that “Nested cross-validation” provides a larger percentage of significant results. The approach using the proposed “Cross-validation and cross-testing” gives significant results more often that the popular “Cross-validation and testing” approach. Thus, while only “Cross-validation and testing” and the proposed approach “Cross-validation and cross-testing” make it possible to report parameters, the latter is more sensitive since more data is effectively used for model fitting. 10.1371/journal.pone.0161788.g004Fig 4 Batches of simulated data of different sizes are analyzed 1000 times with three different approaches. Results show the mean accuracy (upper plot) and proportion of significant results (bottom plot) out of the 1000 runs. More data leads to higher average accuracy and increases the proportion of significant results. “Nested cross-validation” outperforms other approaches and the “cross-validation and testing” gives the worst performance in terms of average accuracy and proportion of significant results. The “Cross-validation and testing” and “Cross-validation and cross-testing” approaches are affected by the size of the separate test set, which must be determined before proceeding to any further analysis. Therefore, we tested the effect of the test set size by fixing the data set size at 100 data points and using different proportions of the 100 data points for testing. The clearly visible downward trend on the first graph in Fig 5 shows that using a larger test set decreases the average accuracy because there is less data for choosing the best parameters. The observed difference in average accuracy is due to the test data resampling. As expected, the higher average accuracy is also reflected in the bottom graph on the Fig 5 by more often leading to significant results. 10.1371/journal.pone.0161788.g005Fig 5 Batches of 100 simulated data points were analyzed 1,000 times with both approaches that contain a separate test set. Results show the mean accuracy (upper plot) and proportion of significant results (bottom plot) out of the 1000 runs. A larger test set leads to smaller average accuracy because there is less data for choosing parameters and fitting a model. “Cross-validation and cross-testing” outperforms “cross-validation and testing” in terms of average accuracy and proportion of significant results as expected. Usually for the popular approach “Cross-validation and testing” it is recommended that the size of the test set should be about 10% to 50%. From the bottom graph on the Fig 5 we can see that for this case it happened that the optimal point is more close to 70%. Interestingly, the proposed approach “Cross-validation and cross-testing” has a different optimal size for the test set. As test data resampling changes the trade-off, we can observe from the bottom graph of Fig 5 that for our proposed approach a test set size around 90% would give a significant result with highest probability. Real data We also analyzed two neuroscience datasets: EEG data from humans and spiking data from rat hippocampus. From Fig 6 we can observe the same trends as above: larger data set sizes imply higher average accuracy and more frequent significant results. We can also see that the approach “Nested cross-validation” gives significant results most often, followed by the proposed approach “Cross-validation and cross-testing”, which gives significant results more often than the popular “Cross-validation and testing” approach. The effect is smaller with EEG data suggesting that more efficient usage of data in the model fitting is not critical for this dataset (eyes open and closed conditions can be clearly distinguishable from few samples of EEG power spectra) and the choice of parameters is actually the main determinant for the classifier performance. 10.1371/journal.pone.0161788.g006Fig 6 Analysis of real data (left: EEG dataset; right: spiking dataset) with three different approaches as a function of data size with test set size fixed at 50%. Results show the mean accuracy (upper graphs) and proportion of significant results (bottom graphs) out of the 1000 runs. More data lead to higher average accuracy and increases the proportion of significant results. “Nested cross-validation” outperforms other approaches while “Cross-validation and testing” gives the worst performance in terms of average accuracy and proportion of significant results. The effect is smaller with EEG data suggesting that more efficient usage of data in the model fitting is not that important and the choice of parameters is actually the main influencer. Fig 7 shows that “Cross-validation and cross-testing” outperforms “Cross-validation and testing” in terms of accuracy and proportion of significant results for a large range of relative test set sizes. Also, as expected, larger test set size reduces the average accuracy because there are less data for choosing the parameters. Interestingly, optimal size for the test set for EEG data is about 30% for both approaches but for the spiking data optimal size for the test set changes from 50% to 70% if approach is changed from “Cross-validation and testing” to “Cross-validation and cross-testing”. We also note that in the limit of small test sets both approaches should converge to the same result on average. This explains the crossover of “Cross-validation and testing” and “Cross-validation and cross-testing” in the average accuracy of the electroencephalogram data (top left) where for very small data sets noise begins to dominate results. 10.1371/journal.pone.0161788.g007Fig 7 Analysis of neuroscience data (left: EEG dataset; right: spiking dataset) with three different approaches as a function of the relative test set size. Results show the mean accuracy (upper graphs) and proportion of significant results (bottom graphs) out of the 1000 runs. Data set size was fixed to 50 for EEG and to 100 for spikes train data set. Larger test set leads to smaller average accuracy because there is less data for choosing parameters and fitting a model. “Cross-validation and cross-testing” outperforms “cross-validation and testing” in terms of average accuracy and proportion of significant results. Random data To validate our whole set-up we also tested it on noisy data which was randomly assigned to 2 classes, and thus no information should be classifiable for this dataset beyond chance level. As seen from Fig 8, the average accuracy is about 50% as expected for random data and the proportion of significant results stays around 0.05 which is also expected as we set of significance threshold to 0.05. 10.1371/journal.pone.0161788.g008Fig 8 Accuracy and proportion of significant results are kept at chance levels when applying the novel approach to random data. Discussion and Conclusion When using machine learning algorithms for making predictions, improving performance of a classifier can be seen as a central goal. At the same time, interpretability of models and parameters beyond the given data are in many cases desirable. Since data are often scarce or expensive to acquire, efficient use of data is another important objective. These three goals—generalization performance, interpretability, and efficient use of data—often lead to a trade-off that is resolved depending on the focus of the researcher. If the focus lies on generating a model that will generalize well to unseen data, this requires an interpretable model and interpretable model parameters. For that purpose, data can be used efficiently by the common approach known as “Cross-validation and testing”. If, however, the goal is to maximize classification performance for finding a statistical dependence between class labels and data, then a better performance can be achieved by using “Nested cross-validation”, which provides a very data efficient approach by repeatedly re-using data. However, this approach does not naturally allow the interpretation of parameters and model weights. Previously, researchers had to choose from these two extremes depending on their goal. In this article, we described a novel approach that uses cross-validation to find and fix the best combination of parameters, but importantly which then resamples test data for augmenting training data, yielding to more accurate estimation of generalization performance. We termed this new approach “Cross-validation and cross-testing” and predicted that it would outperform outperforms “Cross-validation and testing” in terms of accuracy and statistical sensitivity with simulated and empirical data sets. In particular, we tested the effects of different data sets, data set sizes and test set sizes. Indeed, we confirmed in various numerical experiments that too large test sets quickly result in insufficient data for finding the best parameters and not enough data for fitting the best model. On the other hand, too small test sets can imply that there are not enough data for achieving statistically significant results. This trade-off for the size of the test set results in the existence of an optimal range. The proposed “Cross-validation and cross-testing” modifies the range to allow larger test set sizes because with the novel cross-testing part it is still possible to use almost all of the data for model fitting. As predicted, we demonstrate the superiority of “Cross-validation and cross-testing” over “Cross-validation and testing” both in terms of accuracy and in terms of statistical sensitivity. On simulated data where half of the data served as test data we demonstrated that “Cross-validation and cross-testing” performed similar to “Nested cross-validation” in terms of accuracy, but outperformed “Cross-validation and testing” irrespective of data size. This improved performance can be explained by the much larger data set that is available during each testing iteration of “Cross-validation and cross-testing”. In terms of statistical significance, “Cross-validation and cross-testing” constantly outperformed “Cross-validation and testing”, but demonstrated particular sensitivity for intermediate data set sizes, demonstrating the particular usefulness of “Cross-validation and cross-testing” for small to intermediate data sets. Varying the test set size also demonstrated that larger test sets can be superior for the novel approach; however, this depends on the structure of the data and whether parameters can be estimated well even when using very little training data. It is not generally advisable to maximize the test set at the cost of the training set size. Indeed, for empirical neurophysiological data we found similar results as for simulated data in terms of accuracy and statistical sensitivity. Again, “Cross-validation and cross-testing” generally outperformed conventional “Cross-validation and testing”. Here, however, the test set size trade-off did not favor a maximally large test set for the novel approach, but varied between the different data sets. The EEG classification of eyes open and eyes closed resulted in a high accuracy and displayed a sweet spot between the limits of little training data and little test data, with a significant improvement compared with the classical cross-validation and testing approach. The neuronal spiking data from hippocampus is much more variable and probably non-linearly separable with the features given but exactly the same trends were observed with respect to the dependency of the different approaches on training and test sizes. In general, for the empirical data sets used in this article and for maximizing statistical sensitivity, the optimal test set size was around 50%. However, additional simulations and analyses are required before recommendations for maximal performance can be made. This, however, is beyond the scope of the present work. Finally, as a sanity check the analyses were repeated on random data to demonstrate that “Cross-validation and cross-testing” did not introduce any statistical bias. Indeed, the proportion of significant results was around the expected level of significance, demonstrating that the novel approach was not statistically biased. While both “Cross-validation and testing” and the proposed “Cross-validation and cross-testing” make it possible to report parameters, the latter, novel approach uses data more efficiently in the final model fitting phase by reducing the test set size trade-off. This change of the trade-off occurs by reducing the detrimental effect of a larger test set size on the quality of the fitted model by augmenting the size of the training data. In terms of computational cost, our approach is in between the “Cross-validation and testing” and “Nested cross-validation”. Firstly, each model with every combination of model parameters is trained k times in the k-fold cross-validation. And secondly, the model with the chosen combination of parameters is additionally trained k′ times in the k′-fold cross-testing. We would also like to note that the data partitioning scheme proposed here is compatible with most machine learning models (including familiar classifiers and regressions). In the paper, we showed the method using SVM for illustration purposes. It is our hope that the proposed approach can be a useful addition to the toolkit of machine learning approaches. We believe that it might be especially applicable when both data efficiency and parameter interpretations are desired. Supporting Information S1 Public Repository We have included the full source code, graphs and data files for reproducibility. And also source files for the Figs 1, 2 and 3 which schematically explain different approaches. All the material can be accessed via Github repository: https://github.com/kristjankorjus/machine-learning-approaches. (ZIP) Click here for additional data file. We would like to thank Andero Uusberg and Kairi Kreegipuu for sharing the EEG data set. We would also like to thank George Prichard, Ardi Tampuu, Jaan Aru, and Ilya Kuzovkin for reading and commenting the manuscript. RV also thanks the financial support from the Estonian Research Council through the personal research grants program (PUT438 grant). ==== Refs References 1 Alpaydin E . Introduction to machine learning . MIT press ; 2014 8 22 . 2 Hastie T , Tibshirani R , Friedman J , Franklin J . The elements of statistical learning: data mining, inference and prediction . The Mathematical Intelligencer . 2005 6 1 ;27 (2 ):83 –5 . 10.1007/BF02985802 3 Allefeld C , Görgen K , Haynes JD . Valid population inference for information-based imaging: From the second-level t-test to prevalence inference . NeuroImage . 2016 7 20 10.1016/j.neuroimage.2016.07.040 4 Haynes JD , Rees G . Decoding mental states from brain activity in humans . Nature Reviews Neuroscience . 2006 7 1 ;7 (7 ):523 –34 . 10.1038/nrn1931 16791142 5 Larrañaga P , Calvo B , Santana R , Bielza C , Galdiano J , Inza I , Lozano JA , Armañanzas R , Santafé G , Pérez A , Robles V . Machine learning in bioinformatics . Briefings in bioinformatics . 2006 3 1 ;7 (1 ):86 –112 . 10.1093/bib/bbk007 16761367 6 Pereira F , Mitchell T , Botvinick M . Machine learning classifiers and fMRI: a tutorial overview . Neuroimage . 2009 3 31 ;45 (1 ):S199 –209 . 10.1016/j.neuroimage.2008.11.007 19070668 7 Molinaro AM , Simon R , Pfeiffer RM . Prediction error estimation: a comparison of resampling methods . Bioinformatics . 2005 8 1 ;21 (15 ):3301 –7 . 10.1093/bioinformatics/bti499 15905277 8 Kohavi, R. A study of cross-validation and bootstrap for accuracy estimation and model selection. In International joint conference on artificial intelligence, 1995 (pp. 1137-1143). 9 Varma S , Simon R . Bias in error estimation when using cross-validation for model selection . BMC bioinformatics . 2006 2 23 ;7 (1 ):91 10.1186/1471-2105-7-91 16504092 10 Wold S , Esbensen K , Geladi P . Principal component analysis . Chemometrics and intelligent laboratory systems . 1987 8 1 ;2 (1-3 ):37 –52 . 10.1016/0169-7439(87)80084-9 11 Boser BE, Guyon IM, Vapnik VN. A training algorithm for optimal margin classifiers. InProceedings of the fifth annual workshop on Computational learning theory 1992 Jul 1 (pp. 144-152). ACM. 12 Chang CC , Lin CJ . LIBSVM: a library for support vector machines . ACM Transactions on Intelligent Systems and Technology (TIST) . 2011 4 1 ;2 (3 ):27 . 13 Korjus K , Uusberg A , Uusberg H , Kuldkepp N , Kreegipuu K , Allik J , Vicente R , Aru J . Personality cannot be predicted from the power of resting state EEG . Frontiers in human neuroscience . 2015 ;9 10.3389/fnhum.2015.00063 25762912 14 Teeters JL , Harris KD , Millman KJ , Olshausen BA , Sommer FT . Data sharing for computational neuroscience . Neuroinformatics . 2008 3 1 ;6 (1 ):47 –55 . 10.1007/s12021-008-9009-y 18259695 15 Efron B , Tibshirani RJ . An introduction to the bootstrap . CRC press 1994 16 Ernst MD . Permutation methods: a basis for exact inference . Statistical Science . 2004 ;19 (4 ):676 –85 . 10.1214/088342304000000396

PMC005xxxxxx/PMC5001643.txt

==== Front PLoS Comput BiolPLoS Comput. BiolplosploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, CA USA 2756409410.1371/journal.pcbi.1005090PCOMPBIOL-D-16-00561Research ArticleBiology and Life SciencesNeuroscienceCognitive ScienceCognitive PsychologyLearningBiology and Life SciencesPsychologyCognitive PsychologyLearningSocial SciencesPsychologyCognitive PsychologyLearningBiology and Life SciencesNeuroscienceLearning and MemoryLearningBiology and Life SciencesNeuroscienceCognitive ScienceCognitionDecision MakingResearch and Analysis MethodsSimulation and ModelingAgent-Based ModelingComputer and Information SciencesSystems ScienceAgent-Based ModelingPhysical SciencesMathematicsSystems ScienceAgent-Based ModelingResearch and Analysis MethodsSimulation and ModelingBiology and Life SciencesNeuroscienceCognitive ScienceCognitive PsychologyLearningHuman LearningBiology and Life SciencesPsychologyCognitive PsychologyLearningHuman LearningSocial SciencesPsychologyCognitive PsychologyLearningHuman LearningBiology and Life SciencesNeuroscienceLearning and MemoryLearningHuman LearningPhysical SciencesAstronomical SciencesCelestial ObjectsPlanetsPhysical SciencesAstronomical SciencesPlanetary SciencesPlanetsResearch and Analysis MethodsMathematical and Statistical TechniquesMathematical ModelsRandom WalkPhysical SciencesMathematicsProbability TheoryProbability DistributionWhen Does Model-Based Control Pay Off? Reinforcement Learning Trade-Offshttp://orcid.org/0000-0002-4792-7044Kool Wouter 1*http://orcid.org/0000-0002-6929-9982Cushman Fiery A. 1Gershman Samuel J. 121 Department of Psychology, Harvard University, Cambridge, Massachusetts, United States of America2 Center for Brain Science, Harvard University, Cambridge, Massachusetts, United States of AmericaO'Reilly Jill X EditorOxford University, UNITED KINGDOMThe authors have declared that no competing interests exist. Conceived and designed the experiments: WK FAC SJG. Performed the experiments: WK. Analyzed the data: WK. Wrote the paper: WK FAC SJG. * E-mail: wkool@fas.harvard.edu26 8 2016 8 2016 12 8 e10050907 4 2016 1 8 2016 © 2016 Kool et al2016Kool et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Many accounts of decision making and reinforcement learning posit the existence of two distinct systems that control choice: a fast, automatic system and a slow, deliberative system. Recent research formalizes this distinction by mapping these systems to “model-free” and “model-based” strategies in reinforcement learning. Model-free strategies are computationally cheap, but sometimes inaccurate, because action values can be accessed by inspecting a look-up table constructed through trial-and-error. In contrast, model-based strategies compute action values through planning in a causal model of the environment, which is more accurate but also more cognitively demanding. It is assumed that this trade-off between accuracy and computational demand plays an important role in the arbitration between the two strategies, but we show that the hallmark task for dissociating model-free and model-based strategies, as well as several related variants, do not embody such a trade-off. We describe five factors that reduce the effectiveness of the model-based strategy on these tasks by reducing its accuracy in estimating reward outcomes and decreasing the importance of its choices. Based on these observations, we describe a version of the task that formally and empirically obtains an accuracy-demand trade-off between model-free and model-based strategies. Moreover, we show that human participants spontaneously increase their reliance on model-based control on this task, compared to the original paradigm. Our novel task and our computational analyses may prove important in subsequent empirical investigations of how humans balance accuracy and demand. Author Summary When you make a choice about what groceries to get for dinner, you can rely on two different strategies. You can make your choice by relying on habit, simply buying the items you need to make a meal that is second nature to you. However, you can also plan your actions in a more deliberative way, realizing that the friend who will join you is a vegetarian, and therefore you should not make the burgers that have become a staple in your cooking. These two strategies differ in how computationally demanding and accurate they are. While the habitual strategy is less computationally demanding (costs less effort and time), the deliberative strategy is more accurate. Scientists have been able to study the distinction between these strategies using a task that allows them to measure how much people rely on habit and planning strategies. Interestingly, we have discovered that in this task, the deliberative strategy does not increase performance accuracy, and hence does not induce a trade-off between accuracy and demand. We describe why this happens, and improve the task so that it embodies an accuracy-demand trade-off, providing evidence for theories of cost-based arbitration between cognitive strategies. Office of Naval Research (US)N00014-14-1-0800http://orcid.org/0000-0002-6929-9982Cushman Fiery A. National Science Foundation (US)CCF-1231216Gershman Samuel J This research was supported by grant N00014-14-1-0800 from the Office of Naval Research (http://www.onr.navy.mil/) and and based upon work supported by the Center for Brains, Minds and Machines (CBMM, https://cbmm.mit.edu/), funded by NSF STC award CCF-1231216. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll code for reproducing the simulations, the data, and code for the experiments are available at: https://github.com/wkool/tradeoffsData Availability All code for reproducing the simulations, the data, and code for the experiments are available at: https://github.com/wkool/tradeoffs ==== Body Introduction Theoretical accounts of decision making emphasize a distinction between two systems competing for control of behavior [1–6]: one that is fast and automatic, and one that is slow and deliberative. These systems occupy different points along a trade-off between accuracy and computational demand (henceforth demand), making each one suitable for particular task demands. This raises the problem of arbitration: how does the brain adaptively determine which system to use at any given time? Answering this question depends on models and experimental tasks that embody the accuracy-demand trade-off at the heart of dual-system models. Recent research formalizes the dual-system architecture in the framework of reinforcement learning [7, 8], a computational approach to value-guided decision-making that we describe in further detail below. The application of reinforcement learning methods to dual-process models of decision-making sparked an explosion of empirical and theoretical developments over the past decade because it offers a computationally precise characterization of the distinction between “automatic” and “controlled” processes for the task of value guided decision-making. Current research assumes that experimental methods grounded in reinforcement learning also capture a trade-off between accuracy (the proportion of value-maximizing actions) and computational demand (the minimization of computational effort and related costs), but this assumption remains largely untested (cf. [9]). Currently, the dominant method that aims to dissociate mechanisms of behavioral control within the reinforcement learning framework is the “two-step task” introduced by Daw, Gershman, Seymour, Dayan, and Dolan [8] (Fig 1A), which we describe in detail in the next section. This task has proven to be a useful and popular tool to characterize the neural [8, 10–18], behavioral [19–31] and clinical [32–35] implications of dual-process models within the reinforcement learning framework. However, in this paper we argue that the two-step task does not induce a trade-off between accuracy and demand: Our simulations show that the “deliberative” strategy does not increase performance accuracy on the task. These simulations mirror a recent report by Akam, Costa, and Dayan [9], who also show that the two-step task does not embody a trade-off between model-based control and reward. Here, we expand on that result by showing that it holds across an exhaustive range of reinforcement learning parameters. Furthermore, we show that the same shortcoming is present in other, more recent variants of the task that have been reported. We then identify five factors that collectively restrict the accuracy benefits posited to arise from model-based control. Finally, we describe a novel task that induces the accuracy-demand trade-off, while retaining the ability to dissociate formally between distinct processes of behavioral control. 10.1371/journal.pcbi.1005090.g001Fig 1 Design of the Daw two-step task. (A) State transition structure of the original two-step paradigm. Each first-stage choice has a high probability of transitioning to one of two states and a low probability of transitioning to the other. Each second-stage choice is associated with a probability of obtaining a binary reward. (B) To encourage learning, the second-stage reward probabilities change slowly over the course of the experiment. Dual-process models in the reinforcement learning setting The fundamental problem in reinforcement learning is estimation of state-action values (cumulative future reward), which an agent then uses to choose actions. In the dual-system theory, the fast and automatic system corresponds to a “model-free” reinforcement learning strategy, which estimates state-action values from trial-and-error learning [36, 37]. In essence, this strategy is an elaborated version of Thorndike’s Law of Effect: actions that previously led to reward are more likely to be taken in the future. The strategy is “model-free” because it has no representation of the environment’s causal structure (i.e., the transition function between states and the reward function in each state). Instead, it incrementally constructs a look-up table or function approximation from which values can be quickly computed. However, this strategy can lead to errors if the environment changes, because the entire value function must be incrementally updated to accommodate changes. In addition, the strategy can produce sub-optimal credit assignment [8], a property we explore below. These forms of brittleness illustrate how model-free learning gives rise to “habits”—fast but inflexible response tendencies stamped in by repetition. The slow and deliberative system corresponds to a “model-based” learning strategy that possesses operating characteristics complementary to the model-free strategy. This strategy learns an explicit causal model of the environment, which it uses to construct plans (e.g., by dynamic programming or tree search). In contrast to the habitual nature of the model-free strategy, the capacity to plan enables the model-based strategy to flexibly pursue goals. While more computationally expensive (hence slower and more effortful) than the model-free approach, it has the potential to be more accurate, because changes in the environment can be immediately incorporated into the model. The availability of a causal model also allows the model-based strategy to solve the credit-assignment problem optimally. This dual-system framework sketched above can account for important findings in the reinforcement learning literature, such as insensitivity to outcome devaluation following overtraining of an action-reward contingency [7, 38]. Furthermore, the framework has spurred a wealth of new research on the neural [8, 10–13, 39, 40] and behavioral implications of competition and cooperation between reinforcement learning strategies [19–25, 29, 41, 42]. How might the brain arbitrate between model-free and model-based strategies? Since the model-based strategy attains more accurate performance through effortful computation, people can (up to a point) increase reward by engaging this system. However, in time-critical decision making settings, the model-based strategy may be too slow to be useful. Furthermore, if cognitive effort enters into the reward function [43–45], then it may be rational to prefer the model-free strategy in situations where the additional cognitive effort of model-based planning does not appreciably increase reward. It has been hypothesized that this trade-off between accuracy and demand plays a pivotal role in the arbitration between the two strategies [38, 46–50], but so far direct evidence for arbitration has been sparse [51]. Methods The Daw two-step task Here, we will first describe in detail the design of the Daw two-step task, and the reinforcement-learning model of this task [8]. Next, we will show through computational simulations that model-based planning on this task does not yield increased performance accuracy. Finally, we will discuss several factors that contribute to this shortcoming in the current approach in this two-step task, and related paradigms. Experimental design In the two-step task, participants make a series of choices between two stimuli, which lead probabilistically to one of two second-stage states (Fig 1A). These second-stage states require a choice between stimuli that offer different probabilities of obtaining monetary reward. To encourage learning, the reward probabilities of these second-stage choices change slowly and independently throughout the task (Fig 1B), according to a Gaussian random walk (mean = 0, σ = 0.025) with reflecting boundaries at 0.25 and 0.75. Crucially, each first-stage option leads more frequently (70%) to one of the second-stage states (a “common” transition), whereas it leads to the other state in a minority of the choices (a “rare” transition). These low-probability transitions allow for a behavioral dissociation between habitual and goal-directed choice. Since the model-free strategy is insensitive to the structure the task, it will simply increase the likelihood of performing an action if it previously led to reward, regardless whether this reward was obtained after a common or rare transition. Choice dictated by the model-based strategy, on the other hand, reflects an interaction between the transition type and reward on the previous trial (Fig 2A). This strategy will decrease the tendency of repeating a first-stage action after a reward and a rare transition, since the alternative first-stage action is more likely to lead to the previously rewarded second-stage state (Fig 2B). Empirically, behavioral performance on this task reflects a mixture of these two strategies (Fig 2C). That is, the stay probability shows a main effect of reward, increasing when the previous trial was rewarded, but also shows the model-based crossover interaction between the previous transition and reward. 10.1371/journal.pcbi.1005090.g002Fig 2 Probability of repeating the first-stage choice for three agents. (A) For model-free agents, the probability of repeating the previous choice is dependent only on whether a reward was obtained, and not on transition structure. (B) Model-based behavior is reflected in an interaction between previous transition and outcome, increasing the probability of transitioning to the state where reward was obtained. (C) Behavioral performance on this task reflects features of both model-based and model-free decision making, the main effect of previous reward and its interaction with the previous transition. Computational model Behavior on the Daw two-step task can be modeled using an established dual-system reinforcement-learning model [7, 8, 40]. The task consists of three states across two stages (first stage: sA; second stage: sB and sC), all with two available actions (aA and aB). The model consists of model-based and model-free strategies that both learn a function Q(s, a) mapping each state-action pair to its expected discounted future return. On trial t, the first-stage state (always sA) is denoted by s1,t, the second-stage state by s2,t (sB or sC), the first- and second-stage actions by a1,t and a2,t, and the second-stage rewards as r1,t (always zero, there is only reward on the second stage) and r2,t. Model-free strategy. The model-free agent uses the SARSA(λ) temporal difference learning algorithm [52], which updates the value for each state-action pair (s, a) at stage i and trial t according to: QMF(s,a)=QMF(s,a)+αδi,tei,t(s,a) where δi,t=ri,t+QMF(si+1,t,ai+1,t)−QMF(si,t,ai,t) is the reward prediction error, α is the learning rate parameter (which determines to what degree new information is incorporated), and ei,t(s,a) is an eligibility trace set equal to 0 at the beginning of each trial and updated according to ei,t(si,t,ai,t)=ei−1,t(si,t,ai,t)+1 before the Q-value update. The eligibilities of all state-action pairs are then decayed by λ after the update. We now describe how these learning rules apply specifically to the two-step task. The reward prediction error is different for the first two levels of the task. Since r1,t is always zero, the reward prediction error at the first stage is driven by the value of the selected second-stage action QMF(s2,t,a2,t): δ1,t=QMF(s2,t,a2,t)−QMF(s1,t,a1,t) Since there is no third stage, the second-stage prediction error is driven by the reward r2,t: δ2,t=r2,t−QMF(s2,t,a2,t) Both the first- and second-stage values are updated at the second stage, with the first-stage values receiving a prediction error down-weighted by the eligibility trace decay, λ. Thus, when λ = 0, only the values of the current state get updated. Model-based strategy. The model-based algorithm works by learning a transition function that maps the first-stage state-action pairs to a probability distribution over the subsequent states, and then combining this function with the second-level model-free values (i.e., the immediate reward predictions) to compute cumulative state-action values by iterative expectation. In other words, the agent first decides which first-stage action leads to which second-stage state, and then learns the reward values for the second-stage actions. At the second stage, the learning of the immediate rewards is equivalent to the model-free learning, since those Q-values are simply an estimate of the immediate reward r2,t. As we showed above, the SARSA learning rule reduces to a delta-rule for predicting the immediate reward. This means that the two approaches coincide at the second stage, and so we set QMB = QMF at this level. The model-based values are defined in terms of Bellman’s equation [37], which specifies the expected values of each first-stage action using the transition structure P (assumed to be fully known to the agent): QMB(sA,aj)=P(sB|sA,aj)maxa∈{aA,aB}QMF(sB,a)+P(sC|sA,aj)maxa∈{aA,aB}QMF(sC,a) where we have assumed these are recomputed at each trial from the current estimates of the transition probabilities and second-stage reward values. Decision rule. To connect the values to choices, the Q-values are mixed according to a weighting parameter w: Qnet(sA,aj)=wQMB(sA,aj)+(1−w)QMF(sA,aj). Again, at the second stage the decision is made using only the model-free values. We used the softmax rule to translate these Q-values to actions. This rule computes the probability for an action, reflecting the combination of the model-based and model-free action values weighted by an inverse temperature parameter. At both states, the probability of choosing action a on trial t is computed as P(ai,t=a|si,t)=exp(βQnet(si,t,a))∑a′exp⁡(βQnet(si,t,a′)) where the inverse temperature β determines the randomness of the choice. Specifically, when β → ∞ the probability of the action with the highest expected value tends to 1, whereas for β → 0 the probabilities over actions becomes uniform. Simulation of the accuracy-demand trade-off In order to test whether the Daw two-step task embodies a trade-off between goal-directed behavior and reward, we estimated the relationship between control (model based vs. model free) and reward by Monte Carlo simulation. For each simulation, we generated a new set of four series of independently drifting reward probabilities across 201 trials according to a Gaussian random walk (mean = 0, σ = 0.025, the same parameters used by Daw and colleagues [8]) with reflecting boundaries at 0.25 and 0.75 (also used by Daw and colleagues). Then we simulated performance on the task for 11 different values of the weighting parameter w, ranging from 0 to 1, the inverse temperature, ranging from 0 to 10, and the learning rate, ranging from 0 to 10. For each of these, we recorded the reward rate obtained. Next, we ran a linear regression for each combination of inverse temperature and learning rate, predicting the reward rate from the size of the weighting parameter. Note that the data points in each linear regression were generated using the same set of drifting rewards, ensuring that any effect was due to the changes in the weighting parameter and not to random variation across the reward distributions themselves. The eligibility trace parameter was fixed at a value that corresponded approximately with previous reports of this task, λ = 0.5, but we found qualitatively identical results across all simulations when we fixed λ at 0 or 1 (see Supporting Information). We repeated this process 1000 times, computing a surface of regression coefficients across a range of reinforcement learning parameters, and then then averaged across these surfaces. The results of this analysis can be seen in the surface plot in Fig 3A, where we have plotted the average standardized linear effect of the weighting parameter as a function of the learning rate and inverse temperature (the median fit is indicated by the red circle on the surface in Fig 3A). 10.1371/journal.pcbi.1005090.g003Fig 3 Results of simulation of accuracy-demand trade-off in the Daw two-step task. (A) Surface plot of the standardized linear effect of the weighting parameter on reward rate in the original version of the two-step task. Each point reflects the average of 1000 simulations of a dual-system reinforcement-learning model of behavior of this task with different sets of drifting reward probabilities, as a function of the learning rate and inverse temperature of the agents. The red circle shows the median fit. Importantly, across the entire range of parameters, the task does not embody a trade-off between habit and reward. (B) An example of the average relationship between the weighting parameter and reward rate with inverse temperature = 5.0 and α = 0.5 (mirroring the median fits reported by Daw and colleagues [8]) across 1000 simulations. (C) The probabilities of repeating the first-stage action as a function of the previous reward and transition for a purely model-free agent and purely model-based agent. The striking feature of this surface map is that the regression coefficients are uniformly close to zero, indicating that none of the parameterizations yielded a linear relationship between model-based control and reward rate. Fig 3B provides a more fine-grained picture of this relationship for a specific parameterization that follows the median fits reported by Daw and colleagues [8]. Note that even though there is no significant relationship between reward and model-based control, this does not undermine the usefulness of the task for measuring the relative balance of model-based and model-free strategies (see Fig 3C). What we can conclude is that this balance does not embody a trade-off between accuracy and demand. Simulations of related tasks Since its conception, the design of the Daw two-step task has been used in many similar sequential decision making tasks. Given the surprising absence of the accuracy-demand trade-off in the original task, it is important to investigate whether related versions of this paradigm are subject to the same shortcoming. Dezfouli and Balleine two-step task In one of these variants, developed by Dezfouli and Balleine [25, 26], participants also navigate from one first-level stage to two second-level stages, utilizing the same common and rare transition structure as in the Daw two-step task, but the reward probabilities are implemented in different fashion. Instead, choices at the second stage have either a high probability (0.7) or a low probability (0.2) of winning, and on every trial the probability of each second-level action changes randomly to either the high or low probability with a small probability (0.2). This task dissociates model-based and model-free control in a manner similar to the Daw two-step task. We performed the same analysis as reported in the previous section for this task, and the results were strikingly similar (see Fig 4). As before, across the entire range of reinforcement learning parameters, the task did not exhibit a trade-off between demand and accuracy, evidence by the uniformly flat regression coefficients (the median fit is represented by the red circle on the surface of Fig 4). 10.1371/journal.pcbi.1005090.g004Fig 4 Surface plot of the linear relationship between the weighting parameter and reward rate in the Dezfouli and Balleine version of the two-step task. The red circle shows the median fit. Similar to the Daw variant, this task does not capture a trade-off between accuracy and demand across all tested parameterizations. Doll two-step task A second variant, reported by Doll and colleagues [15], uses two first-stage states, but the choices at these states transition deterministically to one of the two second-stage states (Fig 5A). At the second stage the choices have chance of producing reward, with probabilities slowly changing over the course of the experiment according to the same Gaussian walk (mean = 0, σ = 0.025) as in the original task with reflecting bounds at 0.25 and 0.75. 10.1371/journal.pcbi.1005090.g005Fig 5 Results of simulation of the Doll two-step task. (A) Surface plot of the linear relationship between the weighting parameter and reward rate in the Doll version of the two-step task. The red circle shows the median fit. Similar to the Daw variant, this task does not capture a trade-off between accuracy and demand across all tested parameterizations, except for a slightly elevated region of parameter space with high inverse temperature and low learning rate. (B) Behavioral predictions in this task. The model-free system learns separate values for each action in each state, so outcomes only affect choices in the same start state. Our simulation of model-free behavior revealed elevated likelihood of staying after a reward from the other state, since this means there is a current high-probability option that the model-free system has been learning about after transitioning there from both start states. The model-based system (on the right) treats start states as equivalent, since they both afford the same transitions, so choices are not affected by whether the previous start state was the same or different. The dissociation between habit and planning in this task follows a different logic. Here, it is assumed that only model-based learners use the implicit equivalence between the two first-stage states, and can generalize knowledge across them. Therefore, for a model-based learner, outcomes at the second level should equally affect first-stage preferences on the next trial, regardless whether this trial starts with the same state as the previous trial or a different one. For model-free agents, however, rewards that are received following one start state should not affect subsequent choices from the other start state. According to Doll and colleagues [15], this results in a clear dissociation in staying behavior between these two strategies (Fig 5B): The model-based learner shows increased likelihood to stay with the choice made on the previous trial when this led to a reward, regardless of whether the start-state is the same as or different from the start state on the previous trial. (Note that here the term ‘staying’ is used to refer to taking the action that leads to the same previous second-stage state, and not to describe a repetition of the same first-stage action.) The model-free learner, on the other hand, is argued to only show increased likelihood to repeat a choice after a reward when the current start state is the same as that on the previous trial. Behavioral performance on this task is consistent with these predictions, and compared to the original two-step task, seems to reflect a mixture of model-based and model-free strategies [15]. However, our simulations revealed that the behavioral profile for the model-free learner also showed a slightly elevated likelihood to stay with the previous choice after a reward if the start state was different (Fig 5B). On the first glance, this may seem surprising because the model-free system does not have access to the second-stage action values. Note, though, that this system builds up first-stage action values based on the previous reward history. If one particular action has a high chance of producing reward for an extended period of trials, then the model-free system will learn to choose actions in order to transition to the relevant state, resulting in increased stay behavior. It is important to note that in this task the effect is small, and only becomes reliable with large sample sizes (we simulated 1000 reinforcement-learning agents). Regardless, this observation makes the interpretation of raw stay probabilities less clean, since an elevated likelihood to repeat the previous action after a reward with a different start-state is not simply attributable to a contribution of the model-based system. This places extra importance on fitting computational models to behavior in this task, since these incorporate reward histories over the entire experiment, which are omitted when analyzing raw choice behavior on single trials. (This point becomes crucial in subsequent experiments described below.) In order to assess whether the Doll two-step task embodied a trade-off between accuracy and demand, we again estimated the relationship between the weighting parameter and reward rate. This analysis (depicted in Fig 5A) suggested that this relationship was largely similar to the results reported above, and very close to zero for large portions of the parameter range. However, in comparison with the Daw and Dezfouli versions, the Doll task showed increased sensitivity in a parameter space with relatively high inverse temperatures (low randomness) and small learning rates. It is important that even in this elevated part of the coefficient surface the hypothetical effect is small, and that the participants’ parameter fits did not fall in this parameter range (the mean fit is indicated by the red circle on the surface in Fig 5A). Factors contributing to the absence of the trade-off Despite the substantial differences between these variants of the two-step task, we found that none of them encompasses a motivational trade-off between planning and reward. This observation naturally raises a question: Why does planning not produce an increased reward rate in this task? What characteristics of the paradigm distort the accuracy-demand trade-off? We investigate five potential explanations. These are not mutually exclusive; rather, they may have a cumulative effect, and they may also interact with each other. First, we show that the sets of drifting reward probabilities that are most often employed are marked by relatively low distinguishability. Second, we show that the rate of change in this paradigm is slow and does not require fast online (model-based) flexibility. Third, we show that the rare transitions in the Daw two-step task diminish the reward-maximizing effect of a model-based choice. Fourth, we show that the presence of the choice at the second stage decreases the importance of the choice at the first stage, which is the only phase where the model-based system has an influence. Fifth, we show that the stochastic reward observations in this task do not carry enough information about the value of the associated stimuli. We use simulations of performance on novel tasks to demonstrate these five points and, as a result, develop a novel paradigm that embodies an accuracy-demand trade-off. 1. Distinguishability of second-stage probabilities In the two-step task, the difference between model-based and model-free strategies only carries consequences for the first stage, since the second stage values are identical for both strategies. Therefore, the finding that model-based control is not associated with an increased reward rate suggests that the first-stage choices the agent makes do not carry importance, for example, because the reward outcomes at the second stage are too similar. In the original version of the two-step task, the reward probabilities have a lower bound of 0.25 and an upper bound of 0.75. This feature results in a distribution of differences between reward probabilities that is heavily skewed left (Fig 6A). The mean value of this distribution approaches 1/6 (1/3 of the range of 0.5), suggesting that most choices in this task only carry modest consequences, because the associated values have relatively low distinguishability. 10.1371/journal.pcbi.1005090.g006Fig 6 The influence of the range of reward probabilities. (A) Distribution of differences in reward probabilities between the actions of each trial. (B) Increasing the range of probabilities increases the average linear effect between model-based control and reward for a parameter space associated with high inverse temperatures and relatively low learning rate. Average parameter fits in the original report do not lie within this region of increased sensitivity to the accuracy-demand trade-off. One straightforward way to increase the differences between the second-stage options is to maximize the range of reward probabilities by setting the lower bound to 0 and the upper bound to 1 (e.g., [20]). This shifts the mean difference in reward probability between options to 1/3, doubling the consequences of each choice in terms of reward maximization. As can be seen in Fig 6B, this change in the paradigm slightly increases the degree to which behavior in this task reflects an accuracy-demand trade-off (Simulation 1). The regression coefficients of the relationship between reward and w are slightly elevated compared to those of the Daw two-step task (compare with Fig 1), especially in the part of parameter space with high inverse temperature and low learning rate, suggesting that low distinguishability between the options is a factor contributing to the absence of the trade-off. However, the increase in the effect is fairly small, and the parameter space at which the highest increase is observed does not correspond with the average parameter fits reported in the literature. 2. Increased drift rate It is also possible that the changes over time in the second-stage reward probabilities, depicted in Fig 1B, contribute to the absence of the accuracy-demand trade-off in the Daw two-step task. For example, these changes might be too slow, such that model-free learning can adapt to these values at a rate that is proportional to the rate of change over time [53]. Another possibility is that these changes happen too fast, such that the model-based system never accurately reflects the values of the second-stage actions. In order to explore the effect of the drift rate (i.e., the standard deviation of the Gaussian noise that determines the random walks of the reward distributions) on the accuracy-demand trade-off, we performed simulations of the generative reinforcement learning model with inverse temperature parameter β = 5 and learning rate parameter α = 0.5, mirroring the median fits reported by Daw and colleagues [8]. For each of these, we generated a new set of four series of independently drifting reward probabilities across 201 trials according to Gaussian random walks. We simulated performance on the task for 11 values of the weighting parameter, ranging from 0 to 1, and ran a linear regression predicting the reward rate from the size of the weighting parameter. Again, this method ensured that the data points in each linear regression were generated using the same set of drifting reward probabilities, ensuring that any effect was due to the changes in the weighting parameter. This analysis was performed for 7 different drift rate values, ranging from 0 to 0.5, using the narrow range of reward probabilities used in the original report [8] and the broader range described above. The slopes we report will be the average across a large number of iterations (10,000), since we are estimating very subtle effects, especially in the narrow range condition. The results, depicted in Fig 7A, suggest that the drift rate of the Gaussian random walk affects the strength of the accuracy-demand trade-off in the task. Specifically, for the broader range of parameterizations, the strongest relationship was observed for a drift rate of 0.2, and for the parameterization range of the original report the maximum effect occurred with a drift rate of 0.1. Consistent with the section above, both the strength of the relationship between reward and planning, as well as the effect of the drift rate on this relationship was stronger for the task with the broader reward probability range. Note however, that for both probability ranges large drift rates negatively impact the relationship, presumably because for these values there is no learnable stability in the terminal state reward probabilities. We confirmed that this general pattern between the accuracy-demand trade-off, the drift rate, and the range of the reward probabilities occurs not only at the parameterization that most closely matched the estimates of the original report, but also across a broader parameter space (see Supporting Information). Fig 7B depicts the surface map of average regression coefficients for the effect between model-based control and reward rate for a task with the wider reward probability range and a drift rate of 0.2 (Simulation 2). These changes to the task substantially increase the strength of the accuracy-demand relationship, especially for agents with a high inverse temperature. However, as before, the increase in the strength of this relationship primarily occurred in regions of parameter space that do not correspond with the fits reported in studies that employ the Daw two-step task [8, 41], and had a very weak effect size. 10.1371/journal.pcbi.1005090.g007Fig 7 The influence of the drift rate. (A) The effect of the size of the drift rate on the relationship between model-based control and reward, for two-step tasks with a narrow and a broad reward probability range. (B) Increasing the range of probabilities and the drift substantially increases the average linear effect between model-based control and reward when the inverse temperature is high. These analyses show that the rate of change of the reward probabilities in the original Daw two-step task is too slow to promote model-based planning. The relationship between reward and model-based control becomes stronger when the drift rate of the Gaussian random walk governing the reward probabilities is moderately increased, and this effect is especially pronounced when these probabilities are more dissociable. However, even though these two factors contribute substantially to the absence of the accuracy-demand trade-off in the Daw two-step task, we found that a task that adjusted for their shortcomings only obtained a modest trade-off between reward and goal-directed control. 3. Deterministic transition structure Because the Daw two-step task employs rare transitions, model-based choices at the first stage do not always lead to the state that the goal-directed system selected. This feature of the task might lead to a weakening of the relationship between model-based control and reward rate. The task structure employed by Doll and colleagues [15], discussed in the previous section (Fig 5A), avoids this issue by implementing deterministic transitions, such that model-based choices always lead to the desired second-stage state. Indeed, the simulation analysis for this task revealed that a certain range of the simulated parameter space showed an increased relationship between model-based control and reward rate. However, for this analysis the set of drifting rewards still suffered from the factors discussed above–relatively low distinguishability of the second-stage options and a suboptimal drift rate. To assess the influence of the deterministic task structure, we simulated performance on the Doll version of the two-step task with sets of reward probabilities with the wider range and increased drift rate (a bounded Gaussian random walk with μ = 0, σ = 0.2 on a range from 0 to 1 with reflecting bounds). The resulting surface map (Simulation 3a; Fig 8A) showed that, consistent with our predictions, across the entire parameter space the relationship between control and reward was stronger when compared to the Daw two-task with improved sets of reward probabilities in the previous section. 10.1371/journal.pcbi.1005090.g008Fig 8 The influence of a deterministic task structure. (A) Because of the deterministic transitions, model-based choices in the Doll two-step task always result in the desired state outcome. Combined with increased distinguishability and increased drift rate in the reward probabilities, this task results in a substantial increase in the relationship between planning and reward. (B) When this task structure is adapted to include stochastic transitions, the relationship between planning and reward is significantly reduced, indicating an important contribution of the rare transitions in diminishing the accuracy-demand trade-off in the original paradigm. Even though this result is consistent with the assumption that model-based choices in the Daw two-step task lead to the desired state less often than in the deterministic version of the two-step task, it is equally possible that the second task shows an increased accuracy-demand trade-off because it introduces the possibility of generalization across actions, and not because of the elimination of the rare transitions. To disentangle these two possibilities, we simulated reinforcement-learning performance on a hybrid task with two starting states but with rare transitions (Simulation 3b; Fig 8B). The regression coefficients in the resulting surface map were substantially lower than in the deterministic variant of the task, and was comparable to that of the Daw two-step task with broader probability range and a higher drift rate. This indicates a critical role for the rare transitions in diminishing the accuracy-demand trade-off. 4. One choice in second stage As noted above, model-based and model-free strategies make divergent choices only at the first stage of the multi-step paradigms we have considered so far; at the second stage, both strategies perform a biased selection weighted towards the reward-maximizing option. Thus, the advantage of model-based control over model-free control is approximately bounded by the difference between the maximum value of all actions available in one second-stage state and the maximum value of all actions available in the other second-stage state. Intuitively, as the number of actions available within each second-stage state grows, this difference will shrink, because both second-stage states will likely contain some action close to the maximum possible reward value (i.e., a reward probability of 1). Conversely, the difference between the maximum value actions available in each second-stage state will be greatest when only a single action is available in each state. This design should favor the largest possible difference in the rate of return between model-based and model-free strategies. To quantify this, we generated 10,000 sets of reward probabilities in this task (according to a Gaussian random walk with reflecting bounds at 0 and 1 and σ = 0.2). The average difference between any two reward probabilities within a state was equal to 0.33, whereas the average difference between the maximal reward probabilities of the two states was 0.27. Since the model-based system only contributes to the first-stage decision, we simulated performance of the reinforcement-learning model in a deterministic two-step task in which the second-stage states do not contain a choice between two actions. In this task, the average difference in reward probabilities that the model-based system uses to make a choice at the first stage is 33%, an increase in comparison to the task that implements a binary choice at the second stage states. To assess whether this change to the task resulted in a stronger accuracy-demand trade-off, we simulated performance on this task and estimated the strength of the relationship between the weighting parameter and reward rate, across the same range of reinforcement-learning parameters (Simulation 4; Fig 9). This analysis revealed that the elimination of the choices at the second stage indeed strengthened the relationship between w and reward in comparison to the deterministic task with second-stage choice (Fig 8A), because the larger difference between ‘maximal’ reward probabilities between the two second-stage states increased the importance of the model-based contribution to the first-stage choice. 10.1371/journal.pcbi.1005090.g009Fig 9 The influence of reducing the number of second-stage action. Because of the deterministic transitions, model-based choices in the Doll two-step task always result in the desired state outcome. Combined with increased distinguishability and increased drift rate in the reward probabilities, this task results in a substantial increase in the relationship between planning and reward. 5. Informativeness of an observation In order to determine the value of an action in the two-step task, the stochastic nature of the task requires participants to sample the same action repeatedly and integrate their observations. In other words, since each outcome is either a win or a loss, the information contained in one observation is fairly limited. Here, we will test whether the high amount of ambiguity associated with each observation contributes to the absence of the accuracy-demand trade-off in the two-step task. One way to increase the informativeness of an outcome observation is to replace the drifting reward probabilities at the second stage with drifting scalar rewards, so that the payoff of each action is exactly identical to its value [42]. This elimination of uncertainty increases the information obtained from each outcome observation, and thus may lead to a strengthened relationship between model-based control and reward. In order to test whether the reward distributions for the second-stage actions would improve the information obtained from each observation, we performed a series of simulations for two simple reinforcement learning tasks (Fig 10A). Both tasks involved a repeated decision between task options, but the nature of the reward for these two options was different. For one task, these actions were associated with a probability of a reward, which independently drifted across the session according to a Gaussian random walk (μ = 0, σ = 0.2) on a range from 0 to 1 with reflecting bounds. For the other task, the same series of drifting probabilities were treated as series of drifting rewards. Specifically, this meant that if an action afforded a 74% probability for a reward in the first task, the same action in the second task would lead to a payoff of 0.74 points. As before, we performed reinforcement learning simulations (λ = 0.5) on these two tasks across a range of inverse temperatures and learning rates. Because we used the same sets of drifting values as probabilities and payoff (i.e., their expected values are the same), any difference in performance between the two tasks is a function of the increased amount of information available in the payoff condition. 10.1371/journal.pcbi.1005090.g010Fig 10 The influence of the type of reward distribution (points vs probabilities) on choice accuracy. (A) We ran simulations of RL agents on two different two-armed bandit tasks. For one, the reward distributions indicate the reward probability associated with each action. The other task does not include binomial noise, but instead the actions pay off rewards that are directly proportional to its value in the reward distribution. (B) Agents show greater accuracy in choosing the highest-value action on the task the task where the two-armed bandit pays off points instead of affording a probability to win a reward, especially when both the inverse temperature and learning rate were high. (C) The Q-values of each action shows stronger correlations with their objective reward value in the task where the two-armed bandit payed off points instead of affording a probability to win a reward. We first compared the model’s performance on these two tasks by computing the accuracy of its choices, i.e., how often it selected the action with the highest reward probability or reward payoff. Fig 10B displays the average accuracy for each task across the range of inverse temperatures and learning rates. For both tasks, the model showed increased accuracy for higher learning rates and inverse temperatures. That is, agents with less randomness in choice and greater incorporation of new information were more likely to pick the option with the objectively higher chance to win or reward payoff. Consistent with our prediction, this effect was larger in the task with reward payoffs compared to the task with reward probabilities across virtually the entire simulated parameter space. This suggests that the observation of reward outcome in the payoff condition was more informative, leading to overall better performance in the simple two-alternative choice task. As a second metric of the information contained in each outcome observation, we computed the correlation between the model’s action values and the actual payoffs in the simulations reported above. We expected that the increased precision in outcome observations in the payoff condition would lead to a tighter coupling between the Q-values of the model and the objective values as compared to the probabilities. Fig 10C depicts the results of this analysis. In both tasks, the average correlation was strongest for high learning rates, since for these agents, new information was incorporated fully, always reflecting the latest information to the largest extent. Second, the correlation was stronger when the inverse temperature was low, presumably because agents with high randomness in choice sample from both options. Most importantly, the correlations between action and objective values was higher in the reward payoff condition than in the probability range across the entire range of tested reinforcement learning parameters. This observation provides convergent evidence that increased resolution in the outcome observations is associated with enhanced performance in reinforcement-learning tasks. Next, we assessed whether this increased performance in the payoff condition would result in a stronger accuracy-demand trade-off in the deterministic two-step task. We reasoned that if the agent obtained a more accurate estimation of the second-stage action values, then the model-based system would be better positioned to maximize reward. To test this prediction, we again estimated the strength of the relationship between the weighting and reward rate, across the range of reinforcement-learning parameters (Simulation 5; Fig 11). The surface map revealed a marked increase when compared to that of the task with reward probabilities. Across the entire range of inverse temperatures and learning rates, the regression coefficients of the relationship between control and reward were substantially higher in comparison to the deterministic task with reward probabilities (Fig 9). This final analysis revealed that the reward outcomes in the Daw two-step task do not carry enough information about the action’s value, leading to a decrease in accuracy for the model-based system. 10.1371/journal.pcbi.1005090.g011Fig 11 The influence of removing binomial noise from the reward distributions at the second stage. (A) The surface plot of the relationship between model-based control and reward in the novel two-step task with reward payoffs at the second stage. The inclusion of this fifth factor substantially increased the accuracy-demand trade-off in the two-step paradigm. (B) An example of the average relationship between the weighting parameter and reward rate with inverse temperature = 10 and α = 0.4. Comparison with Akam and colleagues In a recent study, Akam and colleagues [9] reported that the original version of the two-step task does not embody a trade-off between control and reward. They simulated performance on the task for a pure model-free and model-based agent, with independently optimized parameters to maximize reward rate, and found no difference between them, consistent with the results of our simulations. Furthermore, they proposed a new task that establishes a trade-off between control and reward. This task is similar to the original Daw version of the task, except for the elimination of the second-stage choice, a reduction of the rare transition probability (20%), and the reward probabilities in the second-stage states alternate between blocks with reward probabilities of 0.8/0.2 and blocks with probabilities of 0.2/0.8 [9]. Using optimized reinforcement learning parameters independently for model-free and model-based agents, they show that reward rate is higher for the model-based agent. This approach—i.e., a comparison of optimal parameter settings under model-free versus model-based control—provides an important existence proof of the potential benefits of model-based control. However, their way of quantifying the accuracy-demand trade-off differs significantly from the current approach. In order get a more comprehensive overview of the accuracy-demand trade-off in the Akam two-step task, we again estimated the strength of the relationship between the weighting parameter and reward rate, across the same range of reinforcement learning parameters (Fig 12). This surface of regression coefficients shows remarkable differences compared to our novel paradigm (presented in Fig 11). Most importantly, high correlation coefficients are restricted to a selective region of parameter space with low learning rate and high inverse temperature. The strength of this relationship drops in the rest of the parameter space. 10.1371/journal.pcbi.1005090.g012Fig 12 Surface plot of the relationship between model-based control and reward in the Akam and colleagues [9] version of two-step task with alternating blocks of reward probabilities at the second-stage states. This feature of the task means that an increase in model-based control, keeping all other RL parameters fixed, is not likely to yield significantly increased total reward, because reinforcement learning parameters tend to vary widely across individuals [54]. However, it has been shown that people adapt RL parameters such as the learning rate and choice randomness to maximize reward in the environment [55], providing alleviation for this concern. A second, distinct advantage of the task we introduce here involves the possibility that humans may identify and exploit higher-level regularities in the structure of reward. Specifically, in the Akam task participants may learn to predict the alternating blocks of reward probabilities, complicating the interpretation of behavior. In contrast, in the task we introduce it is impossible to perfectly anticipate changes in our randomly changing reward distributions. Despite these concerns, both tasks achieve an accuracy-demand trade-off, and in this respect represent a substantial improvement over the Daw two-step task. Future empirical work should compare the empirical correlations between reward and model-based control for our task and the Akam two-step task, so as to gain fuller comprehension of their respective merits. Summary We have identified several key factors that reduce the accuracy-demand trade-off in the Daw two-step task. We found that the sets of drifting reward probabilities that are most often employed in this task are marked by low distinguishability and a rate of change that is too slow to benefit from flexible online adaptation. We also showed that the rare transitions in the original task and the presence of multiple choices in the second-stage states diminished the effect of model-based decisions on reward rate. Finally, we showed that the stochastic reward observations in this task do not carry sufficient information about the value of the associated stimuli. In addition to identifying these factors, we have provided improvements to the paradigm targeting each shortcoming. Fig 13 shows the progression in the average strength of the relationship between reward and control across these changes in the task structure, operationalized as the volume under the surface of each simulation. It reveals a progressive contribution of each change to the task, suggesting that implementing a broader range, increased drift rate, deterministic task structure, one second-stage choice, and reward payoffs all have identifiable contributions to the strength of the relationship between model-based control and reward. 10.1371/journal.pcbi.1005090.g013Fig 13 Comparison of trade-off between model-based control and reward across different paradigms. We calculated the volume under the surface of coefficients of the linear relationship between the weighting parameter and the reward rate for each of the paradigms in the section above. Across these simulations, we progressively included elements that strengthened the relationship, as summarized in this figure. Here, we have presented a progression of five factors that enhance the accuracy-demand trade-off in the two-step task. Which of these factors contributed most to the increase in this strength? Fig 13 represents the increase in this strength as additional factors are layered into the paradigm, and from this figure one might conclude that the conversion from reward probabilities to reward payoffs carried the strongest contribution. So far, however, we have confounded the contribution of each factor with the order in which they were represented. It is possible that the reward payoffs contributed a substantial amount of strength to the trade-off simply because it was the final factor introduced. In order to test the effect of factor order in our analyses, we computed the surface of regression coefficients for all 32 possible combinations of our binary factors (25), using the same procedure as described above (omitting the cases where β = 0, or α = 0). Next, we computed the volume under the surface as an approximation of the average strength of the relationship between model-based control and reward for each these simulations. Fig 14, depicts the volume under the surface of each simulation as a function of the number of factors that were included in the design. These results indicate that the primary cause for the strength in the final paradigm was the inclusion of all five factors, and not necessarily the contribution of one of them. To see this, compare the score of the final task with 5 included factors to the scores of the task with 4 factors. The strength of the effect in the final task was 6.9 standard deviations removed from the tasks with 4 factors, and 5.7 standard deviations from the 4-factor task with the strongest average accuracy-demand trade-off. 10.1371/journal.pcbi.1005090.g014Fig 14 Volume under the surface for all 32 tasks generated by the 5 binary factors discussed in this paper. Each dot represents the volume under the surface of linear regression coefficients for one task, and is plotted as a function of the number of ‘beneficial’ factors that are included in each task’s design. The gray line represents the average increase in the strength of the relationship between model-based control and reward. The converse is also true: all factors had a similar and small individual effect on the original Daw paradigm. To see this, compare the score of the original task with 0 factors to the scores of all tasks with 1 factor. The strength of the effect in the original task was only 1.3 standard deviations removed from the tasks with 1 factor, and even slightly better than the 1-factor task with the smallest effect. Most importantly, even if each individual factor did not substantially increase the total effect compared to the original paradigm, their joint inclusion increased the strength of the relationship between model-based control and reward rate by a factor of approximately 230. At least in theory, we have developed a paradigm that embodies an accuracy-demand trade-off between model-based control and reward rate. Next, we attempt to validate this paradigm by having human participants perform either a novel version of the two-step task with the improved features described above, or the original version of the two-step task as described by Daw and colleagues [8]. We predicted that measures of model-based planning in the novel, but not in the original, paradigm would show a positive correlation with the reward rate. In addition, the comparison between these two paradigms allows us to test whether human participants spontaneously modulate the balance between model-free and model-based control depending on whether a novel task favors model-based control. So far, we have discussed the accuracy-demand trade-off uniquely as it is instantiated in the two-step task. However, if the novel paradigm embodies an empirical accuracy-demand trade-off, then the results of this study allow us to test whether the brain also computes a cost-benefit trade-off between the two systems. We predicted that average model-based control would be elevated in the novel paradigm, since planning was incentivized in this task [56]. Experimental methods Participants Four hundred and six participants (range: 18–70 years of age; mean: 33 years of age; 195 female) were recruited on Amazon Mechanical Turk to participate in the experiment. Participants gave informed consent, and the Harvard Committee on the Use of Human Subjects approved the study. Materials and procedure One hundred and ninety-nine participants completed 125 trials of the novel two-step reinforcement-learning task. The structure of the task was based on the procedure developed in the previous section. The remaining two hundred and seven participants completed 125 trials of the two-step with the original Daw structure [8]. We embedded both tasks in a cover story in order to make it more engaging for participants [31]. Novel paradigm. Every trial in the novel two-step task consisted of two stages (Fig 15A). Each trial would start randomly in one of two possible first-stage states. In both, a pair of ‘spaceships’ appeared side by side on a blue earth-like planet background. Participants were told they had to choose between these two spaceships to fly to one of two different planets. The choice between the left- and right-hand spaceship had to be made using the “F” or “J” button keys within 2000ms. After a choice was made, the selected spaceship was highlighted for the remainder of the response period. The positions of the spaceships were randomly selected on each trial. Depending on the choice of spaceship, the participants would then deterministically transition to one of two second-stage states, a purple or a red planet. The spaceship selected in the first-stage was displayed at the top of the screen in this planet. On each planet, participants found an alien that ‘mines’ from a ‘space mine’. These mines act as the second-stage bandits. Participants were told that sometimes the aliens were in a good part of the mine and they paid off a certain number of points or ‘space treasure’, whereas at other times the aliens were mining in a bad spot, and this yielded negative points or ‘antimatter’. The payoffs of these mines slowly changed over the course of the experiment. Even though there was only one choice available at the second-stage planets, participants were instructed that they were to press the space bar within 2000ms in order to receive the reward. One of these reward distributions was initialized randomly within a range of -4 points to -1 points, and the other within a range of +1 to +5 points. Then, they varied according to a Gaussian random walk (σ = 2) with reflecting bounds at -4 and +5 for the remainder of the experiment. A new set of randomly drifting reward distributions was generated for each participant. At the end of the experiment, participants were given 1¢ for every two points they earned. 10.1371/journal.pcbi.1005090.g015Fig 15 Design of the novel two-step task. (A) State transition structure of the paradigm. At the first stage, participants choose between one of two pairs of spaceships. Each choice deterministically leads to a second-stage state that was associated with a reward payoff that changed slowly according to a random Gaussian walk over the duration of the experiment. Note that the choices in the two different first-stage states are essentially equivalent. (B) Predicted behavior from the generative reinforcement-learning model of this task (using median parameter estimates, and w = 0.5 for the agent with a mixture of strategies). Note that in this task the model does not produce qualitatively different behavior for the different systems as reported in Fig 5. Instead, the differences in behavior are subtler, and therefore differences in strategy arbitration are better captured using model-fitting techniques. The most important feature of the task is that the spaceships at the first states were essentially equivalent. For each pair, one spaceship always led to the red planet and alien, whereas the other always led to the purple planet and alien. Because of this equivalence, we were able to dissociate model-based and model-free contributions to choice behavior, since only the model-based system generalizes across the equivalent start state options by computing each action’s value as its expected future reward. Therefore, model-based and model-free strategies make qualitatively different predictions about how second-stage rewards influence first-stage choices on subsequent trials. Specifically, for a pure model-based learner, each outcome at the second stage should affect first-stage preferences on the next trial, regardless of whether this trial starts with the same or the other pair of spaceships. In contrast, under a pure model-free strategy a reward obtained after one pair of spaceships should not affect choices between the other pair. Daw paradigm. The Daw two-step task used the same buttons, timing, visual appearance, and counter balancing procedures as the novel paradigm, but the structure of the task matched that of the design in the original report (discussed in detail above). At the start of each trial, participants chose between a pair of spaceships. Depending on the choice of spaceship, the participants would then transition to one of two second-stage states, a purple or a red planet. Each spaceship traveled more frequently to one planet than to the other (70% versus 30%), and these transition probabilities were opposite for the two spaceships. On each planet, participants chose between pairs of aliens that mines from a space mine. Participants were told that sometimes the aliens were in a good part of the mine and they were more likely to deliver a piece of space treasure, whereas at other times the aliens were mining in a bad spot, and they were less likely to deliver space treasure. The payoffs of these mines slowly changed over the course of the experiment. One pair of aliens was initialized with probabilities of 0.25 and 0.75, and the other pair with probabilities of 0.4 and 0.6, after which they changed according to a Gaussian random walk (σ = 0.025) with reflecting bounds at 0.25 and 0.75 for the remainder of the experiment. A new set of randomly drifting reward distributions was generated for each participant. To equate average pay-off between conditions, participants were given 1¢ for every point they earned. As explained in detail above, model-based and model-free strategies make qualitatively different predictions about how second-stage rewards influence first-stage choices on subsequent trials. Specifically, choice under a pure model-free strategy should not be affected by the type of transition (common vs. rare) observed on the previous trial (see Fig 2A), whereas pure model-based learners should base their choice on both the type of transition and whether a reward was observed on the previous trial (see Fig 2B). Before completing the full task, participants were extensively trained on different aspects of the task. Participants who completed the novel paradigm first learned about the value of space treasure and antimatter, and the change in payoffs from both space mines by sampling rewards from two different aliens. Next, they learned about the deterministic transitions between spaceships and planets during a phase in which they were instructed to travel to one planet until accurate performance was reached. Participants who completed the Daw paradigm sampled from aliens with different reward probabilities, and were extensively instructed on the transition structure. Finally, both groups of participants practiced the full task for 25 trials. There was no response deadline for any of the sections of the training phase. The color of the planets and aliens in this phase were different from those in the experimental phase. Reinforcement learning model and behavioral predictions We used our reinforcement learning model of the novel task to produce behavioral predictions for a pure model-free and pure model-based decision maker, and an agent with a mixture between model-free and model-based control. This model was largely the same as before, with the exception of how the transition structures were learned. Recall that participants that completed the novel paradigm performed a practice phase in which they were taught a set of deterministic transitions between the four spaceships and two different planets. Next, they were told that in the experimental phase, the rules and spaceships were the same as in the practice phase, but that there would be new planets. Therefore, we assumed that participants would assume equal probability of each spaceship traveling to one of the two planets, until they observed one transition for a first-stage state. After this observation, the model immediately infers the veridical transition structure for that first-stage state. The participants that completed the Daw paradigm of the two-step task learned about the transition structure through instruction and direct experience in a practice phase with two different planets. They were also told that the rules and spaceships would be the same, but that the planets would be new. Therefore, we assumed that participants initially assumed equal probability of transitioning between the spaceships and the planets. Next, we characterized transition learning by assuming that participants chose between three possible transition structures as a function of how many transitions they observed between the states and actions: a flat structure with equal probabilities between all states and actions, or two symmetric transition structures with opposite transition probabilities of 70% and 30% between the two spaceships and planets. As we have argued above, in our novel paradigm the differences in the probability of repeating the previous first-stage choice do not show a major qualitative difference between a purely model-based and model-free strategy, when plotted as a function of whether the previous start state is the same as or different from the current start state and whether a reward was obtained on the previous trial (Fig 15B). In fact, both a model-free and a mixture agent show an interaction between the two factors, start-state similarity and previous reward, with the likelihood of staying being higher if the current start state is similar the start state on the previous compared to when it was different, but still significantly bigger than chance. For the model-free agent, this reflects the presence of a highly rewarding action that the model-free learner learns to approach (for a detailed analysis, see [9]). This erosion of the qualitative predictions afforded by a stay/switch analysis is enhanced in these simulations compared to the original Doll investigation (Fig 5), presumably because reward observations in the current task carry more consequential information for behavior. The lack of qualitative differences in single-trial staying behavior between the model-free and mixture strategies places special importance on model-fitting to quantify the balance between habit and control. Not only does model-fitting incorporate an influence of all previous trials on choice, but it also provides a numerical value for the relative weighting of model-based and model-free strategies (the w parameter). In order to demonstrate that standard model-fitting procedures are sufficient to robustly estimate w on a per-participant basis, we generated data from 200 agents with randomly selected reinforcement learning parameters and then estimated these parameters using the model-fitting procedure described below. This method, described in more detail in the Supporting Information, yielded substantial correlations between the true and estimated parameters (including w, r = 0.68), validating our approach (S1 Text). An alternative way to correct for the influence of reward in the previous trials is by predicting ‘staying’ behavior through a multilevel logistic regression analysis that accounts for this influence with a predictor that incorporates behavior about the outcome of the previous choice [9, 10]. The Supporting Information describes this method in detail; in brief, it produced qualitatively similar results to the model fitting procedure (S2 Text). Model fitting In order to estimate each participant’s weighting parameter, we fitted one of two reinforcement learning models to each participant’s data, dependent on which task they completed. This model was equivalent to the models described above, with the exception for the input into the softmax decision rule: P(ai,t=a|si,t)=exp(β[Qnet(si,t,a)+π∙rep(a)+ρ∙resp(a)])∑a′exp⁡(β[Qnet(si,t,a′)+π∙rep(a′)+ρ∙resp(a′)]) where the indicator variable rep(a) is defined as 1 if a is a first-stage action and is the same one as was chosen on the previous trial, zero otherwise. Multiplied with the ‘stickiness’ parameter π, this captures the degree to which participants show perseveration (π > 0) or switching (π < 0) at the first stage. The indicator variable resp(a) is defined as 1 if a is a first-stage action selecting the same response key as the key that was pressed on the previous trial, zero otherwise. Multiplied with the ‘response stickiness’ parameter ρ, this captures the degree to which participants repeated (ρ > 0) or alternated (ρ < 0) key presses at the first stage. We introduced this parameter since the spaceship’s positions were not fixed, hence participants could show perseveration in spaceship choices, button presses, or both. We used maximum a posteriori estimation with empirical priors, implemented using the mfit toolbox [54] parameters to fit the free parameters in the computational models to observed data for each participant separately. Based on prior work [54], we used weak priors for the distributions for the inverse temperature, β ~ Gamma(4.82, 0.88), and stickiness parameters, π, ρ ~ N(0.15, 1.42), and flat priors for all other parameters. To avoid local optima in the estimation solution, we ran the optimization 25 times for each participant with randomly selected initializations for each parameter. Correlation analysis. In order to assess the relationship between model-based control and reward in our novel paradigm, we computed the Pearson correlation coefficient between the estimated weighting parameter and reward rate obtained in the task. However, since we generated new sets of drifting rewards for each participant, baseline differences in average reward might weaken this correlation. Therefore, we calculated the difference between actual reward and average chance performance for each participant, and used this as the measure of reward obtained to correlate with the weighting parameter. For both tasks, chance performance was computed as the average value across the reward distributions. Exclusion criteria Participants were excluded from analysis if they timed out on more than 20% of all trials (more than 25), and we excluded all trials on which participants timed out (average 2.7%). After applying these criteria, data from 381 participants were submitted to the model-fitting procedure. Results Behavioral performance For the participants who completed the Daw task, we found that a reward on the previous trial increased the probability of staying with the previous trial’s choice [t(196) = 7.70, p < 0.001; Fig 16A], but that this effect interacted with the type of transition on the previous trial [t(196) = 5.38, p < 0.001]. This result replicates the basic finding on the original two-step confirming that participants used both model-based and model-free strategies. 10.1371/journal.pcbi.1005090.g016Fig 16 Behavioral performance on the two-step tasks. (A) Behavioral performance on the Daw task showed both a main effect of previous outcome and an interaction between previous outcome and transition type, suggesting that behavior showed both model-based and model-free strategies. (B) Behavioral performance on the novel paradigm showed a significant difference in stay behavior between same and different start states conditions after a reward, suggesting that behavior was not fully model-based. Error bars indicate within-subject SEM. For the participants who completed the new paradigm, we found that a positive reward on the previous trial significantly enhanced staying behavior from chance for both similar and different current start states, (p < 0.001 for both effects), but this effect was larger for the same compared to the different start state condition [t(183) = 9.64, p < 0.001; Fig 16B]. This pattern of behavior suggests that the participants did not employ a pure model-based strategy (compare with Fig 15B). However, as described above, it does not allow us to assess the relative contributions of model-based and model-free strategies to control based on these raw stay probabilities: both a purely model-free agent and an agent with a mixture of model-based and model-free strategies choices are predicted to show an increased stay probability after a win in a different start state, since a reward is indicative of history of recently reward trials. Model fits The reinforcement learning models described above incorporates the (decayed) experience on all previous trials to choice and is better able to dissociate the contributions of the two strategies. This model consists of a model-free system that updates action values using temporal-difference learning and model-based system that learns the transition model of the task and uses this to compute action values online. The weighting parameter w determines the relative contribution between model-based and model-free control. The stickiness parameters π and ρ capture perseveration on either the response-level or the stimulus-choice. We first investigated whether the inclusion of either stickiness parameter (π and ρ) was justified by comparing both the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC), for models that included none, one, or both parameters for both tasks separately (see S1 Table). For the Daw task, we found that both goodness-of-fit measures favored a model that included both stickiness parameters. For the novel task, the BIC favored a model with response stickiness but not stimulus stickiness included, whereas the AIC favored a model that included both stickiness parameters. We decided to favor the more parsimonious model without stimulus stickiness, and parameter fits from this model will be reported in the following, but the results did not qualitatively change when the stimulus stickiness parameter was included. Second, we used model comparison with both goodness-of-fit measures to analyze whether the hybrid model including the w parameter fit the data better than either a pure model-based or model-free model (see S2 Table). For the Daw task, we found that the AIC favored the hybrid model, but that the BIC favored the pure model-free model. However, there have been many reports in the literature that justify the inclusion of the weighting parameter for this task [8], and so we adopt the hybrid model for consistency with prior work. (Note also that it would be impossible to assess the relationship between model-based control and reward without using the hybrid model). For the new task, we found that both BIC and AIC favored the hybrid model compared to the pure model-based and model-free models. This suggests that human performance in the new paradigm is characterized by a mixture of model-based and model-free strategies. In summary, the model fits presented below used all six free parameters for the participants that completed the Daw paradigm, but omitted the stimulus stickiness parameters for the participants that completed the novel paradigm. These parameter estimates and their quartiles are depicted in Table 1. 10.1371/journal.pcbi.1005090.t001Table 1 Best-fitting parameter estimates shown as median plus quartiles across participants. Paradigm Predictor β α λ π ρ w Daw 25th percentile 2.35 0.11 0.25 0.03 -0.03 0.00 Median 3.35 0.34 0.65 0.21 0.05 0.27 75th percentile 3.88 0.57 1.00 0.4 0.19 0.66 Novel 25th percentile 0.51 0.01 0.07 -0.29 0.04 Median 0.72 0.67 0.62 -0.06 0.48 75th percentile 3.31 1.00 1.00 0.14 0.85 Across participants, we found that the median weighting parameter w was 0.27 for the Daw paradigm and 0.48 for the novel paradigm, indicating that both strategies were mixed in the population for both tasks. However, we found that model-based control was significantly higher for participants in the novel paradigm compared to the Daw paradigm [Wilcoxon two-sample rank sum test, z = 3.31, p < 0.001], suggesting that the existence of the accuracy-demand trade-off in the novel paradigm induced a shift towards model-based control. Of greatest relevance to our present aims, we found that the weighting parameter was positively related to our measure of the reward rate that controlled for average chance performance for the novel task (r = 0.55, p < 0.001), but not for the Daw paradigm (r = 0.10, p = 0.15; Fig 17). A subsequent multiple regression showed that this relationship was significantly different between groups [t(377) = 4.71, p < 0.001]. 10.1371/journal.pcbi.1005090.g017Fig 17 Relationship between the estimated weighting parameters and adjusted reward rate in the Daw and novel two-step paradigms. We found a positive correlation in the novel paradigm, but not in the original paradigm, suggesting that we successfully established a tradeoff between model-based control and reward in the two-step task. Dashed lines indicate the 95% confidence interval. Next, in order to quantify the average gain in points across the entire range of w for both tasks, we ran a set of linear regression analyses predicting the reward rate from the weighting parameters for both tasks. For the Daw task, we found a predicted reward rate of 0.52 for w = 0 (i.e., the intercept), and an increase of 0.002 on top of this for w = 1 (i.e., the slope), indicating a 0.42% increase in points (to 0.525). For the novel task, we found a predicted reward rate of 0.58 for w = 0, and an increase of 0.67 on top of this for reward rate w = 1, indicating a 215% increase in points (to 1.25). When we computed these slopes using the corrected reward rates, subtracting the average value of each participant’s reward distribution from their reward rate, we found an average increase in reward rate of 0.01 across the range of the weighting parameter for the Daw task, and an average increase in reward rate of 0.51 for the novel task. These results validate the accuracy-demand trade-off of the novel two-step paradigm, and also demonstrate that the original Daw two-step paradigm does not embody such a trade-off. Discussion The distinction between planning and habit lies at the core of behavioral and neuroscientific research, and plays a central role in contemporary dual process models of cognition and decision making. Modern reinforcement learning theories formalize the distinction in terms of model-based and model-free control, bringing new computational precision to the long-recognized trade-off between accuracy and demand in decision making. In principle, the model-based strategy attains more accurate performance through increased effort relative to the computationally inexpensive but more inaccurate model-free strategy. Yet, building on prior work [9], we provide an exhaustive demonstration that the hallmark task for dissociating model-based and model-free control—the Daw two-step paradigm (Fig 3A) and several related variants of this task (Figs 4 and 5)—do not embody a trade-off between accuracy and demand across a wide range of parameter space. Using simulations of reinforcement learning agents on variants of the two-step task, we have identified five features that reduce the reward associated with model-based control to such a degree that pure model-based and model-free agents obtain equivalent reward (Fig 13). By systematically eliminating these features from the task, we developed a novel variant that shows a strong relationship between model-based control and reward both in simulations and in experimental data. In addition to providing new insight into the affordances of distinct experimental paradigms, our findings demonstrate that the magnitude of the accuracy-demand trade-off varies greatly with the specific features of any given task. First, we found that the trade-off depends on highly distinguishable reward probabilities. Broadening the range of possible reward probabilities (from 0 to 1) contributed a small, but measurable effect on the relationship between model-based control and reward (Simulation 1, Fig 6B). Second, we found that the trade-off depends on the rate of change of the second-stage reward probabilities. Our analyses indicates that the rate of change in the original paradigm was too slow to elicit a reliable accuracy-demand trade-off, because it allowed the model-free strategy to integrate sufficient information over trials to match the performance of the model-based strategy (Fig 7A). Based on this analysis, we showed that a task with larger drift rate produced a stronger relationship between model-based control and reward (Simulation 2; Fig 7B). Third, the trade-off can be limited by the presence of stochastic transitions. In the original two-step task, model-based choices do not always lead to the desired second-stage state, since this paradigm includes rare transitions from the first to the second stage, reducing the efficacy of model-based control. A new transition structure, using deterministic transitions from two different starting states, avoids this issue, and substantially strengthens the accuracy-demand trade-off (Simulation 3; Fig 8A). Fourth, the trade-off is limited when the environment contains a large number of actions bounded by the same rewarded probabilities. Specifically, by reducing the number of second-stage choice options, the average difference in value between the optimal choices of the two second-stage states is increased, which allows the model-based advantage at the first stage to emerge more distinctly. This change to the paradigm further strengthens the accuracy-demand trade-off (Simulation 4; Fig 9). Fifth, the trade-off is limited under conditions of high uncertainty about the reward value of actions. Specifically, we found that the stochastic reward observations in this task do not carry enough information about the value of the associated stimuli. Subsequently, removing the binomial noise from the reward distributions leads to a substantial increase in the strength of the accuracy-demand trade-off in this paradigm (Simulation 5; Fig 11). Moreover, we find that these factors have a superadditive effect on the relationship between model-based control and reward: All five changes to the task are required to establish a reliable accuracy-demand trade-off. We experimentally confirmed these theoretical predictions, demonstrating that the empirical estimate of model-based control in the new task was correlated with reward rate across participants. It is likely that more than these five factors alone moderate the effect of model-based control on accuracy. For example, in the Akam version of the two-step task, rewards alternate between blocks of opposite reward probabilities, so that one option strictly dominates the other until the next alternation is implemented. As discussed, this change to the paradigm resulted in a strong trade-off between control and reward in a selective region of parameter space. It is plausible that there are alternative versions of the two-step task that embody an even stronger trade-off than those discussed here, and we look forward to a comparison of how those relate to the current paradigm. In addition to the difference in the strength of the accuracy-demand trade-off between paradigms, we also found that novel two-step task elicited greater average model-based control in our participants than the original Daw two-step task. This result is one of the first pieces of behavioral evidence suggesting an adaptive trade-off between model-based and model-free control. Put simply, participants reliably shifted towards model-based control when this was a more rewarding strategy. This may indicate that participants store “controller values” summarizing the rewards associated with model-based and model-free control. However, there are alternative explanations for this result. For example, it is possible the presence of deterministic transition structure or the introduction of negative reward induced increased model-based control triggered by a Pavlovian response to these types of task features. In other words, the increase in planning might not a reflect motivational trade-off, but rather a simple decision heuristic that does not integrate computational demand and accuracy. Future investigations, where task features and reward are independently manipulated, will be able to provide more conclusive evidence that people adaptively weigh the costs and benefits of the two strategies against each other. Although the original Daw two-step task does not embody an accuracy-demand trade-off, choice behavior on this task nonetheless reflects a mixture of model-based and model-free strategies. Furthermore, the degree of model-free control on this task is predicted by individual difference measures such as working memory capacity [23], cognitive control ability [24], processing speed [29], age [20, 31], extraversion [30], and even psychiatric pathology [11, 33, 34]. This discrepancy demands explanation. Why does the original task, without a motivational trade-off, still yield meaningful and interpretable results? One possibility is that, in the absence of a reliable signal from the environment, behavior on this task reflects participants’ belief about how model-based control relates to reward maximization in the real world (where the trade-off is presumably more pervasive). Another possibility is that the extensive training of participants on the transition structure of the experiment induces them to assume they should be using it during task performance. In this sense, the absence of a trade-off is not problematic for mapping out individual differences that co-vary with the use of model-based control. This analysis can help explain the types of experimentally induced shifts in control allocation that have been reported using the two-step task, as well as those that have not. Prior research has demonstrated several factors that increase the control of model-free strategies on decision making. Control shifts to the model-free system with extensive experience [57], under cognitive load [22], and after the induction of stress [23, 28]. Such shifts are rational insofar as there is no advantage to model-based control in the task. Notably, however, few studies report factors that increase the use of model-based control. The exception to this rule is a study in which the underlying neural mechanism was altered by administering dopamine agonists after which control shifted to the model-based system [12]. Apart from this report, no other studies have successfully increased model-based control in the two-step task. Our simulation results suggest an explanation: in the original version of the two-step task, planning behavior does not improve reward, and so there is no incentive to increase the contribution of the model-based system. Our novel paradigm opens up the possibility of studying the neural mechanism underlying the trade-off between model-based and model-free control. The first and most influential neuroimaging study of the two-step task [8] focused on the neural correlates of “reward prediction error” (the difference between expected and observed reward) that is used by both the model-based and model-free controllers. A host of previous research shows that model-free reward prediction errors are encoded in the striatum [36]. The results of Daw and colleagues [8] were in line with this finding; the reward prediction errors of the model-free system correlated with signal in the striatum. However, despite the distinct computational features of the two systems, the model-based reward prediction errors recruited a similar, indistinguishable, region of the striatum (see also [13]). Our recent simulations may shed light on this surprising finding, insofar as model-based system was not appropriately incentivized. An important area for future research is to identify the neural correlates of model-based control under conditions where it obtains a higher average rate of reward than does model-free control. One potential limitation of the current paradigm is that it does not afford a simple qualitative characterization of model-based versus model-free control based exclusively on the relationship between reward (vs. punishment) on one trial and a consistent (vs. inconsistent) behavioral policy on the subsequent trial. As depicted in Fig 15B, both strategies predict an increased likelihood of behavioral consistency after a reward in either start state, but also a higher probability of consistency when the current start state is the same as in the previous trial compared to when the current start state is different. Our results reinforce this point. Even though the raw consistency behavior was not able to distinguish between the pure model-free and mixture strategies, our model-fitting procedure showed that most participants employed both model-based and model-free strategies. Indeed, our exploration of this point revealed an apparent mystery and suggests a potentially illuminating explanation. Although our full model fits of participant data indicate a high degree of model-based control, this trend is not at all evident in their raw stay probabilities, conditioned on reward in the previous trial. Not only do we fail to find the high staying probability we would expect for trials on which the associated stage-one choice was previously rewarded (assuming some influence of model-based control), in fact we find an even lower stay probability than would be expected given a computational model of pure model-free control. How can we explain this divergence between our empirical result and the predictions of our generative model? Recent work on the influence of working memory capacity on reinforcement learning may shed some light on this puzzling finding. Collins and Frank [58] show that the performance accuracy on a reinforcement learning task varied as a function of the number of stimuli that had to be remembered (the load) and the delay between repetitions of the same choice. Behavior in the current task is likely to be subject to similar constraints, since the number of choice options (six) is well above the capacity limit reported by Collins and Frank [58]. Therefore, the smaller-than-predicted probability of staying after a reward in the different start state might be predicted be memory decay, since the average delay of seeing the stimuli in this state is strictly higher than in trials with the same starting state. Exploring these possibilities further, while beyond the scope of the present study, is a key area for further investigation. Finally, we observed a shift in arbitration between model-based and model-free control when comparing the original and novel versions of the two-step paradigm. Specifically, participants in the novel paradigm were more likely to adopt the model-based strategy compared to those who completed the Daw version of the task. This result is one of the first pieces of evidence that the people negotiate an accuracy-demand trade-off between model-based and model-free strategies, and is consistent with a large body of literature that suggests that increased incentives prime more intense controlled processing [56]. Though tantalizing, this result raises several new questions. For example, how does the brain adapt its allocation between model-based and model-free control? At what time scale is this possible? What is the appropriate computational account of arbitration between the two systems? What neural regions are involved in determining whether one should exert more model-based control? Future investigations, using a combination of neural, behavioral, and computational methods will aim at answering these questions. Conclusion In recent years, the Daw two-step task has become the gold standard for describing the trade-off between accuracy (model-based control) and computational demand (model-free control) in sequential decision making. Our computational simulations of this task reveal that it does not embody such a trade-off. We have developed a novel version of this task that theoretically and empirically obtains a relationship between model-based control and reward (a proxy for the accuracy-demand trade-off). The current investigation reveals a critical role for computational simulation of predicted effects, even if these appear to be intuitive and straightforward. It also introduces a new experimental tool for behavioral and neural investigations of cost-benefit trade-offs in reinforcement learning. Finally, it opens new avenues for investigating the features of specific tasks, or domains of task, that favor model-based over model-free control. Supporting Information S1 Fig The influence of the drift rate in the two-step task across a broad range of RL parameters. We found that the size of the drift rate affected the strength between model-based control and reward in a non-monotonic fashion, with the largest effect found at moderate values of the drift rate (0.1–0.3) and with a broad reward probability range. Importantly, the results of this analysis shows that this effect was not only found in the particular parameterization depicted in Fig 8 in the main text, but also across a broad range of learning rates (α) and inverse temperatures (β). (EPS) Click here for additional data file. S2 Fig Volume under the surface for all 32 tasks generated by the 5 binary factors discussed in this paper for agents with eligibility decay parameter λ = 0 and λ = 1. Each dot represents the volume under the surface of linear regression coefficients for one task, and is plotted as a function of the number of ‘beneficial’ factors that are included in each task’s design. The gray line represents the average increase in the strength of the relationship between model-based control and reward. These results are qualitatively identical to those reported in Fig 13, suggesting that λ does not reliably affect the strength of the accuracy-efficiency tradeoff. (EPS) Click here for additional data file. S1 Text Reliability analysis for the model-fitting procedure. (DOCX) Click here for additional data file. S2 Text Multi-level logistic regression analyses. (DOCX) Click here for additional data file. S1 Table Model comparison for the full hybrid model and the hybrid model without choice perseveration parameters. (DOCX) Click here for additional data file. S2 Table Model comparison for the hybrid model and pure model-based and model-free models. (DOCX) Click here for additional data file. We thank Catherine Hartley for generously sharing her stimuli, and the members of the Moral Psychology Research Laboratory and the Computational Cognitive Neuroscience Laboratory for their advice and assistance. ==== Refs References 1 Dickinson A . Actions and habits: The development of behavioural autonomy . Philosophical Transactions of the Royal Society B: Biological Sciences . 1985 ; 308 : 67 –78 . 2 Sloman SA . The empirical case for two systems of reasoning . Psychological Bulletin . 1996 ; 119 : 3 –22 . 3 Kahneman D . A perspective on judgment and choice: Mapping bounded rationality . American Psychologist . 2003 ; 58 : 697 –720 . 14584987 4 Fudenberg D , Levine DK . A dual self model of impulse control . American Economic Review . 2006 ; 96 : 1449 –76 .29135208 5 Balleine BW , O'Doherty J . Human and rodent homologies in action control: Corticostrialtal determinants of goal-directed and habitual action . Neuropsychopharmacology . 2009 ; 35 : 48 –69 . 6 Dolan RJ , Dayan P . Goals and habits in the brain . Neuron . 2013 ; 80 : 312 –25 . 10.1016/j.neuron.2013.09.007 24139036 7 Daw ND , Niv Y , Dayan P . Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control . Nature Neuroscience . 2005 ; 8 : 1704 –11 . 16286932 8 Daw ND , Gershman SJ , Seymour B , Dayan P , Dolan RJ . Model-based influences on humans' choices and striatal prediction errors . Neuron . 2011 ; 69 : 1204 –15 . 10.1016/j.neuron.2011.02.027 21435563 9 Akam T , Costa R , Dayan P . Simple Plans or Sophisticated Habits? State, Transition and Learning Interactions in the Two-Step Task . PLoS computational biology . 2015 ; 11 : e1004648 –25 . 10.1371/journal.pcbi.1004648 26657806 10 Smittenaar P , FitzGerald THB , Romei V , Wright ND , Dolan RJ . Disruption of dorsolateral prefrontal cortex decreases model-based in favor of model-free control in humans . Neuron . 2013 ; 80 : 914 –9 . 10.1016/j.neuron.2013.08.009 24206669 11 Worbe Y , Palminteri S , Savulich G , Daw ND , Fernandez-Egea E , Robbins TW , et al Valence-dependent influence of serotonin depletion on model-based choice strategy . Molecular Psychiatry . 2015 : 1 –6 .25648202 12 Wunderlich K , Smittenaar P , Dolan R . Dopamine enhances model-based over model-free choice behavior . Neuron . 2012 ; 75 : 418 –24 . 10.1016/j.neuron.2012.03.042 22884326 13 Deserno L , Huys QJM , Boehme R , Buchert R , Heinze H-J , Grace AA , et al Ventral striatal dopamine reflects behavioral and neural signatures of model-based control during sequential decision making . Proceedings of the National Academy of Sciences . 2015 ; 112 : 1595 –600 . 14 Doll BB , Bath KG , Daw ND , Frank MJ . Variability in dopamine genes dissociates model-based and model-free reinforcement kearning . Journal of Neuroscience . 2016 ; 36 : 1211 –22 . 10.1523/JNEUROSCI.1901-15.2016 26818509 15 Doll BB , Duncan KD , Simon DA , Shohamy D , Daw ND . Model-based choices involve prospective neural activity . Nature Neuroscience . 2015 ; 18 : 767 –72 . 10.1038/nn.3981 25799041 16 Doll BB , Hutchison KE , Frank MJ . Dopaminergic genes predict individual differences in susceptibility to confirmation bias . Journal of Neuroscience . 2011 ; 31 : 6188 –98 . 10.1523/JNEUROSCI.6486-10.2011 21508242 17 Morris LS , Kundu P , Dowell N , Mechelmans DJ , Favre P , Irvine MA , et al Fronto-striatal organization: Defining functional and microstructural substrates of behavioural flexibility . CORTEX . 2016 ; 74 : 118 –33 . 10.1016/j.cortex.2015.11.004 26673945 18 Smittenaar P , Prichard G , FitzGerald THB , Diedrichsen J , Dolan RJ . Transcranial direct current stimulation of right dorsolateral prefrontal cortex does not affect model-based or model-free reinforcement learning in humans . PLoS ONE . 2014 ; 9 : e86850 –8 . 10.1371/journal.pone.0086850 24475185 19 Economides M , Kurth-Nelson Z , Lübbert A , Guitart-Masip M , Dolan RJ . Model-based reasoning in humans becomes automatic with training . PLOS Computational Biology . 2015 ; 11 : e1004463 –19 . 10.1371/journal.pcbi.1004463 26379239 20 Eppinger B , Walter M , Heekeren HR , Shu-Chen L . Of goals and habits: age-related and individual differences in goal-directed decision-making . Frontiers in Neuroscience . 2013 ; 7 : 253 10.3389/fnins.2013.00253 24399925 21 Gillan CM , Otto AR , Phelps EA , Daw ND . Model-based learning protects against forming habits . Cognitive, Affective, & Behavioral Neuroscience . 2015 ; 15 : 523 –36 . 22 Otto AR , Gershman SJ , Markman AB , Daw ND . The curse of planning: Dissecting multiple reinforcement-learning systems by taxing the central executive . Psychological Science . 2013 ; 24 : 751 –61 . 10.1177/0956797612463080 23558545 23 Otto AR , Raio CM , Chiang A , Phelps E , Daw N . Working-memory capacity protects model-based learning from stress . Proceedings of the National Academy of Sciences USA . 2013 ; 110 : 20941 –6 . 24 Otto AR , Skatova A , Madlon-Kay S , Daw ND . Cognitive control predicts use of model-based reinforcement learning . Journal of Cognitive Neuroscience . 2015 ; 27 : 319 –33 . 10.1162/jocn_a_00709 25170791 25 Dezfouli A , Balleine BW . Actions, action sequences and habits: Evidence that goal-directed and habitual action control are hierarchically organized . PLOS Computational Biology . 2013 ; 9 : e1003364 –14 . 10.1371/journal.pcbi.1003364 24339762 26 Dezfouli A , Lingawi NW , Balleine BW . Habits as action sequences: hierarchical action control and changes in outcome value . Philosophical Transactions of the Royal Society of London Series B, Biological sciences . 2014 ; 369 : 20130482 –. 10.1098/rstb.2013.0482 25267824 27 Friedel E , Koch SP , Wendt J , Heinz A , Deserno L , Schlagenhauf F . Devaluation and sequential decisions: linking goal-directed and model-based behavior . Frontiers in Human Neuroscience . 2014 ; 8 : 587 10.3389/fnhum.2014.00587 25136310 28 Radenbach C , Reiter AMF , Engert V , Sjoerds Z , Villringer A , Heinze H-J , et al The interaction of acute and chronic stress impairs model-based behavioral control . Psychoneuroendocrinology . 2015 ; 53 : 268 –80 . 10.1016/j.psyneuen.2014.12.017 25662093 29 Schad DJ . Processing speed enhances model-based over model-free reinforcement learning in the presence of high working memory functioning . Frontiers in Psychology . 2014 ; 5 : 1450 10.3389/fpsyg.2014.01450 25566131 30 Skatova A , Chan PA , Daw ND . Extraversion differentiates between model-based and model-free strategies in a reinforcement learning task . Frontiers in Human Neuroscience . 2015 ; 7 : 525 . 31 Decker JH , Otto AR , Daw ND , Hartley CA . From creatures of habit to goal-directed learners: Tracking the developmental emergence of model-based reinforcement learning . Psychological Science . in press. 32 Sharp ME , Foerde K , Daw ND , Shohamy D . Dopamine selectively remediates &model-based& reward learning: a computational approach . Brain . 2015 ; 139 : 355 –64 . 10.1093/brain/awv347 26685155 33 Voon V , Baek K , Enander J , Worbe Y , Morris LS , Harrison NA , et al Motivation and value influences in the relative balance of goal-directed and habitual behaviours in obsessive-compulsive disorder . Translational Psychiatry . 2015 ; 5 : e670 –8 . 10.1038/tp.2015.165 26529423 34 Voon V , Derbyshire K , ck CRu , Irvine MA , Worbe Y , Enander J , et al Disorders of compulsivity: a common bias towards learning habits . Molecular Psychiatry . 2014 ; 20 : 345 –52 . 10.1038/mp.2014.44 24840709 35 Gillan CM , Kosinski M , Whelan R , Phelps EA , Daw ND . Characterizing a psychiatric symptom dimension related to deficits in goal-directed control . eLife . in press. 36 Schultz W , Dayan P , Montague PR . A neural substrate of prediction and reward . Science . 1997 ; 275 : 1593 –9 . 9054347 37 Sutton RS , Barto AG . Reinforcement Learning: An Introduction . Cambridge, MA : MIT Press ; 1998 12 22 . 38 Pezzulo G , Rigoli F , Chersi F . The Mixed Instrumental Controller: Using Value of Information to combine habitual choice and mental simulation . Frontiers in Psychology . 2013 ; 4 : 92 10.3389/fpsyg.2013.00092 23459512 39 Deserno L , Wilbertz T , Reiter A , Horstmann A , Neumann J , Villringer A , et al Lateral prefrontal model-based signatures are reduced in healthy individuals with high trait impulsivity . Translational Psychiatry . 2015 ; 5 : e659 10.1038/tp.2015.139 26460483 40 Gläscher J , Daw N , Dayan P , O'Doherty J . States versus rewards: dissociable neural prediction error signals underlying model-based and model-free reinforcement learning . Neuron . 2010 ; 66 : 585 –95 . 10.1016/j.neuron.2010.04.016 20510862 41 Gershman SJ , Markman AB , Otto AR . Retrospective revaluation in sequential decision making: A tale of two systems . Journal of Experimental Psychology: General . 2014 ; 143 : 182 –94 .23230992 42 Cushman F , Morris A . Habitual control of goal selection in humans . Proceedings of the National Academy of Science . 2015 . 43 Kool W , McGuire JT , Rosen ZB , Botvinick MM . Decision making and the avoidance of cognitive demand . Journal of Experimental Psychology: General . 2010 ; 139 : 665 –82 .20853993 44 Kurzban R , Duckworth AL , Kable JW , Myers J . An opportunity cost model of subjective effort and task performance . Behavioral and Brain Sciences . 2013 ; 36 : 661 –726 . 10.1017/S0140525X12003196 24304775 45 Westbrook A , Kester D , Braver TS . What is the subjective cost of cognitive effort? Load, trait, and aging effects revealed by economic preference . PLOS ONE . 2013 ; 22 : e68210 . 46 Keramati M , Dezfouli A , Piray P . Speed/accuracy trade-off between the habitual and the goal-directed processes . PLOS Computational Biology . 2011 ; 7 : e1002055 –21 . 10.1371/journal.pcbi.1002055 21637741 47 Gershman SJ , Horvitz EJ , Tenenbaum JB . Computational rationality: A converging paradigm for intelligence in brains, minds, and machines . Science . 2015 ; 349 : 273 –8 . 10.1126/science.aac6076 26185246 48 Griffiths TL , Lieder F , Goodman ND . Rational use of cognitive resources: Levels of analysis between the computational and the algorithmic . Topics in Cognitive Science . 2015 ; 7 : 217 –29 . 10.1111/tops.12142 25898807 49 Payne JW , Bettman JR , Johnson EJ . Adaptive strategy selection in decision making . Journal of Experimental Psychology: Learning, Memory, and Cognition . 1988 ; 14 : 534 –52 . 50 Rieskamp J , Otto PE . SSL: A theory of how people learn to select strategies . Journal of Experimental Psychology: General . 2006 ; 135 : 207 –36 .16719651 51 Lee SW , Shimojo S , O'Doherty JP . Neural computations underlying arbitration between model-based and model-free Learning . Neuron . 2014 ; 81 : 687 –99 . 10.1016/j.neuron.2013.11.028 24507199 52 Rummery G , Niranjan M . On-line Q-learning using connectionist systems . Cambridge University 1994 . 53 Simon DA, Daw ND. Environmental statistics and the trade-off between model-based and TD learning in humans. In: Shawe-Taylor J, Zemel R, Bartlett P, Pereira F, Weinberger K, editors. Advances in Neural Information Processing Systems. 242011. p. 127–35. 54 Gershman SJ . Empirical priors for reinforcement learning models . Journal of Mathematical Psychology . 2016 ; 71 : 1 –6 . 55 Behrens TEJ , Woolrich MW , Walton ME , Rushworth MFS . Learning the value of information in an uncertain world . Nature Neuroscience . 2007 ; 10 : 1214 –21 . 17676057 56 Botvinick MM , Braver T . Motivation and cognitive control: From behavior to neural mechanism . Annual Review of Psychology . 2015 ; 66 : 83 –113 . 10.1146/annurev-psych-010814-015044 25251491 57 Daw ND , Shohamy D . The cognitive neuroscience of motivation and learning . Social Cognition . 2008 ; 26 : 593 –620 . 58 Collins AGE , Frank MJ . How much of reinforcement learning is working memory, not reinforcement learning? A behavioral, computational, and neurogenetic analysis . European Journal of Neuroscience . 2012 ; 35 : 1024 –35 . 10.1111/j.1460-9568.2011.07980.x 22487033

PMC005xxxxxx/PMC5001644.txt

==== Front PLoS Comput BiolPLoS Comput. BiolplosploscompPLoS Computational Biology1553-734X1553-7358Public Library of Science San Francisco, CA USA 27564311PCOMPBIOL-D-15-0110510.1371/journal.pcbi.1005091Research ArticlePhysical SciencesChemistryChemical CompoundsOrganic CompoundsAmino AcidsAmino Acid SubstitutionPhysical SciencesChemistryOrganic ChemistryOrganic CompoundsAmino AcidsAmino Acid SubstitutionBiology and Life SciencesBiochemistryProteinsAmino AcidsAmino Acid SubstitutionBiology and Life SciencesMolecular BiologyMacromolecular Structure AnalysisProtein StructureBiology and Life SciencesBiochemistryProteinsProtein StructureComputer and Information SciencesData VisualizationInfographicsGraphsResearch and Analysis MethodsDatabase and Informatics MethodsBiological DatabasesProtein Structure DatabasesBiology and Life SciencesMolecular BiologyMacromolecular Structure AnalysisProtein StructureProtein Structure DatabasesBiology and Life SciencesBiochemistryProteinsProtein StructureProtein Structure DatabasesResearch and Analysis MethodsDatabase and Informatics MethodsBiological DatabasesMutation DatabasesBiology and Life SciencesGeneticsMutationMutation DatabasesBiology and Life SciencesMolecular BiologyMacromolecular Structure AnalysisProtein StructureProtein Structure PredictionBiology and Life SciencesBiochemistryProteinsProtein StructureProtein Structure PredictionBiology and Life SciencesGeneticsMutationSubstitution MutationBiology and Life SciencesGeneticsMutagenesisThe Loss and Gain of Functional Amino Acid Residues Is a Common Mechanism Causing Human Inherited Disease Loss and Gain of Functional Residues in Inherited DiseaseLugo-Martinez Jose 1Pejaver Vikas 1Pagel Kymberleigh A. 1http://orcid.org/0000-0002-7169-9483Jain Shantanu 1Mort Matthew 2Cooper David N. 2Mooney Sean D. 3*Radivojac Predrag 1* 1 Department of Computer Science and Informatics, Indiana University, Bloomington, Indiana, United States of America 2 Institute of Medical Genetics, School of Medicine, Cardiff University, Cardiff, United Kingdom 3 Department of Biomedical Informatics and Medical Education, University of Washington, Seattle, Washington, United States of America Bromberg Yana Editor Rutgers University, UNITED STATES The authors have declared that no competing interests exist. Conceived and designed the experiments: JLM PR SDM. Performed the experiments: JLM. Analyzed the data: JLM VP KAP SJ MM DNC SDM PR. Wrote the paper: JLM MM DNC SDM PR. * E-mail: sdmooney@uw.edu (SDM); predrag@indiana.edu (PR)8 2016 26 8 2016 12 8 e10050916 7 2015 2 8 2016 © 2016 Lugo-Martinez et al2016Lugo-Martinez et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Elucidating the precise molecular events altered by disease-causing genetic variants represents a major challenge in translational bioinformatics. To this end, many studies have investigated the structural and functional impact of amino acid substitutions. Most of these studies were however limited in scope to either individual molecular functions or were concerned with functional effects (e.g. deleterious vs. neutral) without specifically considering possible molecular alterations. The recent growth of structural, molecular and genetic data presents an opportunity for more comprehensive studies to consider the structural environment of a residue of interest, to hypothesize specific molecular effects of sequence variants and to statistically associate these effects with genetic disease. In this study, we analyzed data sets of disease-causing and putatively neutral human variants mapped to protein 3D structures as part of a systematic study of the loss and gain of various types of functional attribute potentially underlying pathogenic molecular alterations. We first propose a formal model to assess probabilistically function-impacting variants. We then develop an array of structure-based functional residue predictors, evaluate their performance, and use them to quantify the impact of disease-causing amino acid substitutions on catalytic activity, metal binding, macromolecular binding, ligand binding, allosteric regulation and post-translational modifications. We show that our methodology generates actionable biological hypotheses for up to 41% of disease-causing genetic variants mapped to protein structures suggesting that it can be reliably used to guide experimental validation. Our results suggest that a significant fraction of disease-causing human variants mapping to protein structures are function-altering both in the presence and absence of stability disruption. Author Summary Identifying the molecular changes caused by mutations is a major challenge in understanding and treating human genetic disease. To address this problem, we have developed a wide range of profiling tools designed to predict specific types of functional site from protein 3D structures. We then apply these tools to data sets of inherited disease-associated and putatively neutral amino acid substitutions and estimate the relative contribution of the loss and gain of functional residues in disease. Our results suggest that alterations of molecular function are involved in a significant number of cases of human genetic disease and are over-represented as compared to putatively neutral variants. Additionally, we use experimental data to show that it is possible to computationally identify the loss of specific functional events in disease pathogenesis. Finally, our methodology can be used to reliably identify the potential molecular consequences of disease-causing genetic variants and hence prioritize experimental validation. Ford Foundation Ford Foundation pre-doctoral fellowshipLugo-Martinez Jose Qiagen Inc.Mort Matthew Qiagen Inc.Cooper David N http://dx.doi.org/10.13039/100000002National Institutes of HealthR01LM009722Mooney Sean D http://dx.doi.org/10.13039/100000002National Institutes of HealthR01LM009722Radivojac Predrag http://dx.doi.org/10.13039/100000002National Institutes of HealthR01MH105524Mooney Sean D http://dx.doi.org/10.13039/100000002National Institutes of HealthR01MH105524Radivojac Predrag JLM was supported by the Ford Foundation pre-doctoral fellowship, DNC and MM are grateful to Qiagen Inc for their financial support through a License Agreement with Cardiff University. We gratefully acknowledge the support from the National Institutes of Health through the awards R01LM009722 and R01MH105524. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityHGMD Professional is commercial and can be obtained through the license with Cardiff University. Additionally, there is a public version of HGMD data which validates the conclusions on this manuscript avaiable at http://www.hgmd.cf.ac.uk/docs/register.html. All other data are available at the following link: http://www.cs.indiana.edu/~predrag/data/lugomartinez_ploscb_2016_datasets.zip.Data Availability HGMD Professional is commercial and can be obtained through the license with Cardiff University. Additionally, there is a public version of HGMD data which validates the conclusions on this manuscript avaiable at http://www.hgmd.cf.ac.uk/docs/register.html. All other data are available at the following link: http://www.cs.indiana.edu/~predrag/data/lugomartinez_ploscb_2016_datasets.zip. ==== Body Introduction Spurred by the advances in DNA sequencing, the accumulation of human genetic variation (and with it amino acid substitution data) has over the past two decades been unprecedented. Multiple databases and resources now enumerate and annotate amino acid substitutions, their functional impact, and association with inherited disease [1–3]. However, to further our understanding of human genetic variation and its impact on disease, it is necessary to elucidate the associated molecular alterations [4–6]. Thus, the step of identifying the underlying molecular mechanisms constitutes a serious impediment to understanding and treating human disease. A straightforward approach to integrating genetic and molecular data is to search databases for structural and functional annotations at the variation site or in the neighborhood of interest, and then provide both the possible and the likely effects of mutations on these annotations [7–12]. Although this approach is useful, its major limitation is its dependence on previously observed and curated functional information as well as our inability, except in limited cases [9, 10], to cover mutations that create functional residues. Furthermore, the deterministic nature of data integration does not easily lend itself to a principled strategy of prioritizing many of the possible molecular mechanisms based on their likelihood to impact clinical phenotype, especially when a variant resides in the neighborhood of the functional site. A more comprehensive approach to analyzing the effects of amino acid substitutions involves the use of statistical inference methods that predict functional impact. Although there are many studies that have adopted this strategy using the data from protein sequence or structure [13], most methods make inferences without specifying which functional property has been impacted. Such an approach, however, is feasible if the methodology can be developed to predict a specific function, say a phosphorylation site or a catalytic residue, which is then applied to sequence variants [14–16]. Furthermore, these specific functional predictions can be integrated with general variant effect predictors to provide probabilistic estimates of molecular mechanisms of disease [17]. While many successful machine learning models can be made based on sequence information alone, structural information can provide additional benefits [18, 19]. This suggests that further improvements could be made if specific predictors of protein function could be integrated into this pipeline [20]. Wang and Moult published a seminal work on the impact of germline variants on protein function [7]. They searched the Protein Data Bank (PDB) and used homology modeling to obtain 3D structures of wild-type proteins as a means to characterize the structural and functional effects of both disease and neutral variants. They reported that the majority of disease-causing substitutions affect protein stability, whereas a relatively small proportion directly disrupt molecular function. By contrast, Sahni et al. observed a rather larger fraction of function-impacting variants in their experimental studies into the impact of variants on protein-protein and protein-DNA interactions [21]. Their work therefore challenges the traditional view of the dominance of structure-impacting changes. Finally, Steward et al. examined the structural, functional and physicochemical features of wild-type protein structures where disease-causing variants occur [22]. Unfortunately, the scope of these and several other studies was limited to characterizing the functional effects of amino acid substitutions across a handful of protein functions [23, 24]. There is therefore a need for large-scale studies that use statistical inference methods based on protein structure to explore the relative contributions made by disruption of functional sites in disease pathogenesis. In this work, we carry out a systematic study of the alterations of specific functional sites as the underlying molecular mechanisms of disease over a data set comprising germline disease-causing amino acid substitutions mapped to protein 3D structures. In particular, we develop multiple structure-based functional residue predictors and assess the impact of disease-associated substitutions on catalytic residues, metal-binding sites, macromolecular binding sites, ligand-binding sites, allosteric sites and post-translational modification (PTM) sites. We then quantify the extent to which disruption or introduction of particular types of functional site accounts for the deleterious impact of amino acid substitutions. Our results provide evidence to support the view that the increased and decreased propensity of particular functional activities are common in human inherited disease. Materials and Methods Probabilistic model for alteration of residue function For a given protein structure and a missense variant, we are interested in estimating whether a particular residue functionality, say f, has been impacted. To achieve this, we broadly distinguish between two scenarios resulting in alteration of function: (i) the mutation disrupts protein stability and subsequently impacts residue function and (ii) the mutation does not impact stability and structure yet still leads to an altered function as a consequence of modified functional propensity. The latter scenario might occur, for example, for tyrosine-to-phenylalanine substitutions that result in minimal structural changes, yet the mutation itself may have significant impact on protein phosphorylation and downstream events. This is because phenylalanine cannot be phosphorylated to subsequently create an SH2 binding site [25]. We informally refer to the events of increased or decreased functional propensity as gain and loss of functional activity. We formalize this approach as follows. Let x be a collection of features that encodes a particular mutation in a protein structure and let P(loss of f|x) be the probability of the loss of residue function f consequent to mutation. Then, we can write P(loss of f|x)=P(loss of f|S,x)P(S|x) + P(loss of f|S¯,x)P(S¯|x), where the event S indicates that protein stability is significantly changed and S¯ indicates that the stability is not significantly changed; i.e., P(S|x)=1-P(S¯|x). This expression gives a probabilistic formulation that can be used to estimate loss of a specific function f, with an assumption of a dichotomized impact on protein stability. We use this simple approach in part because of the issues involved in obtaining large amounts of high-quality data to assess the impact of sequence variants on stability. We note, however, that the expression can be generalized to multiple groups of stability disruption and in the limit the sum would turn into an integral. Now we briefly discuss estimating P(S|x), P(loss of f|S, x), and P(loss of f|S¯,x). The posterior probability of stability disruption P(S|x) can be determined by developing a computational model given a representative data set of variants that significantly impact stability of the protein (both stabilizing and destabilizing mutations) and a representative set of variants that do not. The negative data set can also be substituted by a large representative set of variants for which the impact on stability is unknown [26, 27]. This formulation falls into the category of positive-unlabeled learning [28], a version of semi-supervised learning in which the set of negative examples is unavailable or ignored; e.g., because available negative examples are biased. Next, we discuss estimating P(loss of f|S¯,x); i.e., the probability of the loss of function given that there is no significant stability disruption. Let x′ be a collection of features that encodes the structural environment of a residue and let f be the specific residue function of interest; e.g., whether the residue is phosphorylated, DNA-binding, etc. Let P(f|xwt′) denote the probability that the residue at this site is functional in the wild-type protein and P(f|xmt′) the probability that the mutated residue, at the same locus in the protein, is functional. We then define the probability of the loss of residue function f as P(loss of f|S¯,x)=P(f|xwt′)·(1-P(f|xmt′)), where 1-P(f|xmt′) gives the probability that the residue in the mutant protein is not functional. To estimate P(loss of f|S¯,x), we can employ the same functional residue predictor to compute prediction scores on the wild-type and mutant proteins. That is, because there are no structural changes, the same structure-based classifier can be used to compute both P(f|xwt′) and P(f|xmt′), with the only difference being the replaced amino acid in the feature set xmt′. As before, the probabilistic model P(f|x′) can be developed using data sets of positive and negative examples or, in the absence of representative negative examples, using positive and unlabeled data. We will discuss the details of approximating the posterior probability of protein function in the next section. Finally, we can consider that large changes in protein stability and structure always abolish the function of a residue; i.e., P(f|xmt′)≈0. This implies that P(loss of f|S,x)≈P(f|xwt′), that is, the probability of the loss of function at a particular residue is roughly equal to the probability that the residue was functional in the first place. In addition to the loss of protein function, we can also consider the event of the gain of residue function, where P(gain of f|S¯,x)=(1-P(f|xwt))·P(f|xmt). This formulation accounts for the changes in residue microenvironments that increase its functional propensity. While most amino acid substitutions found in nature are neutral or disruptive, there are many examples in which they lead to the gain-of-function events. For example, generation of a sequence motif NX[S/T] has been observed to result in gain of N-linked glycosylation events and disease [9]. Similarly, changes in catalytic residues have been observed to increase the efficiency of catalysis, also with phenotypic implications [29]. Furthermore, assuming that significant stability changes rarely lead to gain of function we can simply take P(gain of f|S,x)≈0. In this paper, we are interested in specific types of residue function f for which sufficiently large data sets could be extracted from biological databases. We organize these residues into catalytic residues, metal-binding residues, macromolecule-binding residues, ligand-binding residues, post-translationally modified sites, and allosteric residues. For the purposes of our study, we consider certain types of residues to be functional although they may also be important for protein stability. For example, a disruption of certain metal-binding sites, say a Zn2+-binding residue, will be considered here as disruption of functional residues that consequently impacts protein stability. We do, however, note that this distinction is somewhat philosophical. Training stability predictors and functional residue predictors All classification models in this work were trained using the positive-unlabeled framework in which we are given a set of positive examples and a set of unlabeled examples. In the case of stability predictors, P(S|x), the positive examples represent mutations that have been experimentally shown to significantly impact protein stability; based on the previous studies we selected these mutations to be either stabilizing or destabilizing with |ΔΔG| > 0.5 kcal/mol [30], although some other studies use higher values [31, 32]. The set of unlabeled examples, on the other hand, was selected using a database of human variants, dbSNP, mapped to available protein structures in PDB. In the case of functional residue predictors P(f|x′), the positive examples were selected by integrating structural and molecular data that provide experimentally observed functional residues, whereas the unlabeled examples were selected from a set of monomeric proteins in PDB. We will describe all data sets precisely at the end of the Methods section. We next discuss how to train a classification model from positive and unlabeled data. Let DL={(xi,yi)}i=1m be a labeled data set, where xi∈X is an input example and yi ∈ {−1, +1} is its class label. Let DU={xi}i=1n be a set of unlabeled examples. In the problem of learning whether a mutation impacts stability, x encodes a set of features corresponding to the mutation and y = +1 indicates large stability disruption. Similarly, in the case of functional site predictors, x encodes a particular residue microenvironment in a protein and y = +1 indicates that the residue is functional. In the positive-unlabeled formulation, all examples in DL have positive class labels, whereas DU is a mixture of positive and negative examples. The probability of positive examples P(y = +1) in the unlabeled set is referred to as the class prior. The task of the predictor is to learn the probability P(y = +1|x) when provided data sets DL and DU. Unfortunately, learning P(y = +1|x) is not straightforward because the negative examples are not available. To address this problem we rely on the body of work in semi-supervised learning that decomposes the problem into the training of a non-traditional classifier [26]; i.e., a model that distinguishes between labeled and unlabeled data, and estimating the class prior P(y = +1). We denote the posterior probability from a non-traditional classifier as P(l = +1|x), where l = +1 refers to the event of data point being labeled. We approximate these probabilities using kernel-based learning with support vector machine (SVM) classifiers as underlying optimization engines. Additionally, we estimate P(y = +1) using the AlphaMax algorithm [27], and point out other available options for an interested reader [26, 33]. Under mild assumptions [27], the output of a non-traditional classifier P(l = +1|x) can be converted into the output of a traditional classifier P(y = +1|x) using P(y=+1|x)=P(y=+1)·nm·P(l=+1|x)1-P(l=+1|x), where m and n are the sizes of labeled and unlabeled data sets, respectively. This predictor can now be applied to any data set D to compute the frequency of the phenomenon using the empirical mean formula 1|D|∑x∈DP(y=+1|x). The probability of alteration for multiple types of function We previously considered the loss and gain of the specific function f at a particular residue of interest. We now extend this definition to multiple types of functional residues as follows. Consider an event of loss of any function f from a set F. We can use previous reasoning to re-write the earlier expression as P(loss of F|x)=P(loss of F|S,x)P(S|x)+P(loss of F|S¯,x)P(S¯|x). To compute this probability, we need to compute probabilities P(loss of F|S,x) and P(loss of F|S¯,x). Because the functional data is too sparse to learn the joint (posterior) models of residue function, we consider two models to approximate this probability using the marginal (posterior) models that the residue is functional. In the first model that we refer to as the independence model, we consider each type of functional residue to be independent of others and write P(loss of F|S¯,x)=1-∏f∈F(1-P(loss of f|S¯,x)). The expression above is the probability that at least one of the functions from F has been lost. Because the functions are not in reality independent, this model may lead to overestimation. The second, more conservative model, approximates the probability of loss as P(loss of F|S¯,x)=maxf∈FP(lossoff|S¯,x). We refer to this model as the max model. Equivalent expressions can be written for P(loss of F|S,x) as well as for the gain-of-function events. We note that F may contain particular groups of functions, say all types of metal binding, or can be used for all functions considered in this work. Graphlet kernels In this section we briefly summarize the graphlet kernel prediction framework and show how these kernels were used to train both stability predictors and functional site predictors. Graphs A graph G is a pair (V, E), where V is a set of vertices (nodes) and E ⊆ V × V is a set of edges. In a vertex-labeled (colored) graph, a labeling function g is defined as g: V → Σ, where Σ is a finite alphabet, commonly referred to as vertex alphabet. A graph without self-loops, i.e. where (v, v) ∉ E, ∀v ∈ V, is said to be simple. An undirected graph is a graph where the order of the vertices in each pair (u, v) ∈ E can be ignored; otherwise, the graph is said to be a directed graph. A rooted graph G is a graph together with a distinguished vertex termed the root. Graphlets A graphlet is a small, simple, connected, rooted graph. We refer to a graphlet with n vertices as an n-graphlet. For more information on graphlets, we direct the reader to [34–38]. Edit distance graphlet kernels Consider a vertex-labeled graph G = (V, E, g, Σ), where |Σ| ≥ 1. Lugo-Martinez and Radivojac [38] defined the m-edit distance representation of vertex v as ϕ(n,m)(v)=(ψ(n1,m)(v),ψ(n2,m)(v),…,ψ(nκ(n,Σ),m)(v)), where ψ(ni,m)(v)=∑nj∈E(ni,m)w(ni,nj)·φnj(v). In the previous expression φnj(v) is the count of the j-th labeled n-graphlet rooted at v, κ(n, Σ) is the total number of vertex-labeled n-graphlets and E(ni, m) is a set of n-graphlets such that for each nj ∈ E(ni, m) there exists an edit distance path of length at most m that transforms ni into nj. That is, the number of edit operations necessary to transform ni into nj is at most m, where edit operations are defined as insertion or deletion of vertices and edges, or in the case of labeled graphs, substitutions of vertex and edge labels. Finally, weights w(ni, nj) ≥ 0 are used to adjust the influence of pseudo counts and control computational complexity; in this study, we set w(ni, nj) = 1 if nj ∈ E(ni, m) and w(ni, nj) = 0 otherwise. The length-m edit distance n-graphlet kernel k(n,m)(u, v) between vertices u and v can be computed as an inner product between the respective count vectors ϕ(n,m)(u) and ϕ(n,m)(v). Hence, the length-m edit distance graphlet kernel function can be expressed as km(u,v)=∑n=1Nk(n,m)(u,v), where N is a small integer; typically defined up to N = 5 for undirected graphs. Additionally, one can define two subclasses of edit distance kernels referred to as (vertex) label-substitution kl (only allows substitutions of vertex labels) and edge-indel kernels ke (only allows insertion or deletion of edges). It is worth noting that if m = 0, then km, kml and kme are all equivalent to the standard graphlet kernel on labeled graphs [37]. In this work, we only considered the normalized kernel calculated as k(u,v)=k*(u,v)k*(u,u)k*(v,v), where k*(u, v) can be km(u, v), kml(u,v), or kme(u,v). The normalized kernel has been previously shown to have favorable performance with respect to non-normalized kernels [37, 38]. Practical aspects of training We used the graphlet kernel framework and SVM classifiers to construct all functional site predictors. First, we modeled protein structures as protein contact graphs, where each amino acid residue was represented as a vertex and two spatially close residues (i.e. 4.5Å or less between any two atoms) were linked by an undirected edge. Fig 1 illustrates a contact graph for a protein kinase rooted at a tyrosine residue at position 148. Next, we computed a set of normalized graphlet kernel matrices K using kml(xi,xj),kme(xi,xj) and km(xi, xj) for all pairs (xi, xj). For each k∈K, we used SVMlight [39] and the default value for the capacity parameter to train a predictor. We incorporated evolutionary information by extending the vertex alphabet Σ from the 20 standard amino acids to 40 based on the median residue conservation observed over the entire data set [38]. For example, the amino acid alanine was split into highly conserved alanines (represented as A) and other alanines (represented as a). Once each predictor was trained, we used Platt’s correction to adjust the outputs of the predictor to the 0-1 range [40]. 10.1371/journal.pcbi.1005091.g001Fig 1 Left: Structure of human aurora kinase A, chain A fragment (PDB entry 2j4z) with highlighted phosphorylation site Tyr148 (denoted as Y148). Right: Corresponding level-3 protein contact graph centered at Tyr148 (denoted with double circles). Nodes represent amino acid residues and edges correspond to spatially neighboring residues (i.e. 6Å or less between Cα atoms). In the case of stability predictors, we augmented the graphlet kernel representation using 33 features previously shown to be informative [41, 42]. In particular, we used 20 features for the 20 different amino acids to encode the mutation information (i.e. −1 for wild-type residue; +1 for mutated residue; 0 otherwise), and 13 features to encode the difference between physicochemical properties between wild-type and mutant amino acid residues [43–50]. Evaluation of in silico predictions The performance of each predictor was first evaluated through a per-chain 10-fold cross-validation. In each iteration of cross-validation, 10% of protein chains were selected for the test set, whereas the remaining 90% were used for training. This enforces that all data points from the same protein sequence belong to either training or test set and, thus, reduces the chance of overestimating the accuracy of the models. We estimated the area under the ROC curve (AUC), which plots the true positive rate as a function of the false positive rate and the Matthews correlation coefficient (MCC). It is impractical to validate, in vitro or in vivo, the functional effects of each amino acid substitution in our data sets. Therefore, we used independent mutagenesis experimental data to additionally evaluate the performance of our functional site predictors and also evaluate predictions of the loss of functional residues. More specifically, we downloaded all human mutagenesis experimental data from UniProt as of September 2014. This data set comprised 14,933 substitutions from 3,044 distinct proteins. We removed all entries associated with more than one substitution. The resulting 11,425 sites were mapped to high-quality PDB structures using the same steps described in the next section. The final data set comprised 3,356 amino acid substitutions from 2,809 different sites in 880 proteins. For each site in this data set, we extracted functional annotations related to metal binding, PTMs, active sites, macromolecular binding, ligand binding and allosteric activity. Then, for each functional site predictor, we built an independent test set such that (i) each site belonged to a chain that was less than 40% identical to any chain in the training data, (ii) there were at least five positive sites in each test set. The resulting set was used to assess the performance of the functional predictors independently of the cross-validation. Similarly, we created a test data set to evaluate loss-of-function predictions as follows: we searched the description of the mutagenesis experiment such that there was an experimentally observed disruption of a functional site. The resulting non-redundant and filtered data sets were then used to estimate the AUC and MCC of all loss-of-function predictors. We attempted to carry out the same steps for gain-of-function events but there were insufficient data. Disease-associated mutation data Missense variants causing inherited disease were obtained from the Human Gene Mutation Database (HGMD) as of June 2013. A set of unlabeled inherited variants was downloaded from dbSNP v.137. All amino acid substitutions were then mapped to protein structures in PDB as follows: (i) a database of amino acid sequences from X-ray crystallographic protein structures with more than 50 amino acids and resolution less than 2.5Å was created, and (ii) for each variant, a 51 residue long sequence centered around the wild-type amino acid at the variant position was aligned using BLAST [51] against the atom sequences in PDB. All alignments without an exact match; i.e., with gaps or sequence identity lower than 100%, were excluded from this study and in the case of multiple exact matches, the structure with the best resolution was selected. This resulted in 10,629 (out of 52,406) disease-causing amino acid substitutions and 8,417 (out of 282,625) unlabeled amino acid substitutions being successfully mapped to high-quality PDB structures. Table 1 summarizes both data sets. There exists an overlap between the HGMD (disease) and dbSNP (unlabeled) variants; the subset of dbSNP variants after the removal of HGMD variants will be referred to as putatively neutral variants (Table 1). 10.1371/journal.pcbi.1005091.t001Table 1 Summary of amino acid substitution (AAS) data sets. Data set name nAAS ns nPDB nc Inherited disease 52406 10629 1177 1387 Unlabeled variants 282625 8417 3121 3585 Putatively neutral 282625 8049 3047 3500 For each data set, we show the total number of amino acid substitutions (nAAS), the number of substitutions mapped to PDB (ns), the number of PDB entries (nPDB) and the number of protein chains (nc). Functional site data sets Metal ions annotated in X-ray structures from PDB as of May 2012 were selected using the HETATM field [52]. A metal-binding residue was defined as the residue that has at least one heavy atom (N, O or S) within 3Å of the metal ion. In order to build an unbiased classifier, we removed chains with: (i) more than 40% sequence identity with any other chain in the data set, (ii) crystallographic resolution greater than or equal to 2.5Å, and (iii) R-values greater than or equal to 30%. We only considered data sets with more than 100 metal-binding residues and we refer to these residues as positives. N-linked glycosylation sites were parsed from PDB and filtered in the same way as the metal-binding sites. In the case of phosphorylation sites, we used the data set assembled previously [20]. Catalytic residues were collected from the Catalytic Site Atlas v2.2.10 [53] and only literature-supported sites were kept as positive examples. As with the previous data sets, we filtered out chains with more than 40% sequence identity. DNA-binding sites were collected from Yan et al. [54], RNA-binding sites were downloaded from ccPDB [55] and protein-protein interaction (PPI) sites were obtained from Chung et al. [56]. Each data set was further filtered to remove redundancy. Protein-protein interaction hot spot residues were obtained from Lise et al. [57]. This data set was created from ASEdb [58] and only sites that mapped to PDB were used. Chains with more than 35% sequence identity were filtered out in the original work. In our analysis, we used ΔΔG greater than 1kcal/mol as the cutoff for hot spots. Ligand binding data sets were collected from ccPDB as of November 2014 and allosteric site data were downloaded from ASD v2.0 [59]. For each data set, we removed chains with resolution greater than or equal to 2.5Å. Protein stability data was collected from Capriotti et al. [41]. We used the S1615 data set which consists of 1615 single site mutations extracted from 42 different proteins in the ProTherm database [60]. The attributes for each data point included solvent accessibility, pH value, temperature and energy change ΔΔG. As previously noted, positive data points comprised mutations with |ΔΔG|>0.5. We further filtered out 112 redundant data points. Table 2 summarizes the protein stability and functional site data sets used in this study. In all situations, the unlabeled data set was constructed using a random sample of 10,000 residues selected from the 40% non-redundant set of monomers in PDB. This set was modified in the case of post-translational modifications to include only modifiable residues; e.g., Asn for N-linked glycosylation and Ser/Thr/Tyr for phosphorylation. It is important to emphasize that the set of negative examples was allowed to contain both buried and surface-exposed residues resulting in somewhat easier downstream classification problems. On the other hand, it allowed us to apply our methods to all PDB-mappable amino acid substitutions and make unbiased inferences related to different data sets. 10.1371/journal.pcbi.1005091.t002Table 2 Performance assessment of structural and functional residue predictors using cross-validation on positive-unlabeled data sets. Category Site type nc n+ AUC sn sp MCC Protein stability Stability (S) 40 1041 0.735 0.118 0.989 0.223 Metal binding Calcium (Ca) 1092 4860 0.895 0.448 0.986 0.561 Cadmium (Cd) 199 1003 0.905 0.233 0.986 0.350 Cobalt (Co) 156 532 0.934 0.596 0.986 0.623 Copper (Cu) 105 440 0.951 0.780 0.985 0.727 Iron (Fe) 187 785 0.976 0.875 0.986 0.843 Potassium (K) 287 978 0.679 0.129 0.986 0.211 Magnesium (Mg) 1348 3282 0.859 0.435 0.986 0.561 Manganese (Mn) 366 1344 0.945 0.665 0.985 0.727 Sodium (Na) 961 2753 0.671 0.105 0.987 0.211 Nickel (Ni) 254 680 0.932 0.565 0.986 0.621 Zinc (Zn) 1307 5778 0.966 0.623 0.987 0.691 PTMs N-glycosylation (Nglyco) 339 736 0.785 0.120 0.986 0.183 Phosphorylation (Phos) 655 1157 0.810 0.375 0.987 0.504 Catalytic activity Catalytic (Cat) 721 2224 0.934 0.433 0.985 0.561 Macromolecular binding DNA-binding (DNA) 139 3791 0.815 0.193 0.987 0.332 RNA-binding (RNA) 83 3436 0.783 0.187 0.985 0.319 Protein-protein interaction (PPI) 112 4350 0.807 0.091 0.987 0.191 PPI hot spots (Hotspot) 35 165 0.803 0.309 0.986 0.278 Ligand binding ADP 162 2589 0.842 0.335 0.985 0.475 ATP 104 1733 0.813 0.242 0.986 0.382 FAD 80 2248 0.840 0.307 0.985 0.448 FMN 42 788 0.824 0.284 0.985 0.384 GDP 45 593 0.843 0.433 0.985 0.502 GTP 22 366 0.716 0.145 0.986 0.181 HEM 83 2246 0.847 0.220 0.986 0.361 NAD 73 1663 0.831 0.259 0.985 0.393 PLP 34 477 0.916 0.505 0.986 0.543 UDP 27 398 0.684 0.080 0.987 0.103 Allosteric regulation Allosteric (Allo) 108 682 0.636 0.041 0.985 0.050 For each data set, we show the number of protein chains (nc) and the number of positive examples (n+). Additionally, we choose a score threshold corresponding to a specificity (sp) of 99% and report sensitivity (sn) and MCC at this threshold, as well as AUC. In each classification problem, the number of unlabeled examples was set to 10,000. S1 Table predictions lists the full name of each ligand code used. For the purposes of this work, structurally important amino acid residues such as specific metal ion binding residues were considered a part of the portfolio of available residue functions. Results In this section we present the development of a stability model and a series of structure-based functional site predictors in order to examine the molecular effects of genetic variants. We evaluate the predictors through cross-validation and using an independent data set. We then summarize our results in relation to the functional impact of disease-causing substitutions and compare them to putatively neutral variants. Assessment of functional site predictors All classifiers developed in this study were constructed using positive and unlabeled data summarized in Table 2. Their performance was estimated via per-chain 10-fold cross-validation and is also shown in Table 2. S2 Table further lists the parameters for the best-performing kernel matrix obtained from a grid search over K, |Σ| = {20, 40}, m = {0, 1} and N = {4, 5}. Each predictor performance was assessed by means of the area under the ROC curve (AUC), sensitivity (sn) at 99% level of specificity (sp), and the Matthews correlation coefficient (MCC). The majority of predictors (26 out of 30) show good performance (≥ 70% AUC); however, we observe that functions related to smaller interfaces such as metal ions and active sites exhibit higher performance than other functional predictors. This result is not unexpected because predictors of macromolecular binding would have benefited from incorporating higher-order structural signatures such as clefts and pockets [20]. We also use an independent data set to evaluate a subset of functional site predictors, as depicted in S3 Table. Interestingly, most predictors, except for macromolecular binding models show similar or improved performance (AUC) values compared to those reported from cross-validation (Table 2). Overall, despite the variability of performance accuracies, limited number of independent data sets and the relatively small size of the validation data, these results provide evidence that functional site predictions are of sufficient quality to identify possible molecular alterations resulting from specific missense mutations. A literature survey suggests that our predictors perform well when compared to established structure-based methods. Extensive comparisons with other work are difficult and were beyond the scope of this study as our main goal was to probabilistically assess molecular mechanisms of disease. A set of predictors built using the same methodology was best suited to this task. Estimating prior and posterior probabilities To use the formal framework laid out in the Methods section, it is important that all methods approximate posterior distributions. Using positive and unlabeled data, we have approached this problem in two steps: (i) by developing classifiers that discriminate between labeled and unlabeled data, and (ii) by estimating the class priors of the positive class in the unlabeled data [27]. Estimated class priors are a particularly useful by-product of learning posterior distributions. For the stability predictor, we estimate up to 13% of unlabeled variants to significantly impact stability using the AlphaMax algorithm [27]. When the stability model was applied to disease variants only, we estimate 14% of these variants to be impactful using the empirical mean formula. It should be noted that when the known disease variants were removed from the unlabeled data set, only 7% of the remaining variants were estimated to severely impact stability. In the case of functional predictors, we applied the AlphaMax algorithm using a set of positive variants and a set of 10,000 variants randomly sampled from a set of non-redundant monomers in PDB (S4 Table). In the case of catalytic residues, we estimate that up to 3% of PDB residues to be catalytic; however, about 5% of disease-causing and 2% of putatively neutral variants were estimated to be catalytic residues, etc. Overall, we generally observe a larger fraction of function-impacting variants in the disease-causing data set as compared with the putatively neutral variants. Applying loss and gain functional site predictors to human variants We applied the structure-based predictors on both the wild-type and mutant structural environments as a means to identify and categorize the functional effects of amino acid substitutions causing inherited disease. The distribution of scores on the putatively neutral variants was used as an empirical null distribution. We then used a particular false positive rate (FPR) value to determine a prediction threshold at which to assess the fraction of disease mutations with loss or gain scores that are as high as or higher than the threshold. Table 3 summarizes the relative contributions of disease mutations that either decrease (loss) or increase (gain) the propensity of functional sites at a conservative threshold of 1% FPR for six different prediction outputs. Fig 2 visualizes a subset of these results for the case when stability is not impacted. Together with Table 3, it provides evidence that the loss and gain of functional sites exist even when protein stability is not disrupted; e.g., in the case of loss of function see columns P(loss|S¯,x) and P(loss,S¯|x) that roughly have the same values as P(loss|x). We mention that when either a loss or a gain of function event is found to be statistically significant, the mutation of this type of functional residue is considered to be an active mechanism of genetic disease. For instance, at 1% FPR, we observe that loss of catalytic residues (Cat; 3.34%; p-value = 1.93 ⋅ 10−28) and iron-binding residues, (Fe; 3.17%; p-value = 2.06 ⋅ 10−25) are among the most significantly affected molecular mechanisms. 10.1371/journal.pcbi.1005091.g002Fig 2 Percentage of disease variants with prediction scores at the 1% false positive rate threshold in putatively neutral variants. For each function f, the bars indicate the percentage (%) of disease mutations that have a greater P(loss|S¯,x) and P(gain|S¯,x) than a conservative threshold at 1% false positive rates (FPR). These thresholds are estimated from the null distributions of P(loss|S¯,x) and P(gain|S¯,x) on the set of dbSNP neutral data, respectively. *Indicates significant p-value measured as a one-tailed Fisher’s Exact test after Bonferroni correction for multiple hypothesis testing (p < 8.62 ⋅ 10−4). The red line indicates the percentage of neutral variants that have greater P(loss|S¯,x) and P(gain|S¯,x) which is exactly 1%. 10.1371/journal.pcbi.1005091.t003Table 3 Percentage of disease variants with prediction scores at the 1% false positive rate threshold in putatively neutral variants. Data set Single type loss events (%) Single type gain events (%) P(loss|S¯,x) P(loss,S¯|x) P(loss|x) P(gain|S¯,x) P(gain,S¯|x) P(gain|x) Ca 2.65* 2.35* 2.69* 3.90* 3.63* 3.63* Cd 1.42 1.31 1.44 1.20 1.02 1.02 Co 1.97* 1.94* 1.98* 0.80 0.70 0.70 Cu 1.87* 1.82* 1.88* 1.29 1.15 1.15 Fe 3.14* 2.97* 3.17* 1.77* 1.62* 1.62* K 2.45* 2.08* 2.45* 2.99* 2.60* 2.60* Mg 3.22* 2.83* 3.14* 3.09* 2.89* 2.89* Mn 2.42* 2.22* 2.45* 2.65* 2.47* 2.47* Na 3.10* 2.55* 3.13* 2.50* 3.05* 3.05* Ni 1.39 1.33 1.40 0.62 0.54 0.54 Zn 2.86* 2.64* 2.89* 1.81* 1.63* 1.63* Nglyco 3.75 1.25 2.81 0.46 0.23 0.23 Phos 1.69 1.47 1.69 0.58 0.46 0.46 Cat 3.18* 2.90* 3.34* 4.45* 3.94* 3.94* DNA 1.58* 1.39 1.61* 1.85* 1.75* 1.75* RNA 0.96 0.89 0.96 1.25 1.05 1.05 PPI 1.53* 1.27 1.65* 1.94* 1.79* 1.79* Hotspot 1.00 0.90 1.00 1.53* 1.45 1.45 ADP 3.12* 2.92* 3.16* 4.03* 3.68* 3.68* ATP 2.73* 2.52* 2.76* 2.76* 2.41* 2.41* FAD 2.77* 2.58* 2.81* 3.21* 2.92* 2.92* FMN 2.15* 2.01* 2.17* 2.40* 2.18 2.18* GDP 2.07* 1.99* 2.08* 2.41* 2.18* 2.18* GTP 1.72* 1.48 1.72* 2.78* 2.15* 2.15* HEM 2.26* 2.05* 2.35* 2.39* 2.05* 2.05* NAD 3.00* 2.64* 3.09* 2.53* 2.15* 2.15* PLP 3.15* 2.98* 3.10* 2.61* 2.37* 2.37* UDP 2.70* 2.41* 2.71* 3.11* 2.46* 2.46* Allo 1.47 1.24 1.49 1.96* 1.74* 1.74* Model Multi-type loss events (%) Multi-type gain events (%) P(loss|S¯,x) P(loss,S¯|x) P(loss|x) P(gain|S¯,x) P(gain,S¯|x) P(gain|x) Independence 3.71* 3.31* 3.89* 4.13* 3.37* 3.37* Max 3.50* 1.93* 3.56* 4.35* 2.63* 2.63* For each of the six prediction outputs and each function f, we show the percentage (%) of disease mutations that have a greater probability of loss and gain of function than a threshold corresponding to a 1% false positive rate (FPR). S1 and S2 Figs show an instance of the inverse cumulative distribution function of P(loss|x) and P(gain|x), respectively. These thresholds were estimated from the empirical null distributions of the probability of loss or gain of function on the set of dbSNP neutral data. *Indicates significant p-value measured by a one-tailed Fisher’s exact test after Bonferroni correction for multiple comparisons. The p-value was separately estimated for each type of posterior distribution, jointly for loss and gain events (p<0.0558=8.62·10-4). The p-values for the combined models were corrected separately (p<0.054=1.25·10-2). Table 3 also summarizes the statistical enrichment of impact on at least one functional site from the entire repertoire of functions using the independence and max models (see Methods). Here we observe a strong enrichment in all categories of loss of function, with or without impact on stability, for both the independence and max models. Additionally, we also see an enrichment in the gain-of-function events. These results provide statistical support for many individual studies that identify loss of function as a signature of human inherited disease. Overall, our results suggest that with some exceptions, the loss of functional residues is enriched and common in human inherited disease; similarly, the gain of functional residues is observed to be an active mechanism in catalytic activity, most types of ligand-binding residues, and majority of metal-binding residues. In contrast to previous studies, our results suggest that the loss and gain of PTM sites do not show statistically significant enrichment in disease (although we observe enrichment for the loss); however, we note that this may be due to a considerable reduction of training data imposed by the availability of protein 3D structures, especially given a relationship between post-translational modifications and intrinsically disordered proteins [61–64]. Table 4 shows the proportions of disease and putatively neutral variants across functional categories for which molecular mechanisms can be computationally hypothesized. In the first part of the table, we compute the fraction of variants for which exactly one of the member predictors reports a score as high or higher than the FPR-value determined threshold. These fractions were then computed separately for disease and neutral variants. For convenience, when a predictor outputs a value as high or higher than the value determined by a 1% FPR, we refer to this prediction as actionable hypothesis of loss or gain of function. On the other hand, when the FPR-based threshold is adjusted using the Bonferroni correction, we refer to these predictions are confident. For example, at a conservative p-value cutoff of p < 8.62 ⋅ 10−4, we find that 1.51% of mutations are likely to alter exactly one metal binding site and 1.43% may alter a single ligand binding site. For all groups of molecular mechanisms, we observe that the probability of observing a high alteration score is more than three times as likely as in the case of putatively neutral variants. 10.1371/journal.pcbi.1005091.t004Table 4 Relative contribution of loss and gain of functional categories from amino acid substitutions. Category Loss (%) Gain (%) Loss or Gain (%) Disease Neutral Disease Neutral Disease Neutral I. Single mechanism Confident biological hypotheses (p-value < 8.62 ⋅ 10−4) Metal binding 1.51 0.27 1.46 0.46 2.88 0.70 PTMs 0 0 0.01 0 0.01 0 Catalytic sites 0.73 0.06 0.40 0.07 1.14 0.14 Macromolecule binding 0.59 0.20 0.64 0.20 1.21 0.40 Ligand binding 1.43 0.47 1.86 0.46 3.07 0.87 Allosteric sites 0.09 0.06 0.25 0.06 0.35 0.12 All 3.03 0.84 3.29 0.97 5.80 1.70 Actionable biological hypotheses (p-value < 0.01) Metal binding 4.90 2.50 5.08 2.99 8.70 5.03 PTMs 0.30 0.17 0.08 0.22 0.39 0.40 Catalytic sites 3.34 1.00 3.94 1.00 7.25 2.00 Macromolecule binding 3.89 2.93 4.03 3.26 7.39 5.75 Ligand binding 9.74 4.65 9.55 4.97 13.72 7.06 Allosteric sites 1.49 1.00 1.74 1.00 3.16 2.00 All 13.28 8.35 12.93 9.21 17.43 13.36 II. Multiple mechanisms Confident biological hypotheses (p-value < 8.62 ⋅ 10−4) Metal binding 0.72 0.16 0.49 0.10 1.23 0.27 Macromolecule binding 0.05 0.04 0.17 0.04 0.22 0.07 Ligand binding 0.24 0.07 0.33 0.09 0.68 0.19 All 1.43 0.34 1.29 0.31 2.78 0.68 Actionable biological hypotheses (p-value < 0.01) Metal binding 5.96 2.31 4.75 2.42 10.37 4.62 Macromolecule binding 0.66 0.51 0.99 0.37 1.72 0.89 Ligand binding 5.55 1.90 5.43 1.85 11.04 4.11 All 12.88 5.42 12.22 5.42 23.57 10.60 For each functional site category, we show the relative contributions (%) of disease and neutral substitutions where at least one function f within a category has a greater P(loss|x)or P(gain|x) than a conservative threshold at 1% FPR. This threshold is estimated from the null distributions of P(loss|x) and P(gain|x) on the putatively neutral polymorphisms data set, respectively. The table is subdivided into two parts: (i) exactly one function (or mechanism) and (ii) two or more mechanisms. In both parts, the relative contributions are assessed at two p-value cutoffs of p < 8.62 ⋅ 10−4 and p < 0.01. Note that in a small number of cases, a loss of one function might result in the gain of another; thus, the sets of residues counted in the loss and gain may overlap. Table 4 also shows situations with two or more functional perturbations consequent to the replacement of a given amino acid residue. The amino acid substitutions disrupting multiple functions may be important in a therapeutic context because addressing a single deficiency (e.g. iron binding) may still not result in a fully corrected phenotype because other deficiencies may still remain (e.g. ligand binding). Here, we have a significantly increased likelihood of observing multi-functional alterations in the disease set compared to the putatively neutral set; i.e., the disease set is several times more likely to contain multi-functional alterations than the putatively neutral set. For instance, 1.23% of disease mutations are likely to affect at least two metal binding sites versus only 0.27% of neutral variants, whereas 0.68% of disease variants may affect more than one ligand binding site as opposed to 0.19% of neutral polymorphisms. If we combine the results for single and multiple mechanisms, we observe that 2.24% of disease variants are predicted, with high confidence, to impair metal-binding sites (1.51% loss of single site and 0.72% loss of multiple sites) and 1.67% probably impair ligand binding sites (1.43% loss of single site and 0.24% loss of more than one site), as depicted in Fig 3. Overall, we believe we can confidently propose molecular mechanisms of disease for 8.6% of all variants in the inherited disease data set whereas we only see about 2.4% of such variants in the neutral set. If we use a p-value cutoff of 0.01 without a Bonferroni correction, then we can computationally hypothesize a molecular mechanism for approximately 40.9% of disease variants. 10.1371/journal.pcbi.1005091.g003Fig 3 Relative contribution of loss and gain of functional categories on each amino acid substitutions data set. For each functional site category, we show the relative contributions (%) of disease and neutral variants where at least one function f within a category has a greater P(loss|x) or P(gain|x) than a conservative threshold at 1% FPR. This threshold is estimated from the null distribution of P(loss|x) and P(gain|x) on the putatively neutral polymorphisms data set, respectively. *Indicates significant p-value measured as a Fisher’s Exact test after Bonferroni correction for multiple hypothesis comparisons (p < 8.62⋅10−4). Validation of loss of function predictions In this study, we have proposed a novel methodology for identifying specific molecular alterations of disease mutations. Given that it is impractical to experimentally validate the predicted functional effects of each individual amino acid substitution, we use mutagenesis experimental data to independently assess the loss of functional site predictions, as shown in S5 Table. To the best of our knowledge, this is the first time a systematic assessment of computationally predicted disruptions of specific types of functional residues has been carried out in the published literature. In general, our loss of function predictors performed as expected. However, more interestingly, if one restricts the loss of function predictions to those with significant p-values (i.e. p < 0.01), then performance (AUC) rises to at least 95% for all predictors. This provides compelling evidence that our methodology can be effectively used to identify molecular mechanisms of disease and hence can be used to prioritize experimental validation. Additionally, Fig 4 depicts two case studies of loss and gain of function predictions which have been experimentally validated. We discuss each case in detail below: 10.1371/journal.pcbi.1005091.g004Fig 4 3D visualization of protein structures with experimentally supported loss and gain of function predictions. Left: SOD1 protein (chain A of PDB entry 2xjl) where residues H63, H71, H80 and D83 form a zinc binding pocket. The substitution D83G gives rise to a loss of zinc binding. Right: CA2 protein (chain A of PDB entry 1fqr) where H94, H96 and H119 are zinc-binding sites. Mutation T198E leads to an increase in zinc affinity. Loss of zinc binding in superoxide dismutase (SOD1) The functional role of SOD1 is to destroy radicals that are normally produced in cells and which are toxic to biological systems. SOD1 forms a zinc-binding pocket consisting of H63, H71, H80 and D83 [65, 66] as shown in Fig 4 (left). Mutations in SOD1 are known to be causative of amyotrophic lateral sclerosis [66–68]. However, the molecular mechanisms underlying these mutations often remain unclear. We predicted a loss of multiple functional activities for mutation D83G and identified zinc binding as the primary underlying molecular mechanism of disease. In particular, D83G has a P(f|xwt′)=0.99 and P(f|xmt′)=3.3·10-3 leading to a P(loss|x) ≈ 1, which is above the 1% FPR threshold of 0.20 with an empirical p-value of 1.2⋅10−3. A literature search for experimental evidence reveals that mutation D83G causes the destabilization of native structure which leads to protein aggregation with the formation of amyloid-like fibrils, and, ultimately, a gain of toxicity [69]. Zinc binding is a known stabilizer of protein structure and, therefore, the loss of the zinc-binding residue D83 appears to be a plausible destabilizing mechanism that ultimately impacts the biological function of SOD1. We note that the quadruple (H63, H71, H80, D83) was not part of the training data for the zinc-binding predictor. This example raises an interesting possibility that the loss of a functionally important residue (zinc-binding residue) results in a loss of stability, and ultimately leads to disease through the loss of the protein’s function. In other words, protein structure and function appears to be intimately and bidirectionally interconnected. At this moment, however, this is only a theoretical possibility because of the lack of data about the structure and stability of the wild-type and mutant proteins in the absence of zinc ions. Gain of zinc affinity in carbonic anhydrase 2 (CA2) CA2 is essential for bone resorption and osteoclast differentiation. CA2 has three zinc-binding residues at H94, H96 and H119 as shown in Fig 4 (right). There are multiple studies that have characterized the effects of variants in CA2 via mutagenesis experiments [70–74]. Among these mutations, we predicted a gain of zinc binding for T198E that was experimentally shown to increase zinc affinity. Specifically, T198E has a P(f|xwt′)=3.8·10-4 and P(f|xmt′)=0.99 leading to a P(gain|x)≈1, which is above the 1% FPR threshold of 0.35 with a p-value of 3.7⋅10−4. The triple (H94, H96, H119) was not part of the training data for the zinc-binding residue predictor. Discussion This study builds on the extensive prior work in structural bioinformatics to provide statistical evidence of the important role that alterations of multiple types of functional residue play in human genetic disease. Most of the existing work has centered around understanding the impact of sequence variants on protein stability or has only considered single types of function such as catalytic residues or protein-interaction sites [7, 20, 23, 24, 30, 75–78]. This work extends these studies by integrating the stability models with a series of functional residue predictors involving metal binding, macromolecular binding, ligand binding and others. Overall, we show and validate the feasibility of computationally predicting mutations that impair specific function using protein 3D structure data. Despite using sophisticated methodology to model loss and gain of functional residues, the nature of this research has limitations involving both data sets and methodology. First, despite major efforts employed by authors and database curators when annotating amino acid substitutions as being causative of a particular disease, it is possible that some amino acid substitutions have been misannotated as disease-causing by the original authors reporting them. Similarly, mutagenesis experimental data are known to be biased toward certain amino acid residues. For example, alanine mutations comprised about 50% of the independent amino acid substitutions data set (due to the frequent use of alanine-scanning mutagenesis). There are also limitations and biases in relation to the protein structures available in PDB as well as in selecting an appropriate set of unlabeled variants. Second, there exist both theoretical and practical limitations in the semi-supervised framework used in this work. The accuracy of our methods is predicated upon the assumption that the computational models are capable of accurately estimating the posterior probability of the class labels. This however could not be guaranteed and thus requires caution when interpreting our results. Furthermore, there are identifiability issues in estimating class priors in the positive-unlabeled framework; i.e., the estimates for the class priors do not have a unique solution and only an upper bound can be estimated [27]. On the practical side, we have been careful to prevent overfitting. We performed only minor parameter selection steps before the final functional predictors were built. Thus, there is the potential to further improve predictor performance through more extensive work. This includes the use of additional features, optimizing the distance threshold used to define an edge between two residues when constructing protein contact graphs, choice of the capacity parameter in SVMlight, among others. Finally, this work was designed to probabilistically reason about molecular mechanisms of disease and not necessarily to develop classifiers that outperform specialized models across the board. If a user needs a tool for a particular prediction task, we recommend that the most accurate predictor for this task be selected. Despite these limitations, we believe this work contributes to an improved understanding of the impact of sequence variants on protein function. We have provided a model that considers functional alteration both when stability of the protein is disrupted and when it is not disrupted (e.g. interestingly, sequence changes can exert a functional effect in disordered regions such as disorder-to-order transition [79]). We believe that our work suggests a new class of approaches to disease studies that might qualify as mechanism-driven and disease-agnostic, where one might be compelled to identify a set of molecular alterations underlying a disease phenotype without necessarily studying a single disease. While each molecular alteration is likely to require an individualized approach to drug design and therapy, we envisage that the next generation of researchers might decide to specialize in addressing particular types of functional deficiencies rather than beginning with a particular disease. Supporting Information S1 Fig Inverse cumulative distribution function (CDF) of P(loss|x). For ATP-binding predictor, we plot the inverse CDF for P(loss|x) on the disease and putatively neutral data sets, respectively. (EPS) Click here for additional data file. S2 Fig Inverse cumulative distribution function (CDF) of P(gain|x). For catalytic residue predictor, we plot the inverse CDF for P(gain|x) on the disease and putatively neutral data sets, respectively. (EPS) Click here for additional data file. S1 Table Mapping between ligand codes and names. (PDF) Click here for additional data file. S2 Table Selected kernel matrix parameters for each structural and functional site predictor. For each data set, we show the best-performing kernel matrix parameters obtained through a per-chain 10-fold cross-validation. The normalized edit distance kernel km(u, v) outperformed both kml(u,v) and kme(u,v) on each data set. Note that the edit distance kernel with m = 0 is equivalent to a standard graphlet kernel. (PDF) Click here for additional data file. S3 Table Performance assessment of functional residue predictors using an independent data set. For each prediction method, we show number of proteins (np), number of positive examples (n+), number of unlabeled examples (nu), AUC and sensitivity (sn) and MCC at score threshold corresponding to specificity (sp) of 99%. In the case of N-linked glycosylation (Nglyco), we only predict if the wild-type residue is an asparagine, whereas for phosphorylation (Phos), we only make predictions on threonine, tyrosine or serine residues. (PDF) Click here for additional data file. S4 Table Class priors for structural and functional predictors. Fraction of residues in a data set estimated to be stability-impacting or functional. Estimates on the unlabeled data were made using the AlphaMax algorithm [27]; minor manual adjustments were made by observing the log-likelihood plots. *Indicates a confident prior estimate assessed by manually observing log-likelihood plots. Estimates on the disease and putatively neutral data were made using the empirical mean formula. (PDF) Click here for additional data file. S5 Table Performance assessment of loss of function predictions using mutagenesis experimental data. Independent validation of predicted loss of functional site events using a set of mutagenesis experimental data mapped to protein structures in PDB. This mutagenesis data set contains 3,356 AAS from 880 human proteins. For each functional feature, we show the number of experimentally determined losses (nl), AUC, sensitivity (sn) and MCC corresponding to a 99% specificity (sp) threshold. Additionally, the last five columns show the number of statistically significant (p < 0.01) loss-of-function predictions (nl*), as well as estimates for AUC (AUC*), sensitivity (sn*), specificity (sp*) and MCC (MCC*) on this filtered set. (PDF) Click here for additional data file. We thank Prof. Charles Dann III for helpful comments. ==== Refs References 1 Sherry ST , Ward MH , Kholodov M , Baker J , Phan L , Smigielski EM , et al dbSNP: the NCBI database of genetic variation . Nucleic Acids Res . 2001 ;29 (1 ):308 –311 . 10.1093/nar/29.1.308 11125122 2 Bairoch A , Apweiler R , Wu CH , Barker WC , Boeckmann B , Ferro S , et al The Universal Protein Resource (UniProt) . Nucleic Acids Res . 2005 ;33 (Database issue ):D154 –159 . 10.1093/nar/gki070 15608167 3 Stenson PD , Mort M , Ball EV , Shaw K , Phillips AD , Cooper DN . The Human Gene Mutation Database: building a comprehensive mutation repository for clinical and molecular genetics, diagnostic testing and personalized genomic medicine . Hum Genet . 2014 ;133 (1 ):1 –9 . 10.1007/s00439-013-1358-4 24077912 4 Mooney S . Bioinformatics approaches and resources for single nucleotide polymorphism functional analysis . Brief Bioinform . 2005 ;6 (1 ):44 –56 . 10.1093/bib/6.1.44 15826356 5 Ng PC , Henikoff S . Predicting the effects of amino acid substitutions on protein function . Annu Rev Genomics Hum Genet . 2006 ;7 (1 ):61 –80 . 10.1146/annurev.genom.7.080505.115630 16824020 6 Karchin R . Next generation tools for the annotation of human SNPs . Brief Bioinform . 2009 ;10 (1 ):35 –52 . 10.1093/bib/bbn047 19181721 7 Wang Z , Moult J . SNPs, protein structure, and disease . Hum Mutat . 2001 ;17 (4 ):263 –270 . 10.1002/humu.22 11295823 8 Yue P , Li Z , Moult J . Loss of protein structure stability as a major causative factor in monogenic disease . J Mol Biol . 2005 ;353 (2 ):459 –473 . 10.1016/j.jmb.2005.08.020 16169011 9 Vogt G , Chapgier A , Yang K , Chuzhanova N , Feinberg J , Fieschi C , et al Gains of glycosylation comprise an unexpectedly large group of pathogenic mutations . Nat Genet . 2005 ;37 (7 ):692 –700 . 10.1038/ng1581 15924140 10 Vogt G , Vogt B , N C , Julenius K , Cooper DN , Casanova JL . Gain-of-glycosylation mutations . Curr Opin Genet Dev . 2007 ;17 (3 ):245 –251 . 10.1016/j.gde.2007.04.008 17467977 11 Li S , Iakoucheva L , Mooney SD , Radivojac P . Loss of post-translational modification sites in disease . Pac Symp Biocomput . 2010 ;15 :337 –347 . 19908386 12 Pan Y , Karagiannis K , Zhang H , Dingerdissen H , Shamsaddini A , Wan Q , et al Human germline and pan-cancer variomes and their distinct functional profiles . Nucleic Acids Res . 2014 ;42 (18 ):11570 –11588 . 10.1093/nar/gku772 25232094 13 Peterson TA , Doughty E , Kann MG . Towards precision medicine: advances in computational approaches for the analysis of human variants . J Mol Biol . 2013 ;425 (21 ):4047 –4063 . 10.1016/j.jmb.2013.08.008 23962656 14 Radivojac P , Baenziger PH , Kann MG , Mort ME , Hahn MW , Mooney SD . Gain and loss of phosphorylation sites in human cancer . Bioinformatics . 2008 ;24 (16 ):i241 –247 . 10.1093/bioinformatics/btn267 18689832 15 Radivojac P , Vacic V , Haynes C , Cocklin RR , Mohan A , Heyen JW , et al Identification, analysis, and prediction of protein ubiquitination sites . Proteins . 2010 ;78 (2 ):365 –380 . 10.1002/prot.22555 19722269 16 Xin F , Myers S , Li YF , Cooper DN , Mooney SD , Radivojac P . Structure-based kernels for the prediction of catalytic residues and their involvement in human inherited disease . Bioinformatics . 2010 ;26 (16 ):1975 –1982 . 10.1093/bioinformatics/btq319 20551136 17 Li B , Krishnan VG , Mort ME , Xin F , Kamati KK , Cooper DN , et al Automated inference of molecular mechanisms of disease from amino acid substitutions . Bioinformatics . 2009 ;25 (21 ):2744 –2750 . 10.1093/bioinformatics/btp528 19734154 18 Saunders CT , Baker D . Evaluation of structural and evolutionary contributions to deleterious mutation prediction . J Mol Biol . 2002 ;322 (4 ):891 –901 . 10.1016/S0022-2836(02)00813-6 12270722 19 Capriotti E , Altman RB . Improving the prediction of disease-related variants using protein three-dimensional structure . BMC Bioinformatics . 2011 ;12 Suppl 4 :S3 10.1186/1471-2105-12-S4-S3 21992054 20 Xin F , Radivojac P . Computational methods for identification of functional residues in protein structures . Curr Protein Pept Sci . 2011 ;12 :456 –469 . 10.2174/138920311796957685 21787297 21 Sahni N , Yi S , Taipale M , Fuxman Bass JI , Coulombe-Huntington J , Yang F , et al Widespread macromolecular interaction perturbations in human genetic disorders . Cell . 2015 ;161 (3 ):647 –660 . 10.1016/j.cell.2015.04.013 25910212 22 Steward RE , MacArthur MW , Laskowski RA , Thornton JM . Molecular basis of inherited diseases: a structural perspective . Trends Genet . 2003 ;19 (9 ):505 –513 . 10.1016/S0168-9525(03)00195-1 12957544 23 De Baets G , Van Durme J , Reumers J , Maurer-Stroh S , Vanhee P , Dopazo J , et al SNPeffect 4.0: on-line prediction of molecular and structural effects of protein-coding variants . Nucleic Acids Res . 2012 ;40 (Database issue ):D935 –939 . 10.1093/nar/gkr996 22075996 24 Yates CM , Sternberg MJ . The effects of non-synonymous single nucleotide polymorphisms (nsSNPs) on protein-protein interactions . J Mol Biol . 2013 ;425 (21 ):3949 –3963 . 10.1016/j.jmb.2013.07.012 23867278 25 Nussinov R , Tsai CJ , Xin F , Radivojac P . Allosteric post-translational modification codes . Trends Biochem Sci . 2012 ;37 (10 ):447 –455 . 10.1016/j.tibs.2012.07.001 22884395 26 Elkan C, Noto K. Learning classifiers from only positive and unlabeled data. In: Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD 2008. New York, NY, USA: ACM; 2008. p. 213–220. 27 Jain S, White M, Trosset MW, Radivojac P. Nonparametric semi-supervised learning of class proportions. arXiv preprint arXiv:160101944. 2016;Available from: http://arxiv.org/abs/1601.01944. 28 Denis F , Gilleron R , Letouzey F . Learning from positive and unlabeled examples . Theor Comput Sci . 2005 ;348 (16 ):70 –83 . 10.1016/j.tcs.2005.09.007 29 Lagace TA , Curtis DE , Garuti R , McNutt MC , Park SW , Prather HB , et al Secreted PCSK9 decreases the number of LDL receptors in hepatocytes and in livers of parabiotic mice . J Clin Invest . 2006 ;116 (11 ):2995 –3005 . 10.1172/JCI29383 17080197 30 Bromberg Y , Rost B . Correlating protein function and stability through the analysis of single amino acid substitutions . BMC Bioinformatics . 2009 ;10 Suppl 8 :S8 10.1186/1471-2105-10-S8-S8 19758472 31 Tokuriki N , Tawfik DS . Stability effects of mutations and protein evolvability . Curr Opin Struct Biol . 2009 ;19 (5 ):596 –604 . 10.1016/j.sbi.2009.08.003 19765975 32 Worth CL , Preissner R , Blundell TL . SDM–a server for predicting effects of mutations on protein stability and malfunction . Nucleic Acids Res . 2011 ;39 (Web Server issue ):W215 –222 . 10.1093/nar/gkr363 21593128 33 du Plessis MC , Sugiyama M . Class prior estimation from positive and unlabeled data . IEICE Transactions on Information and Systems . E97–D(5) :1358 –1362 . 34 Przulj N , Corneil DG , Jurisica I . Modeling interactome: scale-free or geometric? Bioinformatics . 2004 ;20 (18 ):3508 –3515 . 10.1093/bioinformatics/bth436 15284103 35 Przulj N . Biological network comparison using graphlet degree distribution . Bioinformatics . 2007 ;23 (2 ):e177 –e183 . 10.1093/bioinformatics/btl301 17237089 36 Shervashidze N, Vishwanathan SVN, Petri TH, Mehlhorn K, Borgwardt KM. Efficient graphlet kernels for large graph comparison. In: Proc. 12th International Conference on Artificial Intelligence and Statistics. AISTATS’09; 2009. p. 488–495. 37 Vacic V , Iakoucheva LM , Lonardi S , Radivojac P . Graphlet kernels for prediction of functional residues in protein structures . J Comput Biol . 2010 ;17 (1 ):55 –72 . 10.1089/cmb.2009.0029 20078397 38 Lugo-Martinez J , Radivojac P . Generalized graphlet kernels for probabilistic inference in sparse graphs . Network Science . 2014 ;2 (2 ):254 –276 . 10.1017/nws.2014.14 39 Joachims T . Learning to classify text using support vector machines: methods, theory, and algorithms . Kluwer Academic Publishers ; 2002 . 40 Platt J . Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods In: Advances in Large Margin Classifiers . MIT Press ; 1999 p. 61 –74 . 41 Capriotti E , Fariselli P , Casadio R . A neural-network-based method for predicting protein stability changes upon single point mutations . Bioinformatics . 2004 ;20 (1 ):i63 –i68 . 10.1093/bioinformatics/bth928 15262782 42 Cheng J , Randall A , Baldi P . Prediction of protein stability changes for single-site mutations using support vector machines . Proteins . 2006 ;62 (4 ):1125 –1132 . 10.1002/prot.20810 16372356 43 Kyte J , Doolittle RF . A simple method for displaying the hydropathic character of a protein . J Mol Biol . 1982 ;157 (1 ):105 –132 . 10.1016/0022-2836(82)90515-0 7108955 44 Janin J . Surface and inside volumes in globular proteins . Nature . 1979 ;277 (5696 ):491 –492 . 10.1038/277491a0 763335 45 Vihinen M , Torkkila E , Riikonen P . Accuracy of protein flexibility predictions . Proteins . 1994 ;19 (2 ):141 –149 . 10.1002/prot.340190207 8090708 46 Jones S , Thornton JM . Principles of protein-protein interactions . Proc Natl Acad Sci USA . 1996 ;93 (1 ):13 –20 . 10.1073/pnas.93.1.13 8552589 47 Jones S , Thornton JM . Analysis of protein-protein interaction sites using surface patches . J Mol Biol . 1997 ;272 (1 ):121 –132 . 10.1006/jmbi.1997.1234 9299342 48 Zimmerman JM , Eliezer N , Simha R . The characterization of amino acid sequences in proteins by statistical methods . J Theor Biol . 1968 ;21 (2 ):170 –201 . 10.1016/0022-5193(68)90069-6 5700434 49 George RA , Heringa J . An analysis of protein domain linkers: their classification and role in protein folding . Protein Eng . 2002 ;15 (11 ):871 –879 . 10.1093/protein/15.11.871 12538906 50 Nagano K . Logical analysis of the mechanism of protein folding. I. Predictions of helices, loops and beta-structures from primary structure . J Mol Biol . 1973 ;75 (2 ):401 –420 . 10.1016/0022-2836(73)90030-2 4728695 51 Altschul SF , Madden TL , Schäffer AA , Zhang J , Zhang Z , Miller W , et al Gapped BLAST and PSI-BLAST: a new generation of protein database search programs . Nucleic Acids Res . 1997 ;25 (17 ):3389 –3402 . 10.1093/nar/25.17.3389 9254694 52 Xin F. Methods for predicting functional residues in protein structures and understanding molecular mechanisms of disease [Ph.D. Thesis]; 2012. 53 Porter CT , Bartlett GJ , Thornton JM . The Catalytic Site Atlas: a resource of catalytic sites and residues identified in enzymes using structural data . Nucleic Acids Res . 2004 ;32 :D129 –33 . 10.1093/nar/gkh028 14681376 54 Yan C , Terribilini M , Wu F , Jernigan RL , Dobbs D , Honavar V . Predicting DNA-binding sites of proteins from amino acid sequence . BMC Bioinformatics . 2006 ;7 :262 –271 . 10.1186/1471-2105-7-262 16712732 55 Singh H , Chauhan JS , Gromiha MM , Consortium OSDD , Raghava GPS . ccPDB: compilation and creation of data sets from Protein Data Bank . Nucleic Acids Res . 2012 ;40 (D1 ):D486 –D489 . 10.1093/nar/gkr1150 22139939 56 Chung JL , Wang W , Bourne PE . Exploiting sequence and structure homologs to identify protein–protein binding sites . Proteins . 2006 ;62 (3 ):630 –640 . 10.1002/prot.20741 16329107 57 Lise S , Archambeau C , Pontil M , Jones D . Prediction of hot spot residues at protein-protein interfaces by combining machine learning and energy-based methods . BMC Bioinformatics . 2009 ;10 (1 ):365 –381 . 10.1186/1471-2105-10-365 19878545 58 Thorn KS , Bogan AA . ASEdb: a database of alanine mutations and their effects on the free energy of binding in protein interactions . Bioinformatics . 2001 ;17 (3 ):284 –285 . 10.1093/bioinformatics/17.3.284 11294795 59 Huang Z , Mou L , Shen Q , Lu S , Li C , Liu X , et al ASD v2.0: updated content and novel features focusing on allosteric regulation . Nucleic Acids Res . 2014 ;42 (D1 ):D510 –D516 . 10.1093/nar/gkt1247 24293647 60 Gromiha M , An J , Kono H , Oobatake M , Uedaira H , Prabakaran P , et al ProTherm, version 2.0: thermodynamic database for proteins and mutants . Nucleic Acids Research . 2000 ;28 (1 ):283 –285 . 10.1093/nar/28.1.283 10592247 61 Iakoucheva LM , Radivojac P , Brown CJ , O’Connor TR , Sikes JG , Obradovic Z , et al The importance of intrinsic disorder for protein phosphorylation . Nucleic Acids Res . 2004 ;32 (3 ):1037 –1049 . 10.1093/nar/gkh253 14960716 62 Daily KM, Radivojac P, Dunker AK. Intrinsic disorder and protein modifications: building an SVM predictor for methylation. In: Proceedings of the IEEE Symposium on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB). IEEE; 2005. p. 475–481. 63 Gsponer J , Futschik ME , Teichmann SA , Babu MM . Tight regulation of unstructured proteins: from transcript synthesis to protein degradation . Science . 2008 ;322 (5906 ):1365 –1368 . 10.1126/science.1163581 19039133 64 Pejaver V , Hsu WL , Xin F , Dunker AK , Uversky VN , Radivojac P . The structural and functional signatures of proteins that undergo multiple events of post-translational modification . Protein Sci . 2014 ;23 (8 ):1077 –1093 . 10.1002/pro.2494 24888500 65 Roberts BR , Tainer JA , Getzoff ED , Malencik DA , Anderson SR , Bomben VC , et al Structural characterization of zinc-deficient human superoxide dismutase and implications for ALS . J Mol Biol . 2007 ;373 (4 ):877 –890 . 10.1016/j.jmb.2007.07.043 17888947 66 Seetharaman SV , Taylor AB , Holloway S , Hart PJ . Structures of mouse SOD1 and human/mouse SOD1 chimeras . Arch Biochem Biophys . 2010 ;503 (2 ):183 –190 . 10.1016/j.abb.2010.08.014 20727846 67 Alexander MD , Traynor BJ , Miller N , Corr B , Frost E , McQuaid S , et al “True” sporadic ALS associated with a novel SOD-1 mutation . Ann Neurol . 2002 ;52 (5 ):680 –683 . 10.1002/ana.10369 12402272 68 Krishnan U , Son M , Rajendran B , Elliott JL . Novel mutations that enhance or repress the aggregation potential of SOD1 . Mol Cell Biochem . 2006 ;287 (1–2 ):201 –211 . 10.1007/s11010-005-9112-4 16583143 69 Joyce PI , Mcgoldrick P , Saccon RA , Weber W , Fratta P , West SJ , et al A novel SOD1-ALS mutation separates central and peripheral effects of mutant SOD1 toxicity . Hum Mol Genet . 2015 ;24 (7 ):1883 –1897 . 10.1093/hmg/ddu605 25468678 70 Ippolito JA , Christianson DW . Structure of an engineered His3 Cys zinc binding site in human carbonic anhydrase II . Biochemistry . 1993 ;32 (38 ):9901 –9905 . 10.1021/bi00089a005 8399159 71 Krebs JF , Ippolito JA , Christianson DW , Fierke CA . Structural and functional importance of a conserved hydrogen bond network in human carbonic anhydrase II* . J Biol Chem . 1993 ;268 (36 ):27458 –27466 . 8262987 72 Xue Y , Liljas A , Jonsson BH , Lindskog S . Structural analysis of the zinc hydroxide-Thr-199-Glu-106 hydrogen-bond network in human carbonic anhydrase II . Proteins . 1993 ;17 (1 ):93 –106 . 10.1002/prot.340170112 7901850 73 Ippolito JA , Baird TT Jr, M SA , C DW , F CA . Structure-assisted redesign of a protein-zinc-binding site with femtomolar affinity . Proc Natl Acad Sci USA . 1995 ;92 :5017 –5021 . 10.1073/pnas.92.11.5017 7761440 74 Huang S , Sjöblom , Sauer-Eriksson AE , Jonsson BH . Organization of an efficient carbonic anhydrase: Implications for the mechanism based on structure-function studies of a T199P/C206S mutant . Biochemistry . 2002 ;41 (24 ):7628 –7635 . 10.1021/bi020053o 12056894 75 DePristo MA , Weinreich DM , Hartl DL . Missense meanderings in sequence space: a biophysical view of protein evolution . Nat Rev Genet . 2005 ;6 (9 ):678 –687 . 10.1038/nrg1672 16074985 76 Schymkowitz J , Borg J , Stricher F , Nys R , Rousseau F , Serrano L . The FoldX web server: an online force field . Nucleic Acids Res . 2005 ;33 (Web Server issue ):W382 –388 . 10.1093/nar/gki387 15980494 77 Casadio R , Vassura M , Tiwari S , Fariselli P , Luigi Martelli P . Correlating disease-related mutations to their effect on protein stability: a large-scale analysis of the human proteome . Hum Mutat . 2011 ;32 (10 ):1161 –1170 . 10.1002/humu.21555 21853506 78 Gray VE , Kukurba KR , Kumar S . Performance of computational tools in evaluating the functional impact of laboratory-induced amino acid mutations . Bioinformatics . 2012 ;28 (16 ):2093 –2096 . 10.1093/bioinformatics/bts336 22685075 79 Mohan A , Uversky VN , Radivojac P . Influence of sequence changes and environment on intrinsically disordered proteins . PLoS Comput Biol . 2009 ;5 (9 ):e1000497 10.1371/journal.pcbi.1000497 19730682

PMC005xxxxxx/PMC5001645.txt

==== Front PLoS Negl Trop DisPLoS Negl Trop DisplosplosntdsPLoS Neglected Tropical Diseases1935-27271935-2735Public Library of Science San Francisco, CA USA 10.1371/journal.pntd.0004967PNTD-D-16-01476CorrectionCorrection: Prevalence of Lymphatic Filariasis and Treatment Effectiveness of Albendazole/ Ivermectin in Individuals with HIV Co-infection in Southwest-Tanzania Kroidl Inge Saathoff Elmar Maganga Lucas Clowes Petra Maboko Leonard Hoerauf Achim Makunde Williams H. Haule Antelmo Mviombo Prisca Pitter Bettina Mgeni Neema Mabuye Joseph Kowuor Dickens Mwingira Upendo Malecela Mwelecele N. Löscher Thomas Hoelscher Michael 26 8 2016 8 2016 10 8 e0004967© 2016 Kroidl et al2016Kroidl et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Prevalence of Lymphatic Filariasis and Treatment Effectiveness of Albendazole/ Ivermectin in Individuals with HIV Co-infection in Southwest-Tanzania ==== Body The second author's name is spelled incorrectly. The correct name is: Elmar Saathoff. The correct citation is Kroidl I, Saathoff E, Maganga L, Clowes P, Maboko L, et al. (2016) Prevalence of Lymphatic Filariasis and Treatment Effectiveness of Albendazole/ Ivermectin in Individuals with HIV Co-infection in Southwest-Tanzania. PLoS Negl Trop Dis 10(4): e0004618. doi: 10.1371/journal.pntd.0004618 ==== Refs Reference 1 Kroidl I , Saathof E , Maganga L , Clowes P , Maboko L , et al (2016 ) Prevalence of Lymphatic Filariasis and Treatment Effectiveness of Albendazole/ Ivermectin in Individuals with HIV Co-infection in Southwest-Tanzania . PLoS Negl Trop Dis 10 (4 ): e0004618 doi:10.1371/journal.pntd.0004618 27070786

PMC005xxxxxx/PMC5001646.txt

==== Front PLoS GenetPLoS GenetplosplosgenPLoS Genetics1553-73901553-7404Public Library of Science San Francisco, CA USA 10.1371/journal.pgen.1006285PGENETICS-D-16-01781CorrectionCorrection: Curly Encodes Dual Oxidase, Which Acts with Heme Peroxidase Curly Su to Shape the Adult Drosophila Wing Hurd Thomas Ryan Liang Feng-Xia Lehmann Ruth 26 8 2016 8 2016 12 8 e1006285© 2016 Hurd et al2016Hurd et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Curly Encodes Dual Oxidase, Which Acts with Heme Peroxidase Curly Su to Shape the Adult Drosophila Wing ==== Body The following information is missing from the Funding section: This work was supported by NIH grant R01/R37HD41900. ==== Refs Reference 1 Hurd TR , Liang F-X , Lehmann R (2015 ) Curly Encodes Dual Oxidase, Which Acts with Heme Peroxidase Curly Su to Shape the Adult Drosophila Wing . PLoS Genet 11 (11 ): e1005625 doi:10.1371/journal.pgen.1005625 26587980

PMC005xxxxxx/PMC5001647.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756470610.1371/journal.pone.0162049PONE-D-16-02347Research ArticleResearch and Analysis MethodsBiological CulturesOrgan CulturesOrganoidsBiology and Life SciencesAnatomyDigestive SystemGastrointestinal TractColonMedicine and Health SciencesAnatomyDigestive SystemGastrointestinal TractColonBiology and Life SciencesCell BiologyCellular TypesAnimal CellsStem CellsResearch and analysis methodsExtraction techniquesRNA extractionBiology and Life SciencesAnatomyBiological TissueEpitheliumMedicine and Health SciencesAnatomyBiological TissueEpitheliumBiology and Life SciencesDevelopmental BiologyCell DifferentiationPhysical sciencesChemistryChemical compoundsOrganic compoundsVitaminsVitamin APhysical sciencesChemistryOrganic chemistryOrganic compoundsVitaminsVitamin ABiology and Life SciencesMolecular BiologyMolecular Biology TechniquesArtificial Gene Amplification and ExtensionPolymerase Chain ReactionReverse Transcriptase-Polymerase Chain ReactionResearch and Analysis MethodsMolecular Biology TechniquesArtificial Gene Amplification and ExtensionPolymerase Chain ReactionReverse Transcriptase-Polymerase Chain ReactionRetinol Promotes In Vitro Growth of Proximal Colon Organoids through a Retinoic Acid-Independent Mechanism Retinol Promotes Proximal Colon Stem Cell GrowthMatsumoto Taichi 1Mochizuki Wakana 1Nibe Yoichi 1Akiyama Shintaro 1Matsumoto Yuka 1Nozaki Kengo 1Fukuda Masayoshi 1Hayashi Ayumi 1Mizutani Tomohiro 2Oshima Shigeru 1Watanabe Mamoru 1Nakamura Tetsuya 2*1 Department of Gastroenterology and Hepatology, Graduate School, Tokyo Medical and Dental University, 1-5-45 Yushima, Bunkyo-ku, Tokyo 113–8519, Japan2 Department of Advanced Therapeutics for GI Diseases, Graduate School, Tokyo Medical and Dental University, 1-5-45 Yushima, Bunkyo-ku, Tokyo 113–8519, JapanBeekman Jeffrey M. EditorUniversity Medical Center Utrecht, NETHERLANDSCompeting Interests: The authors have declared that no competing interests exist. Conceived and designed the experiments: TN. Performed the experiments: T. Matsumoto WM TN. Analyzed the data: T. Matsumoto WM YN SA YM KN MF T. Mizutani SO MW TN AH. Wrote the paper: TN. * E-mail: nakamura.gast@tmd.ac.jp26 8 2016 2016 11 8 e016204918 1 2016 16 8 2016 © 2016 Matsumoto et al2016Matsumoto et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Retinol (ROL), the alcohol form of vitamin A, is known to control cell fate decision of various types of stem cells in the form of its active metabolite, retinoic acid (RA). However, little is known about whether ROL has regulatory effects on colonic stem cells. We examined in this study the effect of ROL on the growth of murine normal colonic cells cultured as organoids. As genes involved in RA synthesis from ROL were differentially expressed along the length of the colon, we tested the effect of ROL on proximal and distal colon organoids separately. We found that organoid forming efficiency and the expression level of Lgr5, a marker gene for colonic stem cells were significantly enhanced by ROL in the proximal colon organoids, but not in the distal ones. Interestingly, neither retinaldehyde (RAL), an intermediate product of the ROL-RA pathway, nor RA exhibited growth promoting effects on the proximal colon organoids, suggesting that ROL-dependent growth enhancement in organoids involves an RA-independent mechanism. This was confirmed by the observation that an inhibitor for RA-mediated gene transcription did not abrogate the effect of ROL on organoids. This novel role of ROL in stem cell maintenance in the proximal colon provides insights into the mechanism of region-specific regulation for colonic stem cell maintenance. MEXT KAKENHI26112705Nakamura Tetsuya JSPS KAKENHI24390186Nakamura Tetsuya JSPS KAKENHI26221307Watanabe Mamoru http://dx.doi.org/10.13039/501100002241Japan Science and Technology AgencyRegenerative Medicine Realization Base Network ProgramNakamura Tetsuya Ministry of Health, Labor and Welfare of JapanResearch Grants for research on Intractable DiseasesWatanabe Mamoru MEXT KAKENHI (https://www.jsps.go.jp/english/e-grants/) Grant number 26112705 to TN; JSPS KAKENHI (https://www.jsps.go.jp/english/e-grants/) Grant number 24390186 to TN, Grant number 26221307 to MW; the Regenerative Medicine Realization Base Network Program from the Japan Science and Technology Agency (http://www.jst.go.jp/saisei-nw/kadai_05.html) No Number, TN WM and T. Mizutani are receiving this funding; and Health and Labour Sciences Research Grants for research on Intractable Diseases from Ministry of Health, Labor and Welfare of Japan (http://www.mhlw.go.jp/seisakunitsuite/bunya/hokabunya/kenkyujigyou/hojokin-koubo-h26/gaiyo2/11.html) No Number, MW is receiving this grant. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction The inner surface of the colon is lined with simple columnar epithelium structurally organized into crypts. The epithelium continues to self-renew throughout the lifetime, fueled by perpetual and rapid cellular turnover of Lgr5+ adult stem cells located near the base of those crypts [1, 2]. The fate determination of those colonic stem cells is, as is the case with small intestinal stem cells, governed by multiple regulatory signals such as the Wnt, bone morphogenic protein (BMP), Notch signaling pathway, and receptor tyrosine kinases [2, 3]. The proliferative activity of colonic stem cells can be reconstituted in vitro by using the organoid culture system that has recently emerged as a powerful tool for studying stem cell biology. Sato et al. first described the system, which allows long-term expansion of murine small intestinal cells as a structure consisting of crypt domains harboring Lgr5+ stem cells and also a domain lined by differentiated cell types [4]. Importantly, this condition requires Rspo1 (Wnt agonist), Noggin (BMP inhibitor) and Epidermal Growth Factor (EGF), and therefore faithfully recapitulates the functional importance of those signaling pathways for in vivo maintenance of small intestinal stem cells. The organoid culture system was shown to be applicable to colonic epithelial cells when the culture medium is supplemented with Wnt ligands in addition to the factors required for small intestinal cell culture [5]. Using a slightly different combination of growth factors and extracellular matrices, we have also developed a method to culture colonic stem cells [6]. Murine colonic cells were shown to grow almost in perpetuity as spherical organoids in the presence of Wnt3a, Rspo1, Noggin, EGF, Hepatocyte Growth Factor (HGF) and Bovine Serum Albumin (BSA) even under serum-free conditions [6]. The cells expanded by this method were shown to be capable of regenerating normal colonic epithelial tissues when transplanted into mice in which colonic mucosal injuries were induced [6]. This indicates that the cells in organoid cultures preserve many aspects of their original features in vitro, building a rationale for the use of this culture system to assess the response of colonic stem cells to various stimuli and identifying underlying mechanisms that drive those responses. Retinol (ROL) is the alcohol form of vitamin A, which controls proliferation and differentiation in various types of cells [7]. ROL is ubiquitously present in the circulating blood and delivered to different cell types with different specificities [8]. In target cells, ROL is metabolized to retinoic acid (RA) in two steps of oxidation. ROL is first oxidized to retinaldehyde (RAL) by either alcohol dehydrogenases (ADH1, ADH5, and ADH7) [9–11] or retinol dehydrogenases (RDH1 and RDH10) [12, 13]. RAL is then further oxidized to RA by members of the aldehyde dehydrogenase family (ALDH1A1, ALDH1A2, and ALDH1A3) [14–17]. Most of the actions of ROL are generally thought to be mediated primarily by RA, which regulates gene transcription by functioning as a ligand for the retinoic acid receptors (multiple isoforms of RARα, β, and γ) and the retinoid X receptors (multiple isoforms of RXRα, β, and γ) [7, 9, 18]. This ROL-RA pathway is known to control cell fate decisions and the maintenance of various types of stem cells. Although the pathway acts to induce differentiation in many stem cell types [7], it also promotes proliferation in some types of stem cells such as embryonic stem cells [19] and germline stem cells [20]. With regard to colonic epithelial cells, a few studies described the effect of ROL on the growth of human colorectal cancer-derived cell lines [21–23]; however, little is known about whether ROL plays a role in the maintenance of normal colonic epithelia and their stem cell populations. Therefore, we sought to address this question by using the system that we developed [6] to culture primary non-transformed colonic cells in vitro. Materials and Methods Mice C57BL6 male mice at 8 to 12 weeks of age were used in this study. Mice were bred under standard conditions at our university. We used tissue samples that were removed from mice immediately after they were sacrificed by cervical dislocation. All animal experiments were performed with the approval of the Institutional Animal Care and Use Committee of TMDU. Colon Organoid Culture Colonic crypt isolation was performed according to the method described previously [6]. When indicated, crypts were separately isolated from proximal and distal halves of the colon. Briefly, after extensive washing, colonic tissues minced into small pieces were incubated in Dulbecco’s modified eagle medium (DMEM) containing 1% FBS, 500 U/ml collagenase XI (Sigma), 0.4 U/ml dispase (Roche), and 1 mM dithiothreitol (DTT) at 37°C for 20 min. The crypts were further purified by mechanical disruption and density gradient centrifugation. A total of 400 crypts were suspended in 40 μl of the collagen type I solution (Nitta Gelatin Inc.) and placed in 24-well plates. After polymerization, 500 μl of Advanced DMEM/F12 containing 1% BSA (Sigma A9576), 30 ng/ml mWnt3a (R&D Systems), 500 ng/ml mRspo1 (R&D Systems), 20 ng/ml mEGF (Peprotech), 50 ng/ml mHGF (R&D Systems), and 50 ng/ml mNoggin (R&D Systems) were added to each well. The medium was changed every 2 days. The organoids were cultured for 6 days and then used for the following analysis with single-cell passage procedures. Single Cell Passage For single-cell passage, the collagen gel was digested in DMEM containing collagenase type XI (Sigma) at 37°C for 5 min. The released organoids were washed in PBS containing 0.5% BSA. The organoids were digested in 4 ml of TrypLE Express (Thermo Fisher Scientific) at 37°C for 5 min, and then vigorously shaken to obtain disaggregated cells. Single cell preparation was verified by microscopic inspection and viable cells were counted in a hemocytometer by method of Trypan blue exclusion. Cells were then seeded so that 1 x 104 cells were present in 40 μl of the collagen type I gel in each well at the start of culture. When necessary, retinol (Sigma R7632), all-trans-retinal (retinaldehyde) (Sigma R2500) and all-trans-retinoic acid (Sigma R2625) reconstituted in ethanol were added to the culture at indicated doses after passage. AGN 193109 (Santa Cruz Biotechnology, sc210768), reconstituted in DMSO, was added to the medium at 1 μM together with ROL or RA as indicated. The organoids cultured in this way were recovered from the collagen gel on Day 5 of culture and used for another round of culture after single cell passage. Determination of Organoid Forming Efficiency Organoid forming efficiency was evaluated by initiating culture with 1x 104 cells per well after single cell passage. On Day 5 of culture, Z-stacks of phase-contrast images were acquired at Z-steps of 30 μm on a microscope BZ-X710 (KEYENCE). In most cases, depth of observation up to 1 mm was sufficient to cover the entirety of the collagen type I gel in each well. We counted the number of organoids that were clearly visible as round cystic structures on Z-projection images created by the BZ-X710 system. Data were collected for triplicate wells for each of 3 independent experiments (n = 9). The diameter of organoids was analyzed on Z-projection images created as described above. Squares 1.5 mm per side were placed at the center of the images of 9 individual wells (triplicate wells from one of three donor samples). The diameter of each organoid included in the squared area was measured and used for quantitative analysis. Statistical significance was determined by Student’s t test (p < 0.05). Semi-quantitative RT-PCR Semi-quantitative RT-PCR was performed in a standard fashion. Aliquots of 300 ng of total RNA were used for cDNA synthesis in 21 μl of reaction volume. One microliter of cDNA was used for the following RT-PCR. Primer sequences and the detail of reactions are listed in S1 Table. PCR products were separated on agarose gels and visualized using ImageLab (Bio-Rad). In Situ Hybridization (ISH) ISH for Aldh1a1 and Aldh1a3 was performed as described previously [24]. The pcDNA3 plasmid containing a cDNA fragment of mouse Aldh1a1 (nucleotides 65–1892; GenBank NM_013467.3) or Aldh1a3 (nucleotides 2312–3225; GenBank NM053080.3) was constructed. Single-stranded, digoxigenin-labeled RNA probes were generated by an in vitro transcription system (Roche). Frozen sections of colonic Swiss rolls were rehydrated, treated with HCl, digested in proteinase K solution, postfixed, treated in acetic anhydride solution, and hybridized overnight at 65°C with probes. After extensive rinsing and washing, sections were then subjected to the immunohistochemistry by using alkaline phosphatase–conjugated anti-digoxigenin antibody (Roche). Sections were reacted with nitroblue tetrazolium/5-bromo-4-chloro-3-indolyl phosphate solution for color development. Images were acquired on a microscope BZ-X710 (KEYENCE). Images were processed using Adobe Photoshop software. Immunohistochemistry Colonic tissues were fixed overnight at 4°C in 4% paraformaldehyde, sequentially dehydrated in 10, 15 and 20% sucrose in PBS, and embedded in OCT compound (Tissue Tek). Colonic organoids were fixed together with surrounding collagen type I gel and processed in the same manner. Frozen sections of 8-μm thickness were subjected to immunohistochemistry by using antibodies specific for ALDH1a1 (Abcam ab52492), ALDH1a3 (Abjent AP7847a), or Ki-67 (Dako Cytomation). In all immunofluorescence experiments, nuclei were counterstained with DAPI. Fluorescent images of sections were acquired using a microscope BZ-X710 (KEYENCE). If necessary, image processing was carried out using Adobe Photoshop software. For quantification of ALDH1a3+ and Ki-67+ cells, triplicate wells from one of three donor samples were used. On Day 5 after passage, sections of thirty different organoids were chosen from a series of sections obtained from each well. ALDH1a3+ or Ki-67+ cells were counted, and values are presented as a mean percentage to whole cell populations for sections of those 270 organoids. Statistical significance was determined by Student’s t test (p < 0.05). Results We first investigated whether retinol (ROL) has effects on the in vitro growth of epithelial cells that were isolated from the entire length of the mouse colon. We previously reported that, when colonic crypts were isolated and cultured as organoids, Lgr5+ stem cells increased in number over the first several days [6]. To test the effect of ROL on the growth and organoid forming efficiency of the stem cell population, colonic cells were cultured for 6 days after isolation to obtain sufficient number of stem cells. Next, the organoids were dissociated into single cells and the first-passage culture was initiated from this single cell population in the presence or absence of ROL. We found that, when various concentrations of ROL were added to the culture, cells treated with 1 μM or 3 μM of ROL formed a somewhat higher number of organoids as compared to those cultured in the absence of ROL (Fig 1A). In addition, some organoids treated with ROL (1 μM or 3 μM) appeared to have larger diameters than untreated organoids (Fig 1A). Cells treated with ROL at 10 μM (Fig 1A) or greater concentrations (not shown) showed a lower number of organoids than untreated controls. As the levels of ROL in human serum were reported to range from 0.5 to 2 μM [25], we presumed that the treatment with physiological concentrations of ROL would have a positive effect on the organoid forming efficiency of cultured colonic epithelial stem cells. 10.1371/journal.pone.0162049.g001Fig 1 Expression of genes involved in ROL-RA pathway in colonic epithelia. (A) Crypts isolated from the entire length of the colon were cultured for 6 days. Organoids were then dissociated into single cells and further cultured in the presence of vehicle alone, or indicated concentrations of ROL. Phase-contrast images were acquired on Day 5 of culture. Representative images of three independent experiments are shown. Scale bar, 100 μm. (B) Total RNA was extracted from the whole colon or its epithelial compartment isolated as crypts. Semi-quantitative PCR was performed for the indicated genes. Representative data are shown for three independent experiments. ROL is widely known to induce various biological effects in the form of retinoic acid (RA). To assess whether cultured colonic organoids contain the cellular machinery for converting ROL to RA, we investigated the expression of genes involved in the ROL metabolism. Semi-quantitative RT-PCR revealed that mRNAs encoding enzymes capable of catalyzing the first step of the reaction (Adh1, Adh5, Adh7, Rdh1, and Rdh10) were all present in the whole colon tissues as well as in the isolated crypts (Fig 1B). In regard to three ALDH1a family isoenzymes known to catalyze the second step of reaction, mRNAs for Aldh1a1 and Aldh1a3 were found to be present in both samples (Fig 1B). Aldh1a2 mRNA was expressed in the whole tissue, but it was not detected in the epithelial compartment (Fig 1B). These observations indicated that the enzymes that regulate RA synthesis are readily expressed in the colonic epithelium, and this ROL-RA pathway might be related to the seemingly positive but moderate effects of ROL on the organoid forming efficiency of cultured colonic stem cells. Through the further analysis of genes involved in RA synthesis, we noticed intriguing expression patterns of Aldh1a1 and Aldh1a3 in the colon. Consistent with the RT-PCR data (Fig 1B), in situ hybridization (ISH) revealed that both genes were strongly expressed in the epithelial layer of the colon (Fig 2A). Interestingly, Aldh1a1 and Aldh1a3 showed distinct regional differences in distribution along the length of the colon; they were both expressed predominantly in the proximal colon and their expression levels declined in the distal colon and rectum (Fig 2A, top). In addition, these two genes were differently expressed along the crypt axis; Aldh1a1 was preferentially expressed in the surface epithelium, while expression of Aldh1a3 was relatively confined to the crypt bottom (Fig 2A, middle). The mucosa near the distal end of the colon exhibited low or undetectable expression of these two genes (Fig 2A, bottom). The characteristic distribution patterns of ALDH1a1 and ALDH1a3 were also observed at the protein expression level. Immunohistochemistry revealed that ALDH1a1 and ALDH1a3 were predominantly expressed in the proximal colon, while they appeared only in a few cells at low level in the distal part of the colon (Fig 2B). Again, ALDH1a1 was mainly located at the surface epithelium, whereas ALDH1a3 was detected at the lower part of the crypt (Fig 2B). 10.1371/journal.pone.0162049.g002Fig 2 Enzymes involved in ROL-RA pathway are differentially expressed along the length of the colon. (A) Colonic Swiss rolls were assessed for expression of Aldh1a1 (top left) and Aldh1a3 (top right) genes by in situ hybridization. The proximal and distal portions of the colon are located inside and outside, respectively. Magnified views are shown for the proximal (middle) and distal colon (bottom). Representative images of three independent experiments are shown. Scale bars, 1 mm for images on the top and 50 μm for images on the middle and at the bottom. (B) Sections of the proximal (top) and distal (bottom) colon were assayed for protein expression of ALDH1a1 and ALDH1a3. Signals yielded by immunohistochemical staining (red) are shown in merged images with DAPI staining (blue). Representative data are shown for three independent experiments. Arrowheads show positive cells for ALDH1a1 or ALDH1a3 in the distal colon. Scale bars, 50 μm. (C) Colonic crypts were isolated from the entire length (whole colon), or from proximal (prox. colon) or distal (dist. colon) half of the colon. Crypts were then cultured for 6 days. Total RNA was extracted from the isolated crypts before (pre) or after the culture (post). Semi-quantitative PCR was performed for the genes indicated. Representative data are shown for three independent experiments. (D) Colonic crypts were separately isolated from proximal and distal colons and cultured as organoids. Frozen sections of organoids were subjected to immunohistochemistry for ALDH1a3 (left). Merged images with DAPI staining were also shown (right). Scale bars, 50 μm. Sections of 30 organoids were chosen from a series of sections obtained from each well (triplicate wells for each of three independent experiments) and the percentage of Ki-67+ cells to whole cell populations was quantitated (graph on the right). Values are presented as mean ± s.e.m. for those 270 sections. To confirm the unique distribution patterns of Aldh1a1 and Aldh1a3 in the colon, we performed semi-quantitative RT-PCR again with the epithelia separately isolated from the proximal and distal halves of the colon. In addition, in order to assess the gene expression changes during the culture, crypts of the whole colon and those from two colonic regions were independently cultured and their mRNAs were subjected to RT-PCR. In line with the ISH and immunohistochemistry data, the expression levels of Aldh1a1 and Aldh1a3 in pre-culture samples (isolated crypts) were confirmed to be significantly higher in the proximal colon crypts than those in the distal ones (Fig 2C). By contrast, Adh1 and Adh7 genes showed higher expression in pre-culture samples isolated from the distal colon compared with those from proximal one. Other genes involved in ROL metabolism (Adh5, Rdh1 and Rdh10) showed no obvious regional differences between pre-culture crypts (Fig 2C). We also found in this experiment that, during the culture process, expression levels of Adh1, Rdh1 and Aldh1a1 were decreased, while those of Rdh10 and Aldh1a3 were increased in both proximal and distal colonic epithelia (Fig 2C). To investigate whether such mRNA expression changes during culture lead to alterations of protein expression, crypts from proximal and distal halves of the colon were separately cultured and processed for immunohistochemistry for ALDH1a3, which showed a significant increase in mRNA expression during culture (Fig 2C). It was shown that organoids derived from proximal and distal parts clearly contained ALDH1a3+ cells, which represented ~30% of total cells in both samples (Fig 2D). Considering that ALDH1a3 protein expression in the distal colonic tissue was observed in fewer cells than in the proximal one (Fig 2B), this suggested that ALDH1a3 expression increased in the course of culture not only at mRNA level but also at protein level, at least in distal colon organoids. We previously demonstrated that the culture protocol that we developed preferentially expands Lgr5+ stem cells [6]. Since Aldh1a3 was expressed in the proliferative zone at the crypt bottom (Fig 2B & 2C), we assumed that the expression changes of Aldh1a3, and also other genes (Aldh1a1, Adh1, Rdh1 and Rdh10), might be associated with alterations in cell-type composition during the organoid culture. Given the observation that many genes involved in the ROL metabolism, most exemplified by Aldh1a1 and Aldh1a3, exhibited regional differences in their expression along the colon, we thought that ROL might induce different responses between proximal and distal colon organoids. To test this idea, crypts were isolated separately from the two portions of the colon, cultured for 6 days, and passaged as single cells to test the effect of ROL (1 μM). We found that the number of ROL-treated proximal colon organoids was apparently greater than that of untreated organoids (Fig 3A). By contrast, organoids from the distal colon showed no difference between ROL-treated and untreated samples (Fig 3A). To quantitatively assess this finding, we counted the number of organoids that formed clearly discernible round cystic structures on Day 5 after passage. The organoid forming efficiency was significantly higher in ROL-treated proximal colon organoids compared with untreated controls (Fig 3B). Cells from the distal colon showed no difference in this assay between ROL-treated and untreated groups (Fig 3B). We also measured the size of organoids on Day 5. ROL-treated proximal colon organoids had a significantly larger average diameter than did those untreated with ROL (Fig 3C). Such a difference was not observed in distal colon organoids (Fig 3C). To directly assess the expansion of cell populations, we conducted the single-cell passage, initiated organoid culture and then recovered the whole cell population on Day 5 to count them. Under this condition of single-cell passage, the total cell count increased steadily in organoids even in the absence of ROL (Fig 3D). Of note, supplementation of culture medium with ROL (1 μM) significantly enhanced the increase of cell numbers in proximal colon organoids as compared to controls, and this trend continued for another 4 days when the single-cell culture procedure was repeated (Fig 3D, Day 9). To investigate whether the ROL-dependent increase in cell number was due to the enhanced cellular proliferative activity, we fixed the organoids and immunostained the sections for Ki-67 on Day 5 after passage. In proximal colon organoids cultured in the absence of ROL, Ki-67+ cells were scattered sparsely throughout the structure (Fig 3E, top left). By contrast, ROL-treated proximal colon organoids were lined by far more Ki-67+ cells (Fig 3E, bottom left). This was also confirmed by direct counting of Ki-67+ cells in those sections (Fig 3E, right). 10.1371/journal.pone.0162049.g003Fig 3 ROL enhances growth of proximal colon organoids but not of distal ones. (A) Proximal and distal colon organoids were separately subjected to single-cell passage. Respective cells were cultured in the absence (vehicle) or presence of ROL (1 μM) and phase-contrast images were acquired on Day 5. Representative images of three independent experiments are shown. Scale bar, 100 μm. (B) Single-cell passage and the following culture was initiated as in (A) at a density of 1 x 104 cells/well. On day 5, round cystic organoids that formed in each well were counted as described in Materials and Methods. Values are presented as mean ± s.e.m. (n = 9). *, P < 0.05. (C) Proximal and distal colon cells were cultured as in (B). Diameters of organoids were measured on Day 5 of culture as described in Materials and Methods. Total number (n) of organoids analyzed in each group was shown at the bottom. Data are presented as mean ± s.e.m. (n = 9). *, P < 0.05. (D) Proximal and distal colon cells were cultured as in (B) at a density of 1 x 104 cells/well (Day1). Triplicate wells from one of three independent donor samples were separately cultured. On Day 5, organoids were recovered and then the second passage was performed. Cell numbers were counted on Day 5 (at the time point of second passage) and Day 9. Data are presented as mean ± s.e.m. (n = 9). *, P < 0.05. (E) Proximal colon cells were cultured as in (B). On Day 5, frozen sections of organoids were subjected to immunohistochemistry for Ki-67 (left). Merged images with DAPI staining were also shown (right). Scale bars, 50 μm. The percentage of Ki-67+ cells to whole cell populations was quantitated (graph on the right) as described for Fig 2D. Values are presented as mean ± s.e.m. for those 270 sections. *, P < 0.05. We next investigated whether the expression of Lgr5, a marker gene for colonic stem cells, was influenced by ROL treatment. Proximal and distal colon cells were cultured for 5 days after single cell passage, and mRNA extracted from those cells before and after the culture was assessed by RT-PCR. It was shown that, even in the absence of ROL, expression of Lgr5 was increased during culture in both the proximal and distal colon (Fig 4A), as a result of efficient expansion of the stem cell pool by this culture method [6]. Notably, treatment with ROL further augmented the induction of Lgr5 gene expression in the proximal colon organoids, but this phenomenon was not observed in the distal organoids (Fig 4A). These data clearly indicate that the ROL-mediated acceleration of the growth rate of proximal colon organoids involves the stimulation of Lgr5+ stem cells. 10.1371/journal.pone.0162049.g004Fig 4 ROL-dependent growth promotion of proximal colon organoids involves an RA-independent mechanism. (A) Proximal and distal colon organoids were separately subjected to single-cell passage, and the following culture was initiated in the absence (vehicle) or presence of ROL (1 μM). Total RNA was extracted from the isolated cells before (pre) or after the culture (post). Semi-quantitative PCR was performed for Lgr5, a marker gene of colonic stem cells, and Gapdh. (B) After single-cell passage, proximal colon organoids were cultured either with ROL (1 μM), RAL (1 μM) or RA (1 μM) or left untreated (vehicle). Total RNA was extracted after the culture, and semi-quantitative PCR was performed for Lgr5, Rarβ, and Gapdh. (C) Proximal colon cells were cultured as in (B). Representative phase-contrast images on Day 5 are shown. Scale bar, 100 μm. (D) Proximal colon cells were cultured as in (B) either with different concentrations (1, 3, or 10 μM) of ROL or RA, or left untreated (vehicle). Total RNA was extracted on Day 5 of culture and semi-quantitative PCR was performed for MUC2, Lgr5, Rarβ and Gapdh. (E) Proximal colon cells were cultured in the absence (vehicle) or presence of 1 μM ROL. ROL-treated cells were co-treated with different concentrations of RA (1, 3, or 10 μM) or left untreated. Semi-quantitative PCR was performed as described in (B) for Lgr5, Rarβ and Gapdh. (F) Proximal colon cells cultured as in (B) were left untreated (vehicle) or treated with either ROL, ROL+AGN193109, RA, RA+AGN193109 all at a concentration of 1 μM. On Day 5 of culture, total RNA was extracted after the culture, and semi-quantitative PCR was performed for Lgr5, Rarβ and Gapdh. (G) Proximal cells were cultured as in (F) and phase-contrast images on Day 5 were acquired. Scale bar, 100 μm. Experiments for A-G were performed more than twice independently and representative images are shown. To examine whether the ROL-RA pathway mediates the effect of ROL on organoid growth, proximal colon cells were treated either with ROL, RAL or RA (all at a concentration of 1 μM) after single cell passage. Organoids were collected on Day 5 of culture and their mRNA was analyzed. Unexpectedly, in contrast to the significant enhancement of Lgr5 expression in ROL-treated cells, no obvious change was detected in RAL- or RA-treated organoids (Fig 4B). Meanwhile, expression of Rarβ, one of the target genes of RAR-mediated transcription, was clearly induced in organoids treated with any of the three forms of retinoids (Fig 4B). This indicated that RA synthesis from ROL or RAL, and also the RAR-dependent transcription are operational in cultured organoids. Microscopic observation also revealed that ROL promoted organoid growth of proximal colon organoids, while RAL and RA did not show such an effect (Fig 4C). These observations suggested that the ROL-dependent induction of Lgr5 expression and growth enhancement in proximal colon organoids involved a mechanism distinct from the ROL-RA pathway. Several studies have shown that ROL and RA suppress proliferation of human colon cancer cell lines [21–23, 26]. In addition, other reports demonstrated that RA induces expression of MUC2, a differentiation marker gene, in SW480 colon cancer cells [27, 28]. To investigate the effects of retinoids on differentiation of cultured colonic cells, we assessed the expression level of MUC2 gene in organoids treated with various doses of ROL or RA, or left untreated (Fig 4D). Proximal colon organoids did not show changes in MUC2 mRNA expression in response to any concentration (1, 3, or 10 μM) of ROL and RA, indicating that retinoid treatment does not drive differentiation of colonic cells in cultured organoids under this condition (Fig 4D). We also found with this experiment that proximal colon cells treated with 10 μM ROL, or with 3 or 10 μM of RA did not show enhanced growth as judged by microscopic analysis (data not shown). Consistently, Lgr5 expression was up-regulated with 1 μM and 3 μM of ROL at a comparable level, whereas 10 μM ROL did not show such an effect (Fig 4D). In addition, the Lgr5 level remained unchanged when the proximal organoids were treated with 1 μM RA, but was down-regulated by higher concentrations (3 or 10 μM) of RA. The RAR-mediated transcription seemed unlikely to be involved in this phenomenon, as the Rarβ induction seen with stimulation by 1 μM of these retinoids did not increase any further by higher doses (3 or 10 μM) of ROL or RA (Fig 4D). Although the mechanism of these suppressive action of high doses of ROL (10 μM) and RA (3 and 10 μM) on Lgr5 expression remains unclear; however, these results reinforced the notion that the growth-promoting effects of low dose ROL is independent from the direct action of RA. In order to assess hierarchy relationship between opposing actions of low dose ROL and RA, we tested whether co-addition of RA (at 1, 3 or 10 μM) might affect the induction of Lgr5 expression by 1 μM ROL. Since further induction of Rarβ was not observed in the co-presence of ROL and RA (Fig 4E), the transcription activity of RARs appeared to reach the maximum level by 1 μM or lower concentration of ROL or RA, which was also supported by the data shown in Fig 4D. Of note, addition of RA did not counteract the ROL-dependent up-regulation of Lgr5 even when used at 3 or 10 μM (Fig 4E). These data indicated that, when ROL and its metabolite RA coexist, ROL functions as an independent and dominant factor to regulate the expression of Lgr5. To further confirm this independent action of ROL, we treated the proximal colon organoids cultured in the presence of ROL (1 μM) or RA (1 μM) with a pan-RAR antagonist, AGN 193109 [29]. The action of AGN 193109 was verified as it completely abrogated the ROL- and RA-dependent Rarβ gene induction (Fig 4F). By contrast, AGN 193109 did not cancel the ROL-dependent enhancement of Lgr5 expression (Fig 4F). Furthermore, the organoid forming efficiency raised by ROL in proximal colon organoids was not affected by co-incubation of the culture with AGN 193109 (Fig 4G). These collective data indicate that ROL-dependent stimulation of Lgr5+ stem cells of the proximal colonic epithelium is mediated by an RA-independent mechanism. Discussion We showed in this study that several enzymes that are involved in RA synthesis are differentially expressed in epithelia along the length of the mouse colon in vivo. In particular, among these enzymes, we confirmed that Aldh1a1 and Aldh1a3 are predominantly expressed in the proximal colon epithelium by ISH and immunohistochemial analyses. A previous study investigated expression patterns of Aldh1a1, Aldh1a2 and Aldh1a3 in the mouse intestine [30]; however, it only described their expression patterns in the fetal small intestine. Therefore, the present study is the first to show the distribution patterns of ALDH1 genes and proteins in the colonic epithelium of adult mice. Importantly, detection of uneven expression patterns of ALDH1a1 and ALDH1a3 in the colon epithelium triggered our study to assess the effect of ROL on proximal and distal colon organoids separately, which led us to demonstrate that ROL has potent activity to promote the growth of proximal colon organoids, but not of distal ones. This finding suggests that the stem cells located in distinct regions of the colon involve distinct mechanisms for their proliferation. It is generally accepted that ROL exerts its biological effect on differentiation and proliferation of target cells through the action of its active metabolite, RA [9]. Contrary to this, we showed that ROL enhances the growth of proximal colon organoids through an RA-independent mechanism. Treatment of colonic organoids with RAL or RA did not result in enhanced growth and Lgr5 gene induction. In addition, co-incubation of ROL-treated cells with an RAR antagonist did not suppress the enhanced Lgr5 expression or organoid forming efficiency. Previous studies have shown that ROL supports the self-renewal of mouse embryonic stem cells (ESCs) in long-term cultures even in the absence of feeder cells [31]. The same authors further reported that this action of ROL is mediated by direct activation of the phosphoinositide three (PI3) kinase signaling pathway through IGF-1 receptor/insulin receptor substrate 1 (IRS-1), but not by the ROL-RA pathway [32]. Several other reports have also discussed the existence of signaling pathways dependent on ROL, but distinct from that of RA. The hydroxylated forms of ROL were shown to regulate cell fate decisions by enhancing protein kinase C α (PKCα) and its downstream MAP kinase activity [33] or by supporting PDGF signaling [34]. Further studies are underway in our group to address whether these RA-independent mechanisms are also involved in ROL-mediated growth promotion in the proximal colonic stem cells. In addition, although the function of RA was not involved in ROL-dependent growth enhancement of proximal colon cells, RA induced a certain cellular response, an induction of the Rarβ gene, in cultured organoids. Interestingly, we showed that RA did not alter the expression level of MUC2, one of the marker genes that show cellular differentiation of colonic cells (Fig 4D). This argues against previous observations that RA promotes differentiation of colonic epithelial cells when laboratory adapted cell lines derived from human colon cancers were used as model systems [27, 28]. In this regard, it is also important to investigate what is the molecular function of RA in the physiological colonic epithelium and how RA influences cell fate determination of those epithelial cells. Following the development of conditions for growing mouse small intestinal stem cells, the intestinal organoid system has been widely applied for various studies not only on the small intestine but also on other tissues including the colon [35, 36]. This methodological advance has led to new approaches to investigate the mechanism of proliferation of normal colonic epithelial cells [37–39] and also the pathways of their tumorigenesis [40–42]. To our knowledge, however, there has been no study that assessed ex vivo response of colonic epithelial stem cell populations of different parts along the length of the colon and described their different behaviors. In this context, our demonstration of the novel function of ROL on the growth of proximal colon organoids would not only evolve a better system to expand those cells with greater efficiency for many applications, but also stimulate research into the mechanism of region-specific regulation of colonic stem cell maintenance. In summary, we showed that ROL has potent activity to promote the growth of proximal colon organoids by stimulating Lgr5+ stem cell populations. Further study to uncover this RA-independent mechanism would provide insights into the distinct features of proximal and distal colon stem cells. Supporting Information S1 Table Information on primers and reaction conditions for PCR. Gene names, sequences of sense and antisense primers, and the cycle number to semi-quantitatively amplify each gene were presented. (DOCX) Click here for additional data file. We thank Lauren Unik for manuscript editing. This study was supported by MEXT KAKENHI (Grant number 26112705), JSPS KAKENHI (24390186 and 26221307), the Regenerative Medicine Realization Base Network Program from the Japan Science and Technology Agency, and Health and Labour Sciences Research Grants for research on Intractable Diseases from Ministry of Health, Labor and Welfare of Japan. The authors have no conflicting financial interests. ==== Refs References 1 Barker N , van de Wetering M , Clevers H . The intestinal stem cell . Genes Dev . 2008 ;22 (14 ):1856 –64 . 10.1101/gad.1674008 18628392 2 Clevers H . The intestinal crypt, a prototype stem cell compartment . Cell . 2013 ;154 (2 ):274 –84 . 10.1016/j.cell.2013.07.004 .23870119 3 Crosnier C , Stamataki D , Lewis J . Organizing cell renewal in the intestine: stem cells, signals and combinatorial control . Nat Rev Genet . 2006 ;7 (5 ):349 –59 . 10.1038/nrg1840 .16619050 4 Sato T , Vries RG , Snippert HJ , van de Wetering M , Barker N , Stange DE , et al Single Lgr5 stem cells build crypt-villus structures in vitro without a mesenchymal niche . Nature . 2009 ;459 (7244 ):262 –5 . 10.1038/nature07935 .19329995 5 Sato T , Stange DE , Ferrante M , Vries RG , Van Es JH , Van den Brink S , et al Long-term expansion of epithelial organoids from human colon, adenoma, adenocarcinoma, and Barrett's epithelium . Gastroenterology . 2011 ;141 (5 ):1762 –72 . 10.1053/j.gastro.2011.07.050 .21889923 6 Yui S , Nakamura T , Sato T , Nemoto Y , Mizutani T , Zheng X , et al Functional engraftment of colon epithelium expanded in vitro from a single adult Lgr5⁺ stem cell . Nat Med . 2012 ;18 (4 ):618 –23 . 10.1038/nm.2695 .22406745 7 Gudas LJ , Wagner JA . Retinoids regulate stem cell differentiation . J Cell Physiol . 2011 ;226 (2 ):322 –30 . 10.1002/jcp.22417 20836077 8 Kawaguchi R , Yu J , Honda J , Hu J , Whitelegge J , Ping P , et al A membrane receptor for retinol binding protein mediates cellular uptake of vitamin A . Science . 2007 ;315 (5813 ):820 –5 . 10.1126/science.1136244 .17255476 9 Kumar S , Sandell LL , Trainor PA , Koentgen F , Duester G . Alcohol and aldehyde dehydrogenases: retinoid metabolic effects in mouse knockout models . Biochim Biophys Acta . 2012 ;1821 (1 ):198 –205 . 10.1016/j.bbalip.2011.04.004 21515404 10 Molotkov A , Deltour L , Foglio MH , Cuenca AE , Duester G . Distinct retinoid metabolic functions for alcohol dehydrogenase genes Adh1 and Adh4 in protection against vitamin A toxicity or deficiency revealed in double null mutant mice . J Biol Chem . 2002 ;277 (16 ):13804 –11 . 10.1074/jbc.M112039200 11836246 11 Molotkov A , Fan X , Deltour L , Foglio MH , Martras S , Farrés J , et al Stimulation of retinoic acid production and growth by ubiquitously expressed alcohol dehydrogenase Adh3 . Proc Natl Acad Sci U S A . 2002 ;99 (8 ):5337 –42 . 10.1073/pnas.082093299 11959987 12 Zhang M , Hu P , Krois CR , Kane MA , Napoli JL . Altered vitamin A homeostasis and increased size and adiposity in the rdh1-null mouse . FASEB J . 2007 ;21 (11 ):2886 –96 . 10.1096/fj.06-7964com .17435174 13 Sandell LL , Sanderson BW , Moiseyev G , Johnson T , Mushegian A , Young K , et al RDH10 is essential for synthesis of embryonic retinoic acid and is required for limb, craniofacial, and organ development . Genes Dev . 2007 ;21 (9 ):1113 –24 . 10.1101/gad.1533407 17473173 14 Niederreither K , Subbarayan V , Dollé P , Chambon P . Embryonic retinoic acid synthesis is essential for early mouse post-implantation development . Nat Genet . 1999 ;21 (4 ):444 –8 . 10.1038/7788 .10192400 15 Mic FA , Haselbeck RJ , Cuenca AE , Duester G . Novel retinoic acid generating activities in the neural tube and heart identified by conditional rescue of Raldh2 null mutant mice . Development . 2002 ;129 (9 ):2271 –82 . 11959834 16 Fan X , Molotkov A , Manabe S , Donmoyer CM , Deltour L , Foglio MH , et al Targeted disruption of Aldh1a1 (Raldh1) provides evidence for a complex mechanism of retinoic acid synthesis in the developing retina . Mol Cell Biol . 2003 ;23 (13 ):4637 –48 . 12808103 17 Dupé V , Matt N , Garnier JM , Chambon P , Mark M , Ghyselinck NB . A newborn lethal defect due to inactivation of retinaldehyde dehydrogenase type 3 is prevented by maternal retinoic acid treatment . Proc Natl Acad Sci U S A . 2003 ;100 (24 ):14036 –41 . 10.1073/pnas.2336223100 14623956 18 Chambon P . A decade of molecular biology of retinoic acid receptors . FASEB J . 1996 ;10 (9 ):940 –54 . .8801176 19 Khillan JS . Vitamin A/retinol and maintenance of pluripotency of stem cells . Nutrients . 2014 ;6 (3 ):1209 –22 . 10.3390/nu6031209 24662164 20 Zhang S , Sun J , Pan S , Zhu H , Wang L , Hu Y , et al Retinol (vitamin A) maintains self-renewal of pluripotent male germline stem cells (mGSCs) from adult mouse testis . J Cell Biochem . 2011 ;112 (4 ):1009 –21 . 10.1002/jcb.23029 .21308744 21 Dillard AC , Lane MA . Retinol decreases beta-catenin protein levels in retinoic acid-resistant colon cancer cell lines . Mol Carcinog . 2007 ;46 (4 ):315 –29 . 10.1002/mc.20280 .17219422 22 Nicke B , Riecken EO , Rosewicz S . Induction of retinoic acid receptor beta mediates growth inhibition in retinoid resistant human colon carcinoma cells . Gut . 1999 ;45 (1 ):51 –7 . 10369704 23 Lee MO , Han SY , Jiang S , Park JH , Kim SJ . Differential effects of retinoic acid on growth and apoptosis in human colon cancer cell lines associated with the induction of retinoic acid receptor beta . Biochem Pharmacol . 2000 ;59 (5 ):485 –96 . .10660115 24 Fukuda M , Mizutani T , Mochizuki W , Matsumoto T , Nozaki K , Sakamaki Y , et al Small intestinal stem cell identity is maintained with functional Paneth cells in heterotopically grafted epithelium onto the colon . Genes Dev . 2014 ;28 (16 ):1752 –7 . 10.1101/gad.245233.114 25128495 25 Smith FR , Goodman DS . The effects of diseases of the liver, thyroid, and kidneys on the transport of vitamin A in human plasma . J Clin Invest . 1971 ;50 (11 ):2426 –36 . 10.1172/JCI106741 5096525 26 Park EY , Dillard A , Williams EA , Wilder ET , Pepper MR , Lane MA . Retinol inhibits the growth of all-trans-retinoic acid-sensitive and all-trans-retinoic acid-resistant colon cancer cells through a retinoic acid receptor-independent mechanism . Cancer Res . 2005 ;65 (21 ):9923 –33 . 10.1158/0008-5472.CAN-05-1604 .16267017 27 Bengtsson AM , Jönsson G , Magnusson C , Salim T , Axelsson C , Sjölander A . The cysteinyl leukotriene 2 receptor contributes to all-trans retinoic acid-induced differentiation of colon cancer cells . BMC Cancer . 2013 ;13 :336 10.1186/1471-2407-13-336 23829413 28 Ordóñez-Morán P , Dafflon C , Imajo M , Nishida E , Huelsken J . HOXA5 Counteracts Stem Cell Traits by Inhibiting Wnt Signaling in Colorectal Cancer . Cancer Cell . 2015 ;28 (6 ):815 –29 . 10.1016/j.ccell.2015.11.001 .26678341 29 Klein ES , Pino ME , Johnson AT , Davies PJ , Nagpal S , Thacher SM , et al Identification and functional separation of retinoic acid receptor neutral antagonists and inverse agonists . J Biol Chem . 1996 ;271 (37 ):22692 –6 . .8798442 30 Niederreither K , Fraulob V , Garnier JM , Chambon P , Dollé P . Differential expression of retinoic acid-synthesizing (RALDH) enzymes during fetal development and organ differentiation in the mouse . Mech Dev . 2002 ;110 (1–2 ):165 –71 . .11744377 31 Chen L , Khillan JS . Promotion of feeder-independent self-renewal of embryonic stem cells by retinol (vitamin A) . Stem Cells . 2008 ;26 (7 ):1858 –64 . 10.1634/stemcells.2008-0050 .18436859 32 Chen L , Khillan JS . A novel signaling by vitamin A/retinol promotes self renewal of mouse embryonic stem cells by activating PI3K/Akt signaling pathway via insulin-like growth factor-1 receptor . Stem Cells . 2010 ;28 (1 ):57 –63 . 10.1002/stem.251 .19890980 33 Imam A , Hoyos B , Swenson C , Levi E , Chua R , Viriya E , et al Retinoids as ligands and coactivators of protein kinase C alpha . FASEB J . 2001 ;15 (1 ):28 –30 . 10.1096/fj.00-0329fje .11099483 34 Chen Y , Derguini F , Buck J . Vitamin A in serum is a survival factor for fibroblasts . Proc Natl Acad Sci U S A . 1997 ;94 (19 ):10205 –8 . 9294188 35 Sato T , Clevers H . Growing self-organizing mini-guts from a single intestinal stem cell: mechanism and applications . Science . 2013 ;340 (6137 ):1190 –4 . 10.1126/science.1234852 .23744940 36 Date S , Sato T . Mini-gut organoids: reconstitution of the stem cell niche . Annu Rev Cell Dev Biol . 2015 ;31 :269 –89 . 10.1146/annurev-cellbio-100814-125218 .26436704 37 Hirokawa Y , Yip KH , Tan CW , Burgess AW . Colonic myofibroblast cell line stimulates colonoid formation . Am J Physiol Gastrointest Liver Physiol . 2014 ;306 (7 ):G547 –56 . 10.1152/ajpgi.00267.2013 .24481605 38 Davies JM , Santaolalla R , von Furstenberg RJ , Henning SJ , Abreu MT . The Viral Mimetic Polyinosinic:Polycytidylic Acid Alters the Growth Characteristics of Small Intestinal and Colonic Crypt Cultures . PLoS One . 2015 ;10 (9 ):e0138531 10.1371/journal.pone.0138531 26414184 39 Beyaz S , Mana MD , Roper J , Kedrin D , Saadatpour A , Hong SJ , et al High-fat diet enhances stemness and tumorigenicity of intestinal progenitors . Nature . 2016 ;531 (7592 ):53 –8 . 10.1038/nature17173 26935695 40 Onuma K , Ochiai M , Orihashi K , Takahashi M , Imai T , Nakagama H , et al Genetic reconstitution of tumorigenesis in primary intestinal cells . Proc Natl Acad Sci U S A . 2013 ;110 (27 ):11127 –32 . 10.1073/pnas.1221926110 23776211 41 Yum MK , Kang JS , Lee AE , Jo YW , Seo JY , Kim HA , et al AIMP2 controls intestinal stem cell compartments and tumorigenesis by modulating Wnt/β-catenin signaling . Cancer Res . 2016 10.1158/0008-5472.CAN-15-3357 .27262173 42 Zasadil LM , Britigan EM , Ryan SD , Kaur C , Guckenberger DJ , Beebe DJ , et al High rates of chromosome missegregation suppress tumor progression, but do not inhibit tumor initiation . Mol Biol Cell . 2016 10.1091/mbc.E15-10-0747 .27146113

PMC005xxxxxx/PMC5001648.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756441610.1371/journal.pone.0157655PONE-D-16-01644Research ArticleBiology and Life SciencesNeuroscienceBrain MappingMagnetoencephalographyResearch and Analysis MethodsImaging TechniquesNeuroimagingMagnetoencephalographyBiology and Life SciencesNeuroscienceNeuroimagingMagnetoencephalographyPhysical SciencesPhysicsCondensed Matter PhysicsMagnetismMagnetic FieldsBiology and Life SciencesAnatomyHeadScalpMedicine and Health SciencesAnatomyHeadScalpEngineering and TechnologyEquipmentMeasurement EquipmentMagnetometersResearch and Analysis MethodsBioassays and Physiological AnalysisElectrophysiological TechniquesBrain ElectrophysiologyBiology and Life SciencesPhysiologyElectrophysiologyNeurophysiologyBrain ElectrophysiologyMedicine and Health SciencesPhysiologyElectrophysiologyNeurophysiologyBrain ElectrophysiologyBiology and Life SciencesNeuroscienceNeurophysiologyBrain ElectrophysiologyBiology and Life SciencesAnatomyHeadMedicine and Health SciencesAnatomyHeadBiology and Life SciencesNeuroscienceSensory SystemsComputer and Information SciencesNeural NetworksBiology and Life SciencesNeuroscienceNeural NetworksOn the Potential of a New Generation of Magnetometers for MEG: A Beamformer Simulation Study New Generation of Magnetometers for MEGBoto Elena 1Bowtell Richard 1Krüger Peter 2Fromhold T. Mark 2Morris Peter G. 1Meyer Sofie S. 3Barnes Gareth R. 3Brookes Matthew J. 1*1 Sir Peter Mansfield Imaging Centre, School of Physics and Astronomy, University of Nottingham, University Park, NG7 2RD, Nottingham, United Kingdom2 Midlands Ultracold Atom Research Centre, School of Physics and Astronomy, University of Nottingham, University Park, NG7 2RD, Nottingham, United Kingdom3 Wellcome Trust Centre for Neuroimaging, University College London, 12 Queen Square, WC1N 3BG, London, United KingdomJohnson Blake EditorAustralian Research Council Centre of Excellence in Cognition and its Disorders, AUSTRALIACompeting Interests: The authors have declared that no competing interests exist. Conceived and designed the experiments: MJB GRB RB PM TMF PK. Performed the experiments: EB SSM MJB. Analyzed the data: EB SSM MJB. Contributed reagents/materials/analysis tools: EB SSM MJB. Wrote the paper: EB RB PK TMF PM GRB MJB. * E-mail: matthew.brookes@nottingham.ac.uk26 8 2016 2016 11 8 e015765513 1 2016 2 6 2016 © 2016 Boto et al2016Boto et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Magnetoencephalography (MEG) is a sophisticated tool which yields rich information on the spatial, spectral and temporal signatures of human brain function. Despite unique potential, MEG is limited by a low signal-to-noise ratio (SNR) which is caused by both the inherently small magnetic fields generated by the brain, and the scalp-to-sensor distance. The latter is limited in current systems due to a requirement for pickup coils to be cryogenically cooled. Recent work suggests that optically-pumped magnetometers (OPMs) might be a viable alternative to superconducting detectors for MEG measurement. They have the advantage that sensors can be brought to within ~4 mm of the scalp, thus offering increased sensitivity. Here, using simulations, we quantify the advantages of hypothetical OPM systems in terms of sensitivity, reconstruction accuracy and spatial resolution. Our results show that a multi-channel whole-head OPM system offers (on average) a fivefold improvement in sensitivity for an adult brain, as well as clear improvements in reconstruction accuracy and spatial resolution. However, we also show that such improvements depend critically on accurate forward models; indeed, the reconstruction accuracy of our simulated OPM system only outperformed that of a simulated superconducting system in cases where forward field error was less than 5%. Overall, our results imply that the realisation of a viable whole-head multi-channel OPM system could generate a step change in the utility of MEG as a means to assess brain electrophysiological activity in health and disease. However in practice, this will require both improved hardware and modelling algorithms. Medical Research Council (GB) New Investigator GrantMR/M006301/1Brookes Matthew J. Engineering and Physical Sciences Research Council (GB) UK National Quantum Technology Hub for Sensors and Metrology GrantEP/M013294/1Medical Research Council and Engineering and Physical Sciences Research Council (GB) Partnership GrantMR/K6010/86010/1Meyer Sofie S. This work has been funded by the EPSRC UK National Quantum Technology Hub for Sensors and Metrology (EP/M013294/1). MJB is funded by a Medical Research Council New Investigator Research Grant (MR/M006301/1). SSM is funded by a MRC and EPSRC Partnership Grant (MR/K6010/86010/1). Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction Magnetoencephalography (MEG; [1, 2]) is a non-invasive technique to image electrophysiology in the human brain. It is based on assessing the changes in magnetic field outside the head that are generated by synchronised current flow through neuronal assemblies in the brain. These extra-cranial magnetic field fluctuations are of order ~10−14 T and are typically detected via an array of superconducting coils placed around the head, each coupled to a superconducting quantum interference device (SQUID) [3, 4]. Appropriate mathematical modelling of the measured magnetic field facilitates reconstruction of 3-dimensional current density images, which depict moment-to-moment changes in the brain’s electrical activity. Recent years have seen rapid improvement in the utility of MEG, driven primarily by improved algorithms for data analysis. The MEG inverse problem (generating 3D current density images using measurements of extra-cranial magnetic fields) is ill-posed (i.e. no unique solution exists), which makes the direct inference of spatial dynamics mathematically complex. However, many algorithms (for example minimum-norm [5, 6] and beamforming [7–11]) now exist to verifiably reconstruct the spatial, temporal and spectral signature of human brain activity with reasonable (~5–10 mm; [12, 13]) spatial resolution and excellent (<1 ms) temporal resolution. This capability makes MEG one of the most exciting tools for basic neuroscientific and clinical research that is currently available. Multiple studies [14, 15] have shown that appropriately modelled MEG signals are rich in information content and strikingly similar to invasive electrophysiological measurements made in animals and humans. Further, MEG studies have had significant impact on our understanding of neural networks [16–22], allowing elucidation of intrinsic modes of electrophysiological coupling between spatially separate and functionally specific brain regions [23]. Most importantly, MEG is providing new information on the pathophysiology of a range of diseases and conditions including developmental disorders [24], psychoses [25] and neurodegeneration [26]. Despite its excellent and unique potential, MEG is limited by a low signal-to-noise ratio (SNR) and restricted spatial resolution. Magnetic fields generated by the brain are much smaller in magnitude than environmental and biomagnetic interference. This problem is observed directly in comparisons of MEG with invasive electrophysiology; for example, comparing MEG measurements of Zumer et al. [15] to invasive measurements by Mukamel et al. [27], despite good overall agreement, the sensitivity of MEG at high frequency (i.e. measuring neural oscillations above ~50 Hz) decreases due to poor SNR. In principle, SNR could be increased by moving detectors closer to the head (the field from dipoles in the brain follows an inverse square law, meaning that halving the brain to detector separation would quadruple SNR). However, in SQUID-based systems, the requirement for cryogenic cooling of pickup coils and SQUIDs means that a vacuum must be maintained between the head and the coils, thus limiting the minimum separation. Spatial resolution in MEG is governed by our ability to separate magnetic fields generated by sources in spatially separate brain locations; this is affected by the inverse modelling algorithm, but also by sensor coverage, sensor type and SNR. In general, poor spatial resolution means that reconstructed signals at separate locations are spuriously correlated, particularly for locations in close proximity (i.e. nearby sources become linearly mixed). This impacts directly on our ability to adequately characterise brain networks; for example in the study of connectivity (measuring statistical interdependencies between regions; [23, 28, 29]) methodologies [20, 30–33] have to be used for post-hoc correction to reduce such spurious correlations, and the spatial scale of the networks that can be assessed is limited. If these restrictions on SNR and spatial resolution could be mitigated, this would afford a step change in the utility of MEG as a functional neuroimaging modality. Optically-pumped magnetometers (OPMs) measure the transmission of laser light through a vapour of spin-polarised rubidium atoms, which provides a highly sensitive measure of the local magnetic field. These devices are emerging as an alternative to SQUIDs for measuring small magnetic fields and recent studies have begun to show their viability for a number of biomagnetic applications, including MEG. OPMs with sensitivities surpassing SQUIDs’ have been demonstrated [34, 35]; working in the spin-exchange relaxation-free (SERF) regime [36, 37]. Several papers have described the successful OPM detection of phase-locked evoked responses generated by auditory [38–40] or somatosensory [38, 41] stimulation. Other studies have shown that neural oscillations (rhythmic electrical activity in neural networks) can also be detected [41, 42]. In addition, the detection of epileptiform activity in rodents using micro-fabricated OPM sensors [43] has been shown. These empirical demonstrations, coupled with miniaturization of OPM cells to the millimetre scale [44] and the introduction of the first multi-channel measurements [45, 46] have begun to suggest that OPM technology could transform the utility of MEG with chip-scale sensors, in principle, facilitating the use of several hundred detectors, located just millimetres from the scalp surface, with a completely flexible geometry. Qualitatively the advantages afforded by an OPM-based MEG system are obvious—bringing sensors closer to the head will increase the measured signal; assuming noise floors of OPMs and SQUIDs are comparable, this would lead to increased SNR in MEG measurements. Although this result is expected, many critical questions remain unanswered: What quantitative improvement in SNR might be expected and how does this change with cortical location? How would this SNR gain affect source reconstruction accuracy (following inverse modelling)? How would an OPM system be impacted by the accuracy of modelling algorithms? What would the gain in spatial resolution be compared to current systems? In addition, the number of sensors required remains an open question. In general it is well known that increased sensor density improves spatial resolution. However, at high SNR spatial resolution has been shown to be asymptotic [47] with respect to the number of sensors. Given the potential for high SNR when using OPM systems, this potentially means that beyond some critical number, the addition of further sensors may not generate significant improvement in spatial specificity. In the present paper, we quantify the potential advantages of several theoretical MEG systems in which sensors are placed directly on the scalp surface. Note that in principle our theoretical systems could result from any technology in which the external surface of the magnetic field sensor is at room temperature, for example nitrogen-vacancy centres in diamond [48]; however we specifically use the term OPM throughout the paper due to the excellent potential of these devices both in terms of achieving close detector proximity to the head and their size (millimetre scale OPMs mean systems with many hundreds of sensors are plausible). Note also that our paper does not address the issues relating to the practical use of OPMs, rather our aim is to provide a benchmark to which any real system might aspire. Using simulations, we investigate the MEG forward field in order to quantify the improvement in sensitivity of an OPM system compared to a current state-of-the-art SQUID system. Following this, we employ a beamformer spatial filtering approach to quantify the advantages of an OPM system for source reconstruction. We examine reconstruction accuracy of beamforming for single dipoles. In addition, by combining dipolar sources of interest with interference generators, we examine the efficacy of interference rejection in OPM and SQUID systems. We investigate how the accuracy of magnetic field modelling (i.e. the accuracy of the modelled forward field) impacts the OPM and SQUID systems. Finally, we examine the spatial resolution of OPM and SQUID systems, its reliance on the number of channels, and how this might impact connectivity metrics. Our results show that OPM systems, built to the geometries that we suggest, have the potential to offer a step change in the utility of MEG, with significant improvements in sensitivity and spatial resolution well beyond what could ever be expected of current instruments. However, we also show that the advantages of these next generation systems are critically dependent on our ability to generate accurate forward models, meaning that realisation of the true benefits will only result from concerted efforts in both hardware development and mathematical modelling. Methods Simulated systems sSQUID For all simulated SQUID-based measures, we employed the third-order, synthetic gradiometer configuration of a 275-channel CTF whole-head MEG system (MISL, Canada). This system contains 275 radially-oriented, axial gradiometers with a 5 cm baseline, along with a 29-channel reference array that is used to construct third-order synthetic gradiometers. The position of the head within the MEG helmet was based on an experimental recording, and was determined by using three head position indicator coils attached to the subject’s head (nasion, left pre-auricular and right pre-auricular fiducial markers). In this system, the minimum distance from the scalp surface to the closest detector coils is 1.7 cm; this would assume that the scalp is touching the surface of the MEG helmet (which of course is not simultaneously achievable for all scalp locations). In the simulations used here, the subject was positioned with the back of their head touching the back of the helmet, meaning that the largest scalp-to-sensor distance was at the front of the head for the sSQUID system. High resolution anatomical magnetic resonance images (1 mm3 resolution), including fiducial markers, were available for this subject allowing complete registration of the MEG sensor geometry to the brain anatomy. sOPM In the simulated OPM system, measurement locations were 4 mm from the scalp surface; a distance which is realistic for currently available SERF OPMs [49]. The arrangement of sensors was determined by the projection of the 275 sensor locations from the sSQUID system onto the scalp of the same individual used for the sSQUID simulations. Given the size of current micro-fabricated OPM devices (~1 cm2 footprint) such a 275-channel system is conceivable. The scalp surface was extracted from the anatomical MRI using Iso2Mesh (http://iso2mesh.sourceforge.net/; [50]). To ensure equivalence of the sSQUID and sOPM systems for sensitivity comparison, the sOPM system was also configured with its sensors forming radially-oriented axial gradiometers (5 cm baseline). Note however that for beamformer measurements (see later) we consider OPM magnetometers. In both the sOPM and sSQUID simulations, simulated dipoles were located on the cortical surface, which was extracted from the anatomical MRI using Freesurfer (http://surfer.nmr.mgh.harvard.edu/; [51]). This meant that 21,401 individual dipoles could be located at the vertices of the triangular cortical mesh, with each dipole oriented perpendicular to the cortical surface. In order to simulate magnetic fields at the individual sensor locations, a multiple local sphere head model was used [52] in conjunction with the forward model derived by Sarvas [53]. Gradiometers were formed via a subtraction of the radial field components through each gradiometer loop (in the case of the sSQUID) or through two OPMs in the case of sOPM. Third-order gradiometers for the sSQUID were constructed using weighting coefficients from the MEG recording. Comparison of SNR in sOPM and sSQUID systems Assuming an equivalent noise floor (≲10 fT rms/Hz1/2) for the sSQUID [54] and the sOPM systems (both in magnetometer [36] and gradiometer configurations), the difference in field magnitudes at the sensors can be used to calculate the gain in SNR afforded by the sOPM system (thus, in the remainder of this manuscript, all SNR calculations assume equivalent noise floors for the sSQUID and sOPM). Measured fields at each sensor were simulated for each possible dipole location in the cortex, assuming a dipole moment, │Q│ = 1 nAm. For every cortical location, j, we calculated the expected output (i.e. the forward field) lj for all 275 MEG channels for both systems (therefore lj has dimensions of Nch × 1, being Nch the number of channels). The Frobenius norm, fj, of the forward field vector lj was then computed as, fj=|lj|F=Σj=1Nch|lj|2(1) and the comparison of the sSQUID and sOPM systems was derived as, ratioj=fj,OPMfj,SQ(2) This ratio, which effectively represents the change in measured signal (and SNR) afforded by moving from the sSQUID to the sOPM system, was calculated at each source location on the cortical surface. Reconstruction accuracy for single dipoles To examine the accuracy of source reconstruction, we first simulated a number of sets of MEG data. A source time course qj, at cortical location j, containing a total of N time points was simulated as Gaussian-distributed random data with a standard deviation of 5 nAm. Note here that N = f∙Δ, where f is sampling frequency and Δ represents the duration of the simulated recording. This dipole signal was projected through the forward field vector for location j, lj, and simulated MEG data were generated as, m(t)=ljqj+e,(3) with dimensions of Nch × N. Here, e represents noise which, for the purposes of this paper, was generated as normally-distributed random data with a standard deviation of 50 fT. A single set of MEG data, Δ = 600 s (N = 360,000) in duration, was simulated for each of the 21,401 possible dipole locations on the cortical mesh. This was done twice using forward fields computed for the sSQUID and sOPM systems. The SNR of these data was dependent on the cortical location, and varied from 0 to 3 for the sSQUID system and 0 to 24 for the sOPM system (note here SNR is computed as the standard deviation of the signal at the maximally affected channel divided by the mean noise). The difference in SNR between systems was purely a result of the proximity of the sensors to the scalp. To compare the accuracy of source reconstruction in the two systems, a beamformer spatial filter was implemented. Beamforming [7, 9, 11, 55, 56] is a popular solution to the MEG inverse problem that has been used extensively for MEG, particularly in the study of neural oscillations and functional connectivity. An estimate of electrical source strength, q̂j, at cortical location j was given by a weighted sum of sensor measurements so that: q^j=wjTm,(4) Note that we use the ‘hat’ notation to represent a beamformer estimate; i.e. q̂j is the beamformer estimate of qj made using only the MEG data, m. wj is a vector of weighting parameters (dimension Nch × 1) tuned to location j and derived based on variance minimisation; the overall variance in q̂j is minimised with the linear constraint that signals originating from j remain. Mathematically, minwj|〈q^j2〉| subject to wjTlj=1.(5) The most commonly employed solution to this equation is: wjT=ljT{C+μΣ}−1ljT{C+μΣ}−1lj(6) where C = <mT m> and is approximated by the data covariance. Σ is a diagonal matrix representing the noise at each of the MEG channels and μ is a regularisation parameter which for the purposes of this paper was set to 0.001 in order to maximise spatial resolution. Beamforming was applied in this way for all of the 21,401 simulated MEG datasets. In order to determine temporal reconstruction accuracy, the Pearson correlation coefficient was measured between the original simulated dipole time course data, qj, and the beamformer-reconstructed time course data q̂j. These correlation coefficients were computed for the sSQUID and sOPM systems and are denoted as rSQ(qj,q̂j) and rOPM(qj,q̂j), respectively. To examine the spatial performance of the beamformer, 4 dipoles were chosen at random in the frontal, parietal and occipital regions. MEG data were generated separately for each dipole and beamformer-reconstructed time courses were generated at the source location and its 44 nearest neighbours, corresponding to distances of between 0 and 3 cm from the simulated source. The magnitude of temporal correlation between the simulated source time course and the beamformer reconstructions was then plotted as a function of distance from the source. Reconstruction accuracy with ‘brain noise’ Although highly illustrative, beamformer reconstruction of single dipoles, as described in the previous section, is not necessarily representative of real life situations in which the beamformer must reconstruct the source of interest whilst simultaneously removing sources of interference. In order to investigate this, multiple dipole simulations were carried out. Here, a single source of interest, qj, was simulated on the cortical surface as previously. However in addition, 5 further sources, denoted by qInt,k, where k = 1, 2, …, 5, were also simulated as interference generators (brain noise). For both qj and qInt,k, source time courses were simulated as Gaussian-distributed random data with a standard deviation of 5 nAm. The MEG data were then simulated as: m(t)=ljqj+{Σk=15lIntkqIntk}+e(7) where lInt,k represents the forward field vector for interference source k. Two separate cases of interference generators were simulated: Case 1: Interference in close proximity. All 5 interference sources were dipolar, located on the cortical mesh, oriented perpendicular to the cortical surface and defined between 0 cm and 3 cm (Euclidean distance) from the source of interest. Actual interference locations were random, but in all cases were set such that qInt,1 was located at a distance of 0–6 mm from j; qInt,2 was between 6–12 mm from j; qInt,3 was between 12–18 mm from j; qInt,4 was between 18–24 mm from j and finally qInt,5 was between 24-mm from j. The location j, of the source of interest was allowed to vary across 10,400 possible dipole locations within a single hemisphere on the cortical mesh (see Fig 1A). Case 2: Deep sources. Here, all 5 interference sources were again dipolar, oriented perpendicular to the cortical surface and located on the cortical mesh at distances less than 4 cm from the closest sOPM sensor. The source of interest, however, was limited to cortical locations between 4 cm and 6 cm from the closest sOPM sensor. This therefore simulated the situation where the source of interest was located deeper than the sources of interference. The location, j, of the source of interest was allowed to vary across 4,901 possible dipole locations (see Fig 1B). 10.1371/journal.pone.0157655.g001Fig 1 “Brain noise” sources. Schematic showing a single example of the relative locations of the source of interest (black) and the interference generators (green). (A) Case 1 (Interference in close proximity) and (B) Case 2 (Deep sources). In all cases, each MEG dataset was simulated as being 360,000 data samples (600 s) in length. In order to measure reconstruction accuracy we derived correlation coefficients rSQ(qj,q̂j) and rOPM(qj,q̂j) as previously. In order to quantify interference rejection, we also measured correlation between the beamformer-reconstructed estimate, q̂j, and all five interference source time courses; i.e. we measured rIntSQ(qInt,k,q̂j) and rIntOPM(qInt,k,q̂j) for the sSQUID and sOPM systems respectively. Obviously, for perfect beamformer performance we would expect r(qj,q̂j) to be close to one, whilst r(qInt,k,q̂j) should be close to zero. Reconstruction accuracy with forward field error The simulations described in Reconstruction accuracy for single dipoles and with ‘brain noise’ sections assume that the forward model used for beamforming is accurate (i.e. lj in Eqs 3 and 7 which defines the simulated data, and lj in Eq 6 which describes beamforming, were the same). However, in reality, some error in deriving the forward model might be expected due to, for example, inaccurate knowledge of sensor locations or poor physical models of the fields generated. For this reason, it is important to define how forward field accuracy impacts upon beamformer source reconstruction for both sSQUID and sOPM instrumentation. In order to investigate this, we first altered the beamformer weights in Eq 6 such that: wjT=l^jT{C+μΣ}−1l^jT{C+μΣ}−1l^j(8) where l^j=lj+Δlj(9) represents the forward field with added error Δlj. In order to simulate error on the forward model, we employed a geometrical approach depicted in Fig 2A. We first treated the genuine forward field, lj, as a Nch × 1 element vector, and defined a hyperplane in Nch-dimensional space that is perpendicular to lj. In order to perturb lj, we defined the error Δlj as a Nch × 1 element vector that is confined to the hyperplane. The vector sum of lj and Δlj then gave us the modified forward field as in Eq 9. The value ‖Δlj‖/‖lj‖ represents a quantitative measure of the fractional error imposed upon the forward field model. (Note also that the angle α between l̂j and Δlj (see Fig 2A) characterizes the similarity between the accurate and perturbed forward field patterns). 10.1371/journal.pone.0157655.g002Fig 2 Schematic diagram of forward field error simulation. The perfect forward field, lj is perturbed via addition of Δlj to give l̂j. Δlj was constructed by first defining a Nch × 1 vector of random numbers, r, as shown (for the simple 2 dimensional case) in Fig 2B. Projecting this random vector onto the hyperplane gives: Δlj=ξ{r−(lj∥lj∥∥r∥cosβ)}(10) where cosβ=ljTr‖lj‖‖r‖(11) And so Δlj=ξ{r−(lj(ljTlj)−1ljTr)}(12) The magnitude of Δlj is controlled by the parameter ξ which can be used to manipulate the amount by which the forward field is modified. The single source simulations, described in Reconstruction accuracy for single dipoles, were repeated, but this time, sources were reconstructed using beamforming (Eq 8) with added error on the forward field vector. The elements of r were defined as Gaussian-distributed random numbers and ξ was varied through 20 iterations such that ‖Δlj‖/‖lj‖ spanned approximately the range from 0 to 0.5. For each value of ξ in the simulation was repeated 5 times (with 5 realisations of r) and results averaged. This whole process was repeated for 100 dipole locations randomly selected from one cortical hemisphere (performing this calculation for all dipole locations proved too computationally intensive). Results were averaged across locations and differentiated by distance to their closest sOPM sensor, dj. Spatial resolution and channel count To examine the spatial resolution of source-reconstructed MEG data from the sSQUID and sOPM systems, a final set of simulations was employed. Two sources were simulated at separate locations on the cortical surface, k and n. Two source time courses, qk and qn were simulated; both contained a total of N time points and comprised different Gaussian-distributed random data with a standard deviation of 5 nAm, meaning that the sources were temporally orthogonal. MEG data were then generated as m(t)=lkqk+lnqn+e(13) where lk and ln are the forward field vectors for cortical locations k and n respectively, and e represents sensor noise (normally-distributed random data with a standard deviation of 35 fT). This simulation was run multiple times with the source locations (k and n) selected randomly. The Euclidean distance between sources was measured for each source pair. As previously, data were reconstructed using the beamformer spatial filter such that q̂k = wkTm and q̂n = wnTm. In a connectivity measurement in real MEG data, we would look to quantify a relationship between the two dipole time courses, q̂k and q̂n; any statistical interdependency between these quantities would be used to indicate a functional connection between brain regions k and n. However, such a measurement necessarily requires that there is no spurious correlation between q̂k and q̂n caused by the limited spatial resolution. In the present simulation, the underlying time courses qk and qn used to generate the data are orthogonal, so that the correlation coefficient r(qk,qn) = 0. In order to test for the effects of spurious correlations in the beamformer reconstruction, we therefore measured correlation between source reconstructions, r(q̂k,q̂n); if the dipoles simulated at k and n are resolvable, then this quantity should be close to zero. Note that in much of the functional connectivity literature, r(q̂k,q̂n) would be referred to as leakage between sources. This simulation was run 3000 times with N = 360,000. The measured Euclidean distances were categorised into 30 different bins of 2 mm width, the smallest being 0–2 mm, the largest being 58–60 mm. The simulation was designed such that 100 repeats were run for each bin. The spatial resolutions of the sSQUID and sOPM systems were then measured via assessment of the quantity r(q̂k,q̂n) plotted against distance; two sources were considered resolved if r(q̂k,q̂n) < 1/√2 (i.e. less than 50% shared variance between sources). We also investigated the effect of different channel counts on the sOPM system’s spatial resolution. In order to do this, three additional sensor configurations were simulated according: to 1) the international 10–10 EEG system; 2) the 10–5 EEG system [57] and 3) a theoretical 10–2.5 system. The application of these sensor configurations onto the scalp surface [58] resulted in systems with 81, 329 and 1293 sensors respectively. A separate measurement of r(q̂k,q̂n) versus source separation was made for each of these three systems. Finally, a simulation was employed to quantify the spatial resolution of both systems as a function of cortical location. Simulated MEG data were generated for pairs of dipoles as before (Eq 13), with the only difference that k ran for all dipole locations within one cortical hemisphere and n corresponded initially to k’s nearest neighbour on the cortical mesh. In cases where r(q̂k,q̂n) was larger than 1/√2, the simulation was repeated for k’s next nearest neighbour, and so on, until the value of r(q̂k,q̂n) fell below 1/√2. Here, spatial resolution was assessed as the minimum distance between the dipole seed (k) and its neighbour (n) for which both sources can be considered separable. Both this spatial resolution value and the correlation between reconstructed time courses was assessed. Results Fig 3 shows the results of our SNR measurements. Fig 3A shows the basic geometry of our simulations. The upper panel shows the coil arrangement for the sSQUID system which is based upon the axial gradiometer arrangement of a 275-channel CTF MEG instrument. The reference array that is used to form the third-order synthetic gradiometers can also be seen in the figure. The lower panel in Fig 3A shows the sensor arrangement for the sOPM system. Here, again a 5 cm baseline axial gradiometer set-up is employed since this allows a direct comparison of the SNR for the sSQUID and sOPM systems. The sOPM sensor topography across the scalp is based upon the projection of the 275 SQUID channels onto the subject’s scalp. The inner shell of detectors is placed just 4 mm from the scalp surface. Fig 3B shows simulated magnetic fields (i.e. the forward field) for a single dipole located in the parietal lobe. Note that, for this particular dipole, the magnitude of the maximum measured magnetic fields is approximately tenfold higher for the sOPM system than for the sSQUID set-up. As would be expected, the topographic pattern of the magnetic field is also more focal in the sOPM compared to the sSQUID system. Fig 3C shows the ratio of the Frobenius norms of the forward field vectors for the two systems, plotted as a function of dipole location on the cortex. The colours represent the ratio (fj,OPM/fj,SQ) such that a ratio of 5 would indicate a fivefold improvement in SNR of the sOPM compared to the sSQUID system. The figure shows clearly that the sOPM system offers an improvement in SNR for all dipole locations on the cortex; this improvement is approximately a factor of five (yellow) across most of the cortex, increasing to close to ten (red) for some areas of the frontal lobe. The improvement is less dramatic in deeper cortical areas (a consequence of the inverse square law), but the sOPM system still offers twice the SNR of the sSQUID system (blue). Note that the spatial variation of this improvement in SNR depends on the position of the subject’s head in the CTF MEG helmet. As stated above, in the simulations the subject was positioned with the back of their head touching the back of the CTF-system helmet, thus the largest improvements in SNR are observed in the frontal lobes. Note also that these measurements were based on an adult male, with a head size of 21.3 cm anterior to posterior and 14.6 cm left to right, meaning that this individual’s head was well fitted to the size of the CTF helmet: in a subject with a small head, the improvements provided by the sOPM system would be more pronounced. 10.1371/journal.pone.0157655.g003Fig 3 SNR simulation. (A) Simulation set-up; the upper panel shows the coil arrangement for the sSQUID system, based on the axial gradiometer configuration of the 275 channel CTF MEG instrument. The lower panel shows our sOPM system; note 5 cm baseline axial gradiometers are used to allow a direct comparison of the sSQUID and sOPM systems. (B) Simulated magnetic field from a single dipole located in the parietal lobe. Note the increased magnitude and more focal nature of the measured magnetic field patterns. (C) The ratio of the Frobenius norms of the forward fields, plotted as a function of dipole location on the cortex. The colours represent the quantity fj,OPM/fj,SQ; i.e. a ratio of 5 would indicate a fivefold improvement in SNR of the sOPM compared to the sSQUID system (assuming equal noise floors). The left and right panels show different aspects of the same data. Fig 4A shows beamformer reconstruction accuracy in our single dipole simulations (see Reconstruction accuracy for single dipoles). The left panel shows the sSQUID system and right panel shows the sOPM system. In both cases the colour bar indicates temporal correlation between the simulated and reconstructed dipole time courses; i.e. the left and right plots show rSQ(qj,q̂j) and rOPM(qj,q̂j) respectively, plotted against dipole location j on the cortical surface. A value of 1 would indicate perfect reconstruction. Here, the sSQUID system exhibits reasonably good performance with correlation greater than 0.8 across the vast majority of the cortex. However reconstruction accuracy is highest for the occipital lobe and drops in the frontal lobes—most likely due to the position of the subject’s head in the MEG helmet. Reconstruction accuracy is lowest in the temporal lobes and this is likely due to lower sensor coverage for these brain regions. In contrast, for the sOPM system we see excellent reconstruction accuracy across the entire cortex, with correlation greater than 0.9 everywhere on the cortical surface. Note particularly the improvements in reconstruction accuracy in the temporal lobe compared with the sSQUID system. Fig 4B shows spatial reconstruction accuracy for four different locations (shown overlaid in black on the central figure). Correlation coefficients between the original and reconstructed time courses for the simulated dipole location and 44 of its nearest neighbours, are plotted against the relative distance to the simulated dipole location. Results show that both systems generate an accurate spatial reconstruction of the source with maximum correlation at zero distance. However, temporal correlation falls off faster with distance for the sOPM system indicating a sharper (more focal) reconstruction. 10.1371/journal.pone.0157655.g004Fig 4 Reconstruction accuracy. (A) The left hand panel shows beamformer reconstruction accuracy for the sSQUID system. The right hand panel shows beamformer reconstruction accuracy for the sOPM system. In both cases reconstruction accuracy is measured as temporal correlation between the simulated and reconstructed dipole time courses. Colours represent the quantities rSQ(qj,q̂j) and rOPM(qj,q̂j) for left and right panels, respectively. Note the improvement for the sOPM system. (B) Spatial reconstruction accuracy around four cortical locations (in black). The four inset graphs show correlation coefficients between the original simulated time course and reconstructed dipole time courses at the simulated source location and its nearest neighbours (shown in green on the central image). Correlation is plotted as a function of Euclidean distance to the simulated dipole. Note that for both systems, correlation is maximal closest to the simulated source, however correlation falls off more quickly with distance for the sOPM system, indicating a better spatially resolved image. Fig 5, shows beamformer reconstruction accuracy with “brain noise” (see Reconstruction accuracy with ‘brain noise’). Panel 5A shows results for case 1 (interference in close proximity) whilst panel 5B shows results for case 2 (deep sources). In both cases the left hand panel shows average temporal correlation between the simulated and the beamformer-reconstructed source of interest at j (i.e. graphs show the quantity r(qj,q̂j)). For comparison, the cases with and without the interference sources are both shown. Right hand plots represent average correlation between each of the 5 simulated interference sources and the beamformer estimate of the source of interest at j (i.e. the quantity rInt(qInt,k,q̂j)). In Fig 5A (left), it is clear that switching on the interference degrades the quality of the source estimate for both systems, as would be expected; however this degradation is more pronounced for the sSQUID system compared to the sOPM system. Regarding Fig 5A (right), first recall that interference sources were located progressively further away (in terms of Euclidean distance) from the source of interest with increasing index, k; this explains the decreasing temporal correlation with increasing k. Importantly, the beamformer’s ability to reject interference (even in very close proximity) is improved markedly for the sOPM system compared to the SQUID system. Note that the temporal correlation between the beamformer-estimated source of interest and all interfering sources is lower for sOPM than for sSQUID; this is the case for interference in close proximity, and for deep sources. 10.1371/journal.pone.0157655.g005Fig 5 Reconstruction accuracy with “brain noise”. (A) Case 1 (interference in close proximity): left hand panel shows average of correlation coefficients between simulated and reconstructed dipole time courses, r(qj,q̂j), with and without interference. Right hand panel shows average of correlation coefficients between each of the interference sources and the reconstructed main dipole time course, i.e. rInt(qInt,k,q̂j). (B) Case 2 (deep sources): left hand panel represents the average of correlation coefficients between the estimated and the simulated dipole time course, with and without interference sources. Right hand panel shows again average of correlation coefficients with sources of interference. Fig 6 shows how reconstruction accuracy is affected by errors in the forward field model (see Reconstruction accuracy with forward field error). In order to visualise results, dipole locations on the cortical surface were divided according to depth: those whose distance to the closest sOPM sensor, dj, is less than 4 cm (shallow sources—shown in black on the left image) and those whose distance, dj, to the closest sOPM sensor is between 4 cm and 6 cm (deep sources—shown in grey on the left image). To show the influence of forward field accuracy on beamforming, correlation coefficients between simulated and reconstructed time courses (r(qj,q̂j)) were measured and plotted as a function of fractional error on the forward field ‖Δlj‖/‖lj‖. The top graph shows results for deep sources whilst the bottom graph shows results for shallow sources. In both cases note that, with no forward field error, sOPM outperforms sSQUID. This is in agreement with the results given in Figs 4 and 5. However, as forward field error is increased, the reconstruction accuracy of both systems falls, with the sOPM system being more sensitive. This means that, for dipoles close to the scalp surface (shallow dipoles) even an error of less than 5% in the forward field model can negate the significant advantages afforded by the sOPM system. For deeper dipoles, this is less dramatic, but a 25% error on the forward field would again mean that sSQUID and sOPM systems perform approximately equally well, with errors greater than this value meaning the sSQUID system more accurately reconstructs dipole time courses. These results highlight a critical point; although the introduction of an OPM system has the potential to generate a step change in the utility of MEG, the realisation of this potential depends not only on our ability to engineer such a system, but also critically on our ability to model accurately the forward field. This key point will be addressed further in the discussion below. 10.1371/journal.pone.0157655.g006Fig 6 Dependency of reconstruction accuracy on forward field error. Left image shows dipole locations with deep sources shown in grey and shallower sources shown in black. The plots on the right show temporal correlation between the simulated and beamformer reconstructed sources (i.e. r(qj,q̂j)) plotted as a function of fractional error on forward field (‖Δlj‖/‖lj‖). The two separate plots show the case for deep (top) and shallow (bottom) dipoles. The sOPM system is shown in blue and the sSQUID system in red. Note that the improvement in reconstruction accuracy afforded by the OPM based system depends on accurate forward field modelling. Figs 7 and 8 show the results of our spatial resolution measurements. Fig 7A shows how the correlation, r(q̂k,q̂n), between two reconstructed sources (also termed source leakage in connectivity literature—see Spatial resolution and channel count) varies with source separation for the sOPM (blue curve) and sSQUID systems (red curve). The simulated source time courses were orthogonal, so any non-zero value of r(q̂k,q̂n) is purely an artefact of the limited spatial resolution of the beamformer. Two sources are deemed separable if r(q̂k,q̂n) < 1/√2. Note that no forward field error was simulated for results in Fig 7. It is clear that the 275-channel sOPM system offers much improved spatial resolution compared to the sSQUID system. Sources become separable for the sSQUID system when the Euclidean separation is approximately 5 mm. Indeed this value is in approximate agreement with findings from real data [12]. However for the sOPM system, sources are separable at smaller separations, with temporal correlation never reaching more than 0.4 (16% shared variance). This finding shows clearly the significant advantages of an OPM system, particularly in the study of functional connectivity; this will be addressed further in the discussion below. In the inset figure, the correlation coefficients between simulated and reconstructed time courses for each dipole source (r(qk,q̂k) and r(qn,q̂n)) are also plotted for both systems, as a function of distance between sources. This reflects the reconstruction accuracy for each dipole, showing that each dipole time course is better reconstructed as distance between them is increased. Note again improved performance of the sOPM system. Fig 7B shows four different sOPM systems: (from top to bottom) 81 sensors following the 10–10 EEG system; 275 sensors (as shown in Fig 1A); 329 sensors following the 10–5 system; 1293 sensors following a hypothetical 10–2.5 system. These four simulated systems were used to assess the advantages of increased channel counts for spatial resolution. Fig 7C shows results of this measurement: again r(q̂k,q̂n) is plotted as a function of distance between sources for the 81-channel (orange), 275-channel (blue), 329-channel (green) and 1293-channel (purple) OPM systems. A marked improvement in spatial resolution is observed as channel count is increased. 10.1371/journal.pone.0157655.g007Fig 7 Spatial resolution and channel count measurements. (A) Correlation coefficient between two simulated dipole sources after beamforming reconstruction, r(q̂k,q̂n) plotted against Euclidean distance between the sources. The red and blue traces show the results for the sSQUID system and sOPM systems, respectively. Note that the simulated sources were temporally uncorrelated, so a value of zero would indicate perfect reconstruction. Note the improved performance of the sOPM compared to the sSQUID system. The inset shows correlation coefficients between simulated and reconstructed time courses for the two sources separately, i.e. r(qk,q̂k) and r(qn,q̂n). (B) Simulation set-up comparing four different sOPM systems (from top to bottom): 81 sensors following the 10–10 EEG system; 275 sensors (equivalent to that shown in Fig 1A); 329 sensors following the 10–5 system; 1293 sensors following a hypothetical 10–2.5 system. (C) Comparison of spatial resolution for the 81, 275, 329 and 1293 sensor sOPM systems. Again the graphs show correlation between reconstructed time courses plotted as a function of Euclidean distance between sources. Note the improvement in performance as channel count is increased. 10.1371/journal.pone.0157655.g008Fig 8 Spatial resolution (275 channels). (A) Minimum distance between the dipole seed and its neighbours for which two sources are separable (defined when temporal correlation between time courses falls below 1/√2) plotted as a function of seed dipole location. Left hand panel shows results for the sSQUID system and right hand panel shows results for the sOPM system. Importantly, the discrete nature of the cortical mesh means that a dipole separation corresponding to a temporal correlation of precisely 1/√2 is not possible. For this reason, the spatial separations in Fig 8A must be given alongside their corresponding temporal correlation values which are shown in (B). The left and right hand panels correspond to sSQUID and sOPM systems respectively. Note smaller spatial separations can be achieved with the sOPM system compared to the sSQUID system (A), especially in the temporal lobe. Note also that in addition to smaller separation, leakage (temporal correlation) is also reduced dramatically (B). Finally, Fig 8 shows spatial resolution plotted as a function of location on the cortical surface. Fig 8A shows the minimum spatial separation for which two dipoles become resolvable (i.e. the temporal correlation between time courses falls below 1/√2). This in itself does not constitute a measure of spatial resolution because the discrete nature of the cortical mesh means that a dipole separation corresponding to a temporal correlation of precisely 1/√2 is not possible. For this reason, the spatial separations in Fig 8A must be given alongside a corresponding value of temporal correlation at the derived minimum separation. These values are given in Fig 8B. Note that not only is the minimum spatial separation between dipoles reduced from an average of 6.3 mm in the sSQUID system to 2.5 mm in the sOPM system, but also the temporal correlation at those minimum separations is reduced from of 0.49 in the sSQUID to 0.29 in the sOPM. This important result is in agreement with that shown in Fig 7 and further supports the finding that the sOPM system demonstrates dramatically better spatial resolution compared to the sSQUID system. Discussion In this paper, we have used simulations to quantify the potential advantages of an OPM system in terms of SNR, beamformer reconstruction accuracy (with and without interference) and spatial resolution. Our results show that an OPM system, with an equivalent number of channels and equal noise floor to current SQUID systems, would offer approximately a fivefold increase in SNR across most of the cortex, and up to an order of magnitude improvement in SNR in some regions, for an adult brain. This improvement, which results from the closer proximity of the sensors to the cortex, would be even more dramatic for subjects with smaller heads (e.g. infants). Our results also show a clear improvement in the accuracy of beamformer source reconstruction; average correlation between simulated and beamformer-reconstructed sources increased significantly for our simulated OPM system and this was the case with and without interference from other brain sources. However, our simulations also show that such improvement depends critically on the accurate generation of forward field models. Finally, our results show a marked improvement in spatial resolution of the simulated OPM system, which becomes more dramatic when increasing the channel count. Overall, our results imply that the realisation of a multi-channel whole-head OPM system could generate a step change in the utility of MEG as a means to assess electrophysiological brain activity in health and disease. However, such improvements can only result from both improved hardware and improved modelling algorithms. The SNR improvement afforded by OPM systems would revolutionise both the type of neuromagnetic effects that can be measured by MEG, and the brain regions to which MEG is sensitive. At present, MEG measurements are limited, particularly at high frequency. The sensitivity of the measurement falls with frequency since neural networks oscillating at high frequencies tend to be small, and therefore offer low SNR when magnetic fields are measured at a distance. However, invasive recordings show clearly that neuro-electrical effects up to and including frequencies of 200 Hz are of significant importance to brain function [59]. OPM systems, we believe, could offer the best possible route to assess, non-invasively, networks of neurons oscillating at these high frequencies. The improvement in SNR also offers a better coverage of deep brain areas. This would enhance coverage for deep cortical regions such as cingulate and insula, and also sub-cortical grey matter nuclei including basal ganglia and hippocampi; signals from both of these regions have been measured successfully using conventional SQUID-based MEG systems [26, 60], however efforts to characterise their behaviour have been hampered by low SNR. Our simulations (Fig 5B) show that signals acquired from all deep structures using an OPM system would be enhanced in terms of SNR. However, the degree of improvement is a function of depth. Specifically, the non-linear nature of the inverse square law means that, although moving detectors closer to the scalp surface offers increased SNR for all regions, the improvement for shallow sources is proportionally larger than for deep sources. Indeed, Fig 3 shows that whilst improvements for shallow sources could be as high as tenfold, for deeper sources the gain may be closer to twofold. This observation has important consequences. Even in current systems the effective characterisation of deep sources relies on our ability to eliminate interfering signals for shallow sources. When using OPMs, the greater amplification of shallow compared to deep sources makes this problem harder, since the ratio of deep source (of interest) to shallow source (of no interest) signals is markedly reduced. This observation provides more evidence that realising the benefits of an OPM system depends critically on accurate source modelling. Our simulations also showed that, as a result of improved SNR (Fig 3C) and the more focal nature of the forward field (Fig 3B), beamformer reconstruction accuracy was improved significantly when using OPMs compared to a SQUID-based system. Specifically, temporal correlation between simulated and beamformer-reconstructed cortical sources increased significantly (p < 0.05) from 0.87 for a current SQUID-based system, to 0.98 for a 275-channel OPM system (single dipole model). Improvements were also observed in the presence of interference fields from nearby cortical dipoles of no interest and our results in Fig 5 show that the beamformer algorithm, applied to our sOPM system, was better able to cancel out such interference compared to the sSQUID system. Although beamformer performance is enhanced in an OPM system, our results in Fig 6 show clearly that such improvements would rely heavily on accurate forward modelling. In the case of shallow cortical dipoles, simulations showed that a forward field error of as little as 4–10% would negate the significant advantages afforded by OPM instrumentation. This effect was also seen for deeper dipoles although it is somewhat less pronounced, with OPMs still outperforming SQUIDs up to approximately a 25% error on the forward field. Whilst perhaps counter-intuitive, this observation is well known in beamforming; in short, as SNR increases, inverse modelling becomes increasingly sensitive to the accuracy of the forward model. In the case of OPMs, the increase in SNR afforded by the shift in sensor locations towards the scalp surface drives this necessity for increased forward field accuracy. In the simulations presented, we use a simple mathematical method in order to artificially generate an error on the forward solution. This simulated error is an abstract mathematical formulation, but nevertheless illustrates the problem. Experimentally, these forward field errors come from 1) inaccurate co-registration between the head location and the sensor geometry, 2) subject movement (the forward field is derived based on average head position) and 3) inaccuracy in the forward model itself (e.g. deviation from the dipole model). In recent years, significant improvements in the latter have been introduced via, for example, the use of boundary element methods (BEM) based upon high resolution MR images. However, the co-registration problem is still heavily reliant on techniques such as surface digitisation and matching; which often leads to significant systematic errors in estimating MEG sensor locations relative to the brain anatomy. In fact, previous work suggests that the largest source of forward modelling error in MEG comes from co-registration [61]. Also, for OPMs, as a consequence of subject movement we might expect independent errors if the sensors are not mounted rigidly with respect to one another. In addition, the OPM’s aperture (sensitive volume to magnetic field) will have a smaller area than that of the SQUID coils, and also have a variable gain (in terms of magnitude and directional sensitivity) which will depend not only on fabrication but also on the proximity (and cross-talk) of other sensors. These errors could also impact significantly on forward field accuracy. Here we demonstrate that an abstract modelling error causes a significant problem that could easily negate any advantages afforded by an OPM system, particularly in modelling shallow cortical dipoles. Future work should include accurate characterisation (in terms of systematic, independent, gain and orientation errors) given a specific OPM sensor type. It is noteworthy that a potential solution to co-registration and subject movement problems is the use of 3D-printed head casts which have been demonstrated in recent work [62] and their use in current helmet-based (SQUID) systems facilitates accurate placement of the head relative to the scanner. If OPM-based systems could be fabricated, with bespoke head casts to hold the individual sensors in accurately known locations relative to the subject’s anatomical MRI, this might represent one way to ensure that forward modelling is not impacted by systematic error in sensor locations. This methodology, coupled with accurate (e.g. BEM) models, will likely allow forward fields to be derived with a sufficient accuracy to realise the potential of OPM-based instrumentation. However this may come at the expense of subject comfort. Assuming that solutions to the forward field problem can be found, our spatial resolution measurements in Figs 7 and 8 show the clear advantages of an OPM system in separating dipoles in close proximity in the brain. This is important for all studies, but would be particularly beneficial in the study of functional connectivity between brain regions. The study of connectivity (functional and effective) has caused a paradigm shift in the way in which neuroimaging experiments are conducted and, particularly given the recent advances in dynamic connectivity (temporal changes in connectivity between regions over time) MEG now offers the most attractive means to characterise the way in which the brain forms and dissolves a hierarchy of electrophysiological networks of communication, in order to support current processing demand [21, 22]. Although new MEG methodologies allow for unique insights into such network formation, at present, functional connectivity can only be measured reliably via the use of techniques for post-hoc correction of source leakage between brain regions. This source leakage is a direct result of the limited spatial resolution of the MEG beamformer, and multiple techniques including the use of the imaginary part of the coherence [30], the phase-lag index [32] or linear regression [20, 63] must be employed. Whilst these techniques work, they all eliminate genuine zero-phase lag correlation between electrical signals from separate regions. Invasive measures show clearly that such zero-phase lag effects exist in the brain [64, 65] and are currently missed by MEG. Further, MEG is currently also blind to functional connectivity at high frequency due to poor SNR (see also above) meaning that important electrophysiological effects (for example those observed invasively [66]) cannot be probed. Our results imply that post-hoc correction methods would not be required for OPM systems, for the typical networks observed and high frequency connectivity could be measured. This would prove of key importance in future characterisation of brain network behaviour and dynamics. It is important to point out the limitations of the simulations that have been undertaken. First, these simulations are simple: they are based on small numbers of dipoles. Second, the noise is assumed to be Gaussian-random and therefore uncorrelated across MEG channels. Although these assumptions are unrealistic compared to experimental measurement, it is worth noting that our source reconstruction results are approximately in agreement with what is observed in real data from current SQUID-based systems [12]. Furthermore, simple simulations of this type have been used extensively in a wide variety of published research [54, 67] and these papers provided inspiration to the field of MEG source reconstruction for more than a decade. The quantitative observations made are therefore likely to be representative, and provide a benchmark to which any real system might aspire. There are also many aspects of OPM system design which we have not touched upon. Most importantly, OPM sensors offer a means to measure the magnetic field in two orthogonal orientations perpendicular to the laser beam. This means that in addition to measuring magnetic field in the radial direction (the approach taken by all conventional systems), one additional tangential direction may also be measured using the same sensor. Practically, this requires extra instrumentation (an extra field modulation coil on each sensor is required, as well as an increased number of readout channels). From a modelling point of view, the tangential component of magnetic field is mainly affected by volume currents and so its measurement would necessitate the use of more complex finite element forward models. Third, from a system design point of view, it is not immediately obvious how one might orient the additional axis (i.e. at each measurement location, which tangential field component should be measured). This latter point in particular is the reason that, in the present paper, we limited ourselves to assessing the radial field only. In fact, design of a system which measure vector field components would necessitate significant simulations to identify an optimal sensor arrangement. However, future simulation work should begin to assess this potentially significant advantage of OPMs compared to SQUIDs. There are other significant design problems with OPMs; for example they operate around a zero field resonance meaning that compensation coils are often mounted on the OPM itself to cancel stray Earth’s field. In addition, the operation of a multi-channel array may be complicated by cross-talk between sensors from both the internal field zeroing coils and from field modulation used for lock-in detection. Interfering fields between sensors in close proximity would significantly alter both the sensor gain and potentially the orientation of the sensitive axis of detection. Our paper does not claim to solve these practical issues relating to the practical use of OPMs, rather it was our intention to present a vision of the potential advantages of an OPM system as well as to elucidate the potential modelling problems that the realisation of such a system might generate. Finally, note that, in principle, the theoretical systems described here could result from any technology in which the external surface of the magnetic field sensor is at room temperature. However, recent demonstrations of micro-fabricated, OPM sensors suggest that this technology represents the basis of viable MEG hardware. Ultimately, the potential for OPM sensors to replace SQUIDs as the fundamental building block of MEG devices depends upon more than simple arguments relating to SNR. OPMs have the potential to become relatively inexpensive to construct. Furthermore, because OPMs operate free from cryogenic cooling, there is no reliance on liquid helium which is both costly and in short supply. These factors combined make OPM-MEG potentially cheaper in terms of both installation and running costs compared to current machines. More importantly, the flexibility to place sensors anywhere on the scalp surface not only increases SNR, but further increases the patient populations that can be studied, and the types of paradigms that can be employed. The most obvious potential benefit of this flexibility is in the imaging of infants; however, this flexible sensor placement will also allow better coverage of different regions, for example allowing measurement of electrical activity in the cerebellum, or even the spinal cord. Further, OPM array geometry can be optimised via modelling to specify bespoke sensor array patterns to target specific structures. Such flexible determination of the sensor array is a significant consideration in the argument to replace SQUIDs with OPMs. These arguments coupled with the observations made here suggest that OPMs have a bright future in in the development of a new generation of MEG hardware. Conclusion MEG exhibits great potential as a tool for generating a new understanding of human brain function. However, it is limited by low SNR which is caused by the inherently small magnetic fields generated by the brain and by the distance from the scalp surface to the MEG sensors. In recent years, OPMs have become a viable alternative to superconducting devices for MEG, bringing the advantage that they can be brought to within ~4 mm of the scalp surface, thus offering increased SNR. Here we have quantified the advantages of an OPM system in terms of SNR, beamformer reconstruction accuracy and spatial resolution. Our results show that such an OPM system could offer up to an order of magnitude improvement in SNR for an adult brain. Further, our results show clear improvement in reconstruction accuracy and a marked improvement in spatial resolution. Our results imply that the realisation of a multi-channel whole-head OPM system could generate a step change in the utility of MEG as a means to assess human electrophysiological brain activity and connectivity in health and disease. However such a change is critically dependent, not only on the generation of new multi-channel OPM systems, but also on improved algorithms modelling for MEG data. Supporting Information S1 Data File containing the data used to replicate the figures. (ZIP) Click here for additional data file. S1 Supporting Information Description of the contents in S1_data.zip file. (PDF) Click here for additional data file. ==== Refs References 1 Cohen D . Magnetoencephalography: evidence of magnetic fields produced by alpha-rhythm currents . Science . 1968 ;161 (3843 ):784 –6 . 5663803 2 Cohen D . Magnetoencephalography: detection of the brain's electrical activity with a superconducting magnetometer . Science . 1972 ;175 (4022 ):664 –6 . 5009769 3 Hämäläinen M , Hari R , Ilmoniemi RJ , Knuutila J , Lounasmaa OV . Magnetoencephalography—theory, instrumentation, and applications to noninvasive studies of the working human brain . Reviews of Modern Physics . 1993 ;65 (2 ):413 . 4 Vrba J , Anderson G , Betts K , Burbank M , Cheung T , Cheyne D , et al 151-Channel whole-cortex MEG system for seated or supine positions In: Yoshimoto T , Kotani M , Kuriki S , Karibe H , Nakasato N , editors. Recent Advances in Biomagnetism . Sendai, Japan : Tohoku University Press ; 1999 p. 93 –6 . 5 Hämäläinen MS , Ilmoniemi RJ . Interpreting magnetic fields of the brain: minimum norm estimates . Medical & Biological Engineering & Computing . 1994 ;32 (1 ):35 –42 .8182960 6 Dale AM , Liu AK , Fischl BR , Buckner RL , Belliveau JW , Lewine JD , et al Dynamic statistical parametric mapping: combining fMRI and MEG for high-resolution imaging of cortical activity . Neuron . 2000 ;26 (1 ):55 –67 . 10798392 7 Gross J , Kujala J , Hämäläinen M , Timmermann L , Schnitzler A , Salmelin R . Dynamic imaging of coherent sources: studying neural interactions in the human brain . Proceedings of the National Academy of Sciences of the United States of America . 2001 ;98 (2 ):694 –9 . 11209067 8 Sekihara K , Nagarajan SS , Poeppel D , Marantz A , Miyashita Y . Reconstructing spatio-temporal activities of neural sources using an MEG vector beamformer technique . IEEE Transactions on Biomedical Engineering . 2001 ;48 (7 ):760 –71 . 11442288 9 Van Veen BD , Van Drongelen W , Yuchtman M , Suzuki A . Localization of brain electrical activity via linearly constrained minimum variance spatial filtering . IEEE Transactions on Biomedical Engineering . 1997 ;44 (9 ):867 –80 . 9282479 10 Barnes GR , Furlong PL , Singh KD , Hillebrand A . A verifiable solution to the MEG inverse problem . NeuroImage . 2006 ;31 (2 ):623 –6 . 16480896 11 Robinson SE , Vrba J . Functional neuroimaging by synthetic aperture magnetometry (SAM) . In: Yoshimoto T , Kotani M , Kuriki S , Karibe H , Nakasato N , editors. Recent Advances in Biomagnetism . Sendai, Japan : Tohoku University Press ; 1999 p. 302 –5 . 12 Brookes MJ , Zumer JM , Stevenson CM , Hale JR , Barnes GR , Vrba J , et al Investigating spatial specificity and data averaging in MEG . NeuroImage . 2010 ;49 (1 ):525 –38 . 10.1016/j.neuroimage.2009.07.043 19635575 13 Troebinger L , López JD , Lutti A , Bestmann S , Barnes GR . Discrimination of cortical laminae using MEG . NeuroImage . 2014 ;102 :885 –93 . 10.1016/j.neuroimage.2014.07.015 25038441 14 Hall SD , Holliday IE , Hillebrand A , Singh KD , Furlong PL , Hadjipapas A , et al The missing link: analogous human and primate cortical gamma oscillations . NeuroImage . 2005 ;26 (1 ):13 –7 . 10.1016/j.neuroimage.2005.01.009 .15862200 15 Zumer JM , Brookes MJ , Stevenson CM , Francis ST , Morris PG . Relating BOLD fMRI and neural oscillations through convolution and optimal linear weighting . NeuroImage . 2010 ;49 (2 ):1479 –89 . 10.1016/j.neuroimage.2009.09.020 19778617 16 de Pasquale F , Della Penna S , Snyder AZ , Lewis C , Mantini D , Marzetti L , et al Temporal dynamics of spontaneous MEG activity in brain networks . Proceedings of the National Academy of Sciences of the United States of America . 2010 ;107 (13 ):6040 –5 . 10.1073/pnas.0913863107 20304792 17 Brookes MJ , Woolrich MW , Luckhoo H , Price D , Hale JR , Stephenson MC , et al Investigating the electrophysiological basis of resting state networks using magnetoencephalography . Proceedings of the National Academy of Sciences of the United States of America . 2011 ;108 (40 ):16783 –8 . 10.1073/pnas.1112685108 21930901 18 de Pasquale F , Della Penna S , Snyder AZ , Marzetti L , Pizzella V , Romani GL , et al A cortical core for dynamic integration of functional networks in the resting human brain . Neuron . 2012 ;74 (4 ):753 –64 . 10.1016/j.neuron.2012.03.031 22632732 19 Hillebrand A , Barnes G , Bosboom J , Berendse H , Stam C . Frequency-dependent functional connectivity within resting-state networks: an atlas-based MEG beamformer solution . NeuroImage . 2012 ;59 (4 ):3909 –21 . 10.1016/j.neuroimage.2011.11.005 22122866 20 Hipp JF , Hawellek DJ , Corbetta M , Siegel M , Engel AK . Large-scale cortical correlation structure of spontaneous oscillatory activity . Nature Neuroscience . 2012 ;15 (6 ):884 –90 . 10.1038/nn.3101 22561454 21 Baker AP , Brookes MJ , Rezek IA , Smith SM , Behrens T , Smith PJP , et al Fast transient networks in spontaneous human brain activity . Elife . 2014 ;3 :e01867 10.7554/eLife.01867 24668169 22 O'Neill GC , Bauer M , Woolrich M , Morris PG , Barnes GR , Brookes MJ . Dynamic recruitment of resting state sub-networks . NeuroImage . 2015 ;115 :85 –95 . 10.1016/j.neuroimage.2015.04.030 25899137 23 Engel AK , Gerloff C , Hilgetag CC , Nolte G . Intrinsic coupling modes: multiscale interactions in ongoing brain activity . Neuron . 2013 ;80 (4 ):867 –86 . 10.1016/j.neuron.2013.09.038 24267648 24 Taylor MJ , Donner EJ , Pang EW . fMRI and MEG in the study of typical and atypical cognitive development . Clinical Neurophysiology . 2012 ;42 (1–2 ):19 –25 . 10.1016/j.neucli.2011.08.002 .22200338 25 Brookes MJ , Hall EL , Robson SE , Price D , Palaniyappan L , Liddle EB , et al Complexity measures in magnetoencephalography: measuring "disorder" in schizophrenia . PloS One . 2015 ;10 (4 ):e0120991 10.1371/journal.pone.0120991 25886553 26 Timmermann L , Gross J , Dirks M , Volkmann J , Freund HJ , Schnitzler A . The cerebral oscillatory network of parkinsonian resting tremor . Brain . 2003 ;126 (Pt 1 ):199 –212 . .12477707 27 Mukamel R , Gelbard H , Arieli A , Hasson U , Fried I , Malach R . Coupling between neuronal firing, field potentials, and FMRI in human auditory cortex . Science . 2005 ;309 (5736 ):951 –4 . 10.1126/science.1110913 .16081741 28 Schoffelen JM , Gross J . Source connectivity analysis with MEG and EEG . Human Brain Mapping . 2009 ;30 (6 ):1857 –65 . 10.1002/hbm.20745 .19235884 29 Scholvinck ML , Leopold DA , Brookes MJ , Khader PH . The contribution of electrophysiology to functional connectivity mapping . NeuroImage . 2013 ;80 :297 –306 . 10.1016/j.neuroimage.2013.04.010 23587686 30 Nolte G , Bai O , Wheaton L , Mari Z , Vorbach S , Hallett M . Identifying true brain interaction from EEG data using the imaginary part of coherency . Clinical Neurophysiology . 2004 ;115 (10 ):2292 –307 . 10.1016/j.clinph.2004.04.029 .15351371 31 Ewald A , Marzetti L , Zappasodi F , Meinecke FC , Nolte G . Estimating true brain connectivity from EEG/MEG data invariant to linear and static transformations in sensor space . NeuroImage . 2012 ;60 (1 ):476 –88 . 10.1016/j.neuroimage.2011.11.084 .22178298 32 Stam CJ , Nolte G , Daffertshofer A . Phase lag index: assessment of functional connectivity from multi channel EEG and MEG with diminished bias from common sources . Human Brain Mapping . 2007 ;28 (11 ):1178 –93 . 17266107 33 Colclough GL , Brookes MJ , Smith SM , Woolrich MW . A symmetric multivariate leakage correction for MEG connectomes . NeuroImage . 2015 ;117 :439 –48 . 10.1016/j.neuroimage.2015.03.071 25862259 34 Kominis I , Kornack T , Allred J , Romalis M . A subfemtotesla multichannel atomic magnetometer . Nature . 2003 ;422 (6932 ):596 –9 . 12686995 35 Dang H , Maloof A , Romalis M . Ultrahigh sensitivity magnetic field and magnetization measurements with an atomic magnetometer . Applied Physics Letters . 2010 ;97 (15 ):151110 . 36 Allred J , Lyman R , Kornack T , Romalis M . High-sensitivity atomic magnetometer unaffected by spin-exchange relaxation . Physical Review Letters . 2002 ;89 (13 ):130801 12225013 37 Happer W , Tang H . Spin-exchange shift and narrowing of magnetic resonance lines in optically pumped alkali vapors . Physical Review Letters . 1973 ;31 (5 ):273 –6 . 38 Johnson CN , Schwindt PDD , Weisend M . Magnetoencephalography with a two-color pump-probe, fiber-coupled atomic magnetometer Applied Physics Letters . 2010 ;97 (24 ). 39 Xia H , Ben-Amar Baranga A , Hoffman D , Romalis MV . Magnetoencephalography with an atomic magnetometer . Applied Physics Letters . 2006 ;89 (21 ):Article number 211104. 40 Johnson CN , Schwindt P , Weisend M . Multi-sensor magnetoencephalography with atomic magnetometers . Physics in Medicine and Biology . 2013 ;58 (17 ):6065 –77 . 10.1088/0031-9155/58/17/6065 23939051 41 Sander TH , Preusser J , Mhaskar R , Kitching J , Trahms L , Knappe S . Magnetoencephalography with a chip-scale atomic magnetometer . Biomedical Optics Express . 2012 ;3 (5 ):981 –90 . 10.1364/BOE.3.000981 22567591 42 Kamada K , Sato D , Ito Y , Natsukawa H , Okano K , Mizutani N , et al Human magnetoencephalogram measurements using newly developed compact module of high-sensitivity atomic magnetometer . Japanese Journal of Applied Physics . 2015 ;54 (2 ):Article number 026601. 43 Alem O , Benison AM , Barth DS , Kitching J , Knappe S . Magnetoencephalography of Epilepsy with a Microfabricated Atomic Magnetrode . Journal of Neuroscience . 2014 ;34 (43 ):14324 –7 . 10.1523/JNEUROSCI.3495-14.2014 25339745 44 Schwindt PDD , Knappe S , Shah V , Hollberg L , Kitching J , Liew L-A , et al Chip-scale atomic magnetometer . Applied Physics Letters . 2004 ;85 (26 ):6409 –11 . 45 Johnson CN , Schwindt PDD , Weisend M . Multi-sensor magnetoencephalography with atomic magnetometers . Physics in Medicine and Biology . 2013 ;58 (17 ):6065 –77 .table 10.1088/0031-9155/58/17/6065 23939051 46 Kim K , Begus S , Xia Hc , Lee S-K , Jazbinsek V , Trontelj Z , et al Multi-channel atomic magnetometer for magnetoencephalography: A configuration study . NeuroImage . 2014 ;89 :143 –51 . 10.1016/j.neuroimage.2013.10.040 24185014 47 Koga R, Hiyama E, Matsumoto T, Sekihara K, editors. Quantitative performance assessments for neuromagnetic imaging systems. 35th Annual International Conference of the IEEE EMBS; 2013 3–7 July 2013; Osaka, Japan: IEEE. 48 Taylor J , Cappellaro P , Childress L , Jiang L , Budker D , Hemmer P , et al High-sensitivity diamond magnetometer with nanoscale resolution . Nature Physics . 2008 ;4 (10 ):810 –6 . 49 Sander T , Preusser J , Mhaskar R , Kitching J , Trahms L , Knappe S . Magnetoencephalography with a chip-scale atomic magnetometer . Biomedical Optics Express . 2012 ;3 (5 ):981 –90 . 10.1364/BOE.3.000981 22567591 50 Fang Q, Boas DA, editors. Tetrahedral mesh generation from volumetric binary and grayscale images. Biomedical Imaging: From Nano to Macro, 2009 ISBI'09 IEEE International Symposium on; 2009: IEEE. 51 Fischl B , Liu A , Dale AM . Automated manifold surgery: constructing geometrically accurate and topologically correct models of the human cerebral cortex . IEEE Transactions on Medical Imaging . 2001 ;20 (1 ):70 –80 . 10.1109/42.906426 .11293693 52 Huang M , Mosher J , Leahy R . A sensor-weighted overlapping-sphere head model and exhaustive head model comparison for MEG . Physics in Medicine and Biology . 1999 ;44 (2 ):423 10070792 53 Sarvas J . Basic mathematical and electromagnetic concepts of the biomagnetic inverse problem . Physics in Medicine and Biology . 1987 ;32 (1 ):11 –22 . 3823129 54 Vrba J . Magnetoencephalography: the art of finding a needle in a haystack . Physica C-Superconductivity and Its Applications . 2002 ;368 (1 ):1 –9 . 55 Sekihara K , Hild KE , Nagarajan SS . A novel adaptive beamformer for MEG source reconstruction effective when large background brain activities exist . IEEE Transactions on Biomedical Engineering . 2006 ;53 (9 ):1755 –64 . 16941831 56 Sekihara K , Nagarajan SS . Adaptive spatial filters for electromagnetic brain imaging . Nagel JH , editor: Springer ; 2008 . 57 Oostenveld R , Praamstra P . The five percent electrode system for high-resolution EEG and ERP measurements . Clinical Neurophysiology . 2001 ;112 (4 ):713 –9 . 11275545 58 Giacometti P , Perdue KL , Diamond SG . Algorithm to find high density EEG scalp coordinates and analysis of their correspondence to structural and functional regions of the brain . Journal of Neuroscience Methods . 2014 ;229 :84 –96 . 10.1016/j.jneumeth.2014.04.020 24769168 59 Crone NE , Sinai A , Korzeniewska A . High-frequency gamma oscillations and human brain mapping with electrocorticography . Progress in Brain Research . 2006 ;159 :275 –95 . 17071238 60 Luckhoo H , Hale JR , Stokes MG , Nobre AC , Morris PG , Brookes MJ , et al Inferring task-related networks using independent component analysis in magnetoencephalography . NeuroImage . 2012 ;62 (1 ):530 –41 . 10.1016/j.neuroimage.2012.04.046 22569064 61 Hillebrand A , Barnes G . The use of anatomical constraints with MEG beamformers . NeuroImage . 2003 ;20 (4 ):2302 –13 . 14683731 62 Troebinger L , López JD , Lutti A , Bradbury D , Bestmann S , Barnes GR . High precision anatomy for MEG . NeuroImage . 2014 ;86 :583 –91 . 10.1016/j.neuroimage.2013.07.065 23911673 63 Brookes MJ , Woolrich MW , Barnes GR . Measuring functional connectivity in MEG: a multivariate approach insensitive to linear source leakage . NeuroImage . 2012 ;63 (2 ):910 –20 . 10.1016/j.neuroimage.2012.03.048 22484306 64 Singer W . Neuronal synchrony: a versatile code for the definition of relations? Neuron . 1999 ;24 (1 ):49 –65 , 111 –25 . .10677026 65 Engel AK , Fries P , Singer W . Dynamic predictions: oscillations and synchrony in top-down processing . Nature Reviews: Neuroscience . 2001 ;2 (10 ):704 –16 . 10.1038/35094565 .11584308 66 Ossandón T , Jerbi K , Vidal JR , Bayle DJ , Henaff M-A , Jung J , et al Transient suppression of broadband gamma power in the default-mode network is correlated with task complexity and subject performance . Journal of Neuroscience . 2011 ;31 (41 ):14521 –30 . 10.1523/JNEUROSCI.2483-11.2011 21994368 67 Barnes GR , Hillebrand A . Statistical flattening of MEG beamformer images . Human Brain Mapping . 2003 ;18 (1 ):1 –12 . 10.1002/hbm.10072 .12454907

PMC005xxxxxx/PMC5001649.txt

==== Front PLoS Negl Trop DisPLoS Negl Trop DisplosplosntdsPLoS Neglected Tropical Diseases1935-27271935-2735Public Library of Science San Francisco, CA USA 2756486310.1371/journal.pntd.0004961PNTD-D-16-00236Research ArticleMedicine and Health SciencesTropical DiseasesNeglected Tropical DiseasesDengue FeverMedicine and Health SciencesInfectious DiseasesViral DiseasesDengue FeverBiology and Life SciencesBiochemistryBiomarkersMedicine and Health SciencesDiagnostic MedicineSigns and SymptomsHemorrhageMedicine and Health SciencesPathology and Laboratory MedicineSigns and SymptomsHemorrhageMedicine and Health SciencesVascular MedicineHemorrhageResearch and Analysis MethodsMathematical and Statistical TechniquesStatistical MethodsRegression AnalysisPhysical SciencesMathematicsStatistics (Mathematics)Statistical MethodsRegression AnalysisMedicine and Health SciencesDiagnostic MedicineSigns and SymptomsFeversMedicine and Health SciencesPathology and Laboratory MedicineSigns and SymptomsFeversMedicine and Health SciencesInfectious DiseasesBacterial DiseasesMedicine and Health SciencesDiagnostic MedicineClinical Laboratory SciencesClinical LaboratoriesBiology and Life SciencesAnatomyBody FluidsBloodBlood PlasmaMedicine and Health SciencesAnatomyBody FluidsBloodBlood PlasmaBiology and Life SciencesPhysiologyBody FluidsBloodBlood PlasmaMedicine and Health SciencesPhysiologyBody FluidsBloodBlood PlasmaMedicine and Health SciencesHematologyBloodBlood PlasmaSerum Procalcitonin and Peripheral Venous Lactate for Predicting Dengue Shock and/or Organ Failure: A Prospective Observational Study PCT and PVL to Identify Dengue Shock/Organ FailureThanachartwet Vipa 1Desakorn Varunee 1Sahassananda Duangjai 2Jittmittraphap Akanitt 3Oer-areemitr Nittha 4Osothsomboon Sathaporn 5Surabotsophon Manoon 6Wattanathum Anan 4*1 Department of Clinical Tropical Medicine, Faculty of Tropical Medicine, Mahidol University, Bangkok, Thailand2 Information Technology Unit, Faculty of Tropical Medicine, Mahidol University, Bangkok, Thailand3 Department of Microbiology and Immunology, Faculty of Tropical Medicine, Mahidol University, Bangkok, Thailand4 Pulmonary and Critical Care Division, Department of Medicine, Phramongkutklao Hospital, Bangkok, Thailand5 Hospital for Tropical Diseases, Faculty of Tropical Medicine, Mahidol University, Bangkok, Thailand6 Pulmonary and Critical Care Division, Department of Medicine, Ramkhamhaeng Hospital, Bangkok, ThailandMesser William B EditorOregon Health and Science University, UNITED STATESAll authors have declared that no competing interests exist in this study. Roche Diagnostics (Thailand) Ltd. provided Elecsys BRAHMS PCT as a gift for this study. All authors have declared that the Roche Diagnostics (Thailand) Ltd. had no role in study design, data collection and/or analysis, the decision to publish, or preparation of the manuscript. Conceived and designed the experiments: VT VD DS NOa AW. Performed the experiments: VT VD DS AJ NOa SO MS. Analyzed the data: VT VD DS AW. Wrote the paper: VT VD DS MS AW. Literature review: VT VD AW. * E-mail: wattanathum_anan@hotmail.com26 8 2016 8 2016 10 8 e000496112 2 2016 8 8 2016 © 2016 Thanachartwet et al2016Thanachartwet et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Background Currently, there are no biomarkers that can predict the incidence of dengue shock and/or organ failure, although the early identification of risk factors is important in determining appropriate management to reduce mortality. Therefore, we sought to determine the factors associated with dengue shock and/or organ failure and to evaluate the prognostic value of serum procalcitonin (PCT) and peripheral venous lactate (PVL) levels as biomarkers of dengue shock and/or organ failure. Methodology/Principal Findings A prospective observational study was conducted among adults hospitalized for confirmed viral dengue infection at the Hospital for Tropical Diseases in Bangkok, Thailand between October 2013 and July 2015. Data, including baseline characteristics, clinical parameters, laboratory findings, serum PCT and PVL levels, management, and outcomes, were recorded on pre-defined case report forms. Of 160 patients with dengue, 128 (80.0%) patients had dengue without shock or organ failure, whereas 32 (20.0%) patients developed dengue with shock and/or organ failure. Using a stepwise multivariate logistic regression analysis, PCT ≥0.7 ng/mL (odds ratio [OR]: 4.80; 95% confidence interval [CI]: 1.60–14.45; p = 0.005) and PVL ≥2.5 mmol/L (OR: 27.99, 95% CI: 8.47–92.53; p <0.001) were independently associated with dengue shock and/or organ failure. A combination of PCT ≥0.7 ng/mL and PVL ≥2.5 mmol/L provided good prognostic value for predicting dengue shock and/or organ failure, with an area under the receiver operating characteristics curve of 0.83 (95% CI: 0.74–0.92), a sensitivity of 81.2% (95% CI: 63.6–92.8%), and a specificity of 84.4% (95% CI: 76.9–90.2%). Dengue shock patients with non-clearance of PCT and PVL expired during hospitalization. Conclusions/Significance PCT ≥0.7 ng/mL and PVL ≥2.5 mmol/L were independently associated with dengue shock and/or organ failure. The combination of PCT and PVL levels could be used as prognostic biomarkers for the prediction of dengue shock and/or organ failure. Author Summary Dengue is a major global health concern, particularly in tropical countries, and affects all age groups. Mortality rates among patients who have been hospitalized with severe dengue are 1.6–10.9%, and death in adults is mainly due to the development of dengue shock and organ dysfunction. In states of poor tissue circulation or shock, lactate is produced. Additionally, procalcitonin is a highly specific biomarker of systemic inflammation. Therefore, we assessed whether procalcitonin and peripheral venous lactate could be used to predict the incidence of dengue shock and/or organ failure in patients with dengue. Our study showed that a combination of serum procalcitonin levels ≥0.7 ng/mL and peripheral venous lactate levels ≥2.5 mmol/L at admission could discriminate between patients who did and did not develop shock and/or organ failure, with high sensitivity and specificity. These parameters may therefore be useful as prognostic biomarkers. Our results suggest that serum procalcitonin is indicative of an extensive early inflammatory response, which may occur during the systemic phase of dengue. Peripheral venous lactate may be produced as a result of the poor tissue circulation that precedes dengue shock. Our findings may help clinicians to predict dengue shock and/or organ failure earlier to reduce in-hospital mortality. http://dx.doi.org/10.13039/501100004156Mahidol University2013the Royal College of Physicians of ThailandRoche Diagnostics (Thailand)This study was supported by the Dean's Fund Research 2013, Faculty of Tropical Medicine, Mahidol University; Research Grant 2013, the Royal College of Physicians of Thailand; and Roche Diagnostics (Thailand) Ltd. for providing Elecsys BRAHMS PCT for this study. The funders had no role in study design, data collection and/or analysis, the decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction Dengue is the most important arthropod-borne viral disease, and it exerts a high burden on populations and public health systems in most tropical countries [1,2]. The incidence has dramatically increased during the last 50 years (by 30-fold) for all four dengue virus serotypes (DENV 1–4) in more than 100 countries, including those in Southeast Asia, Central and South America, the Western Pacific, Africa, and the Eastern Mediterranean [2,3]. A previous report estimated that 390 million people are infected with DENV per year worldwide, of which 96 million show clinical manifestations of dengue [4]. Clinical manifestations range from acute febrile illness to severe dengue, which is a life-threatening condition [1]. In-hospital mortality is observed among 1.6–10.9% of patients with severe manifestations of dengue, including dengue hemorrhagic fever and/or dengue shock syndrome [5–7]. The World Health Organization (WHO) has implemented a goal of reducing dengue mortality by at least 20% and morbidity by 25% by the year 2020 [2]. Early recognition of severe dengue would help clinicians achieve close monitoring and provide proper fluid resuscitation in order to prevent severe disease, which would reduce mortality and morbidity. The revised 2009 WHO case definition was introduced in order to improve early recognition of severe dengue by increasing awareness of warning signs [1]. However, a recent systematic review showed that the definition had a wide range of sensitivity (59–98%) and specificity (41–99%) in the prediction of severe dengue [8]. The pathophysiology of severe dengue is complex, and involves an interplay of host immune and genetic factors with virulent strains of DENV [9,10]. The critical phase of severe dengue usually occurs as viremia declines [1]. DENV replication occurs within cells, particularly hepatocytes, monocytes, and macrophages, during systemic infection and the immune-mediated response following DENV infection, which is proportional to the viral load [11,12]. Immune-mediated pathogenesis has been considered a major cause of the increased vascular permeability of endothelial cells, leading to plasma leakage [9–11]. Delayed recognition and improper management of patients with plasma leakage can lead to shock and/or organ failure [11]. The prevalence of dengue shock among adults is approximately 18%, and it is the most common cause of death from DENV [13]. A previous systematic review and meta-analysis showed that several clinical factors, including age, female sex, neurological signs, nausea/vomiting, abdominal pain, gastrointestinal bleeding, hemoconcentration, ascites, pleural effusion, hypoalbuminemia, hypoproteinemia, hepatomegaly, high levels of aspartate aminotransferase (AST) and alanine aminotransferase (ALT), abnormal coagulators, primary/secondary infection, and DENV-2, were independently associated with the development of dengue shock [13]. Procalcitonin (PCT) is a functional immune modulating protein consisting of 114–116 amino acids, and is currently used as a novel biomarker for diagnostic and prognostic purposes [14]. PCT is produced and released into the bloodstream in response to infection and/or inflammation in various tissues. In particular, hepatocytes and peripheral blood mononuclear cells are potent PCT secretors [15]. A recent systematic review and meta-analysis showed that PCT was a useful biomarker for the early diagnosis of sepsis in critically ill patients, with a sensitivity and specificity of 77% and 79%, respectively [16]. The area under the receiver operating characteristics curve (AUROC) was 0.85, indicating moderate diagnostic accuracy [16]. In Southeast Asia, an endemic area for tropical infectious diseases, the AUROC for discrimination between bacterial and viral infections using PCT was 0.74, which was also indicative of moderate diagnostic accuracy [17]. Of the patients with dengue, 72% had a PCT level ≥0.1 ng/mL and 25% had a PCT level ≥0.5 ng/mL, which was higher than that of patients with influenza (34% at a PCT level ≥0.1 ng/mL and 16% at a PCT level ≥0.5 ng/mL [17]. Previous reports have also shown that PCT levels in patients with sepsis are associated with the severity of organ dysfunction [18], and that PCT could be used as a prognostic marker for discrimination between patients with and without septic shock, in addition to survival [19]. A previous study showed that PCT levels on admission were significantly higher among patients who died following infection with the 2009 H1N1 strain of influenza, compared with those who survived (14.5 vs. 1.7 ng/mL) [20]. In addition, arterial or venous lactate may be used as a biomarker for tissue hypoperfusion, regardless of organ failure or shock, particularly among patients with sepsis [21]. Our previous prospective study showed that peripheral venous lactate (PVL) concentration was independently associated with severe dengue [22]. In clinical practice, it can be difficult to identify the early stages of dengue shock and/or organ failure using clinical data. PCT and/or PVL may provide a superior prognostic method for predicting dengue severity at the time of hospital admission, particularly in the identification of patients at high risk of developing dengue shock and/or organ failure. At present, there have been no studies assessing the capacity of PCT and/or PVL to predict dengue shock and/or organ failure. Thus, we hypothesized that PCT and/or PVL may discriminate between patients who develop dengue shock and/or organ failure and those who do not. Therefore, we undertook a prospective observational study among hospitalized adults with dengue and determined the factors associated with dengue shock and/or organ failure. The prognostic values of PCT and PVL as biomarkers for predicting dengue shock and/or organ failure were evaluated. Methods Ethical considerations The study design was approved by the Ethics Committee of the Faculty of Tropical Medicine, Mahidol University in Bangkok, Thailand. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement (S1 Checklist) and the Standards for the Reporting of Diagnostic (STARD) accuracy (S2 Checklist) were followed in this study [23,24]. Patients aged ≥15 years with clinical dengue, defined as acute fever and ≥2 of the following symptoms were included: 1) headache, 2) ocular pain, 3) myalgia, 4) arthralgia, 5) rash, 6) a positive tourniquet test (≥20 petechiae per square inch), or 7) leukopenia (white blood cell [WBC] counts <5.0 × 103 cells/μL). Patients had been admitted to hospital for treatment, and the broad criteria allowed physicians to invite all potential patients to participate in the study at the outpatient and emergency department. Written informed consent was obtained from all patients, or the patient's guardians if the patient was 15–18 years old, before participation in the study. Study design and population This prospective observational study was performed among patients who were admitted to the Hospital for Tropical Diseases (Faculty of Tropical Medicine, Mahidol University in Bangkok, Thailand) between October 2013 and July 2015. The inclusion criteria were (i) age ≥15 years, (ii) clinical dengue, and (iii) confirmed dengue viral infection by reverse-transcriptase polymerase chain reaction (RT-PCR) from a serum sample obtained at admission, and/or positive micro-neutralization test results from serum samples obtained at admission and 2 weeks after admission, and/or dengue-specific immunoglobulin M (IgM) and immunoglobulin G (IgG) detected using enzyme-linked immunosorbent assays (ELISAs) in paired serum samples taken at admission and 2 weeks after admission. Patients with an underlying medical illness, mixed infection, current pregnancy, current use of any non-topical antibiotic, or current fluid therapy were excluded from this study. Laboratory tests were conducted at admission, including a complete blood count and blood chemistry assessment, and samples for the measurement of PCT and PVL were collected. Blood samples for PCT and PVL analysis were collected every 24 h until the patient exhibited a body temperature of <37.8°C for 48 h. Treating physicians and investigators were blinded to the PCT and PVL results. All patients received standard management from their treating physicians, according to the 2009 WHO guidelines for dengue [1]. In order to exclude other infections, two blood samples for microbiological cultures were obtained, urinalysis was performed, and plain radiography of the chest was routinely performed at admission. Diagnostic tests for other infectious diseases were also performed when indicated by clinical findings at admission or during hospitalization. Dengue severity and outcomes were summarized on discharge. All patient data, including baseline characteristics, clinical parameters, laboratory findings, management, and outcomes, were recorded on a pre-defined case report form. At a 2-week follow-up appointment, blood samples were collected for complete blood counts and serum creatinine assessment. Subsequent follow-up was required within the following 2 months until the laboratory results reached reference ranges in order to serve as a baseline. Case definitions for dengue The WHO 2009 dengue definition was used to classify dengue shock and organ failure in this study [1]. Dengue shock was defined as plasma leakage with shock. Plasma leakage was defined as ≥20% increase in hematocrit above baseline or clinical fluid accumulation manifested by pleural effusion, ascites, or serum albumin <3.5 g/dL. Shock was defined as (1) a rapidly weak pulse with pulse pressure <20 mmHg, or (2) a systolic blood pressure of <90 mmHg with tissue hypoperfusion evidenced by one of the following criteria: (i) decreased urine output (<0.5 mL/kg/h), (ii) impaired consciousness, (iii) AST >1000 IU/L, (iv) ALT >1000 IU/L, (v) cold skin, or (vi) clammy skin. Organ failure was defined as the presence of one of the following criteria: (i) respiratory distress (a respiratory rate of ≥24 breaths/min with <95% oxygen saturation in room air and/or the need for oxygen therapy), (ii) serum creatinine increased ≥3-fold from baseline, (iii) AST >1000 IU/L, (iv) ALT >1000 IU/L, (v) myocarditis, (vi) encephalitis, or (vii) spontaneous gastrointestinal bleeding requiring blood transfusion. The WHO 2009 dengue definitions for warning signs (WSs) were also used in this study; WSs included (1) abdominal pain; (2) vomiting; (3) clinical fluid accumulation defined as the presence of pleural effusion determined by plain radiography of the chest or a serum albumin level <3.5 g/dL; (4) lethargy; (5) a liver span of >15 cm; (6) bleeding from a mucosal area, including the nose, gums, gastrointestinal tract, or vagina; and (7) an increase in hematocrit of 2% above the sex-specific reference range for a healthy Thai adult with a platelets of ≤100 × 103/μL. Reverse-transcriptase polymerase chain reaction Dengue viral RNA was detected from patient serum at admission using a two-step PCR method, as described by Lanciotti et al. [25], and modified using the methods of Reynes et al. [26]. Viral RNA was detected from acute serum samples using a PureLink Viral RNA/DNA Mini Kit (Invitrogen, Grand Island, NY, USA), according to the manufacturer’s instructions. Micro-neutralization test Serum samples collected at admission and 2 weeks after admission were assayed for serotype-specific DENV using the micro-neutralization test described by Vorndam et al. [27], with the slightly modified protocol of Putnak et al. [28]. The micro-neutralization test based on the principle of the plaque reduction neutralization test was used to measure serotype specific anti-DENV neutralizing antibodies against all 4 serotypes. Serum samples were tested in triplicate and sera were serially diluted 2-fold from 1:20 to 1:5120 in a 96-well microplate. Each microplate included media only (negative control), a virus control and sera of known specific DENV serotypes (positive controls). The average number of virus foci were counted, and only assays with a virus control in the range of 50–60 foci per well and a media only control with no foci were included. For control sera of known DENV serotypes, at least 50% inhibition by viral replication was required (25–30 foci per well), compared with the virus control. The virus neutralization titer was defined as the reciprocal of the serum dilution providing 50% inhibition of viral replication compared with the virus control. A positive serotype specific anti-DENV test was defined as a 4-fold rise in neutralizing antibody titer in paired samples for 1 of the 4 DENV serotypes. Dengue viral infection serology All sera collected at admission and 2 weeks after admission were tested using four separate capture ELISA assays for IgM and IgG against dengue virus and Japanese encephalitis virus, as described by Innis et al. [29]. The assay was performed using serum samples diluted 1:100. Assay results for test samples were expressed as units calculated by the following formula: units = 100 × (A492test sample–A492NS)/(A492PS–A492NS), where A492 was an absorbance at 492 nm, NS was a normal human serum negative standard, and PS was pooled sera from flavivirus infected patients. Both acute and convalescent sera were used for the assay. Only sera with either anti-dengue IgM or anti-Japanese encephalitis IgM levels ≥40 units were evaluated. To discriminate between dengue and other flavivirus infections, we determined the ratio of dengue IgM to Japanese encephalitis virus IgM, with a ratio ≥1.0 indicating dengue virus infection and a ratio <1.0 indicating other flavivirus infection. To discriminate primary from secondary dengue infection, the ratio of anti-dengue IgM to anti-dengue IgG was also calculated, with a ratio ≥1.8 indicating primary dengue infection and a ratio <1.8 indicating secondary dengue infection. This cut-off value was applied for either acute or convalescent samples as long as either an anti-dengue IgM or anti-dengue IgG response could be detected. Pooling the results for both acute and convalescent sera using the same cut-off value allowed more accurate classification of primary and secondary dengue. Measurement of serum procalcitonin PCT was measured using an electrochemiluminescence method (Elecsys BRAHMS PCT, Roche Diagnostic, Mannheim, Germany) according to the manufacturer’s instructions using a Cobas e 411 immunoassay analyzer (Roche Diagnostic, Mannheim, Germany). Prior to assessment, frozen serum samples were stored at –80 °C by laboratory personnel blinded to patient status. The detection limit for the PCT assay was 0.02 ng/mL. The coefficients of variation for low and high concentrations were 1.7% and 1.4%, respectively. Measurement of peripheral venous lactate Blood samples were collected from a vein in an upper extremity without the use of a tourniquet. A 2 mL blood sample was collected into a vacutainer tube containing sodium fluoride, immediately placed on ice, sent to the laboratory, and analyzed for lactate within 10 min. Lactate levels were measured by a colorimetric assay with an enzymatic reaction using an auto-analyzer (Roche/Hitachi Cobas C Systems, USA), according to the manufacturer’s protocol. The laboratory personnel were blinded to the sample sources. The coefficient of variation for the assay in our laboratory was 1.1%. Sample size calculation A previous prospective study at the Hospital for Tropical Diseases (Bangkok, Thailand) indicated that the incidence of dengue shock and/or organ failure was 21.0% among hospitalized adults with dengue [30]. Based on this information, we calculated that a sample size of at least 122 patients was needed for this study, using a specificity of 90% with a confidence interval (CI) of ±6%. Statistical analysis All data were analyzed using SPSS software (version 18.0; SPSS Inc., Chicago, IL). Numerical variables were tested for normality using Kolmogorov-Smirnov tests. Variables with non-normal distribution were summarized as medians and interquartile ranges (IQRs), and were compared using Mann-Whitney U tests for two-group comparisons. Categorical variables were expressed as frequencies and percentages, and were analyzed using chi-squared or Fisher’s exact tests, as appropriate. A univariate logistic regression analysis was performed with each potential factor included as an independent variable, and the presence or absence of dengue shock and/or organ failure as the dependent variable. Any variable with a p-value ≤0.2 was considered potentially significant and was further analyzed in a stepwise multivariate logistic regression analysis using a backward selection method for determining significant independent factors. The optimal cut-off values of factors predictive of dengue shock and/or organ failure were determined using ROC curves. Prognostic parameters were evaluated using 2 × 2 tables, and 95% CIs were calculated to determine sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), positive likelihood ratio (LR+), and negative likelihood ratio (LR–). The optimal PCT and PVL cut-off values were then combined in a single “bioscore”, as described by Gibot et al, 2012 [31]. The bioscore attributed one point per biomarker with a value above or equal to the optimal cut-off value. The bioscore was defined as 0 (both biomarkers below their respective cut-off value), 1 (any one of the two biomarkers above/equal to the cut-off value), or 2 (both biomarkers above/equal to the cut-off value). The bioscore was then further tested for prognostic value in predicting dengue shock and/or organ failure by logistic regression analysis. All tests of significance were two-sided, with a p-value <0.05 indicating statistical significance. Results Study population A total of 189 adults with suspected dengue were admitted to the Hospital for Tropical Diseases (Bangkok, Thailand) between October 2013 and July 2015. Of 189 hospitalized adults with suspected dengue viral infection, 29 patients were excluded due to an underlying illness (17 patients, 58.6%), mixed infection (10 patients, 34.5%), or a negative RT-PCR/micro-neutralization/ELISA for dengue (2 patients, 6.9%). Thus, 160 hospitalized adults with confirmed dengue viral infection were finally recruited for this study. Among the 160 patients, 32 (20.0%) patients had dengue shock (23 patients [71.9%]) and/or organ failure (26 patients [81.2%]), whereas 128 (80.0%) patients had dengue without shock or organ failure (Fig 1). In the 26 patients with organ failure, respiratory distress (11 patients [42.3%]), AST levels >1000 IU/L and/or ALT >1000 IU/L (9 patients [34.6%]), serum creatinine concentration ≥3-fold greater than baseline (6 patients [23.1%]), spontaneous gastrointestinal bleeding requiring blood transfusion (6 patients [23.1%]), myocarditis (4 patients [15.4%]), and encephalitis (3 patients [11.5%]) were observed. 10.1371/journal.pntd.0004961.g001Fig 1 Flow diagram showing the recruitment of study patients. ELISA, enzyme-linked immunosorbent assay; RT-PCR, reverse-transcriptase polymerase chain reaction. Comparison of baseline characteristics and clinical and laboratory parameters between dengue patients with and without shock and/or organ failure At admission, patients with dengue shock and/or organ failure were significantly more likely to have a longer duration of fever (p = 0.031), skin bleeding (p = 0.012), mucosal bleeding (p <0.001), vomiting (p = 0.024), a liver span of >15 cm (p = 0.001), decreased breathing sounds (p <0.001), and increased respiratory rate (p = 0.010). When numerical parameters were categorized, patients aged >40 years (p = 0.023), with a fever duration ≥5 days (p = 0.041), respiratory rate ≥24 breaths/min (p = 0.005), mean arterial pressure <70 mmHg (p = 0.030), or pulse pressure <30 mmHg (p = 0.005) were more likely to have dengue shock and/or organ failure (Table 1 and S1 Table). 10.1371/journal.pntd.0004961.t001Table 1 Baseline characteristics and clinical parameters at admission among 160 hospitalized adults with dengue. Characteristic With dengue shock and/or organ failure (n = 32) n (%) No dengue shock or organ failure (n = 128) n (%) p-value Age >40 years 11 (34.4) 19 (14.8) 0.023 Fever ≥5 days 19 (59.4) 48 (37.5) 0.041 Skin bleeding 25 (78.1) 66 (51.6) 0.012 Mucosal bleeding 25 (78.1) 52 (40.6) <0.001 Vomiting 22 (68.8) 57 (44.5) 0.024 Liver span >15 cm 22 (68.8) 44 (34.4) 0.001 Respiratory rate ≥24 breaths/min 11 (34.4) 15 (11.7) 0.005 Pulse pressure <30 mmHg 11 (34.4) 15 (11.7) 0.005 Mean arterial pressure <70 mmHg 4 (12.5) 3 (2.3) 0.030 Regarding laboratory parameters (Table 2 and S2 Table), patients with dengue shock and/or organ failure had significantly higher hemoglobin concentrations (p = 0.045), increased hematocrit values above baseline (p <0.001), higher WBC counts (p = 0.044), higher absolute bands (p = 0.022), higher absolute atypical lymphocyte counts (p = 0.007), higher AST levels (p <0.001), higher ALT levels (p <0.001), higher PCT levels (p = 0.001), and higher PVL levels (p <0.001) (Fig 2). However, patients with dengue shock and/or organ failure had significantly lower platelet counts (p <0.001) and albumin levels (p <0.001). When laboratory parameters were categorized based on the reference ranges (Table 2), patients with WBC counts >5.0 × 103 cells/μL (p = 0.004), absolute bands >200 cells/μL (p = 0.049), absolute atypical lymphocyte counts >300 cells/μL (p = 0.006), AST >120 IU/L (p = 0.002), ALT >120 IU/L (p = 0.002), PCT ≥0.7 ng/mL (p = 0.002), and PVL ≥2.5 mmol/L (p <0.001) were more likely to have dengue shock and/or organ failure. In addition, patients with platelet counts <50.0 × 103 cells/μL (p = 0.012) and albumin <3.5 g/dL (p = 0.001) were also more likely to have dengue shock and/or organ failure. 10.1371/journal.pntd.0004961.g002Fig 2 Serum procalcitonin and peripheral venous lactate at admission in dengue patients. (A) Serum procalcitonin levels among patients with and without dengue shock and/or organ failure. (B) Peripheral venous lactate levels among patients with and without dengue shock and/or organ failure. Data are presented as box and whisker plots with median (horizontal line), interquartile range (box), maximum value within 1.5 of interquartile range (whiskers), outliers (circles), and extreme outliers (asterisks). 10.1371/journal.pntd.0004961.t002Table 2 Laboratory parameters categorized based on reference ranges among 160 hospitalized adults with dengue. Characteristic With dengue shock and/or organ failure No dengue shock or organ failure p-value (n = 32) n (%) (n = 128) n (%) WBC >5.0 × 103 cells/μL 15 (46.9) 26 (20.3) 0.004 Absolute bands >200 cells/μL 16 (50.0) 38 (29.7) 0.049 Absolute atypical LYM >300 cells/μL 17 (53.1) 33 (25.8) 0.006 Platelet counts <50.0 × 103/μL 14 (43.8) 26 (20.3) 0.012 Albumin <3.5 g/dL 8 (25.0) 5 (3.9) 0.001 AST >120 IU/L 21 (65.6) 43 (33.6) 0.002 ALT >120 IU/L 15 (46.9) 24 (18.8) 0.002 Procalcitonin ≥0.7 ng/mL 13 (40.6) 18 (14.1) 0.002 Lactate ≥2.5 mmol/L 18 (56.3) 5 (3.9) <0.001 ALT, alanine aminotransferase; AST, aspartate aminotransferase; LYM, lymphocytes; WBC, white blood cell counts. Assessment of patient management and outcomes during hospitalization demonstrated that a significant proportion of patients with dengue shock and/or organ failure received albumin as fluid resuscitation (p <0.001) and antibiotics (p = 0.017). Of the 32 patients with dengue shock and/or organ failure, 4 (12.5%) received mechanical ventilation, 3 (9.4%) received renal replacement therapy, and 2 (6.2%) received vasopressors. Patients with dengue shock and/or organ failure had significantly longer durations of hospitalization (p = 0.006). However, only two patients expired during hospitalization, both due to multi-organ failure (S2 Table). Univariate and multivariate analyses for the prediction of dengue shock and/or organ failure A univariate logistic regression analysis was used to determine which of the baseline characteristics, clinical parameters, and laboratory findings were associated with the occurrence of dengue shock and/or organ failure. All clinical factors potentially associated with the occurrence of dengue shock and/or organ failure were included in the univariate logistic regression analysis. The following variables were identified as clinical parameters associated with dengue shock and/or organ failure: (1) age >40 years, (2) fever duration ≥5 days, (3) absolute bands >200 cells/μL, (4) absolute atypical lymphocyte counts >300 cells/μL, (5) PCT ≥0.7 ng/mL, and (6) PVL ≥2.5 mmol/L (Table 3). 10.1371/journal.pntd.0004961.t003Table 3 Univariate analysis for the prediction of dengue shock and/or organ failure using clinical and laboratory parameters. Characteristic n Odds ratio (95% CI) p-value Age >40 years 160 3.00 (1.25–7.22) 0.014 Fever ≥5 days 160 2.44 (1.10–5.37) 0.027 Absolute bands >200 cells/μL 160 2.37 (1.08–5.22) 0.032 Absolute atypical LYM >300 cells/μL 160 3.26 (1.47–7.26) 0.004 Procalcitonin ≥0.7 ng/mL 160 4.15 (1.76–9.92) 0.001 Lactate ≥2.5 mmol/L 160 31.63 (10.17–98.36) <0.001 CI, confidence interval; LYM, lymphocytes. All parameters with a p-value ≤0.2 in the univariate logistic regression analysis were then further analyzed by a stepwise multivariate logistic regression analysis using a backward selection method, in order to determine the independent factors significantly associated with the occurrence of dengue shock and/or organ failure. The following clinical and laboratory parameters were found to be independently associated with the occurrence of dengue shock and/or organ failure: (1) PCT ≥0.7 ng/mL (odds ratio [OR]: 4.80; 95% CI: 1.60–14.45; p = 0.005) and (2) PVL ≥2.5 mmol/L (OR: 27.99, 95% CI: 8.47–92.53; p <0.001) (Table 4). The two biomarkers PCT ≥0.7 ng/mL and PVL ≥2.5 mmol/L were assessed as a combined bioscore using a logistic regression model to evaluate the prognostic capacity in predicting the occurrence of dengue shock and/or organ failure. Higher bioscores were associated with increased occurrence of dengue shock and/or organ failure, with ORs of 22.23 (95% CI 7.85–63.00) and 30.00 (95% CI 5.76–156.31) for a bioscore 1 and 2, respectively (p <0.001) (Table 4). 10.1371/journal.pntd.0004961.t004Table 4 Multivariate logistic regression analysis for the prediction of dengue shock and/or organ failure using laboratory parameters. Characteristic n Odds ratio (95% CI) p-value Procalcitonin ≥0.7 ng/mL 160 4.80 (1.60–14.45) 0.005 Lactate ≥2.5 mmol/L 160 27.99 (8.47–92.53) <0.001 Combined procalcitonin and lactate (bioscore)     0 114 1.00 (Reference)     1 38 22.23 (7.85–63.00) <0.001     2 8 30.00 (5.76–156.31) <0.001 CI, confidence interval. Prognostic value of serum procalcitonin and peripheral venous lactate for predicting dengue shock and/or organ failure The AUROC for PCT in the prediction of dengue shock and/or organ failure was 0.69 (95% CI: 0.59–0.80) (Fig 3A). The AUROC for PVL in the prediction of dengue shock and/or organ failure was 0.78 (95% CI: 0.68–0.88) (Fig 3B). The prognostic values of PCT and PVL at admission for predicting dengue shock and/or organ failure are shown in Table 5. The sensitivities for nearly all PCT and PVL categories were low, except for PVL ≥1.5 mmol/L. The specificities for PCT and PVL categories were high, except for PCT ≥0.5 ng/mL and PVL ≥1.5 mmol/L. The PPVs and LR+ values for PCT and PVL categories were low, except for PVL ≥2.5 mmol/L and ≥3.0 mmol/L. The NPVs for all PCT and PVL categories were high. The LR–values for PCT and PVL were high, indicating possible prediction of dengue shock and/or organ failure. 10.1371/journal.pntd.0004961.g003Fig 3 Receiver operating characteristic curves for serum procalcitonin and peripheral venous lactate in the prediction of dengue shock and/or organ failure at admission. (A) The area under the receiver operating characteristic curve (AUROC) for serum procalcitonin at admission was 0.69 (95% confidence interval [95% CI]: 0.59–0.80). (B) The AUROC for peripheral venous lactate at admission was 0.78 (95% CI: 0.68–0.88). (C) The AUROC for a combined bioscore at admission was 0.83 (95% CI: 0.74–0.92). (D) The AUROC for the number of warning signs at admission was 0.77 (95% CI: 0.68–0.87). 10.1371/journal.pntd.0004961.t005Table 5 The prediction of dengue shock and/or organ failure at admission using serum procalcitonin and peripheral venous lactate. Cut-off value Confirmed dengue viral infection Sensitivity Specificity PPV NPV LR+ LR– With shock and/or organ failure (n = 32) No shock or organ failure (n = 128) (95% CI) (95% CI) (95% CI) (95% CI) (95% CI) (95% CI) PCT (ng/mL) ≥0.5 18 31 56.2 (37.6–73.6) 75.8 (67.4–82.9) 36.7 (23.4–51.7) 87.4 (79.7–92.9) 2.3 (1.2–3.6) 0.6 (0.4–0.9) ≥0.6 14 20 43.8 (26.4–62.3) 84.4 (76.9–90.2) 41.2 (24.6–59.3) 85.7 (78.4–91.3) 2.8 (1.6–4.9) 0.7 (0.5–0.9) ≥0.7 13 18 40.6 (23.7–59.4) 85.9 (78.7–91.4) 41.9 (24.6–60.9) 85.3 (78.0–90.9) 2.9 (1.6–5.3) 0.7 (0.5–0.9) ≥0.8 10 15 31.2 (16.1–50.0) 88.3 (81.4–93.3) 40.0 (21.1–61.3) 83.7 (76.4–89.5) 2.7 (1.3–5.4) 0.8 (0.6–1.0) PVL (mmol/L) ≥1.5 26 57 81.2 (63.6–92.8) 55.5 (46.4–64.2) 31.3 (21.6–42.4) 92.2 (83.8–97.1) 1.8 (1.4–2.4) 0.3 (0.2–0.7) ≥2.0 18 21 56.2 (37.7–73.6) 83.6 (76.0–89.6) 46.2 (30.1–62.8) 88.4 (81.4–93.5) 3.4 (2.1–5.6) 0.5 (0.4–0.8) ≥2.5 18 5 56.2 (37.7–73.6) 96.1 (91.1–98.7) 78.3 (56.3–92.5) 89.8 (83.4–94.3) 14.4 (5.8–35.8) 0.5 (0.3–0.7) ≥3.0 12 2 37.5 (21.1–56.3) 98.4 (94.5–99.8) 85.7 (57.2–98.2) 86.3 (79.6–91.4) 24.0 (5.6–101.9) 0.6 (0.5–0.8) PCT (ng/mL) and/or PVL (mmol/L) PCT ≥0.7 and/or PVL ≥2.5 26 20 81.2 (63.6–92.8) 84.4 (76.9–90.2) 56.5 (41.1–71.1) 94.7 (88.9–98.0) 5.2 (3.4–8.0) 0.2 (0.1–0.5) CI, confidence interval; LR+, positive likelihood ratio; LR-, negative likelihood ratio; NPV, negative predictive value; PCT, procalcitonin; PPV, positive predictive value; PVL, peripheral venous lactate. In order to accurately predict dengue shock and/or organ failure in a greater number of patients, the optimal levels of PCT ≥0.7 ng/mL and PVL ≥2.5 mmol/L were combined as a bioscore. The AUROC for a combined bioscore in the prediction of dengue shock and/or organ failure was 0.83 (95% CI: 0.74–0.92) (Fig 3C). The combined bioscore provided good prognostic value for the prediction of dengue shock and/or organ failure among hospitalized adults with dengue, giving an optimal sensitivity of 81.2% (95% CI: 63.6–92.8%); specificity of 84.4% (95% CI: 76.9–90.2%); PPV of 56.5% (95% CI: 41.1–71.1%); NPV of 94.7% (95% CI: 88.9–98.0%); LR+ of 5.2 (95% CI: 3.4–8.0); and LR–of 0.2 (95% CI: 0.1–0.5) (Table 5 and S3 Table). Diagnostic values of the WHO 2009 warning signs for identifying dengue shock and/or organ failure In order to evaluate the use of WHO 2009 WSs for identifying dengue shock and/or organ failure at admission, the diagnostic values of individual WSs and number of WSs were evaluated (S4 Table and Table 6). The sensitivities for all individual WSs were low, except for the following: lethargy, 87.5% (95% CI: 71.0–96.5%); mucosal bleeding, 78.1% (95% CI: 60.0–90.7%); and hematocrit >2% with platelets ≤100 ×103/μL, 79.4% (95% CI: 40.6–76.3%). Similarly, the specificities for all WSs were low, except that for clinical fluid accumulation (89.8%; 95% CI: 83.3–94.5%). The PPVs were low, but the NPVs for the WSs were high. The LR+ values for the WSs were low, except that for clinical fluid accumulation (5.2; 95% CI: 2.8–9.6). The LR–values for the WSs ranged from 0.4 to 0.8 (S4 Table). When the number of WSs was used to identify dengue shock and/or organ failure, this resulted in an AUROC of 0.77 (95% CI: 0.68–0.87) (Fig 3D). WSs ≥4 had an optimal sensitivity of 75.0% (95% CI: 56.6–88.5%) and a specificity of 71.1% (95% CI: 62.4–78.8%). A low PPV of 39.3% (95% CI: 27.1–52.7%) and an LR+ of 2.6 (95% CI: 1.8–3.6), but a high NPV of 91.9% (95% CI: 84.7–96.4%) and an LR–of 0.4 (95% CI: 0.2–0.6), were obtained for WSs ≥4, which indicated a small decrease in the likelihood of developing dengue shock and/or organ failure when the number of WSs was <4 (Table 6). 10.1371/journal.pntd.0004961.t006Table 6 Diagnostic values of the number of WHO 2009 warning signs for identifying dengue shock and/or organ failure at admission. Number of warning signs Confirmed dengue viral infection Sensitivity Specificity PPV NPV LR+ LR– With shock and/or organ failure (n = 32) No shock or organ failure (n = 128) (95% CI) (95% CI) (95% CI) (95% CI) (95% CI) (95% CI) ≥ 2 31 102 96.9 (83.8–99.9) 20.3 (13.7–28.3) 23.3 (16.4–31.4) 96.3 (81.0–99.9) 1.2 (1.1–1.4) 0.2 (0.0–1.1) ≥ 3 26 72 81.2 (63.6–92.8) 43.8 (35.0–52.8) 26.5 (18.1–36.4) 90.3 (80.1–96.4) 1.4 (1.2–1.8) 0.4 (0.2–0.9) ≥ 4 24 37 75.0 (56.6–88.5) 71.1 (62.4–78.8) 39.3 (27.1–52.7) 91.9 (84.7–96.4) 2.6 (1.8–3.6) 0.4 (0.2–0.6) ≥ 5 21 24 65.6 (46.8–81.4) 81.2 (73.4–87.6) 46.7 (31.7–62.1) 90.4 (83.5–95.1) 3.5 (2.3–5.4) 0.4 (0.3–0.7) ≥ 6 12 7 37.5 (21.1–56.3) 94.5 (89.1–97.8) 63.2 (38.4–83.7) 85.8 (79.0–91.1) 6.9 (2.9–16.0) 0.7 (0.5–0.9) CI, confidence interval; LR+, positive likelihood ratio; LR-, negative likelihood ratio; NPV, negative predictive value; PPV, positive predictive value; WHO, World Health Organization. Daily changes in serum procalcitonin and peripheral venous lactate during hospitalization among patients with or without dengue shock and/or organ failure In order to evaluate the average changes in PCT and PVL during hospitalization, PCT and PVL levels were measured at admission and every 24 h during hospitalization until the patient exhibited a body temperature <37.8°C for 48 h. Of the 32 patients who developed dengue shock and/or organ failure, only two (6.2%) expired during hospitalization, while 30 (93.8%) patients survived with complete recovery of organ function. All 128 patients without dengue shock or organ failure survived. Among the patients who survived, those with dengue shock and/or organ failure had higher PCT and PVL levels at admission and during hospitalization than those without (Fig 4). However, the patients who expired during hospitalization showed a trend toward increased PCT (>2 ng/mL) and PVL (>10 mmol/L) levels 24 h after admission (Fig 4). 10.1371/journal.pntd.0004961.g004Fig 4 Daily changes in serum procalcitonin and peripheral venous lactate during hospitalization. (A) Changes in average serum procalcitonin (ng/mL) levels among patients with and without dengue shock and/or organ failure by survival status. (B) Changes in peripheral venous lactate (mmol/L) levels among patients with and without dengue shock and/or organ failure by survival status. Data are presented as box and whisker plots with median (horizontal line), interquartile range (box), and maximum value within 1.5 of interquartile range (whiskers). Discussion A previous study showed that several factors were associated with dengue shock, including neurological signs, gastrointestinal bleeding, ascites, pleural effusion, hypoalbuminemia, hypoproteinemia, hepatomegaly, high levels of liver enzymes, and abnormal coagulators. It was postulated that this may have been due to the multiple organ involvement characteristic of dengue shock [13]. Recently, understanding of the pathogenesis of severe dengue has improved. It is now understood that a complex interplay between host factors and DENV is involved, resulting in vascular endothelial cell damage, which is the key factor leading to plasma leakage among patients with dengue [9–11,32]. Failure to recognize plasma leakage or inappropriate fluid administration in dengue patients with plasma leakage can lead to shock and organ failure [11,32]. Therefore, this prospective observational study was conducted among hospitalized adults with dengue in order to determine the independent factors associated with dengue shock and/or organ failure. Our results showed that PCT ≥0.7 ng/mL and PVL ≥2.5 mmol/L were independently associated with dengue shock and/or organ failure. The combination of PCT ≥0.7 ng/mL and PVL ≥2.5 mmol/L as a bioscore of 1 or 2 effectively predicted dengue shock and/or organ failure with ORs of 22.23 and 30.00, respectively. In addition, the combined bioscore provided good prognostic value in the prediction of dengue shock and/or organ failure, with an AUROC of 0.83 and an optimum sensitivity of 81.2%, specificity 84.4%, PPV 56.5%, NPV 94.7%, LR+ 5.2, and LR– 0.2. Furthermore, the combined bioscore provided a better diagnostic value for predicting dengue shock and/or organ failure compared to the WHO 2009 WSs in our study. Previous studies also showed the WSs have low sensitivities but high NPVs for identifying severe dengue in adults [33–35]. To date, PCT has been assessed as a biomarker for local and systemic inflammatory responses, disease severity, and necrosis related to organ failure, particularly in patients with bacterial infection [14,16,36]. However, a number of previous studies have shown that patients with viral diseases had PCT levels <0.5 ng/mL [37,38]. In our study, PCT ≥0.7 ng/mL was independently associated with dengue shock and/or organ failure. Previous studies have shown that patients with severe manifestations of dengue, including dengue hemorrhagic fever and dengue shock syndrome, had significantly higher viral titers than patients with dengue fever alone [39,40]. An increase in infected cells results in elevated acute-phase response proteins, cytokines, chemokines, generation of immune complexes, and consumption of complement, leading to damage of the vascular endothelium and increased vascular permeability [11,32]. PCT is an immunologically active protein induced through different steps of activation. Unlike various cytokines, PCT is activated in a time-dependent process, followed by adhesion and intercellular contact between injured cells and monocytes facilitated by adhesion molecule expression [14,15]. It is probable that increased PCT levels during DENV infection might be due to widespread inflammation in multiple organs. After cellular injury, PCT can be detected rapidly in the bloodstream within 2–6 h and reaches significant concentrations after 6 h, with peak values occurring at 12–48 h. The half-life of PCT is approximately 20–24 h with a daily clearance rate of 30% [14,41]. The clearance of PCT is not influenced by age, sex, or renal function [41]. A previous report showed that a PCT level of 0.79 ng/mL was observed among patients with localized bacterial infections [42]. In dengue, concurrent bacterial infections were observed after a median duration of 6.5 days in 4–25% of adults, and were particularly prevalent among patients with severe plasma leakage [43–45]. Sources of concurrent bacterial infection among patients with dengue included urinary tract infection (39.1%), pneumonia (38.2%), and primary bacteremia (22.7%) [43]. At admission, no patients with dengue in our study revealed bacterial growth in either of the two hemoculture samples, and there were no clinical symptoms indicative of any mixed infections. During hospitalization, four patients developed hospital-acquired infections, including two (1.6%) without dengue shock or organ failure who developed urinary tract infections, and two (6.2%) with dengue shock and/or organ failure who developed catheter-related infections during management at the intensive care unit. However, a significant proportion of patients with dengue shock and/or organ failure received antibiotics during hospitalization. In our clinical practice, antibiotics were prescribed to dengue patients suspected of having a concurrent bacterial infection, such as those with signs of peritonitis, fever with elevated bands of neutrophils, or a fever duration of more than 6 days. It is probable that bacterial translocation from the gastrointestinal or respiratory tract among patients with severe plasma leakage resulted in inflammatory responses, and increased PCT levels in the early stages of concurrent bacterial infection. A previous systematic review and meta-analysis showed that PCT ≥0.5 ng/mL could be used as a prognostic biomarker for bacterial infection, and elevated PCT levels with non-clearance were strongly associated with the in-hospital mortality of septic patients [36]. Regarding viral diseases, a previous report showed that two patients who died from 2009 H1N1 influenza infection had significantly higher mean PCT levels on admission compared to those who survived (14.5 vs. 1.7 ng/mL) [20]. Similarly, our study showed that patients who died tended to have increased PCT levels of >2 ng/mL 24 h after admission. However, the PCT levels among patients with dengue shock and/or organ failure ranged from mild to moderate elevation. The PCT assay using the Roche system in our study was previously evaluated for analytical performance in line with Clinical and Laboratory Standards. The PCT assay demonstrated acceptable precision, no evidence of nonlinearity, sample carryover or drift, and achieved high recovery from serum samples taken from patients with lower respiratory tract infections [46]. As the assay is marketed to reliably detect PCT concentrations as low as 0.02 ng/mL, the clinical laboratories providing testing are required to participate in regular international quality assurance programs to validate the test, particularly for standard hospital care. In our study, PVL ≥2.5 mmol/L was independently associated with dengue shock and/or organ failure, regardless of hypotension. In sepsis, lactate is a biomarker of anaerobic tissue metabolism resulting from hypoxemia and hypoperfusion, and aerobic mitochondrial dysfunction termed cytopathic hypoxia [14,47,48]. Like sepsis, cytopathic hypoxia has been demonstrated during DENV infection [49,50]. In combination, PCT and PVL provided the optimal prognostic value, with a sensitivity of 81.2% and a specificity of 84.4%. These findings may have resulted from the early stage of an extensive inflammatory response during the systemic phase of dengue, or possibly the early stage of a concurrent bacterial infection. In contrast, PVL is indicative of the condition of tissue hypoperfusion before the occurrence of shock in dengue. Our study also showed that expired patients exhibited a trend toward increased PVL levels of >10 mmol/L 24 h after admission. A previous study showed that lactate clearance ≥50% during the first 6 h was an independent predictor of survival in patients with septic shock [51]. Conclusions The strengths of our study were the prospective observational design and assessment of serial samples for PCT and PVL analysis. In addition, treating physicians and investigators were blinded to results in order to reduce missing data and minimize bias. All participants in this study were enrolled during the febrile phase of dengue; thus, the predictive parameters determined in this study could help physicians with early prediction of dengue shock and organ failure during the critical phase of dengue. However, our study had some limitations, as follows: (1) this study was conducted in a single center in Thailand, which was the referral center for tropical infectious diseases; (2) we could not perform cultures from sites requiring invasive investigation, such as peritoneal fluid or pleural fluid, as patients with dengue are at risk of bleeding; (3) empiric antibiotics were prescribed after hemocultures were taken, and (4) although all adult patients with clinical dengue were enrolled as described in the inclusion criteria, a number of older patients with dengue do not exhibit the full range of symptoms and may therefore have been inadvertently excluded. Therefore, our study focused on the assessment of younger adults with dengue. The utility of PCT and PVL in older patients with dengue remains unknown. Nonetheless, this study was the first to demonstrate that PCT levels ≥0.7 ng/mL and PVL levels ≥2.5 mmol/L were independently associated with dengue shock and/or organ failure, and that their combination provided good prognostic value for predicting dengue shock and/or organ failure. Dengue shock patients with non-clearance of PCT or PVL expired during hospitalization. These finding may help clinicians to predict dengue shock and/or organ failure earlier among hospitalized adults with dengue, leading to improved patient management and reduced in-hospital mortality and morbidity among patients with dengue. Supporting Information S1 Checklist STROBE checklist. (DOCX) Click here for additional data file. S2 Checklist STARD checklist. (DOCX) Click here for additional data file. S1 Table Baseline characteristics and clinical parameters at admission among 160 hospitalized adults with dengue. Data are presented as median (interquartile range) unless otherwise noted. (DOCX) Click here for additional data file. S2 Table Laboratory parameters, management, and outcomes among 160 hospitalized adults with dengue. Data are presented as median (interquartile range) unless otherwise noted. aPositive serological response to infection (n = 130). ALT, alanine aminotransferase; AST, aspartate aminotransferase; LYM, lymphocytes; PMN, polymorphonuclear leukocytes; RT-PCR, reverse-transcriptase polymerase chain reaction; WBC, white blood cell counts. (DOCX) Click here for additional data file. S3 Table Prediction of dengue shock and/or organ failure by serum procalcitonin and peripheral venous lactate levels at admission. CI, confidence interval; LR+, positive likelihood ratio; LR-, negative likelihood ratio; NPV, negative predictive value; PCT, procalcitonin; PPV, positive predictive value; PVL, peripheral venous lactate. (DOCX) Click here for additional data file. S4 Table Diagnostic values of the WHO 2009 warning signs for identifying dengue shock and/or organ failure at admission. CI, confidence interval; LR+, positive likelihood ratio; LR-, negative likelihood ratio; NPV, negative predictive value; PPV, positive predictive value; WHO, World Health Organization. (DOCX) Click here for additional data file. The authors thank all patients who participated in this study, and the staff and nurses in the emergency departments and referral centers (Ramathibodi Hospital, Mahidol University; Phramongkutklao Hospital; Rajavithi Hospital; and other private hospitals in Bangkok, Thailand). We also thank the nurses in Private Ward 1 and the General Female Ward at the Hospital for Tropical Diseases (Faculty of Tropical Medicine, Mahidol University) for their valuable help in patient care, as well as the staff of the central laboratory and Ms. Boongong Noochan (Clinical Infectious Diseases Research Unit, Department of Clinical Tropical Medicine, Faculty of Tropical Medicine, Mahidol University) for their help in performing this study. ==== Refs References 1 World Health Organization . Dengue: Guidelines for diagnosis, treatment, prevention and control Geneva : World Health Organization ; 2009 Available at: http://www.who.int/rpc/guidelines/9789241547871/en/ 2 World Health Organization . Global strategy for dengue prevention and control 2012–2020 Geneva : World Health Organization ; 2012 Available at: http://www.who.int/denguecontrol/9789241504034/en/ 3 Guzman MG , Halstead SB , Artsob H , Buchy P , Farrar J , Gubler DJ , et al Dengue: a continuing global threat . Nat Rev Microbiol . 2010 ;8 : S7 –S16 . 10.1038/nrmicro2460 21079655 4 Bhatt S , Gething PW , Brady OJ , Messina JP , Farlow AW , Moyes CL , et al The global distribution and burden of dengue . Nature . 2013 ;496 : 504 –507 . 10.1038/nature12060 23563266 5 Halstead SB . Is there an inapparent dengue explosion? Lancet . 1999 ;353 : 1100 –1101 . 6 Wali JP , Biswas A , Handa R , Aggarwal P , Wig N , Dwivedi SN . Dengue haemorrhagic fever in adults: a prospective study of 110 cases . Trop Doct . 1999 ;29 : 27 –30 . 10418277 7 Guzmán MG , Kourí G , Valdés L , Bravo J , Vázquez S , Halstead SB . Enhanced severity of secondary dengue-2 infections: death rates in 1981 and 1997 Cuban outbreaks . Rev Panam Salud Publica . 2002 ;11 : 223 –227 . 12049030 8 Horstick O , Jaenisch T , Martinez E , Kroeger A , See , Farrar J , et al Comparing the usefulness of the 1997 and 2009 WHO dengue case classification: a systematic literature review . Am J Trop Med Hyg . 2014 ;91 : 621 –634 . 10.4269/ajtmh.13-0676 24957540 9 Yacoub S , Mongkolsapaya J , Screaton G . The pathogenesis of dengue . Curr Opin Infect Dis . 2013 ;26 : 284 –289 . 10.1097/QCO.0b013e32835fb938 23449140 10 St. John AL , Abraham SN , Gubler DJ . Barriers to preclinical investigations of anti-dengue immunity and dengue pathogenesis . Nat Rev Microbiol . 2013 ;11 : 420 –426 . 10.1038/nrmicro3030 23652323 11 Guzman MG , Harris E . Dengue . Lancet 2015 ;385 : 453 –465 . 10.1016/S0140-6736(14)60572-9 25230594 12 Halstead SB . Controversies in dengue pathogenesis . Paediatr Int Child Health . 2012 ;32 Suppl 1 : 5 –9 . 10.1179/2046904712Z.00000000045 22668442 13 Huy NT , Van Giang T , Thuy DH , Kikuchi M , Hien TT , Zamora J , et al Factors associated with dengue shock syndrome: a systematic review and meta-analysis . PLoS Negl Trop Dis . 2013 ;7 : e2412 10.1371/journal.pntd.0002412 24086778 14 Meisner M . Update on procalcitonin measurements . Ann Lab Med . 2014 ;34 : 263 –73 . 10.3343/alm.2014.34.4.263 24982830 15 Liu HH , Guo JB , Geng Y , Su L . Procalcitonin: present and future . Ir J Med Sci . 2015 ;184 : 597 –605 . 10.1007/s11845-015-1327-0 26159431 16 Wacker C , Prkno A , Brunkhorst FM , Schlattmann P . Procalcitonin as a diagnostic marker for sepsis: a systematic review and meta-analysis . Lancet Infect Dis . 2013 ;13 : 426 –435 . 10.1016/S1473-3099(12)70323-7 23375419 17 Lubell Y , Blacksell SD , Dunachie S , Tanganuchitcharnchai A , Althaus T , Watthanaworawit W , et al Performance of C-reactive protein and procalcitonin to distinguish viral from bacterial and malarial causes of fever in Southeast Asia . BMC Infect Dis . 2015 ;15 : 511 10.1186/s12879-015-1272-6 26558692 18 Meisner M , Tschaikowsky K , Palmaers T , Schmidt J . Comparison of procalcitonin (PCT) and C-reactive protein (CRP) plasma concentrations at different SOFA scores during the course of sepsis and MODS . Crit Care . 1999 ;3 : 45 –50 . 11056723 19 Clec'h C , Ferriere F , Karoubi P , Fosse JP , Cupa M , Hoang P , et al Diagnostic and prognostic value of procalcitonin in patients with septic shock . Crit Care Med . 2004 ;32 : 1166 –1169 . 15190968 20 Ingram PR , Inglis T , Moxon D , Speers D . Procalcitonin and C-reactive protein in severe 2009 H1N1 influenza infection . Intensive Care Med . 2010 ;36 : 528 –532 . 10.1007/s00134-009-1746-3 20069274 21 Mikkelsen ME , Miltiades AN , Gaieski DF , Goyal M , Fuchs BD , Shah CV , et al Serum lactate is associated with mortality in severe sepsis independent of organ failure and shock . Crit Care Med . 2009 ;37 : 1670 –1677 . 10.1097/CCM.0b013e31819fcf68 19325467 22 Thanachartwet V , Oer-Areemitr N , Chamnanchanunt S , Sahassananda D , Jittmittraphap A , Suwannakudt P , et al Identification of clinical factors associated with severe dengue among Thai adults: a prospective study . BMC Infect Dis . 2015 ;15 : 420 10.1186/s12879-015-1150-2 26468084 23 von Elm E , Altman DG , Egger M , Pocock SJ , Gøtzsche PC , Vandenbroucke JP ; STROBE Initiative . Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies . BMJ . 2007 ;335 : 806 –808 . 17947786 24 Korevaar DA , van Enst WA , Spijker R , Bossuyt PM , Hooft L . Reporting quality of diagnostic accuracy studies: a systematic review and meta-analysis of investigations on adherence to STARD . Evid Based Med . 2014 ;19 : 47 –54 . 10.1136/eb-2013-101637 24368333 25 Lanciotti RS , Calisher CH , Gubler DJ , Chang GJ , Vorndam AV . Rapid detection and typing of dengue viruses from clinical samples by using reverse transcriptase-polymerase chain reaction . J Clin Microbiol . 1992 ;30 : 545 –551 . 1372617 26 Reynes JM , Ong S , Mey C , Ngan C , Hoyer S , Sall AA . Improved molecular detection of dengue virus serotype 1 variants . J Clin Microbiol . 2003 ;41 : 3864 –3867 . 12904404 27 Vorndam V , Beltran M . Enzyme-linked immunosorbent assay-format microneutralization test for dengue viruses . Am J Trop Med Hyg . 2002 ;66 : 208 –212 . 12135295 28 Putnak JR , de la Barrera R , Burgess T , Pardo J , Dessy F , Gheysen D , et al Comparative evaluation of three assays for measurement of dengue virus neutralizing antibodies . Am J Trop Med Hyg . 2008 ;79 : 115 –122 . 18606774 29 Innis BL , Nisalak A , Nimmannitya S , Kusalerdchariya S , Chongswasdi V , Suntayakorn S , et al An enzyme-linked immunosorbent assay to characterize dengue infections where dengue and Japanese encephalitis co-circulate . Am J Trop Med Hyg . 1989 ;40 : 418 –427 . 2540664 30 Aung KL , Thanachartwet V , Desakorn V , Chamnanchanunt S , Sahassananda D , Chierakul W , et al Factors associated with severe clinical manifestation of dengue among adults in Thailand . Southeast Asian J Trop Med Public Health . 2013 ;44 : 602 –612 . 24050093 31 Gibot S , Béné MC , Noel R , Massin F , Guy J , Cravoisy A , et al Combination biomarkers to diagnose sepsis in the critically ill patient . Am J Respir Crit Care Med . 2012 ;186 : 65 –71 . 10.1164/rccm.201201-0037OC 22538802 32 Simmons CP , Farrar JJ , Nguyen VV , Wills B . Dengue . N Engl J Med 2012 ;366 : 1423 –1432 . 10.1056/NEJMra1110265 22494122 33 Thein TL , Gan VC , Lye DC , Yung CF , Leo YS . Utilities and limitations of the World Health Organization 2009 warning signs for adult dengue severity . PLoS Negl Trop Dis . 2013 ;7 : e2023 10.1371/journal.pntd.0002023 23350013 34 Leo YS , Gan VC , Ng EL , Hao Y , Ng LC , Pok KY , et al Utility of warning signs in guiding admission and predicting severe disease in adult dengue . BMC Infect Dis . 2013 ;13 : 498 10.1186/1471-2334-13-498 24152678 35 Thanachartwet V , Wattanathum A , Oer-areemitr N , Jittmittraphap A , Sahassananda D , Monpassorn C , Surabotsophon M , Desakorn V . Diagnostic accuracy of peripheral venous lactate and the 2009 WHO warning signs for identifying severe dengue in Thai adults: a prospective observational study . BMC Infect Dis . 2016 ;16 : 46 10.1186/s12879-016-1386-5 26832147 36 Liu D , Su L , Han G , Yan P , Xie L . Prognostic value of procalcitonin in adult patients with sepsis: a systematic review and meta-analysis . PLoS One . 2015 ;10 : e0129450 10.1371/journal.pone.0129450 26076027 37 Branche AR , Walsh EE , Vargas R , Hulbert B , Formica MA , Baran A , et al Serum procalcitonin measurement and viral testing to guide antibiotic use for respiratory infections in hospitalized adults: a randomized controlled trial . J Infect Dis . 2015 ;212 : 1692 –1700 . 10.1093/infdis/jiv252 25910632 38 Gendrel D , Raymond J , Coste J , Moulin F , Lorrot M , Guérin S , et al Comparison of procalcitonin with C-reactive protein, interleukin 6 and interferon-alpha for differentiation of bacterial vs. viral infections . Pediatr Infect Dis J . 1999 ;18 : 875 –881 . 10530583 39 Vaughn DW , Green S , Kalayanarooj S , Innis BL , Nimmannitya S , Suntayakorn S , et al Dengue viremia titer, antibody response pattern, and virus serotype correlate with disease severity . J Infect Dis . 2000 ;181 : 2 –9 . 10608744 40 Murgue B , Roche C , Chungue E , Deparis X . Prospective study of the duration and magnitude of viraemia in children hospitalised during the 1996–1997 dengue-2 outbreak in French Polynesia . J Med Virol . 2000 ;60 : 432 –438 . 10686027 41 Meisner M , Schmidt J , Hüttner H , Tschaikowsky K . The natural elimination rate of procalcitonin in patients with normal and impaired renal function . Intensive Care Med . 2000 ;26 Suppl 2 : S212 –S216 . 10.1007/BF02900740 18470722 42 Rey C , Los Arcos M , Concha A , Medina A , Prieto S , Martinez P , et al Procalcitonin and C-reactive protein as markers of systemic inflammatory response syndrome severity in critically ill children . Intensive Care Med . 2007 ;33 : 477 –484 . 17260130 43 See KC , Phua J , Yip HS , Yeo LL , Lim TK . Identification of concurrent bacterial infection in adult patients with dengue . Am J Trop Med Hyg . 2013 ;89 : 804 –810 . 10.4269/ajtmh.13-0197 23980129 44 Premaratna R , Dissanayake D , Silva FH , Dassanayake M , de Silva HJ . Secondary bacteraemia in adult patients with prolonged dengue fever . Ceylon Med J . 2015 ;60 : 10 –12 . 10.4038/cmj.v60i1.7165 25804911 45 Leo YS , Thein TL , Fisher DA , Low JG , Oh HM , Narayanan RL , et al Confirmed adult dengue deaths in Singapore: 5-year multi-center retrospective study . BMC Infect Dis . 2011 ;11 : 123 10.1186/1471-2334-11-123 21569427 46 de Wolf HK , Gunnewiek JK , Berk Y , van den Ouweland J , de Metz M . Comparison of a new procalcitonin assay from roche with the established method on the brahms kryptor . Clin Chem . 2009 ;55 (5 ): 1043 –1044 . 10.1373/clinchem.2008.117655 19264851 47 Fink MP . Cytopathic hypoxia. Is oxygen use impaired in sepsis as a result of an acquired intrinsic derangement in cellular respiration? Crit Care Clin . 2002 ;18 : 165 –175 . 11910729 48 Belikova I , Lukaszewicz AC , Faivre V , Damoisel C , Singer M , Payen D . Oxygen consumption of human peripheral blood mononuclear cells in severe human sepsis . Crit Care Med . 2007 ;35 : 2702 –2708 . 18074472 49 El-Bacha T , Midlej V , Pereira da Silva AP , Silva da Costa L , Benchimol M , Galina A , et al Mitochondrial and bioenergetic dysfunction in human hepatic cells infected with dengue 2 virus . Biochim Biophys Acta . 2007 ;1772 : 1158 –1166 . 17964123 50 Hottz ED , Oliveira MF , Nunes PC , Nogueira RM , Valls-de-Souza R , Da Poian AT , et al Dengue induces platelet activation, mitochondrial dysfunction and cell death through mechanisms that involve DC-SIGN and caspases . J Thromb Haemost . 2013 ;11 : 951 –962 . 10.1111/jth.12178 23433144 51 Puskarich MA , Trzeciak S , Shapiro NI , Albers AB , Heffner AC , Kline JA , et al Whole blood lactate kinetics in patients undergoing quantitative resuscitation for severe sepsis and septic shock . Chest . 2013 ;143 : 1548 –1553 . 10.1378/chest.12-0878 23740148

PMC005xxxxxx/PMC5001650.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 10.1371/journal.pone.0161762PONE-D-16-05298Research ArticleBiology and Life SciencesPlant SciencePlant AnatomyPollenBiology and Life SciencesOrganismsPlantsTreesOaksBiology and Life SciencesOrganismsPlantsTreesPeople and PlacesGeographical LocationsEuropePolandBiology and Life SciencesTaxonomyComputer and Information SciencesData ManagementTaxonomyBiology and Life SciencesPlant SciencePlant AnatomyInflorescencesBiology and Life SciencesPopulation BiologyPopulation DynamicsGeographic DistributionResearch and Analysis MethodsMicroscopyElectron MicroscopyScanning Electron MicroscopyComparative Pollen Morphological Analysis and Its Systematic Implications on Three European Oak (Quercus L., Fagaceae) Species and Their Spontaneous Hybrids Pollen Morphological Analysis on Three European Oak Species and Their Spontaneous Hybridshttp://orcid.org/0000-0003-2431-6192Wrońska-Pilarek Dorota 1*Danielewicz Władysław 1Bocianowski Jan 2Maliński Tomasz 1Janyszek Magdalena 31 Department of Forest Botany, Poznan University of Life Sciences, Poznań, Poland2 Department Mathematical and Statistical Methods, Poznan University of Life Sciences, Poznań, Poland3 Department of Botany, Poznan University of Life Sciences, Poznań, PolandAravanopoulos Filippos A. EditorAristotle University of Thessaloniki, GREECECompeting Interests: The authors have declared that no competing interests exist. Conceived and designed the experiments: DWP WD. Performed the experiments: DWP. Analyzed the data: DWP JB. Contributed reagents/materials/analysis tools: TM MJ. Wrote the paper: DWP WD JB. * E-mail: pilarekd@up.poznan.pl26 8 2016 2016 11 8 e01617625 2 2016 11 8 2016 © 2016 Wrońska-Pilarek et al2016Wrońska-Pilarek et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Pollen morphology of three parental Quercus species (Q. robur L., Q. petraea (Matt) Liebl, Q. pubescens Willd.) and two spontaneous hybrids of these species (Q. ×calvescens Vuk. = Q. petraea × Q. pubescens and Q. ×rosacea Bechst. = Q. robur × Q. petraea) was investigated in this study. The pollen originated from 18 natural oak sites and 67 individuals (oak trees). Each individual was represented by 30 pollen grains. In total, 2010 pollen grains were measured. They were analysed for nine quantitative and four qualitative features. Pollen size and shape were important features to diagnosing Quercus parental species and hybrids. On the basis of exine ornamentation, it was possible to identify only Q. pubescens, while the remaining species and hybrids did not differ significantly with respect to this feature. The determination of the diagnostic value of endoaperture features requires further palynological studies. On the basis of pollen size and shape Q. robur × Q. petraea was clearly separated. Grouping of 67 oak trees on the basis of pollen grain features has shown that individuals from different as well as same taxa occurred in the same groups. Likewise, with respect to natural sites, oak trees originating from the same places as well as from geographically distant ones, grouped together. Pollen morphological features allow to distinguish a part of the studied Quercus taxa. Therefore, it can be used as an auxiliary feature in the taxonomy. The author(s) received no specific funding for this work. Data AvailabilityAll relevant data are within the paper.Data Availability All relevant data are within the paper. ==== Body Introduction Quercus genus is the largest genus in the Fagaceae family. Depending on different authorities, the number of oak species varies between 300 and 400 [1–5], through 500–531 [6, 7] up to 600 [8, 9]. The members of this genus occupy territories of the Northern Hemisphere in Asia, North America, Europe and Africa, with a few species extending to the equator [10–14]. The taxonomy of Quercus genus is extremely complex as a result of high species numbers, wide geographical distribution, great morphological variability, as well as widespread hybridization between infrageneric taxa and changes in morphological features [7, 8, 15–21]. Therefore, the classification of oaks has been a matter of debate, from De Candolle [22] to Nixon [13], and more than 20 classifications were proposed [16, 23]. Up to date, the most taxonomically valuable morphological features of Quercus species are foliar and fruit characteristics. Hence, the taxonomic classifications of oaks are usually based on these features [5, 13, 15, 24, 25]. According to Menitski’s [11] classification, the three species analyzed for the present study belong to the Quercus L., subgenus Quercus, section Quercus, series Quercus and the subseries: Quercus—Quercus robur L., Quercus petraea (Matt.) Liebl. and Galliferae (Spach) Guerke—Quercus pubescens Willd. According to the infrageneric groups recognized by Denk and Grimm [26] the three species belong to the Quercus, Group Quercus. Occurrence and frequency of interspecific hybrids in natural European oak populations has not been clarified satisfactorily so far [27]. In the botanical literature, interspecific hybrids in natural European oak populations have been distinguished for a long time, primarily on the basis of morphological traits of indirect characters between the alleged parental species [1, 24, 28]. They comprise, among others, such taxons as: Quercus ×calvescens Vuk. (Q. petraea × Q. pubescens), Q. ×kerneri Simk. (Q. robur × Q. pubescens) and Q. ×rosacea Bechst. (Q. robur × Q. petraea). According to Rushton [20], in mixed populations of Q. robur and Q. petraea, the proportion of trees morphologically “truly intermediate” between the above species in north-western Europe, estimated on the basis of data quoted by different researchers, ranges from 2 to 29%. Many authors think that hybrids between these species are rare, and frequency of their appearance in natural population is approximately 2–3% [15, 29–35]. However, detailed comparative, palynological studies of pollen grain characters between oak hybrids and parental species have not been performed. These characters were described for taxonomic purposes in numerous studies using scanning electron microscopy (SEM) [17, 36–48]. Many researchers maintain, that pollen morphology can be a valuable source of information for Quercus taxonomy [17, 36–40, 43, 45, 46], whereas others contest the possibility of differentiating species through characteristics of pollen grains, since intraspecific variability may appear to be the same or greater than interspecific differences [49–52]. Palynological studies on interspecific hybrids concern artificial hybrids [53–59], or spontaneous hybrids found in nature [60–65]. In these studies only a few, selected quantitative pollen features are usually compared; most commonly they include pollen size [55, 57–60, 62–66], more rarely—exine ornamentation [38, 40, 53, 54, 56] or aperture numbers [54, 55]. Only few researchers investigated also the different degrees of deformed pollen grains deformation in parental species and hybrids [60, 64]. A recent pollen morphological study focusing on pollen ornamentation [38] showed that this pollen character is diagnostic for six major infrageneric groups within Quercus. These six infrageneric groups are also recognized by molecular phylogenetic studies [26, 67, 68]. Therefore, the aim of our study was to evaluate whether and how pollen grains of the three studied parental species (Q. robur L., Q. petraea (Matt) Liebl., Q. pubescens Willd.) differ from those of their spontaneous hybrids: Quercus ×calvescens Vuk. (Q. petraea × Q. pubescens), and Q. × rosacea Bechst. (Q. robur × Q. petraea). This has not been the object of palynological studies so far. In addition, pollen grain variability of the investigated oak taxa has not yet been comprehensively analysed. We think that one of the strengths of this study is that we take advantage of a relatively large sample from several individuals, in order to capture much of intraspecific variability, in contrast to other studies, that focus in distinguishing different phylogenetic groups [44–46, 49]. Material and Methods While gathering sufficiently large samples from typical Q. robur, Q. petraea and Q. pubescens individuals was not difficult, the collection of inflorescences from trees morphologically intermediate between them (assumed hybrids) was considerably limited, because of their rare occurrence. The assumed hybrids collected from Bielinek, the sole Q. pubescens site in Poland (52°56'26"N, 14°8'54"E) comprised mainly Q. petraea × Q. pubescens hybrids, although single specimens of Q. robur × Q. pubescens cannot be excluded. Plant materials in the form of fresh inflorescences was selected and verified by Professor Władysław Danielewicz (Department of Forest Botany, Poznań University of Life Sciences), whereas, that from the Herbarium of the Institute of the Dendrology of Polish Academy of Sciences in Kórnik was verified by Professor Jerzy Zieliński. In this study, pollen morphology of three parental Quercus species (Q. robur L., Q. petraea (Matt) Liebl, Q. pubescens Willd.) and two spontaneous hybrids of these species (Quercus ×calvescens Vuk. = Q. petraea × Q. pubescens and Q. ×rosacea Bechst. = Q. robur × Q. petraea) were analysed (Table 1). 10.1371/journal.pone.0161762.t001Table 1 List of localities of the studied Quercus taxa. No Species-Taxon Localities Coordinates Collector, herbarium No of trees (samples) for each location 1 Q. petraea (Matt) Liebl. Dębno, zachodniopomorskie, Poland 52°44'20"N, 14°41'52"E Danielewicz; PZNF 5 2 Q. petraea (Matt) Liebl. Poznań - Marcelin, wielkopolskie, Poland 52°24'23"N, 16°51'86"E Danielewicz; PZNF 2 3 Q. petraea (Matt) Liebl. Puszcza Bukowa, zachodniopomorskie, Poland 53°20'12"N, 14°40'29"E Danielewicz; PZNF 5 4 Q. petraea (Matt) Liebl. Rokita, zachodniopomorskie, Poland 53°45'53"N, 14°50'27"E Danielewicz; PZNF 3 5 Q. petraea (Matt) Liebl. Różańsko, zachodniopomorskie, Poland 52°50'56"N, 14°46'54"E Danielewicz; PZNF 5 6 Q. petraea (Matt) Liebl. Wielkopolski National Park, wielkopolskie, Poland Province 52°16'94"N, 16°47'57"E Danielewicz; PZNF 5 7 Q. petraea (Matt) Liebl. Vitosha Mts., Bulgaria, 950 m 42°34'00"N, 23°17'00"E Vihodcevsky; KOR 1 8 Q. petraea (Matt) Liebl. Vall de Ribes, Pyrenees Mts., Spain, 1500 m 42°37′56"N, 0°39′28"E Boratyński; KOR 1 9 Q. pubescens Willd. Bielinek, zachodniopomorskie, Poland 52°56′26"N, 14°8'54"E Maliński; PZNF 10 10 Q. pubescens Willd. Hafnerberg, Austria 48°01′00"N, 16°01′00"E Browicz; KOR 1 11 Q. pubescens Willd. Perithia, Korfu, Greece 39°76′46"N, 19°87′54"E Boratyński, Browicz; KOR 1 12 Q. pubescens Willd. Crimea, Russia 45°18′00"N, 34°24′00"E Dzewanowskaya; KOR 1 13 Q. robur L. Białowieża, podalskie, Poland 52°45′76″N 23°52′45″E Danielewicz; PZNF 5 14 Q. robur L. Jarocin, wielkopolskie, Poland 51°58'23"N, 17°30'12"E Danielewicz; PZNF 2 15 Q. robur L. Promno, wielkopolskie, Poland 52°27'37"N, 17°14'44"E Danielewicz; PZNF 5 16 Q. robur L. Puszcza Bukowa, zachodniopomorskie, Poland 53°20'12"N, 14°40'29"E Danielewicz; PZNF 1 17 Q. robur L. Rogalin, wielkopolskie, Poland 52°14'43"N, 16°56'27"E Danielewicz; PZNF 2 18 Q. robur L. Rokita, zachodniopomorskie, Poland 53°45'53"N, 14°50'27"E Danielewicz; PZNF 5 19 Q. robur L. Wielkopolski National Park, wielkopolskie, Poland 52°16'94"N, 16°47'57"E Danielewicz; PZNF 2 20 Q. robur L Andrusul de Sus, Moldova 46°0′20"N, 28°14′51"E Poneasheskij; KOR 1 21 Q. ×calvescens Vuk. = Q. petraea × Q. pubescens Bielinek, zachodniopomorskie, Poland 52°56'26"N, 14°8'54"E Maliński; PZNF 3 22 Q. ×rosacea Bechst. = Q. robur × Q. petraea Poznań, Dendrological Garden of Poznan University of Life Sciences, wielkopolskie, Poland 52°25'40"N, 16°53'57"E Danielewicz; PZNF 1 KOR—Herbarium of the Institute of Dendrology, Polish Academy of Sciences, Kórnik, Poland; PZNF—Herbarium of Department of Forest Botany, Poznan University of Life Sciences, Poland. Male inflorescences investigated for this study originate from 18 natural oaks sites, located in Austria, Bulgaria, Greece (Corfu), Spain, Crimea, Moldova and Poland. Except for the Polish material, male inflorescences were obtained from herbarium material stored in the Herbarium of the Institute of Dendrology of the Polish Academy of Sciences in Kórnik (52°14'12''N, 17°05'55''E)–KOR (Poland) (Table 1). Several, randomly selected inflorescences were collected from each of 67 oak individuals. Each individual is represented by 30 correctly formed, mature pollen grains [69]. In total, 2010 pollen grains were measured. Malformed pollen grains were also noticed in the samples, and their percentage was determined considering 1000 pollen grains in Q. petraea and Q. robur (five randomly selected samples of 200 pollen grains), and 600 grains in Q. pubescens and Q. petraea × Q. pubescens (three samples) and 200 pollen grains in the rare hybrid Q. robur × Q. petraea. For the measurements, samples were acetolysed according to Erdtman’s method [70]. The acetolysing mixture was made up of nine parts of acetic acid anhydride and one part of concentrated sulphuric acid and the process of acetolysis lasted 2.5 minutes. The measurements were made on acetolysed grains with light microscope (Biolar 2308) and observations of qualitative features were carried out with scanning electron microscope (Hitachi S-3000N) on acetolysed grains to. Pollen grains were prepared in glycerine jelly and measured using the eyepiece (ocular) with scale. Next, the pollen grains were analysed for nine quantitative features, i.e. length of polar axis (P), equatorial diameter (E), length of ectocolpi (Le), thickness of exine along polar axis (Exp) and equatorial diameter (Exe) and four ratios: P/E, Exp/P, Exe/E, Le/P; and the following qualitative ones: exine ornamentation, endoaperture type, pollen outline and shape. The palynological terminology follows Punt et al. [71] and Hesse et al. [72]. Firstly, the normality of the distributions of the studied traits (P, E, P/E, Exp, Exe, Exp/P, Exe/E, Le and Le/P) was tested using the Shapiro-Wilk’s normality test [73]. Multivariate analysis of variance (MANOVA) was performed on the basis of the following model using a procedure MANOVA in GenStat 17th edition: Y = XT+E, where: Y is (n×p)–dimensional matrix of observations, n is number of all observations, p is number of traits (in this study p = 9), X is (n×k)–dimensional matrix of design, k is number of taxa (in this study k = 5), T is (k×p)—dimensional matrix of unknown effects,—is (n×p)–dimensional matrix of residuals. Next, one-way analyses of variance (ANOVA) were performed in order to verify the zero hypothesis on a lack of taxon effect in terms of values of observed traits, i.e. P, E, P/E, Exp, Exe, Exp/P, Exe/E, Le and Le/P for each trait independent, on the basis of the following model: yij = μ+τi+εij, where: yij is jth observation of ith taxon, μ is grand mean, τi is effect of ith taxon and εij is an error observation. The minimal and maximal values of traits as well as arithmetical means and coefficients of variation—CV (in %)—were calculated. Moreover, the Fisher’s least significant differences (LSDs) were also estimated at the significance level α = 0.001. The relationships between the observed traits were assessed on the basis of Pearson’s correlation coefficients using the FCORRELATION implementation in GenStat 17th edition. The parallel coordinate plot has been proposed as an efficient tool for visualization of species and their hybrids visualization [74, 75]. Results were also analysed using multivariate methods. The analysis of canonical variables was applied in order to present multitrait assessment of similarity of tested genotypes (two separate analyses: first for species and hybrids and second for trees) in a lower number of dimensions with the least possible loss of information [76]. This makes it possible to illustrate variation in genotypes in terms of all observed traits a graphic way. Mahalanobis’ distance was suggested as a measure of “polytrait” genotypes similarity [77], whose significance was verified by means of critical value Dα called “the least significant distance” [78]. Mahalanobis’ distances were calculated for taxa and trees, independently. All the analyses were conducted using the GenStat 17th edition statistical software package [79]. Results General pollen morphology Quantitative features of pollen grains are summarized in Table 2 and illustrated with scanning electron micrographs (Figs 1A–1M and 2A–2H). Pollen grains of the examined taxa are monads, isopolar, radially symmetrical. The pollen grains are tricolporoidate or tricolpate. 10.1371/journal.pone.0161762.t002Table 2 Range (min-max), mean values and coefficient of variation (cv) of studied features. One-way ANOVA’s were performed separately for each of traits. Same letters indicate a lack of statistically significant differences between analyzed taxa according to Tukey’s post hoc test (p < 0.001). Feature P E P/E Species Min-Max Mean CV (%) Min-Max Mean CV (%) Min-Max Mean CV (%) Q. petraea 22–40 31.37 b 8.01 20–40 30.26 bc 9.60 0.8125–1.5 1.043 b 9.92 Q. pubescens 24–42 31.78 b 8.69 22–38 30.31 bc 7.94 0.8–1.545 1.052 b 9.62 Q. robur 24–42 30.79 b 9.32 20–38 29.26 c 9.69 0.75–1.636 1.060 b 11.56 Q. petraea × Q. pubescens 28–38 32.00 b 5.85 24–36 31.24 ab 7.65 0.8333–1.417 1.030 b 9.34 Q. robur × Q. petraea 32–42 37.07 a 7.05 30–40 32.80 a 6.52 1–1.312 1.133 a 7.71 LSD0.001 1,63 1,69 0.067 F statistic 46.04*** 27.94*** 7.19*** Feature Exp Exe Exp/P Species Min-Max Mean CV (%) Min-Max Mean CV (%) Min-Max Mean CV (%) Q. petraea 0.4–2 1.025 bc 41.96 0.4–2 1.017 bc 40.68 0.0111–0.0909 0.0329 a 43.17 Q. pubescens 0.6–2 1.066 abc 37.94 0.6–2 1.033 bc 35.61 0.0158–0.0769 0.0337 a 38.39 Q. robur 0.6–2 1.182 ab 38.45 0.6–2 1.150 ab 37.66 0.0158–0.0769 0.0387 a 39.33 Q. petraea × Q. pubescens 0.6–2 1.304 a 37.37 0.6–2 1.291 a 37.17 0.0167–0.0714 0.0408 a 37.39 Q. robur × Q. petraea 0.6–1.4 0.880 c 27.09 0.6–1.4 0.853 c 26.07 0.0150–0.0368 0.0237 b 25.36 LSD0.001 0.266 0.253 0.009 F statistic 19.59*** 19.02*** 24.89*** Features Exe/E Le Le/P Species Min-Max Mean CV (%) Min-Max Mean CV (%) Min-Max Mean CV (%) Q. petraea 0.012–0.09091 0.03402 ab 43.12 16–36 26.18 b 11.92 0.5714–1 0.835 b 8.83 Q. pubescens 0.01579–0.07692 0.03427 ab 36.39 18–38 26.97 b 11.62 0.4737–1 0.849 ab 7.62 Q. robur 0.0117–0.1 0.0399 a 40.43 18–38 26.32 b 13.04 0.6000–1 0.854 ab 8.25 Q. petraea × Q. pubescens 0.01765–0.08333 0.04194 a 40.82 20–32 27.31 b 9.89 0.6875–1 0.853 ab 7.87 Q. robur × Q. petraea 0.015–0.04375 0.02620 b 27.86 28–36 32.07 a 6.85 0.7778–1 0.867 a 5.85 LSD0.001 0.009 1.964 0.023 F statistic 23.44*** 28.66*** 8.89*** P—length of polar axis; E—equatorial diameter; Le—length of ectocolpi; Exp—thickness of exine along polar axis; Exe—thickness of exine along equatorial diameter. *** P<0.001. 10.1371/journal.pone.0161762.g001Fig 1 Q. petraea, A-E. A, numerous pollen grains in polar and equatorial view, spheroidal or prolate-spheroidal in shape; B, equatorial view; C, polar and equatorial view of two pollen grains; D, polar view with two ectocolpi; E, ectocolpus and granulate-verrucate exine ornamentation with the biggest verrucae (>1 μm), slightly smaller granules (0.5–1μm), and the smallest and low granules (0.1–0.3 μm); Q. pubescens, F-H. F, equatorial view; G, polar view with three ectocolpi; H, ectocolpus and granulate-verrucate exine ornamentation without small granules; Q. robur, I-M. I, equatorial view; J, polar view with three, closed ectocolpi; K, four pollen in polar and equatorial view; L, ectocolpus and granulate-verrucate exine ornamentation; M, exine ornamentation details—see (Fig 1E). 10.1371/journal.pone.0161762.g002Fig 2 Q. petraea × Q. pubescens, A-D. A, equatorial view; B, polar view with one ectocolpus; C, ectocolpus and exine granulate-verrucate ornamentation; D, exine ornamentation details—see (Fig 1E); Q. robur × Q. petraea, E-H. E, equatorial view, F, polar view with two ectocolpi; G, ectocolpus and exine granulate-verrucate ornamentation; H, exine ornamentation details—see (Fig 1E). According to Erdtman’s [80] pollen size classification, the studied pollen grains of parental species were medium—99% (25.1–50 `m), very rarely (1%) small-sized (10.0–25.0 μm). The size of all pollen of hybrids was medium (Table 2). Parental species exhibited smaller pollen grains as compared to hybrids. The average length of the polar axis (P) in parental Quercus species was 31.24 (22.00–42.00) μm and in hybrids—33.27 (28.00–42.00) μm. In parental species, the shortest polar axis (P) occurred in the pollen of Q. petraea (22.00 μm), while the longest one—in Q. robur and Q. pubescens (42.00 μm). For hybrids, the shortest polar axis (P) was found in Q. petraea × Q. pubescens (28.00 μm) and the longest one—in Q. robur × Q. petraea (42.00 μm; Table 2). The mean length of the equatorial diameter (E) in parental Quercus species amounted to 29.90 (20.00–40.00) μm and in hybrids—31.63 (24.00–40.00) μm. In all studied taxa, the outline in polar view was mostly circular, more rarely triangular or elliptic, whereas in equatorial view it was mostly elliptic or circular. The mean P/E ratio in parental Quercus species was 1.05 and ranged from 0.75 to 1.64 in Q. robur and in hybrids it was 1.06 (range 0.83–1.42) in Q. petraea × Q. pubescens (Table 2). With respect to features, pollen shapes of parental species and hybrids were different (Table 3). In the case of parental species, most frequent pollen grains were prolate-spheroidal (36.2%), spheroidal (26.6%) and oblate-spheroidal (20.4%), while subprolate ones occurred more rarely (14.2%), prolate (1.5%) and suboblate (1.1%) pollen were found only sporadically. In hybrids, spheroidal (32.5%) and prolate-spheroidal (30.8%) pollen grains were most common, while oblate-spheroidal and subprolate (17.5% each) were not so frequent and prolate and suboblate pollen were encountered only in single grains (0.8% each). Slightly different results were obtained when analysing the distribution of pollen shape class in individual taxa. In the case of parental species, the results were similar to those reported above, but in hybrids—they differed significantly both from one another and from parental species (Table 3). Quercus robur × Q. petraea was distinguished by the highest number of elongated pollen grains (subprolate—40% and prolate-spheroidal—46.7%). In Q. robur × Q. petraea—spheroidal pollen were not numerous (13.3%), while oblate-spheroidal pollen—fairly frequent in other taxa (17.2–24.4%)—did not occur at all. On the other hand, Q. petraea × Q. pubescens, in contrast to Q. ×rosacea, exhibited most frequently spheroidal (38.9%) and oblate-spheroidal (24.4%) pollen accompanied by the lowest proportion of prolate-spheroidal pollen (25.6%) among the studied taxa. 10.1371/journal.pone.0161762.t003Table 3 The percentage participation of pollen grains in shape classes (P/E ratio) according to Erdtman’s (1952) classification. Taxon Pollen shape classes suboblate oblate-spheroidal spheroidal prolate-spheroidal subprolate prolate Q. petraea 1.2 20.9 27.8 36.9 12.1 1.1 Q. pubescens 0.8 16.4 29.0 42.1 10.5 1.3 Q. robur 1.2 22.0 23.8 32.0 18.8 2.2 Q. petraea × Q. pubescens - 24.44 38.89 25.56 10.00 1.11 Q. robur × Q. petraea - - 13.33 46.67 40.00 - Suboblate (0.75–0.88); oblate-spheroidal (0.89–0.99); spheroidal (1.00); prolate-spheroidal (1.01–1.14); subprolate (1.15–1.33); prolate (1.34–2.00). The mean exine thickness was 1.08 (0.4–2.0) μm (parental species) and 1.19 (0.60–2.0) μm (hybrids; Table 2). Exine was thinnest in Q. petraea (0.4 μm) and thickest (2.0 μm) among all studied species. In hybrids, exine thickness varied the least, from 0.6 μm in both studied taxa to 2.0 μm—in Q. petraea × Q. pubescens. The relative thickness of the exine (Exp/P ratio) was similar for parental species and hybrids amounting to 0.04 (0.01–0.1) and 0.04 (0.02–0.08), respectively. Pollen grains had three apertures and were tricolpate or tricolporoidate (this is in angiosperm pollen a rare character, when ectoaperture consists of a colpus with an indistinct endoaperture). Colpi were arranged meridionally, regularly. They were very narrow, with acute to narrowly obtuse ends. Commonly colpi were covered at the equator by a geniculum—a bulge formed by sexine extensions. Colpus membranes were usually smooth. Colpi were long; mean length in parental species—26.40 (16–38) μm and in hybrids—28.50 (20–36) μm (Table 2). On average, the length of colpi in parental species comprised 84% of the polar axis length and in hybrids—86%. Therefore, parental species, on average, exhibited slightly shorter colpi in comparison with hybrids. Their width was variable and usually greatest in the equatorial region. An endoaperture was absent to clearly-developed. In all studied taxa, exine ornamentations in SEM were granulate or granulate-verrucate, because they were made up, primarily of granules less than 1μm in size (usually measuring from 0.5 to 1μm), and less frequently greater than 1μm verrucae (wart-like elements, broader than high; Hesse et al. [72]) (Figs 1E, 1H, 1M, 2D and 2H). In all the examined taxa, with the exception of Q. pubescens (Fig 1H), small and low granules, usually measuring from 0.1 to 0.3 μm, also occurred profusely. Perforations of varying diameters were minor, scarce to numerous, and sometimes not observed. On average, the percentage share of deformed pollen grains in the samples (from 200 to 1000 grains per taxon) was similar and ranged from 15% in Q. robur and Q. robur × Q. petraea to 25% in Q. petraea × Q. pubescens (Fig 3). The highest frequency of deformed pollen was found in samples of Q. petraea and Q. petraea × Q. pubescens (30%) and the lowest in Q. robur (10%). In parental species, the lowest percentage of deformed pollen grains was observed in samples of Q. robur (10%); 20% in Q. pubescens and 30% in Q. petraea. On average, deformed pollen occurred at frequencies of 15, 20 and 22%, respectively, in the three species. In hybrids, the percentages of deformed pollen grains were: 15% in Q. robur × Q. petraea and 25% in Q. petraea × Q. pubescens. Many well-preserved pollen grains were found in the majority of the samples. The deformations consisted mainly in local ruptures of pollen grains, nearly always in the aperture area, and their slight flattening due to reduced turgor. A small number of pollen grains were burst in the area of apertures and strongly deformed to the extent that they had unusual shapes and outlines caused by almost complete loss of turgor and strong flattening. Our observations in LM and SEM were made on acetolysed grains. The experience of the authors of the article allows for a statement that pollen prepared in such manner are subject to deformation in the process of acetolysis, under the influence of high temperature or impact of concentrated acids, as well as in the course of coating with gold target prior to observations in SEM, and in vaccum in SEM, when stream of electrons falls upon them. Due to such actions pollen burst and in consequence lose turgor. 10.1371/journal.pone.0161762.g003Fig 3 Percentage of deformed pollen grains. Interspecific variability of pollen grains Results of the performed MANOVA indicate, that all taxa were significantly (Wilk’s λ = 0.7984; F36,7482 = 12.87; P < 0.0001) different for all nine traits. The analysis of variance for nine biometric traits [P (F4,2004 = 46.04), E (F4,2004 = 27.94), P/E (F4,2004 = 7.19), Exp (F4,2004 = 19.59), Exe (F4,2004 = 19.02), Exp/P (F4,2004 = 24.89), Exe/E (F4,2004 = 23.44), Le (F4,2004 = 28.66) and Le/P (F4,2004 = 8.89)] confirmed the variability of the tested taxa at the significance level α = 0.001 (Table 2). Mean values and coefficients of variations (CV) for the observed traits indicate high variability among the tested taxa for which significant differences were found in terms of all analysed morphological traits (Table 2, Fig 4). 10.1371/journal.pone.0161762.g004Fig 4 Parallel coordinate plots (PCPs) for five studied taxa and nine traits (P, E, P/E, Exp, Exp/P, Le, Le/P). The performed correlation analysis indicates statistically significant correlation coefficients of 29 out of 36 coefficients (Table 4). In the case of seven pairs of traits, no significant correlation was established of: Exp with P, Exp/P with P/E, Le with Exp, Le with Exe, Le/P with Exp, Le/P with Exe and Le/P with Exp/P. Seventeen out of 29 significantly correlated pairs of traits were characterised by positive correlation coefficients. This means that a value increase of one trait in a given pair leads to a value increase of the second trait. 10.1371/journal.pone.0161762.t004Table 4 The correlation matrix for the observed features. Feature P E P/E Exp Exe Exp/P Exe/E Le Le/P P 1 E 0.394*** 1 P/E 0.492*** -0.598*** 1 Exp 0.006 -0.111*** 0.110*** 1 Exe -0.045* -0.099*** 0.055* 0.851*** 1 Exp/P -0.216*** -0.192*** -0.007 0.971*** 0.842*** 1 Exe/E -0.133*** -0.331*** 0.201*** 0.828*** 0.966*** 0.843*** 1 Le 0.742*** 0.349*** 0.308*** -0.010 -0.038 -0.178*** -0.122*** 1 Le/P 0.045* 0.101*** -0.060** -0.026 -0.014 -0.041 -0.045* 0.701*** 1 See explanations to Table 2. * P<0.05; **P<0.01; ***P<0.001. The greatest differentiation of all the analysed phenotypic traits expressed jointly with the greatest Mahalanobis distance was recorded for the pollen grain of Q. robur × Q. petraea (Table 5). Pollen grains of Q. robur × Q. petraea differed significantly with respect to all the examined traits from the remaining taxa. In turn, the greatest phenotypic similarity was observed for Q. robur and Q. petraea × Q. pubescens (0.915), Q. petraea × Q. pubescens and Q. petraea (0.881) as well as for Q. petraea × Q. pubescens and Q. pubescens (0.821). 10.1371/journal.pone.0161762.t005Table 5 Phenotypic distance between the taxa calculated on the basis P, E, P/E, Exp, Exp/P, Le, Le/P by Mahalanobis distance. Taxon Q. petraea Q. pubescens Q. robur Q. petraea × Q. pubescens Q. robur × Q. petraea Q. petraea 0 Q. pubescens 0.344 0 Q. robur 0.881 0.821 0 Q. petraea × Q. pubescens 0.619 0.599 0.915 0 Q. robur × Q. petraea 2.931* 2.782* 3.089** 3.122** 0 * P<0.05; ** P<0.01. The first two canonical variables accounted for 84.38% of total multivariate variability between species and hybrids (Fig 5). This diagram of the first two canonical variables was used to divide the studied taxa into three groups. The first group comprised Q. petraea, Q. pubescens and Q. petraea × Q. pubescens, the second one included one taxon—Q. robur and the last group also embraced just one taxon Q. ×rosacea, which was very distant from the remaining groups (Fig 5). 10.1371/journal.pone.0161762.g005Fig 5 Distribution of five Quercus taxa studied in the space of two first canonical variables. Interesting results were obtained by the contrast analysis between parental species and their hybrids (Table 6). With respect to P, Le and E features, and to a lesser degree, also P/E, pollen grains of Q. robur × Q. petraea exhibited significantly and considerably higher means in comparison with the mean value of its parental forms (negative value of the contrast). In the case of Exp, Exe, Exp/P, Exp/E traits, mean values for Q. robur × Q. petraea were statistically significantly smaller, than the mean value of Q. robur and Q. petraea. Only for Le/P, there were no statistically significant differences between the mean values for hybrids and the parental forms (Table 6). The Q. pubescens × Q. petraea hybrid was characterised by statistically significantly higher mean values of P, E, Exp, Exe, Exp/E and Exe/E traits than its parental forms. Only for P/E, the Q. petraea × Q. pubescens hybrid outlined a lower mean from parental species (positive contrast value). 10.1371/journal.pone.0161762.t006Table 6 Results of contrasts analysis between parental species and their hybrids. Feature Contrast Q. robur, Q. petraea and Q. robur × Q. petraea Q. petraea, Q. pubescens and Q. petraea × Q. pubescens P -5.97 *** -0.61 * E -3.00 *** -1.38 *** P/E -0.082 *** 0.027 * Exp 0.217 ** -0.187 *** Exe 0.225 ** -0.204 *** Exp/P 0.0119 *** -0.0050 ** Exe/E 0.0105 *** -0.0051 ** Le -5.82 ** -0.57 Le/P -0.023 -0.0016 See explanations to Table 2. * P<0.05; ** P<0.01; *** P<0.001. Fig 6 shows the variability of pollen grain traits of 67 studied Quercus individuals in the configuration of the first two canonical variables. On the graph, the coordinates of the point for particular trees are values of the first and second canonical variables, respectively. The first two canonical variables accounted for 61.75% of the total multivariate variability between individual trees. The goal of the study was to establish whether pollen grains collected from various oak trees growing in different habitat conditions (soil, climate) would differ from one another. Six groups of trees were distinguished (Fig 6). The majority of the examined individuals was found in the first group (I). To the other five groups (II-VI) belongs a few trees (II—14, 22, 23 36, 48 and 63, III—19–21, 24, 25, 30, 31, 49, 61, IV—44, 45, V—7, 46, 58, VI—64 (Fig 6). The analysis of the sites, from which flowers (pollen grains) from individual oak trees were collected, has shown, that in individual groups, both trees derived from the same sites [e.g. in group I, occur all analysed Q. petraea trees from Rokita (41–43) or nearly all Q. pubescens trees from Bielinek (51–57, 59–60)] as well as from places geographically distant from one another [e.g. from Austria—Q. pubescens (62) or from Poland—Q. robur from distant Białowieża and Bukowa Primeval Forest. A similar situation occurred also in smaller groups, for example, in group V—each of the three trees represents a different species derived from a different place (Q. robur from Wielkopolski National Park—7, Q. petraea from Bukowa Primeval Forest—46, Q. pubescens from Bielinek—58). Only group IV is made up of two oaks derived from the same place—Bukowa Primeval Forest. 10.1371/journal.pone.0161762.g006Fig 6 Distribution of 67 Quercus trees studied in the space of two first canonical variables. Discussion Palynological investigations on pollen grain features of parental species and their interspecific hybrids focus on comparing pollen size and much rarely involve proportions of deformed pollen grains in both these groups. According to some palynologists, hybrids have significantly larger pollen grains than those of their parents [58, 64, 81–83]. Also Quercus taxa investigated in the present study belong to this group because—as in the case of mean values of P and E features (pollen size), as well as for individual taxa—hybrids had greater pollen grains than parental species. Other researchers proved that hybrids can have pollen size similar or smaller than their parents [53, 55, 57, 60, 62, 66, 84, 85]. Last but not the least, there are also cases where some hybrids are characterised by pollen grains larger than parental species, while others—smaller [59, 61, 64]. Among the studied parental Quercus species, it was found that Q. petraea and Q. pubescens exhibited pollen grains most similar to each other. Q. robur differed from them on average, smaller pollen size and greater exine thickness (Table 2, Fig 5). In hybrids, the dissimilarity of Q. robur × Q. petraea pollen grain features was more conspicuous than in all the remaining taxa. It is worth emphasising, that the oak from which the pollen grains derived, exhibited quite distinct hybrid morphological features. On the basis of contrast analysis, this taxon had the largest pollen grains of longest colpi, significantly bigger with respect to P, E, P/E and Le traits than the mean value of its parental forms. At the same time, it exhibited a fairly thin exine; therefore, mean values of traits associated with it (Exp, Exe, Exp/P, Exp/E) for Q. robur × Q. petraea were smaller in comparison with Q. robur and Q. petraea. Hybrid Q. petraea × Q. pubescens, even though, did not distinguish itself so clearly as Q. robur and Q. petraea (Fig 5). It was also characterised by larger mean values of nearly all analysed traits than in parental forms, including exine features (P, E, Exp, Exe, Exp/E and Exe/E) (Table 6). The hybrids derived from Bielinek on the Oder (NW Poland) to hybrids between Q. petraea and Q. pubescens. The phenomenon of crossing of Q. pubescens mainly with Q. petraea in a peculiar, strongly isolated as well as most distant population of this species from its dense range in Bielinek on the Oder was stressed by Staszkiewicz [86], Danielewicz et al. [87] as well as Krzakowa et al. [88]. However, in recent years, on the basis of genetic analyses employing 14 nuclear microsatellites as markers, it was found that degree of relationship between Q. pubescens individuals was considerable. It implies that crossing in the population occurs, to a large extent, between related individuals and, to a lesser degree, with other species [89]. This, by no means, indicates that interspecific hybrids do not occur there at all, but shows their smaller frequency. Despite numerous palynological studies, descriptions of several important morphological features of Quercus pollen grains are not clear. This refers, in particular, to a very diverse, among representatives of this genus, exine ornamentation but also to endoaperture types. Also, data regarding perforation numbers are not accurate [38, 43, 62]. Palynologists give different types of exine ornamentation in SEM in different representatives of the Quercus genus. It can be either micro-rugulate, scabrate or scabrate-verrucate with verrucae beset with small, rounded processes [36], scabrate, microscabrate or microverrucate-scabrate [90], granulate, scabrate and microgranulate [62], granulate [72], verrucate or microverrucate [38], psilate-verrucate, verrucate, scabrate, scabrate-verrucate and psilate-scabrate [43] or microverrucate to verrucate, rarely regulate-granulate [90]. The three study species were either granulate or granulate-verrucate. This type of exine ornamentation has been selected because it is made primarily, of granules which are less than 1 μm (usually—05–1 μm), whereas verrucae exceed 1 μm [72]. In all studied taxa, with the exception of Q. pubescens, besides larger granules and verrucae also numerous smaller granules, commonly measuring 0.1–0.3 μm, occurred. Dissimilarity of Q. pubescens exine ornamentation compared to all studied Quercus taxa is also corroborated by Benthem et al. [36], Colombo et al. [17] and Smit [44]. The results of the present study agree with Benthem et al. [36] as well as Hesse et al. [72] that colpate or colporoidate pollen grains occurred in the investigated Quercus taxa. The latter ones were composed of a colpus (ectoaperture) with an indistinct endoaperture. Some researchers maintain, that both poroides as well as pori [43], or only pori [17, 44, 91–92] occur here, while others—in some part of the species (e.g. Q. pubescens Willd., Q. aristata Hook. & Arn., Q. dumosa Nutt., Q. laurina Bonpl.)—also find absence of the endoapertures [17, 47, 48]. The number of perforations differs, depending on authors; some report their total absence or only a few and others mention many with differing diameters and distributions on pollen grains [17, 36, 43, 47, 48, 91]. Results of this study corroborate the above observations; perforations were small, scarce or numerous, sometimes they could not be seen at all. They had different diameters and were usually distributed irregularly. The results of statistical analyses are not unequivocal both with respect to the share of the 67 individuals (oak trees) in 6 groups to which they were assigned as well as to places of their collection. The majority of the investigated individuals belonged to the first, large group, while the remaining ones occurred from single to several oak trees in the other five groups. In these groups, both individuals from different taxa (e.g. group 5 is made up of three individuals—Q. petraea, Q. pubescens and Q. robur) as well as trees representing the same taxa (e.g. group 4—Q. petraea) were found. Flowers from five Q. pubescens trees with traits typical for this species (51–55) were collected from the site in Bielinek Reserve (Poland) and, for comparison, from five other Q. pubescens trees (56–60) derived from Bielinek and growing in the Dendrological Garden of Poznań University of Life Sciences. Pollen grains of all these trees were similar to one another, because almost all of them were found in the same group—group 1 and only one tree (58) belonged to group 5. A similar situation was observed with geographical distribution. In the same group occurred both trees growing in the same site (e.g. in group 1—three Q. petraea trees from Rokita) as well as oaks derived from geographically distant sites (in group 2—Q. robur from Spain and from Dębno in Poland). Not much information can be found in the literature on the subject concerning deformed pollen grains. Some palynologists maintain that the share of deformed pollen grains is greater in hybrids, than in parental species. Karlsdóttir et al. [60] reported that the in natural birch hybrids it was found two to three times more abnormal pollen than in parental species Betula nana and B. pubescens. However, in three parental Crataegus species and in their three natural hybrids, deformed pollen grains occurred with similar frequencies (20–40%) [64]. The results on Quercus pollen grains reported here are similar. The proportions of deformed pollen in parental Quercus species and hybrids were similar and, on average, represented by ap to 15–25%. Recapitulating, it was to be expected that not all of the closely related species of oaks can be safely distinguished using pollen morphology. In spite of such close relationships of the examined Quercus taxa, it was, nevertheless, possible to identify two (Q. robur × Q. petraea and Q. pubescens) from among five taxa on the basis of several analysed pollen features. Pollen size can be used as an auxiliary feature when diagnosing Quercus parental species and hybrids. Pollen shape is an interesting, hitherto omitted trait, which distinguishes both hybrids, especially Q. robur × Q. petraea characterised by the most elongated pollen grains. On the basis of exine ornamentation, it was possible to identify only Q. pubescens; the remaining species as well as hybrids did not differ significantly with regard to this feature. Only a greater number of such studies, based on large pollen samples, will show if there really is signal in pollen shape or exine ornamentation to tell species and hybrids apart. The determination of the diagnostic value of endoapreture features, i.e. their type (pori, poroides or both of these aperture types) as well as their presence or absence requires further comprehensive palynological investigations. The authors are grateful to the employees of Hajnówka, Dębno, Jarocin, Rokita, and Różańsko Forest Districts for support in the collection of plant materials. We would like to thank Professor Jean Bernard Diatta for linguistic support. We would like to thank the Reviewers for their detailed and valuable comments on the manuscript. ==== Refs References 1 Camus A . Les chênes: monographie du genre Quercus Encyclopédie économique de silviculture, 6–8 . Paris : Paul Lechevalier and Fils ; 1936 –1954 . 2 Elias TS . The genera of Fagaceae in the southeastern United States . J Arnold Arbor . 1971 ; 52 : 159 –195 . 3 Lawrence GHM . Taxonomy of Vascular Plants . New York : The Macmillan Company ; 1951 . 4 Nixon CK , Jensen RJ , Manos P , Muller CH . Quercus In: Flora of North America, North of Mexico. 3 Magnoliophyta: Magnoliidae and Hamamelidae . Oxford : University Press Nueva York ; 1997 pp. 445 –447 . 5 Trelease W . The American oaks . Mem Natl Acad Sci . 1924 ; 20 : 1 –255 . 6 Govaerts R , Frodin DG . World Checklist and Bibliography of Fagales (Betulaceae, Corylaceae, Fagaceae and Ticodendraceae) . Kew : Royal Botanical Gardens ; 1998 . 7 Manos PS , Doyle JJ , Nixon KC . Phylogeny, biogeography, and processes of molecular differentiation in Quercus subgenus Quercus (Fagaceae) . Mol Phylogenet Evol . 1999 ; 12 : 333 –349 . 10413627 8 Jones JH . Evolution of the Fagaceae: the implications of foliar features . Ann Missouri Bot Gard . 1986 ; 73 : 228 –275 . 9 Soepadmo E . Fagaceae . Flora Malenesia 1972 ; 1 (7 ): 265 –403 . 10 Axelrod DI . Biogeography of oaks in the Arcto-Tertiary province . Ann Miss Bot Gard . 1983 ; 70 : 629 –657 . 11 Menitsky YL . Oaks of Asia . USA : Science Publishers, Enfield ; 2005 . 12 Miller H , Lamb SH . Oaks of North America . USA : Naturegraph Publishers ; 1985 . 13 Nixon KC . Infrageneric classification of Quercus (Fagaceae) and typification of sectional names . Ann For Sci . 1993 ; 50 : 25 –34 . 14 Valencia SA . Diversidad del genero Quercus (Fagaceae) en Mexico . Bol Soc Bot Méx . 2004 ; 75 : 33 –54 . 15 Aas G . Taxonomical impact of morphological variation in Quercus robur and Q. petraea: a contribution to the hybrid controversy . Ann For Sci . 1993 ; 50 (1 ): 107 –113 . 16 Burger WC . The species concept in Quercus . Taxon . 1975 ; 24 (1 ): 45 –50 . 17 Colombo PM , Lorenzoni FC , Grigoletto F . Pollen grain morphology supports the taxonomical discrimination of Mediterranean oaks (Quercus, Fagaceae) . Plant Syst Evol . 1983 ; 141 : 273 –284 . 18 Palmer EJ . Hybrid oaks of North America . J. Arnold Arbor . 1948 ; 29 : 1 –48 . 19 Petit R , Bodenes C , Ducousso A , Roussel G , Kremer A . Hybridization as a Mechanism of Invasion in Oaks . New Phytol . 2004 ; 161 : 151 –164 . 20 Rushton BS . Natural hybridization within the genus Quercus L . Ann For Sci . 1993 ; 50 (1 ): 73 –90 . 21 Tucker JM . Patterns of parallel evolution of leaf form in new world oaks . Taxon 1974 ; 23 (1 ): 129 –154 . 22 De Candolle AP . Prodromus Systematis Naturalis Regni Vegetabilis . Part 16 (2 ). Paris : Sumptibus Victoris Masson et Filii ; 1868 . 23 Van Valen L . Ecological species, multispecies, and oaks . Taxon 1976 ; 2/3 : 233 –239 . 24 Schwarz O . Monographie der Eichen Europas und des Mittelmeergebietes Repertorium specierum novarum regni vegetabilis . Berlin-Dahlem : Sonderbeiheft D. ; 1937 . 25 Schwarz O . Quercus L. In: Tutin TG , Heywood VH , Burges NA , Valentine DH , Walters SM , Webb DA , editors. Flora Europaea . Cambridge : Cambridge University Press ; 1964 pp. 61 –64 . 26 Denk T , Grimm GW . The oaks of western Eurasia: traditional classifications and evidence from two nuclear markers . Taxon 2010 ; 59 : 351 –366 . 27 Muir G , Fleming CC , Schltterer C . Taxonomy: Species status of hybridizing oaks . Nature 2000 ; 405 : 1016 doi: 10.1038/35016640 10890434 28 Rehder A . Manual of cultivated trees and shrubs. Second edition revised and enlarged . MacMillan Company , New York 1940 pp 996 . 29 Viscosi V , Lepais O , Gerber S , Fortini P . Leaf morphological analyses in four European oak species (Quercus) and their hybrids: A comparison of traditional and geometric morphometric methods . Plants Biosystems 2009 ; 143 : 564 –574 . 30 Dering M , Lewandowski A . Unexpected disproportion observed in species composition between oak mixed stands and their progeny populations . Ann For Sci . 2007 ; 64 : 413 –418 . 31 Dupouey JL , Badeau V . Morphological variability of oaks (Quercus robur L., Q. petraea (Matt) Liebl., Q. pubescens Willd.) in northeastern France. Preliminary results . Ann Sci Forest . 1993 ; 50 (1 ): 35 –40 . 32 Kremer A , Dupouey JL , Deans JD , Cottrell J , Csaikl U , Finkeldey R , et al Leaf morphological differentiation between Quercus robur and Quercus petraea is stable cross western European mixed oak stands . Ann Sci Forest . 2002 ; 59 : 777 –787 . 33 Curtu AL , Gailing O , Finkeldey R . Evidence for hybridization and introgression within a species-rich oak (Quercus spp.) community . BMC Evol Biol . 2007 ; 7 : 218 17996115 34 Jedináková-Schmidtová J , Paule L , Magic D , Gömöry D . Morphological and genetic differentiation among the Central European white oaks . Forest Genetics 2004 ; 11 (3–4 ): 263 –271 . 35 Gömöry D , Schmidtová J . Extent of nuclear genome sharing among white oak species (Quercus L. subgen. Lepidobalanus (Endl.) Oerst.) in Slovakia estimated by allozymes . Plant Syst Evol . 2007 ; 266 : 253 –264 . 36 Benthem F , Clarke GCS , Punt W . The Northwest European pollen flora 33. Fagaceae . Rev Palaeobot Palynol . 1984 : 42 (1/4 ): 87 –110 . 37 Cao M , Zhou ZK . Pollen morphology and its systematic significance of the Quercus from China . Guihaia 2002 ; 22 (1 ): 14 –18 . 38 Denk T , Grimm GW . Significance of pollen characteristics for infrageneric classification and phylogeny in Quercus (Fagaceae) . Intern J Plant Scie . 2009 ; 170 : 926 –940 . 39 Denk T , Tekleva MV . Pollen morphology and ultrastructure of Quercus with focus on Group Ilex (= Quercus subgenus Heterobalanus (Oerst.) Menitsky): implications for oak systematics and evolution . Grana 2014 ; 53 (4 ): 255 –282 . 40 Makino M , Hayashi R , Takahara H . Pollen morphology of the genus Quercus by scanning electron microscope . Sci Rep Kyoto Pref Univ . 2009 ; 61 : 53 –81 . [In Japanese] 41 Miyoshi N . Pollen morphology of Japanese Quercus (Fagaceae) by means of scanning electron microscope . Jap J Palyn . 1981 ; 27 : 45 –54 . 42 Ostrolucka MG . Morphological characteristics of pollen grains of the species of genus Quercus . Fol. Dendro . 1984 ; 11 : 233 –261 . 43 Panahi PP , Pourmajidian MR , Fallah A , Pourhashemi M . Pollen morphology of Quercus (subgenus Quercus, section Quercus) in Iran and its systematic implication . Acta Soc Bot Pol . 2012 ; 81 ,1 : 33 –41 . 44 Smit A . A scanning electron microscopical study of the pollen morphology in the genus Quercus . Acta Bot Neerl . 1973 ; 22 (6 ): 655 –665 . 45 Solomon AM . Pollen morphology and plant taxonomy of white oaks in eastern North America . Am J Bot . 1983a ; 70 : 481 –494 . 46 Solomon AM . Pollen morphology and plant taxonomy of red oaks in eastern North America . Am J Bot . 1983b ; 70 : 495 –607 . 47 Medus J , Flores GG . Pollen morphology of some Mexican oaks . Grana 1984 ; 23 (2 ): 77 –84 . 48 Diethart B. Quercus robur. In: PalDat—a palynological database. 2005; Available: https://www.paldat.org/pub/Quercus_robur/108787. 49 Deng M , Song YG , Li QJ , Li QS . Pollen morphology of Quercus subg. Cyclobalanopsis (Fagaceae) and its systematic implication . Guihaia 2013 ; 33 (3 ): 368 –375 [in Chinese]. 50 Dupont P , Dupont S . Etude de pollens de chênes (genre Quercus L.) en microscopie électronique à balayage . Compt Rend Acad des Sciences . 1972 ; 274 (17 ): 2503 –2506 . 51 Olsson U . On the size and microstructure of pollen grains of Quercus robur and Q. petraea (Fagaceae) . Bot Notiser . 1975 ; 128 (2 ): 256 –264 . 52 Savitskii VD , Martinyuk OO , Shumik MI . Palynomorphological features of species of the genus Quercus L. in Ukraine . Ukr Bot Zhurn . 1999 ; 56 (1 ): 33 –36 . 53 Chaturvedi M , Ram T , Pal M . Pollen morphology in Chorisia species and their hybrid . Phytomorphology 1993 ; 43 (1/2 ): 25 –28 . 54 Datta K , Chaturvedi M ., Ram T . Pollen exine ornamentation in the F2 generation of an interspecific hybrid of Chorisia (Bombacoideae, Malvaceae) in relation to inheritance pattern . Grana 2006 ; 45 (2 ): 109 –114 . 55 Franssen AS , Skinner DZ , Al-Khatib K , Horak MJ . Pollen morphological differences in Amaranthus species and interspecific hybrids . Weed Science 2001 ; 49 (6 ): 732 –737 . 56 Kumar CR , Nair PKK . Inheritance of exine ornamentation and pollen shape in the interspecific tetraploid hybrids of Gloriosa . Can J Bot . 2011 ; 64 : 3134 –3140 . 57 Lu XM , Chen JY , Zhang LM , Zheng SQ , Yu D , Liao RY . Observation and comparison on pollen morphology of a new hybrid loquat variety 'Zaozhong 6' and its parentals . Acta Hort Sin . 2002 ; 29 (3 ): 271 –273 . 58 Rhee HK , Cho HR , Kim KJ , Kim KS . Comparison of pollen morphology in interspecific hybrid lilies after in vitro chromosome doubling . Acta Hort . 2005 ; 673 : 639 –643 . 59 Rushton BS . Pollen grain size in Quercus robur L. and Quercus petraea (Matt. Leibl.) . Watsonia 1976 ; 11 : 137 –140 . 60 Karlsdóttir L , Hallsdóttir M , Thórsson AT , Anamthawat-Jónsson K . Characteristics of pollen from natural triploid Betula hybrids . Grana 2008 ; 47 (1 ): 52 –59 . 61 Lazarević M , Siljak-Yakovlev S , Lazarević P , Stevanović B , Stevanović V . Pollen and seed morphology of resurrection plants from the genus Ramonda (Gesneriaceae): relationship with ploidy level and relevance to their ecology and identification . Turk J Bot . 2013 ; 37 : 872 –885 . 62 Scareli-Santos C , Herrera-Arroyo ML , Sánchez-Mondragón ML , González-Rodríguez A , Bacon J , Oyama K . Comparative analysis of micromorphological characters in two distantly related Mexican oaks, Quercus conzattii and Q. eduardii (Fagaceae), and their hybrids . Brittonia 2007 ; 59 (1 ): 37 –48 . 63 Srivastava V , Pal M , Nair PKK . A study of the pollen grains of Amaranthus spinosus Linné and A. dubius Mart ex Thellung and their hybrids . Rev Palaeobot Palynol . 1977 ; 23 (4 ): 287 –291 . 64 Wrońska-Pilarek D , Bocianowski J , Jagodziński AM . Comparison of pollen grain morphological features of selected species of the genus Crataegus L. (Rosaceae and their spontaneous, interspecific hybrids . Bot J Linn Soc . 2013 ; 172 : 555 –571 . 65 Yang Qing Hua . Pollen morphology of Osmanthus decorus and O. ×burkwoodii . J Hubei Univ . 2010 ; 28 (3 ): 286 –288 . [in Chinese]. 66 Ohashi H , Hoshi H ., Iketani H . Taxonomy and pollen morphology of hybrids between Sorbus and Micromeles in the genus Sorbus (Rosaceae subfamily Maloideae) . J Jap Bot . 1991 ; 66 (2 ): 110 –124 . 67 Manos PS , Zhou Z , Cannon CH . Systematics of Fagaceae: phylogenetic tests of reproductive trait evolution . Int J Plant Sci 2001 ; 162 :1361 –1379 . 68 Hubert F , Grimm GW , Jousselin E , Berry V , Franc A , Kremer A . Multiple nuclear genes stabilize the phylogenetic backbone of the genus Quercus . Syst. biodivers . 2014 ; 12 : 405 –423 . 69 Wrońska-Pilarek D , Jagodziński AM , Bocianowski J , Janyszek M . The optimal sample size in pollen morphological studies using the example of Rosa canina L.–Rosaceae . Palynology 2015 ; 39 (1 ): 56 –75 . 70 Erdtman G. The Acetolysis Method. A Revised Description . Svensk Bot Tidskr . 1960 ; 54 : 561 –564 . 71 Punt W , Hoen PP , Blackmore S , Nilsson S , Le Thomas A . Glossary of pollen and spore terminology . Rev Palaeobot Palynol . 2007 ; 1431 (2 ): 1 –81 . 72 Hesse M , Halbritter H , Zetter R , Weber M , Buchner R , Frosch-Radivo A , et al Pollen Terminology An illustrated handbook . Vienna : Springer ; 2009 . 73 Shapiro SS , Wilk MB . An analysis of variance test for normality (complete samples) . Biometrika 1965 ; 52 : 591 –611 . 74 Bocianowski J , Szulc P , Nowosad K . Parallel coordinate plots of maize traits under different magnesium applications . J Integrat Agric . 2015 ; 14 (3 ): 593 –597 . 75 Kozak M . Use of parallel coordinate plots in multi-response selection of interesting genotypes . Comm Bio Crop Science . 2010 ; 5 : 83 –95 . 76 Rencher AC . Interpretation of canonical discriminant functions, canonical variates, and principal components . Am Stat . 1992 ; 46 : 217 –225 . 77 Seidler-Łożykowska K , Bocianowski J . Evaluation of variability of morphological traits of selected caraway (Carum carvi L.) genotypes . Industrial Crops Prod . 2012 ; 35 : 140 –145 . 78 Mahalanobis PC . On the generalized distance in statistics . Proc Natl Inst Sci India . 1936 ; 12 : 49 –55 . 79 Payne RW , Murray DA , Harding SA , Baird DB , Soutar DM . Introduction to GenStat for Windows (15th Edition ). UK : VSN International ; 2012 . 80 Erdtman G . Pollen morphology and plant taxonomy. Angiosperms An introduction to palynology 1 Stockholm : Almquist and Wiksell ; 1952 . 81 Rizaeva SM , Akhmedova MZh , Abdullaev AA . Pollen viability and pollen grain morphology in interspecific cotton hybrids differing in origin and ploidy . Selsk Biol . 1985 ; 9 : 63 –66 [in Russian]. 82 Polevova SV , Kosenko YV , Leunova VM , Romanova ES , Severova EE , Tekleva MV . Pollen morphology of apple species and cultivars (Malus, Rosaceae) . Bot Zhurn . 2014 ; 99 (12 ): 1317 –1335 [in Russian]. 83 Van der Walt ID , Littlejohn GM . Pollen morphology, male hybrid fertility and pollen tube pathways in Protea . S Afr J Bot . 1996 ; 62 : 236 –246 . 84 Delaporte K , Conran J , Sedgley M . Morphological analysis to identify the pollen parental of an ornamental interspecific hybrid Eucalyptus . Hort Sci . 2001 ; 89 (1 ): 55 –72 . 85 Olsson U . A biometric study of the pollen morphology of Linaria vulgaris L. Miller and L. repens L. Miller (Scrophulariaceae) and their hybrid progeny in F1 and F2 generations . Grana 1974 ; 14 (2–3 ): 92 –99 . 86 Staszkiewicz J . The systematical position of Quercus pubescens (pubescent oak) from the wood-land- steppe reserve in Bielinek on the Odra river basing it on the biometric analysis of the leaves . Fragm. Flor. et Geobot . 1977 ; 23 (304 ): 259 –273 [in Polish]. 87 Danielewicz W , Bąkowska K , Krzak M . Variability of downy oak (Quercus pubescens Willd.) marginal population in Bielinek (north-western Poland) in marginal traits of leaves . Rocz Dendr . 2002 ; 50 : 43 –48 . 88 Krzakowa M , Bąkowska K , Danielewicz W . Genetic variation in a marginal population of pubescent oak (Quercus pubescens Willd.) in Bielinek, on the Odra riverside . Eco Quest . 2004 ; 4 : 77 –82 . 89 Chybicki IJ , Oleksa A , Kowalkowska K , Burczyk J . Genetic evidence of reproductive isolation in a remote enclave of Quercus pubescens in presence of cross-fertile species . Plant Syst Evol . 2012 ; 298 : 1045 –1056 . 90 Nakagawa T , Yasuda Y , Tabata H . Pollen morphology of Himalayan Pinus and Quercus and its importance in palynological studies in Himalayan area . Rev Palaeobot Palynol . 1996 ; 91 : 317 –329 . 91 Tekleva MV , Naryshkina NN , Evstigneeva TA . Fine structure of Quercus pollen from the Holocene sediments of the Sea of Japan . Plant Syst Evol . 2014 ; 300 : 1877 –1893 . 92 Yamazaki T , Takeoka M . Electronmicroscope investigations on the surface structure of the pollen membrane, based on the replica method. V. Especially on the pollen of genus Quercus . J Jap Forest Soc . 1959 ; 41 : 125 –129 .

PMC005xxxxxx/PMC5001651.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756372510.1371/journal.pone.0161050PONE-D-15-44668Research ArticleBiology and Life SciencesOrganismsAnimalsInvertebratesMolluscsGastropodsSnailsEarth SciencesMarine and Aquatic SciencesBodies of WaterLakesEcology and Environmental SciencesAquatic EnvironmentsFreshwater EnvironmentsLakesEarth SciencesMarine and Aquatic SciencesAquatic EnvironmentsFreshwater EnvironmentsLakesBiology and Life SciencesEvolutionary BiologyPopulation GeneticsPloidyBiology and Life SciencesGeneticsPopulation GeneticsPloidyBiology and Life SciencesPopulation BiologyPopulation GeneticsPloidyBiology and Life SciencesZoologyMalacologyBiology and Life SciencesOrganismsBacteriaBiology and Life SciencesMicrobiologyMedical MicrobiologyMicrobiomeBiology and Life SciencesGeneticsGenomicsMicrobial GenomicsMicrobiomeBiology and Life SciencesMicrobiologyMicrobial GenomicsMicrobiomeResearch and Analysis MethodsDatabase and Informatics MethodsBiological DatabasesSequence DatabasesBiology and Life SciencesMolecular BiologyMolecular Biology TechniquesSequencing TechniquesSequence AnalysisSequence DatabasesResearch and Analysis MethodsMolecular Biology TechniquesSequencing TechniquesSequence AnalysisSequence DatabasesBiology and Life SciencesOrganismsBacteriaCyanobacteriaDistinct Bacterial Microbiomes in Sexual and Asexual Potamopyrgus antipodarum, a New Zealand Freshwater Snail The Microbiota of Sexual and Asexual Potamopyrgus antipodarumTakacs-Vesbach Cristina 1*King Kayla 2Van Horn David 1Larkin Katelyn 3Neiman Maurine 31 Department of Biology, University of New Mexico, Albuquerque, New Mexico, United States of America2 Department of Zoology, University of Oxford, Oxford, United Kingdom3 Department of Biology, University of Iowa, Iowa City, Iowa, United States of AmericaLarsen Peter E. EditorArgonne National Laboratory, UNITED STATESCompeting Interests: The authors have declared that no competing interests exist. Conceived and designed the experiments: CTV DVH KL MN. Performed the experiments: CTV DVH KL MN. Analyzed the data: CTV KK DVH MN. Contributed reagents/materials/analysis tools: CTV MN. Wrote the paper: CTV KK KL MN. * E-mail: cvesbach@unm.edu26 8 2016 2016 11 8 e016105010 10 2015 29 7 2016 © 2016 Takacs-Vesbach et al2016Takacs-Vesbach et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Different reproductive strategies and the transition to asexuality can be associated with microbial symbionts. Whether such a link exists within mollusks has never been evaluated. We took the first steps towards addressing this possibility by performing pyrosequencing of bacterial 16S rRNA genes associated with Potamopyrgus antipodarum, a New Zealand freshwater snail. A diverse set of 60 tissue collections from P. antipodarum that were genetically and geographically distinct and either obligately sexual or asexual were included, which allowed us to evaluate whether reproductive mode was associated with a particular bacterial community. 2624 unique operational taxonomic units (OTU, 97% DNA similarity) were detected, which were distributed across ~30 phyla. While alpha diversity metrics varied little among individual samples, significant differences in bacterial community composition and structure were detected between sexual and asexual snails, as well as among snails from different lakes and genetic backgrounds. The mean dissimilarity of the bacterial communities between the sexual and asexual P. antipodarum was 90%, largely driven by the presence of Rickettsiales in sexual snails and Rhodobacter in asexual snails. Our study suggests that there might be a link between reproductive mode and the bacterial microbiome of P. antipodarum, though a causal connection requires additional study. http://dx.doi.org/10.13039/1000001521122176Neiman Maurine This work was supported by NSF-MCB grant 1122176 and the Iowa Center for Research by Undergraduates. Data AvailabilityAll raw sequence data from this study are available through the NCBI Sequence Read Archive. The individual sff files from this study were assigned the accession numbers SAMN03276406 through SAMN03276465 under Bioproject PRJNA271685 (http://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA271685).Data Availability All raw sequence data from this study are available through the NCBI Sequence Read Archive. The individual sff files from this study were assigned the accession numbers SAMN03276406 through SAMN03276465 under Bioproject PRJNA271685 (http://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA271685). ==== Body Introduction The production of offspring is one of the most important determinants of organismal fitness. Despite the tight links between reproduction and fitness, the mechanisms of offspring production are characterized by remarkable inter- and intraspecific variation. It is increasingly evident that different reproductive strategies, such as host mating behaviour [1–3] and the production of viable offspring [4], can be associated with the composition of an organism’s microbiome. The transition from sexual to asexual reproduction can be driven by infection of sexual females with endosymbiotic bacteria [5–10], and links between asexuality and bacterial endosymbionts have been established in many arthropod taxa (reviewed in [5–7,9–11]) and in nematodes [12] (and reviewed in [5]). Together, these studies demonstrate that the microbiome is an integral component of host reproductive biology for at least some animals, though connections between microbial communities and reproductive mode in other animal taxa remain largely unexplored. Here, we compare the microbiome community structure between obligately sexual and obligately asexual Potamopyrgus antipodarum, a New Zealand freshwater snail, to determine if reproductive mode variation in this system might be associated with a particular bacterial community. Potamopyrgus antipodarum is a powerful system for evaluating potential connections between reproductive mode and microbiota because natural populations of this snail vary widely in the relative frequency of obligately sexual individuals (male and female) and obligately asexual individuals (nearly always female; males are produced by asexual females at a rate of 1–2% of all offspring [13–15]). Like many mixed sexual/asexual animal taxa (reviewed in [16]), asexual P. antipodarum are typically polyploid (triploid and tetraploid [17]), while sexuals are diploid [18]. The rare occurrence of diploid asexual P. antipodarum suggests that polyploidy is a consequence rather than a cause of asexuality in P. antipodarum, with polyploidy arising via fertilization of the unreduced asexually produced eggs produced by asexual females [17]. Phylogeographic [17,19] and population genetic [20,21] data indicate that asexual lineages of P. antipodarum have evolved multiple times from sexual populations of P. antipodarum, thereby providing a replicated system to evaluate links between reproductive mode and microbiome composition within a single species. Materials and Methods Host Tissue Collection, Identification, and Preparation We used barcoded amplicon pyrosequencing of 16S rRNA genes to characterize the bacterial communities associated with a diverse array of laboratory-cultured and field-collected P. antipodarum. Laboratory-cultured snails had been housed at the University of Iowa for at least 2–3 generations; the eight laboratory lineages that we used (two sexual diploid, “2x”; three asexual triploid, “3x”; three asexual tetraploid, “4x”) were each descended from the offspring produced by a single field-collected female that originated from the littoral zone (maximum depth 1m) of the South Island New Zealand lakes Alexandrina, Gunn, Kaniere, Poerua, Rotoiti, and Taylor (Table 1). Field-collected P. antipodarum samples in this study were collected in January 2011 from the South Island New Zealand lakes Ianthe, Kaniere, and Sarah (maximum depth 1m). We made an effort to select laboratory lineages and field collections that represented genetically and geographically distinct source populations and that allowed us to compare separately derived asexual lineages [17,21]. We hereafter refer to each distinct laboratory lineage and field collection as a “population”. 10.1371/journal.pone.0161050.t001Table 1 Characteristics of the 11 populations and sample types used in this study. Lakea,b Latitude Longitude Sample Type Reproductive Mode Ploidy # ♀ headb # ♀ bodyb #♂ headb #♂ bodyb Juvenilec Total Alexandrina 43.9500°S 170.4500°E Laboratory lineage Sex 2x 1 1 1 1 0 4 Gunn 44.8833°S 168.0833°E Laboratory lineage Asex 4x 1 1 1 1 0 4 Gunn 44.8833°S 168.0833°E Laboratory lineage Asex 4x 1 1 1 1 0 4 Ianthe 43.0500°S 170.6167°E Field collection Sex 2x 4 4 4 4 0 16 Kaniere 42.8333°S 171.1500°E Field collection Sex 2x 1 1 1 1 0 4 Kaniere 42.8333°S 171.1500°E Laboratory lineage Sex 2x 1 1 1 1 0 4 Poerua 42.7000°S 171.5000°E Laboratory lineage Asex 3x 1 1 1 1 0 4 Poerua 42.7000°S 171.5000°E Laboratory lineage Asex 3x 1 1 1 1 0 4 Rotoiti 38.0390°S 176.4277°E Laboratory lineage Asex 4x 1 1 1 1 0 4 Sarah 43.0500°S 171.7667°E Field collection Asex 3x 4 0 0 0 4 8 Taylor 42.4500°S 172.1600°E Laboratory lineage Asex 3x 1 1 1 1 0 4 60 aRepresents lake of origin for founding female of laboratory lineage and lake of collection for field-collected samples. bEach sample except those from Ianthe and Sarah represents tissue from 3 pooled individual snails; head and body tissue subsamples were analyzed separately except for juveniles, which are so small that we combined head and body tissue for DNA extraction. Ianthe and Sarah snails were analyzed individually. cJuveniles are too young to sex. These juveniles were triploid and as such were presumed to be female. One head half and the body tissue were combined for each juvenile to ensure that we had enough tissue for DNA extraction. We arbitrarily selected three males and three females (i.e., six snails per lineage) from each of the eight laboratory lineages (two sexual and six asexual, including two distinct asexual lineages from Gunn and two distinct asexual lineages from Poerua) and the Kaniere (sexual) field-collected population. Similarly, we arbitrarily selected four males and four females from the Ianthe (sexual) field-collected sample. We were unable to find any males in the several hundred adults from the sample collected from lake Sarah (suggesting that this sample is wholly asexual; also see [17]), and thus instead included four arbitrarily selected females from this sample. We followed the same procedures for four juvenile (<2 mm in shell length) snails from the Sarah field collection, with the exception that juvenile P. antipodarum cannot be reliably sexed (e.g., [22]). Because nearly all offspring produced by asexual polyploid female P. antipodarum are female [15], and because the lake Sarah population appears to be comprised only of triploid asexuals [17] (also see Table 1), it is likely (>92%, conservatively assuming that 2% of all triploids are male; [15]) that all of these Sarah-collected juveniles were female. Each snail was sexed, with males distinguished from females by the presence of a penis. We then removed the shell and dissected each snail into “head” (anterior half, i.e., no reproductive tissue) and “body” (containing reproductive tissue along with digestive tissue) tissue subsamples because asexuality-causing endosymbionts such as Wolbachia typically reside in tissues associated with vertical transmission (e.g., [23]; reviewed in [9]; but see [24]). This dissecting approach allowed us to compare microbiome profiles in tissue with gonadal subcomponents to non-gonadal tissue. The head samples were dissected again in half, such that we then had three tissue samples that together encompassed one whole snail: two head halves and one body. Each tissue sample was then snap-frozen in liquid nitrogen and stored at -80°C until further processing. We dissected, extracted, and sequenced individual snails in three batches, each of which included both sexuals and asexuals and multiple asexual lineages, minimizing the potential that batch effects figured in the patterns we detected among our samples. We next used flow cytometric analysis (following the protocol outlined in [17,21,25]) of the tissue from one head half per each of the field-collected snails to determine ploidy and thus infer reproductive mode of the samples [26]. As expected from earlier studies [17,21], all of the samples from Ianthe and Kaniere were diploid (sexual) and all of the samples from Sarah were triploid (asexual). Finally, we pooled tissue samples by sex and body section from three individuals in a single 1.5 mL Eppendorf tubes, such that each tube contained three tissue samples of the same type (head half vs. body) from each of three snails of the same sex, of the same ploidy level, and from the same population. While we pooled most tissue samples in order to ensure that we obtained enough DNA for sequencing, we did not pool tissue samples from the Ianthe and Sarah field collections (i.e., each sample contained the head or body tissue from one individual), which allowed us to investigate if any individual-level differences were detectable in P. antipodarum bacterial communities. Additional deviations from the standard protocol for the Sarah P. antipodarum included the use of only head tissue for the adults (no body tissue was available, which was used for another project) and the fact that juveniles cannot be sexed and can thus not be separated by sex. Because the Sarah juveniles were so small, we combined the remaining head half and body tissue for DNA extractions for each individual juvenile to ensure that we had enough tissue for DNA extraction. With the exception of these four juveniles, all samples were from adult snails. A detailed description of each sample is provided in Table 1 and S1 Table. DNA Sequencing and Analysis We used the Qiagen DNeasy Plant kit (Qiagen, Valencia, CA, USA) following the manufacturer-specified protocol (with the exception of using nanopure water for elution) to extract DNA from each of the 60 tissue samples. Pyrosequencing was performed as described previously [27], using the universal bacterial primers 104F 5’-GGCGVACGGGTGAGTAA-3’ and 530R 5’-CCGCNGCNGCTGGCAC-3’ to target the V2-V3 region of the 16S rRNA gene [28]. For each sample, PCR was performed in triplicate with 100 ng of DNA by a single-step PCR to create 16S rRNA gene amplicons containing the Roche-specific sequencing adapters and a barcode unique to each sample. Amplicons were purified using Agencourt Ampure beads and combined in equimolar concentrations. Pyrosequencing was performed on a Roche 454 FLX instrument using Roche titanium reagents and procedures. The 16S rRNA gene sequences were quality filtered, denoised, screened for PCR errors, and checked for chimeras using AmpliconNoise and Perseus to minimize potential methodological artifacts [29]. The Quantitative Insights into Microbial Ecology (QIIME) pipeline was used to analyze alpha and beta diversity of the bacterial DNA sequence data [30]. Unique operational taxonomic units (OTUs, i.e., DNA sequences or amplicon types) were identified by the 97% DNA identity criterion using the uclust OTU picker [31,32] in QIIME. A set of representative DNA sequences was chosen for each unique OTU in QIIME (pick_rep_set.py) and used for all subsequent analyses. Taxonomic affiliation was assigned to OTUs by comparing the rep-set DNA sequences to the Greengenes database (gg8.15.13 [33]). Rep-set DNA sequences were aligned using MUSCLE [32], and a phylogenetic tree necessary for the downstream alpha and beta diversity analysis was constructed using FastTree [34]. Measures of alpha diversity (Chao1, Shannon, dominance, equitability, Faith’s phylogenetic diversity [35], Good’s coverage) and beta diversity (Bray-Curtis and Unifrac distances [36]) were determined on a randomly selected subset of 400 sequences from each sample to standardize for varying sequencing effort across samples. Non-parametric t-tests with 1000 Monte Carlo permutations were used to determine if alpha diversity differed among the samples according to their reproductive mode (2x sexuals vs. pooled 3x and 4x asexuals) or other sample factors including lake of origin, population, population source (laboratory lineage or field-collected), sex, body section, or ploidy level. Bray-Curtis and weighted and unweighted Unifrac [36] distance matrices were generated, and principal coordinate analysis (PCoA) was performed as implemented in QIIME by the beta_diversity_through_plots.py script (-e 400). Select OTUs were compared to their nearest neighbor DNA sequence from Greengenes by aligning with MUSCLE, and a distance matrix was generated using the Jukes-Cantor model. Statistical Analysis Statistical significance of community structure similarity was evaluated on Bray-Curtis distance matrices in the Community Analysis Package v.5 using analysis of similarity (ANOSIM). ANOSIM was performed on Unifrac distances in the vegan package [37] in R. In addition to testing individual sample factors, we nested samples within lake of origin when possible to control for variance due to origin from the same lake (e.g., for population, population source (laboratory lineage or field-collected), and body section). We were unable to control for lake of origin for reproductive mode or ploidy level because the P. antipodarum populations found in many of the lakes that we sampled represented only one or two ploidy levels and/or were entirely sexual or asexual. We also investigated the degree to which bacterial community profiles identified sample factors by using random forests, a robust machine-learning technique for classification that is appropriate for a wide diversity of data types, including microbial community data [38]. We used the QIIME implementation of random forests with 10-fold cross-validation on a rarefied (-e 400) OTU level table (97% DNA identity). The OTU table was filtered before rarefaction to remove any OTUs that were observed in the dataset fewer than 10 times (using the filter_otus_from_otu_table.py script). SIMPER (Similarity Percentage) was used to determine the relative contribution of OTUs to the observed similarity/dissimilarity within sample type. Indicator Species Analyses (ISA; [39]) in PC-ORD (v 5.33) was used to determine which OTUs, if any, were statistically associated with reproductive mode and other sample factors and to confirm SIMPER results. Finally, the sensitivity of our results to specific lineages, sample types, bacterial phyla, and singletons was investigated by reanalyzing the original dataset after filtering samples (e.g., Ianthe snails, which comprised 57% of sexual snails) or specific OTUs (e.g., Cyanobacteria were likely an artifact of ingestion of cyanobacteria in New Zealand lakes (field-collected snails) and/or of our laboratory snail food, dried Spirulina cyanobacteria) from the data. Thus, we repeated all analyses described above after filtering the original OTU tables in QIIME (filter_samples_from_otu_table.py and filter_otus_from_otu_table.py). DNA Sequence Availability All raw sequence data from this study are available through the NCBI Sequence Read Archive. The individual sff files from this study were assigned the accession numbers SAMN03276406 through SAMN03276465 under Bioproject PRJNA271685 (http://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA271685). Results Barcoded amplicon pyrosequencing targeting the bacterial 16S rRNA gene for the 60 P. antipodarum samples resulted in 258412 16S rRNA gene sequences. After quality filtering, denoising, and chimera removal, the mean number of sequences per sample was 4307 (range = 430 to 13333; SD = 3422). Good’s coverage estimates for the individual samples computed at an even depth of 400 sequences per sample ranged from 83% to 97% (mean = 90%; SD = 3%), indicating that a majority of the species richness that was PCR-amplified in the samples was detected in this DNA sequencing survey. We identified 2624 unique OTUs (97% DNA similarity) distributed across approximately thirty phyla (Fig 1 and S1 Fig). In general, samples were dominated by the Proteobacteria and Cyanobacteria, which comprised 58% and 23%, respectively, of the entire dataset. Proteobacteria sequences were primarily distributed among the alpha and beta classes (40% each), but were also represented by Gammaproteobacteria (10% of the Proteobacteria sequences). Sequences classified as Cyanobacteria were largely from chloroplast (66% of the cyanobacterial sequences) and likely represent algae inside snail intestinal tracts (P. antipodarum graze on algae growing on lake bottoms and vegetation surfaces and laboratory-cultured P. antipodarum are fed dried Spirulina, a cyanobacterium). Additional phyla detected among the samples include the Actinobacteria, Verrucomicrobacteria, Planctomycetes, and Firmicutes. Overall, bacterial communities varied among the individual samples; no OTUs were shared among all the samples, and only 11 of the total 2624 OTUs detected were shared among at least 75% of all samples (Table 2). 10.1371/journal.pone.0161050.g001Fig 1 Phylum-level taxonomy of P. antipodarum populations expressed as the percentage of total sequences (prior to rarefaction and filtering Cyanobacteria OTUs). We detected 2624 unique OTUs (97% DNA similarity) distributed across approximately thirty phyla. Figure includes only the eight most abundant phyla; remaining OTUs are compressed into the “other” category. 10.1371/journal.pone.0161050.t002Table 2 Core OTUs across 75% of all 60 snail samples. OTU Taxonomic Assignment 96 Bacteria;Actinobacteria;Actinobacteria;Actinomycetales;Propionibacteriaceae;Propionibacterium 253 Bacteria; Proteobacteria; Alphaproteobacteria 362 Bacteria; Acidobacteria; Acidobacteria-6; iii1-15; mb2424 529 Bacteria; Proteobacteria; Alphaproteobacteria; Rhizobiales; Hyphomicrobiaceae 549 Bacteria; Proteobacteria; Alphaproteobacteria; Rhizobiales 608 Bacteria; Proteobacteria; Alphaproteobacteria; Sphingomonadales; Sphingomonadaceae 651 Bacteria; Proteobacteria; Betaproteobacteria; Burkholderiales; Comamonadaceae; Methylibium 726 Unknown Bacteria 969 Bacteria; Proteobacteria; Alphaproteobacteria; Rhodobacterales; Rhodobacteraceae; Rhodobacter 1813 Bacteria; Proteobacteria; Alphaproteobacteria; Rhizobiales; Hyphomicrobiaceae; Hyphomicrobium 2113 Bacteria; Proteobacteria; Betaproteobacteria; Burkholderiales; Comamonadaceae Alpha diversity metrics used to estimate richness, dominance/equitability, phylogenetic diversity, and coverage among the samples were most variable when compared by lake, although the only significant differences detected for any pairwise lake comparisons were between the Chao1 estimates of Gunn vs. Ianthe samples and Alexandrina vs. Ianthe field collections (p = 0.028). In both cases, the Ianthe snails had significantly lower values of Chao1 diversity. In addition, dominance was significantly greater (p = 0.007) and equitability was significantly lower (p = 0.047) in the sexual snails compared to the asexual snails. No other significant alpha diversity differences were detected for any of the other sample factors (population source (laboratory lineage vs. field-collected), sex, body section, and ploidy level). Alpha diversity results summarized by sample factor are presented in S2 Table. There were significant differences in bacterial community composition and structure between sexual and asexual P. antipodarum (Fig 2A): Sexual and asexual snail samples tended to cluster separately in PCoA ordinations of distance matrices based on OTU presence/absence (unweighted Unifrac) and abundance (Bray-Curtis and weighted Unifrac). Community composition and structural differences among the sexual and asexual snails were supported by ANOSIM (Bray-Curtis R-statistic = 0.274, p = 0.001). Within the adult asexual snails, the 3x samples were marginally different than the 4x samples (Bray-Curtis R-statistic = 0.114, p = 0.056), though the global R-statistic for ploidy was not significant for Bray-Curtis distances. We did observe significant clustering of samples by ploidy among the Unifrac distances (ANOSIM R-statistic for unweighted Unifrac = 0.228 and weighted Unifrac = 0.149, p = 0.001 for both unweighted and weighted Unifrac distances), reinforcing the potential that these analysis results do reflect ploidy effects. 10.1371/journal.pone.0161050.g002Fig 2 PCoA of unweighted Unifrac distances of bacterial communities associated with P. antipodarum determined from the entire dataset (panel A) and when Cyanobacteria OTUs were excluded from the analysis (panel B). The community composition and structural differences among the sexual (red symbols, 2x ploidy) and asexual (light blue open symbols, 3x ploidy; dark blue filled symbols, 4x ploidy) snails shown here were supported by ANOSIM (R-statistic = 0.274, p = 0.001). Within the adult asexual snails, the 3x samples were marginally different than the 4x samples (R-statistic = 0.114, p = 0.056). When Cyanobacteria OTUs were excluded, similarly significant or even stronger clustering was observed among the samples (R-statistic = 0.240, p = 0.001 for reproductive mode and R-statistic = 0.260, p = 0.001 for ploidy). Lake and population were the only factors to explain as much or more variance in snail bacterial community structure as reproductive mode. Because the Ianthe snails represented over half of the sexual samples (16/28; ~57%), it is possible that Ianthe-specific effects rather than sex per se are driving the significant differences between sexual and asexual-associated bacterial communities. We addressed this possibility by removing the Ianthe snails from the analysis, which revealed that samples continued to cluster by ploidy (S2 Fig, Bray-Curtis R-statistic = 0.219, p = 0.002), though the ANOSIM R-statistic (0.09) of the Bray-Curtis distance for reproductive mode was not significant (p = 0.102). One potential explanation for our failure to detect a reproductive mode effect on clustering in the Ianthe-filtered dataset is that removal of the 16 Ianthe samples meant that our reproductive mode-focused comparisons were now imbalanced and suffering from low statistical power, with only 12 sexual samples relative to the 32 asexual samples. The likelihood of this explanation is reinforced by a power analysis of the mean distances between samples within the same reproductive mode compared to samples from different reproductive modes using a Mann-Whitney test of two independent samples (non-parametric; performed in XLSTAT v.2016.01.26437), which indicated that for an alpha of 0.05 and a sample size of 562 observations, the type 2 error is 0.949 and the power is 0.051. Significant clustering by both reproductive mode and ploidy were also observed for Unifrac distances (p <0.03 for weighted and unweighted distances). Taken together, these results at least suggest that the reproductive mode and ploidy effects we have observed are not driven entirely by the Ianthe samples. All significant differences among pairwise comparisons of reproductive mode by sex in the original dataset were between sexual and asexual snails; there were no significant differences between males and females within the same reproductive modes (S3 Table). We did detect significant differences in the bacterial communities of head vs. body tissue in the sexual snails (Bray-Curtis R-statistic = 0.111, p = 0.03, S3 Table), but not within the asexual snails (p = 0.476). There was significant clustering by other sample factors (Table 3), but lake (Bray-Curtis R-statistic = 0.389, p = 0.001) and population (Bray-Curtis R-statistic = 0.716, p = 0.001) were the only factors to explain as much or more variation in snail bacterial community structure as reproductive mode. Similar trends were observed for all factors of ANOSIM results of Unifrac distances, but global R-statistics were always greater than the R-statistics for Bray-Curtis distances. Thus, the ANOSIM results we report for the Bray-Curtis distances are conservative estimates of the relationships we observed among snail bacterial community composition and sample factors. 10.1371/journal.pone.0161050.t003Table 3 Analysis of similarity (ANOSIM) of Bray-Curtis distances among snail bacterial communities by sample factor. Sample Factor Factor Groups Global R-statistica p value Reproductive mode Sexual vs. Asexual 0.274 0.001 Ploidy level 2x vs. 3x vs. 4x 0.083 0.058 Lake Gunn vs. Taylor vs. Ianthe etc. 0.389 0.001 Population Gunn10 vs. Gunn14 vs. Ianthe etc. 0.716 0.001 Source Laboratory lineage vs. field-caught 0.154 0.095 Sex Male vs. Female 0.002 0.434 Body section Head vs. Body 0.089 0.013 aSignificant within-factor pairwise comparisons within sample factor are given in S2 Table (Reproductive mode) and S3 Table (Population). We investigated bacterial community composition variation among individuals of the same lineage by comparing a set of the Ianthe samples that were not pooled before PCR amplification, but did not detect significant differences among ANOSIM pairwise comparisons (p = 0.128–0.728). This result suggests that individual snail differences do not contribute significantly to factor differences. We did find that juvenile snails from Sarah contained significantly different communities than adults from the same lake (Sarah R-statistic = 0.916, p = 0.014). Because the juvenile samples from Sarah included pooled heads and bodies while the adult samples from Sarah only included heads, we cannot exclude the possibility that this difference between lake Sarah adults and juveniles may be attributed to the tissue type used for DNA extraction. Despite being maintained under identical conditions, there was significant variation in the bacterial communities among the laboratory lineages (ANOSIM R-statistic = 0.716, p = 0.001). This effect extended even to lineages that originated from the same lake but were maintained in separate tanks (e.g., the two replicate lineages from lake Gunn, R-statistic = 1, p = 0.014). Although global R-statistics did not indicate that laboratory lineages harbored significantly different bacterial communities than field-collected snails (R-statistic = 0.154, p = 0.095), the Kaniere laboratory-cultured snail communities were significantly different than the Kaniere field-collected snails (R-statistic = 0.989, p = 0.014). ANOSIM results by population are summarized in S4 Table. We also used the classification algorithm random forests to test how well the bacterial communities we detected predicted sample factors. ANOSIM can sometimes fail to detect significant clustering among samples in studies with an unbalanced design and heterogeneous distances among groups [40] or when gradients in community composition exist among groups [41]. In general, distance-based tests such as ANOSIM can confound community composition differences among samples, depending on dispersion homogeneity among factors and the relative location of the centroid of the sample clusters [42,43]. We were particularly interested in confirming factors with low ANOSIM R-statistics (<0.400) but significant p values (e.g., reproductive mode and ploidy level). Overall, random forests results were similar to ANOSIM; factors related to lake, population, reproductive mode, and ploidy level were the strongest predictors of community composition, whereas a sample’s source (field vs. laboratory), sex, and body section were much weaker factors (Table 4). Ratios of random error to model error were 29.33 and 66 for lake and population, respectively (a minimum ratio of 2 is expected for factors that can be accurately predicted [38]). For reproductive mode and ploidy level of the samples, the ratios of random error to model error were 14 and 16, respectively. The predictive accuracy was 95% to 99% for lake, population, reproductive mode, and ploidy level (only 1–3 of 60 samples were misclassified), whereas the predictive accuracy of the sample’s source, sex, and body section ranged from 30% to 92% and had correspondingly low ratios of random error to model error (0.7 to 2.37). Thus, random forest classifiers indicated that source, sex, and body section are poor predictors of snail community composition, compared to factors related to lake and reproductive mode. 10.1371/journal.pone.0161050.t004Table 4 Results from random forests classifiers. Factors Ratioa Predictive Accuracy (%) Sex 0.7 30 Section 1.16 55 Source 2.37 92 Reproductive Mode 14 97 Ploidy 16 97 Population 29.33 95 Lake 66 99 aRatio of random error to model error. The mean dissimilarity of the bacterial communities between the sexual and asexual samples was 90% (SIMPER analysis, Table 5). We identified twenty OTUs that cumulatively accounted for 50% of the community compositional differences between sexual and asexual snails. The three OTUs that explained the most variance between sexual and asexual snails were identified as Cyanobacteria (2 OTUs) and a member of the Alphaproteobacteria. These three OTUs were most abundant in the sexual snails, specifically the head section, which differed from asexual heads (ANOSIM R-statistic = 0.345, p = 0.001) and sexual bodies (ANOSIM R-statistic = 0.111, p = 0.03). The Alphaproteobacterium OTU is a sequence that was classified as a member of the Rickettsiales by QIIME, which uses a limited reference database. In order to provide a more specific taxonomic description of this sequence, we used blastn against the NCBI non-redundant nucleotide (nr) database (a much larger sequence database than QIIME), to find that the DNA sequence of this OTU is 99% similar (over 388 nucleotides) to various 16S rRNA gene sequences in Genbank from uncultured organisms originated from aquatic and soil environments. These taxa include several Rickettsiales members, including some Rickettsiales that were affiliated with putative endosymbionts of Acanthomoeba according to their sequence metadata in Genbank. 10.1371/journal.pone.0161050.t005Table 5 Similarity percentage analysis (SIMPER) between sexual and asexual snails. Mean dissimilarity between sexual and asexual snails = 90%. OTU Mean abundance in sexual snails Mean abundance in asexual snails Mean Dissimilarity % Contribution Cumulative % Taxonomic Assignment (quality score)a Dominant in sexual snails 2137 649.79 218.25 7.67 8.48 8.48 Bacteria;Cyanobacteria (1.0) 2283 514.39 62.31 5.63 6.22 14.7 Bacteria;Cyanobacteria;Chloroplast;Rhodophyta (1.0) 985 529.89 0.16 5.15 5.69 20.39 Bacteria;Proteobacteria;Alphaproteobacteria;Rickettsiales (0.96) 2302 341.82 6.47 3.19 3.53 28.07 Bacteria;Cyanobacteria;Chloroplast;Stramenopiles (1.0) 270 272 21.5 2.58 2.85 30.92 Bacteria;Proteobacteria;Betaproteobacteria;Burkholderiales;Comamonadaceae (1.0) 549 157.57 56.81 1.81 2 37.63 Bacteria;Proteobacteria;Alphaproteobacteria;Rhizobiales (0.96) 182 60.14 38.06 0.96 1.06 46.8 Bacteria;Planctomycetes;OM190;CL500-15 (1.0) 362 72.82 41.22 0.91 1.01 47.8 Bacteria;Acidobacteria;Acidobacteria-6;iii1-15;mb2424 (0.99) 228 76.32 23.81 0.89 0.98 49.79 Bacteria;Proteobacteria;Betaproteobacteria;Burkholderiales;Comamonadaceae;Rubrivivax (0.98) Dominant in asexual snails 1972 11.54 375.47 3.76 4.16 24.54 Bacteria;Cyanobacteria;Chloroplast;Stramenopiles (1.0) 2217 0.14 314.19 2.33 2.58 33.49 Bacteria;Cyanobacteria;Chloroplast;Stramenopiles (0.99) 1691 4 251.09 1.93 2.13 35.63 Bacteria;Proteobacteria;Alphaproteobacteria;Rhodobacterales;Rhodobacteraceae;Rhodobacter (1.0) 821 8.43 96.41 1.52 1.68 39.31 Bacteria;Proteobacteria;Gammaproteobacteria;Alteromonadales;[Chromatiaceae];Rheinheimera (1.0) 1014 48.82 72.31 1.31 1.45 40.76 Bacteria;Proteobacteria;Gammaproteobacteria;Pseudomonadales;Pseudomonadaceae (1.0) 969 54.5 79.28 1.29 1.43 42.19 Bacteria;Proteobacteria;Alphaproteobacteria;Rhodobacterales;Rhodobacteraceae;Rhodobacter (1.0) 104 2.93 82.38 1.08 1.2 43.39 Bacteria;Proteobacteria (0.99) 1327 19.25 49.72 1.07 1.18 44.56 Bacteria;Proteobacteria;Alphaproteobacteria;Caulobacterales;Caulobacteraceae;Phenylobacterium (1.0) 2405 0 70.16 1.06 1.18 45.74 Bacteria;Proteobacteria;Betaproteobacteria (0.85) 634 7.68 40.69 0.91 1.01 48.81 Bacteria;Proteobacteria;Betaproteobacteria;Burkholderiales;Comamonadaceae (1.0) aConsensus taxonomic assignment by assign_taxonomy.py in QIIME (-m uclust), quality scores represents the proportion of sequences that match the assignment. The three OTUs that were the most enriched in asexual vs. sexual snails were also assigned to the Cyanobacteria and Alphaproteobacteria. Rather than being affiliated with the Rickettsiales, the Alphaproteobacterium OTU from the asexual snails was classified as a Rhodobacteriales by QIIME and was 99% similar to uncultured Rhodobacter DNA sequences in Genbank. We confirmed that the abundance of singletons in our OTU table (94% of the cells were empty) was not obscuring any patterns in our data by repeating our analyses on an OTU table that had been filtered for rare OTUs (any OTUs that were represented by <25 sequences). Overall, results were unchanged by this filtering process, although the difference between sexual and asexual snails calculated by SIMPER decreased by 3% (from 90% to 87%). Random forests analysis also confirmed the importance of the SIMPER OTUs in sexual compared to asexual samples, and results from the indicator species analysis were qualitatively similar to the outcome of the SIMPER analysis. For example, Cyanobacteria and Alphaproteobacteria still contributed significantly to observed differences associated with reproductive mode in the indicator species analysis, although there were some differences in the specific OTUs that differed between sexual and asexual snails in these two analyses (Table 6). 10.1371/journal.pone.0161050.t006Table 6 Indicator species analysis by reproductive mode. Reproductive Mode OTU Indicator Value p value Taxonomic Assignment (quality score)a Sexual 1813 83.5 0.0002 Bacteria;Proteobacteria;Alphaproteobacteria;Rhizobiales;Hyphomicrobiaceae;Hyphomicrobium (0.94) 2302 80.6 0.0002 Bacteria;Cyanobacteria;Chloroplast;Stramenopiles (1.0) 270 76.1 0.0002 Bacteria;Proteobacteria;Betaproteobacteria;Burkholderiales;Comamonadaceae (1.0) 726 74.6 0.0002 Bacteria (1.0) 1634 74.4 0.0002 Bacteria;Proteobacteria;Alphaproteobacteria;Rhodobacterales;Hyphomonadaceae (0.98) 1091 67.2 0.0002 Bacteria;Proteobacteria;Alphaproteobacteria;Sphingomonadales (1.0) 2161 52.4 0.0002 Bacteria;Actinobacteria;Actinobacteria;Actinomycetales (0.85) 659 46.4 0.0002 Bacteria;Proteobacteria (0.90) 1322 46 0.0002 Bacteria;Proteobacteria;Alphaproteobacteria;Sphingomonadales (1.0) 1798 61.7 0.0004 Bacteria;Proteobacteria;Betaproteobacteria;Burkholderiales;Burkholderiaceae (0.91) 1891 42.5 0.0004 Bacteria;Bacteroidetes;Flavobacteriia;Flavobacteriales;Cryomorphaceae (0.89) 985 39.3 0.0004 Bacteria;Proteobacteria;Alphaproteobacteria;Rickettsiales (0.96) 160 35.5 0.0004 Bacteria;Cyanobacteria;Synechococcophycideae;Synechococcales (0.95) 608 71.8 0.0008 Bacteria;Proteobacteria;Alphaproteobacteria;Sphingomonadales;Sphingomonadaceae (0.91) 2388 65.7 0.0008 Bacteria;Proteobacteria;Betaproteobacteria;Burkholderiales;Burkholderiaceae (0.90) Asexual 1691 86.1 0.0002 Bacteria;Proteobacteria;Alphaproteobacteria;Rhodobacterales;Rhodobacteraceae;Rhodobacter (1.0) 1972 81.9 0.0002 Bacteria;Cyanobacteria;Chloroplast;Stramenopiles (1.0) 2217 81.2 0.0002 Bacteria;Cyanobacteria;Chloroplast;Stramenopiles (0.99) 2405 81.2 0.0002 Bacteria;Proteobacteria;Betaproteobacteria (0.85) 1494 59 0.0002 Bacteria;Actinobacteria;Actinobacteria;Actinomycetales;Actinomycetaceae;Actinomyces (1.0) 1483 50 0.0002 Bacteria;Proteobacteria;Betaproteobacteria;Methylophilales;Methylophilaceae (1.0) 2075 49.7 0.0002 Bacteria;Proteobacteria;Deltaproteobacteria;Myxococcales (0.98) 941 43.7 0.0002 Bacteria;Bacteroidetes;Sphingobacteriia;Sphingobacteriales;Saprospiraceae (1.0) 821 77.6 0.0004 Bacteria;Proteobacteria;Gammaproteobacteria;Alteromonadales;[Chromatiaceae];Rheinheimera (1.0) 347 37.5 0.0004 Bacteria;WPS-2 (0.97) 516 42.9 0.0006 Bacteria;Proteobacteria;Betaproteobacteria;Burkholderiales;Comamonadaceae;Limnobacter (1.0) 48 42.6 0.0008 Bacteria;Verrucomicrobia;Verrucomicrobiae;Verrucomicrobiales;Verrucomicrobiaceae;Prosthecobacter;debontii (0.91) aConsensus taxonomic assignment by assign_taxonomy.py in QIIME (-m uclust), quality scores represent the proportion of sequences that match the assignment. The importance of the Cyanobacteria OTUs in the SIMPER and Indicator Species analysis is concerning because we suspect that these sequences are derived from snail food and are thus unlikely to be involved in reproductive mode. Therefore, we performed all analyses described above with an OTU table that was filtered to remove any OTUs assigned to the phylum Cyanobacteria. Removal of these sequences did not qualitatively change any of the results described above, but did result in similarly significant or even stronger clustering among the samples (Fig 2B, R-statistic = 0.240, p = 0.001 for reproductive mode; R-statistic = 0.260, p = 0.001 for ploidy; Bray-Curtis distances). The non-cyanobacterial indicator taxa detected by SIMPER were also the same taxa revealed by analyses of the original complete dataset. All OTUs identified by SIMPER (before and after filtering Cyanobacteria DNA sequences) were unique (<97% similar to each other in their DNA sequence) and were 64 to 100% similar to their nearest relatives identified from the NCBI nr database. A distance matrix comparing the genetic distance among the OTUs identified by SIMPER and their nearest relatives in Genbank is given in S5 Table. A heatmap showing the relative abundance of the SIMPER OTUs across all samples highlights the differences among factors (Fig 3), including the absence of the Rickettsiales OTU from nearly all of the asexual snail samples and the presence of some OTUs (e.g., OTU2405, an unassigned Betaproteobacteria) that were strictly limited to the asexual snails. All of these results must formally be considered from the perspective of contamination, which cannot be ruled out in any study using sensitive techniques such as PCR and next-generation sequencing. Because all of our samples were treated similarly, we would expect any contaminating DNA sequence to manifest in our dataset as low-abundance OTUs distributed uniformly across all samples. The results we report here do not represent this type of pattern, allowing us to conclude that contamination is not likely to be a major driver of any of our findings. 10.1371/journal.pone.0161050.g003Fig 3 Heatmap showing relative abundance of non-Cyanobacteria OTUs identified by SIMPER as contributing significantly to the dissimilarity between sexual and asexual snails. Vertical lines delineate individual populations. Samples were normalized and relative abundance of each OTU was log transformed. Color legend on right indicates the relative abundance of each OTU where lighter colors indicate greater OTU abundance. Discussion Microbes that reside in the animal body have the potential to play an important role in host reproduction. Here, we found evidence suggesting distinct differences in the microbiome composition between sexual and asexual freshwater snails sampled from multiple laboratory lineages and New Zealand lakes. In particular, the sexual and asexual P. antipodarum included in our study differed with respect to the presence of Rickettsiales bacteria in the somatic tissue of sexual individuals and Rhodobacter in asexual individuals. We also detected strong phylogeographic signatures in the bacterial community structure associated with snails from populations from different lakes and from different populations within lakes. These results mirror the outcomes of phylogeographic analyses of P. antipodarum itself, which reveals substantial genetic variation among P. antipodarum of the same ploidy level and reproductive mode but from different lakes, as well as among P. antipodarum that differ in ploidy level and reproductive mode [21]. Because the comparisons between sexual and asexual snails required samples from different populations and different lakes, and because we could not control for phylogenetic nonindependence, it is possible that the significant differences in microbiota community structure between 2x, 3x, and 4x snails revealed by the principal coordinate analysis could be due to a population and/or lake effect rather than a reproductive mode effect per se. Even so, the community compositional differences between samples of different ploidy levels (Fig 2) suggest that the effects of reproductive mode that we report here are not a simple artifact of phylogeography. For example, the variance between the bacterial communities harbored by sexual vs. asexual populations (ANOSIM R-statistic = 0.274, p = 0.001) was more than two-fold greater than between the bacterial communities harbored by triploid and tetraploid populations (ANOSIM R-statistic = 0.114, p = 0.056), despite the fact that all four groupings (sexual, asexual, triploid, tetraploid) represented populations from multiple lakes. While teasing apart the effects of ploidy from population in P. antipodarum will require additional study (and in particular, the inclusion of more populations), that the 3x and 4x snails in our study did possess different microbiota is intriguing in light of the distinct possibility that ploidy level can influence immune function and host resistance [44]. The case for a connection between reproductive mode and microbiome community in sexual P. antipodarum is strengthened by the absence of Rickettsiales in the triploid and tetraploid asexuals and their presence in both sexual males and females (Fig 3), regardless of lake origin. Members of the group Rickettsiales have a wide host range and cause a diversity of host effects across the parasite-mutualist continuum [45]. These symbionts can be transmitted through blood-sucking arthropods such as ticks and are commonly found as parasitic bacteria in herbivorous arthropods [46] and vertebrates [45]. Rickettsiales have also been found associated with established non-arthropod hosts like leeches and amoeba [47,48], though their effects are largely unknown. Here, the presence of Rickettsiales in both male and female sexual P. antipodarum from both field-collected and laboratory-cultured sexual individuals (see Fig 3) hints at its effects and transmission routes. In particular, the first result suggests that males inherit or acquire Rickettsiales in a similar manner to females, while the second result suggests that these symbionts are persistent and might be vertically transmitted within sexual lineages. The apparent absence of Rickettsiales from the body tissue of some of the sexual P. antipodarum studied here could highlight a role for horizontal transmission or suggest that symbionts develop in one part of the body before transferring to germ tissue [24]. As suggested by Frost et al. [24], a tropism in non-reproductive tissue may rule out any specialization for reproductive parasitism (e.g., sex-ratio distortion, cytoplasmic incompatibility, parthenogenesis induction) and suggest alternative parasitic or even mutualistic phenotypic effects. Across populations of field collections and laboratory-cultured lineages, asexual snails consistently possessed bacteria closely related to the Proteobacteria genus Rhodobacter in their microbiome. Rhodobacter bacteria are phototrophic in aquatic environments and have been found to be symbionts of marine sponges [49] and Daphnia [50], although their symbiotic effects remain unknown. The transmission route of Rhodobacter bacteria into P. antipodarum individuals is also not clear, and may only reflect their abundance in aquatic environments. The fact that Rhodobacter was found in both adults and juveniles from Lake Sarah (and at a much higher abundance in juveniles) suggests that this bacterium might be an inherited endosymbiont. Several established endosymbiotic bacteria belonging to the genera Burkholderia, Aeromonas, Brevibacillus, and Ideonella were present in both adults and juveniles from lake Sarah, indicating that some or all of these bacteria might also be inherited or acquired soon after birth. Additional comparisons of more juveniles and adults from additional populations will be required to more fully evaluate this possibility. Genetic differentiation between host strains has been implicated in the variation of the microbiota associated with laboratory-reared animals [51,52]. We used comparisons among genetically distinct asexual populations descended from snails originally collected from Lakes Taylor, Rotoiti, Poerua, and Gunn and maintained in the laboratory over several generations to detect multiple genotype-specific effects on microbiome composition. Our results indicate that these different P. antipodarum populations are associated with different bacterial communities even when raised under common garden conditions. Previous studies of the bacterial communities of various Drosophila species indicated that there is a large difference in the diversity and community composition of laboratory lineages vs. wild populations of this important model organism. Given these discrepancies, it might be difficult to justify the use of laboratory cultures as appropriate models of host—microbe interactions that occur in the wild [53]. By contrast, we found no significant difference in the diversity or community composition between our laboratory lineages and field-caught populations, except between the Kaniere samples. While this laboratory/field comparison requires more rigorous sampling and testing to be considered definitive, the lack of evidence for markedly different bacterial communities between our laboratory and field-collected snails raises the interesting possibility that P. antipodarum has the potential to be a good model for the laboratory study of host-microbe interactions in situ. One particularly exciting potential implication of intraspecific variation in P. antipodarum microbiota communities is the possibility that across-lineage microbial variation could mediate the already well-documented coevolutionary interactions between P. antipodarum and Microphallus, a sterilizing trematode worm [13,54–56]. Microbial host-protection against infection has been demonstrated across animal and plant species [57,58]. In bumblebees, genotype-specific microbial communities have been shown to facilitate infection specificity [59]. Infection by Microphallus in lake populations of P. antipodarum snails is host genotype-specific [20,60], hinting that the genotype-specific variation in the microbiome could play a role in mediating this host-parasite interaction. Altogether, multiple lines of analyses suggest that lake, population, ploidy level, and reproductive mode might be linked to differences in bacterial community structure of a diverse array of sexual and asexual P. antipodarum. Although we do not yet know the exact nature of these associations, the consistent differences in Rickettsiales and Rhodobacter presence between sexual and asexual P. antipodarum set the stage for future studies that more directly address whether these microbes play a role in host reproduction. Critical next steps in addressing this possibility include experimental manipulations aimed at determining whether exposure to and elimination of particular bacterial taxa can influence P. antipodarum reproductive mode (e.g., [61]) as well as direct comparisons of the bacterial communities of sexual and asexual P. antipodarum originating from the same New Zealand lakes. Supporting Information S1 Fig Barchart showing the percent composition of OTUs identified at the 97% DNA identity level across all samples included in this study. Only OTUs contributing more than 2% of total DNA sequence data are annotated in the legend, which includes all non-cyanobacterial OTUs identified by SIMPER. (EPS) Click here for additional data file. S2 Fig PCoA of unweighted Unifrac distances of bacterial communities associated with P. antipodarum from 10 of the populations used in this study. Ianthe samples, which comprised 57% of the 2x samples, are not included in this analysis to determine if Ianthe snails were driving sexual vs. asexual clustering patterns observed in the entire dataset. Cyanobacteria OTUs were not filtered. Sexual samples are in red and asexual samples are in blue. Ploidy (2x, 3x, 4x) of each population is given in the legend. (EPS) Click here for additional data file. S1 Table List and description of samples. (PDF) Click here for additional data file. S2 Table Summary of alpha diversity results and non-parametric t-test with 1000 Monte Carlo permutations. (PDF) Click here for additional data file. S3 Table Analysis of similarity (ANOSIM) of Bray-Curtis distances among snail bacterial communities by reproductive mode. (PDF) Click here for additional data file. S4 Table Analysis of similarity (ANOSIM) of Bray-Curtis distances among snail bacterial communities by snail population. (PDF) Click here for additional data file. S5 Table Estimates of 16S rRNA gene divergence between OTU sequences identified as significant in SIMPER and nearest BLAST matches. (PDF) Click here for additional data file. This work was supported by NSF-MCB grant 1122176 and the Iowa Center for Research by Undergraduates. The authors have no conflict of interests to state. We thank several anonymous reviewers for helpful feedback on earlier versions of the manuscript. ==== Refs References 1 Dillon RT . The Ecology of Freshwater Molluscs . Cambridge : Cambridge University Press ; 2000 10.1017/CBO9780511542008 2 Matsuura K . Sexual and asexual reproduction in termites In Biology of Termites: A Modern Synthesis . Dordrecht : Springer Netherlands ; 2010 pp. 255 –277 . 3 Lizé A , McKay R , Lewis Z . Kin recognition in Drosophila: the importance of ecology and gut microbiota . ISME J . 2014 ; 8 : 469 –477 . 10.1038/ismej.2013.157 24030598 4 Breeuwer JAJ , Werren JH . Microorganisms associated with chromosome destruction and reproductive isolation between two insect species . Nature . 1990 ; 346 : 558 –60 .2377229 5 Braig HR , Turner BD , Normark BB , Stouthamer R . Microorganism-induced parthenogenesis In Reproductive Biology of Invertebrates , Volume XI John Wiley and Sons ; 2002 . 6 Huigens ME , Stouthamer R . Parthenogenesis associated with Wolbachia In Insect Symbiosis . Boca Raton : CRC Press ; 2003 1 : 247 –266 . 7 Koivisto RKK , Braig HR . Microorganisms and parthenogenesis . Biol J Linn Soc . 2003 ; 79 : 43 –58 . 10.1046/j.1095-8312.2003.00185.x 8 Duron O , Bouchon D , Boutin S , Bellamy L , Zhou L , Engelstadter J , et al The diversity of reproductive parasites among arthropods: Wolbachia do not walk alone . BMC Biol . 2008 ; 6 : 27 10.1186/1741-7007-6-27 18577218 9 Werren JH , Baldo L , Clark ME . Wolbachia: master manipulators of invertebrate biology . Nature Reviews Microbiology . 2008 ; 6 : 741 –751 . 10.1038/nrmicro1969 18794912 10 Engelstadter J , Hurst GDD . What use are male hosts? The dynamics of maternally inherited bacteria showing sexual transmission or male killing . Am Nat . 2009 ; 173 : E159 –E170 . 10.1086/597375 19272014 11 Leach IM , Pannebakker BA , Schneider MV , Driessen G , van de Zande L , Beukeboom LW . Thelytoky in Hymenoptera with Venturia canescens and Leptopilina clavipes as case studies In Lost Sex . Dordrecht : Springer Netherlands ; 2009 pp. 347 –375 . 10.1007/978-90-481-2770-2_17 12 Vandekerckhove T , Willems A , Gillis M , Coomans A . Occurrence of novel verrucomicrobial species, endosymbiotic and associated with parthenogenesis in Xiphinema americanum-group species (Nematoda, Longidoridae) . IJSEM . 2000 ; 50 : 2197 –2205 . 10.1099/00207713-50-6-2197 11155997 13 Lively CM . Evidence from a New Zealand snail for the maintenance of sex by parasitism . Nature . 1987 ; 328 : 519 –521 . 10.1038/328519a0 14 Lively CM , Jokela J . Temporal and spatial distributions of parasites and sex in a freshwater snail . Evol Ecol Res . 2002 ; 4 : 219 –226 . 15 Neiman M , Larkin K , Thompson AR , Wilton P . Male offspring production by asexual Potamopyrgus antipodarum, a New Zealand snail . Heredity . 2012 ; 109 : 57 –62 . 10.1038/hdy.2012.13 22491063 16 Neiman M , Schwander T . Using parthenogenetic lineages to identify advantages of sex . Evol Biol . 2011 ; 38 : 115 –123 . 10.1007/s11692-011-9113-z 17 Neiman M , Paczesniak D , Soper DM , Baldwin AT , Hehman G . Wide variation in ploidy level and genome size in a New Zealand freshwater snail with coexisting sexual and asexual lineages . Evolution . 2011 ; 65 : 3202 –3216 . 10.1111/j.1558-5646.2011.01360.x 22023586 18 Wallace DC . Mitochondrial genetics: a paradigm for aging and degenerative diseases? Science . 1992 ; 256 : 628 –632 . 10.1126/science.1533953 1533953 19 Neiman M , Lively CM . Pleistocene glaciation is implicated in the phylogeographical structure of Potamopyrgus antipodarum, a New Zealand snail . Mol Ecol . 2004 ; 13 : 3085 –3098 . 10.1111/j.1365-294X.2004.02292.x 15367122 20 Dybdahl MF , Lively CM . Host-parasite interactions: infection of common clones in natural populations of a freshwater snail (Potamopyrgus antipodarum) . Proc R Soc Lond B . 1995 ; 260 : 99 –103 . 21 Paczesniak D , Jokela J , Larkin K , Neiman M . Discordance between nuclear and mitochondrial genomes in sexual and asexual lineages of the freshwater snail Potamopyrgus antipodarum . Mol Ecol . 2013 ; 22 : 4695 –4710 . 10.1111/mec.12422 23957656 22 Jokela J , Lively CM . Parasites, sex, and early reproduction in a mixed population of freshwater snails . Evolution . 1995 ; 49 : 1268 –1271 . 10.2307/2410451 28568513 23 Zchori-Fein E , Roush RT , Rosen D . Distribution of parthenogenesis-inducing symbionts in ovaries and eggs of Aphytis (Hymentoptera: Aphelinidae) . Curr Microbiol . 1998 ; 36 : 1 –8 . 10.1007/s002849900270 9405738 24 Frost CL , Pollock SW , Smith JE , Hughes WOH . Wolbachia in the flesh: symbiont intensities in germ-line and somatic tissues challenge the conventional view of Wolbachia transmission routes . PLoS ONE . 2014 ; 9 : e95122 10.1371/journal.pone.0095122 24988478 25 Krist AC , Kay AD , Larkin K , Neiman M . Response to phosphorus limitation varies among lake populations of the freshwater snail Potamopyrgus antipodarum . PLoS ONE . 2014 ; 9 : e85845 10.1371/journal.pone.0085845 24454936 26 Osnas EE , Lively CM . Host ploidy, parasitism and immune defence in a coevolutionary snail-trematode system . J Evol Biol . 2006 ; 19 : 42 –48 . 10.1111/j.1420-9101.2005.00994.x 16405575 27 Dowd SE , Sun Y , Wolcott RD , Domingo A , Carroll JA . Bacterial tag—encoded FLX amplicon pyrosequencing (bTEFAP) for microbiome studies: bacterial diversity in the ileum of newly weaned Salmonella-infected pigs . Foodborne Pathog Dis . 2008 ; 5 : 459 –472 . 10.1089/fpd.2008.0107 18713063 28 Baker GC , Smith JJ , Cowan DA . Review and re-analysis of domain-specific 16S primers . J Microbiol Methods . 2003 ; 55 : 541 –555 . 10.1016/j.mimet.2003.08.009 14607398 29 Quince C , Lanzen A , Davenport RJ , Turnbaugh PJ . Removing noise from pyrosequenced amplicons . BMC Bioinformatics . 2011 ; 12 : 38 10.1186/1471-2105-12-38 21276213 30 Caporaso JG , Kuczynski J , Stombaugh J , Bittinger K , Bushman FD , Costello EK , et al QIIME allows analysis of high-throughput community sequencing data . Nat Methods . 2010 ; 7 : 335 –336 . 10.1038/nmeth.f.303 20383131 31 Edgar RC . Search and clustering orders of magnitude faster than BLAST . Bioinformatics . 2010 ; 26 : 2460 –2461 . 10.1093/bioinformatics/btq461 20709691 32 Edgar RC . MUSCLE: multiple sequence alignment with high accuracy and high throughput . Nucleic Acids Res . 2004 ; 32 : 1792 –1797 . 10.1093/nar/gkh340 15034147 33 DeSantis TZ , Hugenholtz P , Larsen N , Rojas M , Brodie EL , Keller K , et al Greengenes, a chimera-checked 16S rRNA gene database and workbench compatible with ARB . Appl Environ Microbiol . 2006 ; 72 : 5069 –5072 . 10.1128/AEM.03006-05 16820507 34 Price MN , Dehal PS , Arkin AP . FastTree: computing large minimum evolution trees with profiles instead of a distance matrix . Mol Biol Evol . 2009 ; 26 : 1641 –1650 . 10.1093/molbev/msp077 19377059 35 Faith DP . Conservation evaluation and phylogenetic diversity . Biol Cons . 1992 ; 61 : 1 –10 . 10.1016/0006-3207(92)91201-3 36 Lozupone C , Knight R . UniFrac: a new phylogenetic method for comparing microbial communities . Appl Environ Microbiol . 2005 ; 71 : 8228 –8235 .16332807 37 Oksanen J, Blanchet FG, Kindt R. Package Vegan: Community Ecology Package, version 2.0. 10. 2013 http://CRAN.R-project.org/package=vegan. 38 Knights D , Costello EK , Knight R . Supervised classification of human microbiota . FEMS Microbiol Rev . 2011 ; 35 : 343 –359 . 10.1111/j.1574-6976.2010.00251.x 21039646 39 Dufrêne M , Legendre P . Species assemblages and indicator species: the need for a flexible asymmetrical approach . Ecol Monogr . 1997 ; 67 : 345 –366 . 40 Anderson MJ , Walsh DCI . PERMANOVA, ANOSIM, and the Mantel test in the face of heterogeneous dispersions: What null hypothesis are you testing? Ecol Monogr . 2013 ; 83 : 557 –574 . 10.1890/12-2010.1 41 Somerfield PJ , Clarke KR . A comparison of the power of categorical and correlational tests applied to community ecology data from gradient studies . J Anim Ecol . 2002 ; 71 : 581 –593 . 42 Anderson MJ . A new method for non-parametric multivariate analysis of variance . Austral Ecol . 2001 ; 26 : 32 –46 . 10.1111/j.1442-9993.2001.01070.pp.x 43 Warton DI , Wright ST , Wang Y . Distance-based multivariate analyses confound location and dispersion effects . Methods Ecol Evol . 2012 ; 3 : 89 –101 . 10.1111/j.2041-210X.2011.00127.x 44 King KC , Seppälä O , Neiman M . Is more better? Polyploidy and parasite resistance . Biol Lett . 2012 ; 8 : 598 –600 . 10.1098/rsbl.2011.1152 22258448 45 Perlman SJ , Hunter MS , Zchori-Fein E . The emerging diversity of Rickettsia . Proc R Soc Lond B . 2006 ; 273 : 2097 –2106 . 10.1098/rspb.2006.3541 46 Azad AF , Beard CB . Rickettsial pathogens and their arthropod vectors . Emerging Infect Dis . 1998 ; 4 : 179 –186 . 10.3201/eid0402.980205 9621188 47 Kikuchi Y , Sameshima S , Kitade O , Kojima J , Fukatsu T . Novel clade of Rickettsia spp. from leeches . Appl Env Microbiol . 2002 ; 68 : 999 –1004 .11823253 48 Dyková I , Veverková M , Fiala I , Machácková B , Pecková H . Nuclearia pattersoni sp. n. (Filosea), a new species of amphizoic amoeba isolated from gills of roach (Rutilus rutilus), and its rickettsial endosymbiont . Folia Parasitol . 2003 ; 50 : 161 –170 . 14535340 49 Althoff K , Schütt C , Steffen R , Batel R , Müller WEG . Evidence for a symbiosis between bacteria of the genus Rhodobacter and the marine sponge Halichondria panicea: harbor also for putatively toxic bacteria? Marine Biol . 1998 ; 130 : 529 –536 . 50 Qi W , Nong G , Preston JF , Ben-Ami F , Ebert D . Comparative metagenomics of Daphnia symbionts . BMC Genomics . 2009 ; 10 : 172 10.1186/1471-2164-10-172 19383155 51 Deloris Alexander A , Orcutt RP , Henry JC , Baker J Jr, Bissahoyo AC , Threadgill DW . Quantitative PCR assays for mouse enteric flora reveal strain-dependent differences in composition that are influenced by the microenvironment . Mamm Genome . 2006 ; 17 : 1093 –1104 . 10.1007/s00335-006-0063-1 17091319 52 Roeselers G , Mittge EK , Stephens WZ , Parichy DM , Cavanaugh CM , Guillemin K , et al Evidence for a core gut microbiota in the zebrafish . ISME J . 2011 ; 5 : 1595 –1608 . 10.1038/ismej.2011.38 21472014 53 Chandler JA , Lang JM , Bhatnagar S , Eisen JA , Kopp A . Bacterial communities of diverse Drosophila species: ecological context of a host—microbe model system . PLoS Genet . 2011 ; 7 : e1002272 10.1371/journal.pgen.1002272 21966276 54 Lively CM , Dybdahl MF , Jokela J , Osnas EE . Host sex and local adaptation by parasites in a snail-trematode interaction . Am Nat . 2004 ; 164 : S6 –S18 . 10.1086/424605 15540142 55 Jokela J , Dybdahl MF , Lively CM . The maintenance of sex, clonal dynamics, and host-parasite coevolution in a mixed population of sexual and asexual snails . Am Nat . 2009 ; 174 : S43 –S53 . 10.1086/599080 19441961 56 King KC , Delph LF , Jokela J , Lively CM . The geographic mosaic of sex and the Red Queen . Curr Biol . 2009 ; 19 : 1438 –1441 . 10.1016/j.cub.2009.06.062 19631541 57 Dillon RJ , Vennard CT , Buckling A , Charnley AK . Diversity of locust gut bacteria protects against pathogen invasion . Ecol Lett . 2005 ; 8 : 1291 –1298 . 10.1111/j.1461-0248.2005.00828.x 58 Mendes R , Kruijt M , de Bruijn I , Dekkers E , van der Voort M , Schneider JHM , et al Deciphering the rhizosphere microbiome for disease-suppressive bacteria . Science . 2011 ; 332 : 1097 –1100 . 10.1126/science.1203980 21551032 59 Koch H , Schmid-Hempel P . Socially transmitted gut microbiota protect bumble bees against an intestinal parasite . Proc Nat Acad Sci USA . 2011 ; 108 : 19288 –19292 . 10.1073/pnas.1110474108 22084077 60 Dybdahl MF , Lively CM . Host-parasite coevolution: evidence for rare advantage and time-lagged selection in a natural population . Evolution . 1998 ; 52 : 1057 –1066 . 10.2307/2411236 28565221 61 Li YY , Floate KD , Fields PG , Pang BP . Review of treatment methods to remove Wolbachia bacteria from arthropods . Symbiosis . 2014 ; 62 : 1 –15 . 10.1007/s13199-014-0267-1

PMC005xxxxxx/PMC5001687.txt

==== Front Chem Sci Chem Sci Chemical Science 2041-6520 2041-6539 Royal Society of Chemistry 27574558 c6sc00666c 10.1039/c6sc00666c Chemistry Lesion orientation of O4-alkylthymidine influences replication by human DNA polymerase η† †Electronic supplementary information (ESI) available: Materials, experimental procedures, compound characterization, and additional discussion. The atomic coordinates and structure factors (codes 5DLF, 5DLG, 5DQG, 5DQH and 5DQI) have been deposited in the Protein Data Bank (http://www.wwpdb.org/). See DOI: 10.1039/c6sc00666c O'Flaherty D. K. a Patra A. b Su Y. b Guengerich F. P. b Egli M. b Wilds C. J. a a Department of Chemistry and Biochemistry , Concordia University , Montréal , Québec H4B1R6 , Canada . Email: chris.wilds@concordia.ca b Department of Biochemistry , Vanderbilt Institute of Chemical Biology , Center for Structural Biology , School of Medicine , Vanderbilt University , Nashville , Tennessee 37232 , USA . Email: martin.egli@vanderbilt.edu 1 8 2016 26 4 2016 7 8 48964904 12 2 2016 22 4 2016 This journal is © The Royal Society of Chemistry 2016 2016 This article is freely available. This article is licensed under a Creative Commons Attribution 3.0 Unported Licence (CC BY 3.0) Conformation of the α-carbon of O4-alkylthymidine was shown to exert an influence on human DNA polymerase η (hPol η) bypass. Crystal structures of hPol η·DNA·dNTP ternary complexes reveal a unique conformation adopted by O4-methylthymidine, where the nucleobase resides nestled at the active site ceiling where hydrogen-bonding with the incoming nucleotide is prevented. DNA lesions that elude repair may undergo translesion synthesis catalyzed by Y-family DNA polymerases. O4-Alkylthymidines, persistent adducts that can result from carcinogenic agents, may be encountered by DNA polymerases. The influence of lesion orientation around the C4–O4 bond on processing by human DNA polymerase η (hPol η) was studied for oligonucleotides containing O4-methylthymidine (O4MedT), O4-ethylthymidine (O4EtdT), and analogs restricting the O4-methylene group in an anti-orientation. Primer extension assays revealed that the O4-alkyl orientation influences hPol η bypass. Crystal structures of hPol η·DNA·dNTP ternary complexes with O4MedT or O4EtdT in the template strand showed the nucleobase of the former lodged near the ceiling of the active site, with the syn-O4-methyl group engaged in extensive hydrophobic interactions. This unique arrangement for O4-methylthymidine with hPol η, inaccessible for the other analogs due to steric/conformational restriction, is consistent with differences observed for nucleotide incorporation and supports the concept that lesion conformation influences extension across DNA damage. Together, these results provide mechanistic insights on the mutagenicity of O4MedT and O4EtdT when acted upon by hPol η. ==== Body Introduction DNA alkylation results from a variety of endogenous and/or exogenous agents that can interfere with vital cellular processes, i.e. replication and transcription.1 The addition of alkyl appendages on the DNA scaffold can have adverse consequences such as DNA polymerase (Pol) blockage, nucleotide misincorporation, chromosomal instability, and activation of the cellular apoptotic pathway.1,2 However, organisms have various repair pathways to restore damaged DNA. In the event that a lesion evades the process of DNA repair, translesion synthesis (TLS) by Y-family DNA Pols can occur, allowing bypass of the DNA lesion in an error-free or error-prone manner.3 Y-family DNA Pols are described as more “promiscuous” given their larger active sites when compared with replicative DNA Pols, which accounts for their ability to bypass damaged nucleotides that induce blockage. DNA Pol η in humans (hPol η) plays a pivotal role in the bypass of certain UV-induced DNA damage, which impedes DNA replication.4 hPol η activity has also been correlated with chemotherapeutic resistance to platinum-based agents such as cisplatin and the efficient bypass of the oxidative DNA lesion 7,8-dihydro-8-oxo-2′-deoxyguanosine.5,6 The O4-position of thymidine is susceptible to alkylation by agents such as N-nitroso alkylamines in certain foods, water, air, and particularly tobacco products.7,8 Albeit a minor site of alkylation, lesions such as O4-methylthymidine (O4MedT) and O4-ethylthymidine (O4EtdT) are poorly processed by mammalian repair pathways, making them persistent in the genome.9,10 O4MedT and O4EtdT hinder high fidelity replicative DNA Pol activity, resulting in misinsertion of dGTP in the daughter DNA strands.10–12 Correlations between the mutagenicity of O4MedT and cancer have been established,13,14 highlighting the importance of investigating the structural properties and biological outcomes associated with this type of DNA damage. The current understanding of the mechanism of Y-family DNA Pol misincorporation during TLS depends on a number of factors, including the nature of the DNA damage, the DNA Pol and the incoming nucleoside triphosphate. The geometrical array of the ternary complex formed (involving DNA, protein and nucleoside triphosphate) is the key characteristic that governs efficient bypass of a DNA lesion. Structural investigation by NMR and X-ray crystallography of duplexes containing an O4MedT insert has revealed that the methyl group preferentially adopts a syn conformation around the C4–O4 bond (Fig. 1).15,16 We hypothesized that the conformation of the O4-alkyl lesion could affect the base pair geometry during the primer extension reaction catalyzed by DNA Pol η. To address this possibility, we probed hPol η processivity with thymidine analogs that link the C5 and O4 atoms by a dimethylene or trimethylene group, which limits the O4-lesion to adopt an anti-conformation (Fig. 1) to relate the structural features of O4-alkylated dT with the bypass activity of hPol η. hPol η was selected as the model Y-family DNA polymerase, given previous studies concerning bypass of O4MedT and O4EtdT.17,18 Results of these studies indicated that Pol η, from yeast or human, were most efficient in extending across and past O4MedT17 and O4EtdT,18 respectively. Fig. 1 Structures of the (a) unmodified dT control, (b) O4MedT, (c) O4EtdT, (d) DFP and (e) TPP adducts. In bold are the bonds between the atoms shown to have a syn (b and c) or anti (d and e) orientation around the O4–C4 bond. Full primer extension assays for the control undamaged substrate (dT), O4MedT, O4EtdT, DFP, and TPP-bearing substrates with hPol η. The template strand sequence is 3′-AGCATTCGCAGTAXTACT-5′ where X denotes the modification and the 5′-FAM labeled primer strand sequence is 5′-TCGTAAGCGUCAT-3′. We investigated bypass profiles opposite all four lesions by hPol η (steady-state single nucleotide incorporation and LC-MS/MS analysis of full-length extension products). Crystal structures of ternary hPol η·DNA·dATP and hPol η·DNA·dGTP with template strands containing O4MedT or O4EtdT reveal a distinct orientation of the former lesion that stacks atop a tryptophan residue near the ceiling of the active site instead of pairing with the incoming nucleotide. Conversely, O4EtdT pairs with both incoming dA and dG nucleotides via bifurcated H-bonds in the insertion complexes and displays the same configuration opposite primer dG in the crystal structure of an extension complex adjacent to the nascent dG:dCTP pair. The structures provide a better understanding of the different behavior of the O4MedT or O4EtdT lesions in hPol η-catalyzed error-prone bypass reactions and suggests a unique intermediate step in the bypass of O4MedT. Results Synthesis and characterization of modified oligonucleotides The structures of O4MedT, O4EtdT and the modified pyrimidyl nucleosides 3-(2′-deoxypentofuranosyl)-5,6-dihydrofuro[2,3-d]pyrimidin-2(3H)-one (DFP) and 3-(2′-deoxypentofuranosyl)-3,5,6,7-tetrahydro-2H-pyrano[2,3-d]pyrimidin-2-one (TPP) are shown in Fig. 1 (methods describing the preparation of nucleosides and oligonucleotides can be found in the ESI†). UV thermal denaturation studies of duplexes containing single inserts of the DFP or TPP modification revealed a comparable destabilizing effect to O4MedT and O4EtdT with complementary strands containing adenine or any mismatched base pairing partner (Fig. S38†). Circular dichroism spectra of duplexes containing the DFP or TPP inserts revealed little deviation from a B-form structure (see Fig. S39†). Steady-state kinetics Steady-state kinetic assays of individual nucleotide incorporations opposite O4MedT, O4EtdT, DFP, TPP and unmodified dT were carried out with the catalytic core construct of hPol η (amino acids 1-432). In all cases, these pyrimidyl modifications blocked DNA synthesis by the polymerase relative to the unmodified control (Fig. 2a, values shown in Table S1†). Incorporation of the correct dAMP nucleotide by hPol η opposite to O4MedT, O4EtdT, DFP, and TPP was reduced approximately 6.5-, 12-, 4.5-, and 5-fold, respectively, relative to dT (see Fig. 2a). Fig. 2 (a) Steady-state incorporation efficiencies opposite dT, O4MedT, O4EtdT, DFP, and TPP by hPol η with individual dNTPs. (b) Incorporation frequencies based on ESI-MS/MS analysis of primer extension products opposite the dT, O4MedT, O4EtdT, DPF, and TPP containing templates (“_” indicates frameshift adduct formation. Tabulated kinetic parameters and fragments identified by LC-MS/MS analysis of full-length extension products can be found in ESI†). hPol η incorporated dCMP and dTMP opposite all the pyrimidyl modifications and dT with similar catalytic efficiencies (ranging from approximately 0.002–0.015 μM–1 s–1). However, a strong preference for either dAMP or dGMP incorporation opposite the modified pyrimidines was observed. Other than O4MedT, hPol η preferentially incorporated the correct dAMP nucleotide opposite all the pyrimidyl modifications. The significant incorporation of dGMP when hPol η encountered these pyrimidyl modifications, compared to the unmodified control, indicates a clear loss in substrate specificity by the polymerase (Fig. 2a). In the case of O4MedT, dGMP was slightly preferred as the nucleotide incorporated by hPol η (0.19 ± 0.01 vs. 0.18 ± 0.03 μM–1 s–1 for dGMP and dAMP, respectively). LC-MS/MS analysis of full-length extension products produced by hPol η Analysis of single insertions by a DNA polymerase is useful for kinetic evaluation but may not reflect incorporation fidelity in the presence of all four dNTPs across the damage and beyond this site. The fidelity of hPol η and its processivity past the damage site was addressed by the use of a full extension assay coupled with LC-MS/MS analysis.5,19,20 The optimal reaction times to observe the full extension products from the template strands containing the modifications and the unmodified control were evaluated (Fig. 1). Full extension was achieved for the unmodified control at 30 min, whereas templates containing the modifications required longer reaction times (60–90 min). The time course assay revealed that hPol η had difficulty in extending past O4MedT and O4EtdT and displayed a significant “S + 1” band at reaction times of 30 and 60 min. UPLC separation of the cleaved products and mass spectrometry analysis of their sequence identities revealed that dGMP was incorporated most efficiently opposite all the modifications (see Fig. 2b). The incorporation frequency opposite dT, O4MedT, O4EtdT, DFP, and TPP for the full extension products was evaluated (see Fig. 2b and Table S2†). The presence of O4MedT increased the level of frameshift formation by hPol η relative to the control (9.5 vs. 3.5%). Comparable levels of frameshifts were observed opposite O4EtdT and the dT control. However, the templates containing the bicyclic pyrimidine adducts did not induce a similar increase in frameshift formation with levels that were approximately one-half, relative to the control. The correct dAMP nucleotide was incorporated by hPol η at a frequency of approximately 30, 24, 22, 16, and 93% opposite O4MedT, O4EtdT, DFP, TPP, and dT, respectively. Out of the lesions investigated, hPol η exhibited the highest fidelity opposite O4MedT and lowest opposite the TPP. Incorporation of dGMP was observed to occur in the extension products with overall frequencies of 60, 70, 72, 82, 3% opposite O4MedT, O4EtdT, DFP, TPP, and dT, respectively. The accuracy of bypass varied for the O4-alkylthymidine modifications by approximately 2 : 1 in favor of dGMP opposite O4MedT to 5 : 1 in favor of dGMP opposite TPP. An increased adduct size, from O4MedT to O4EtdT and DFP to TPP, resulted in a 10% increase of dGMP misinsertion at the expense of a 10% decrease of the correct dAMP incorporation. Similarly, the conformationally restrained analogues (DFP and TPP) induced an increase in dGMP misinsertion (10%) by hPol η compared to O4MedT and O4EtdT, respectively. Pre-steady-state kinetics The pre-steady-state kinetic assays of dATP and dGTP incorporations opposite O4MedT, O4EtdT, DFP, and TPP, and dATP incorporation opposite unmodified dT were carried out with the catalytic core of hPol η. The burst rates for dATP insertion were 3.1-, 4.2-, and 1.8-fold higher compared to dGTP opposite O4MedT, DFP, and TPP, respectively (Fig. S48 and Table S3†). The burst rates were low in the case of dATP and dGTP opposite O4EtdT. The burst amplitudes for the extensions were 15–35% opposite O4MedT, O4EtdT, DPF, and TPP, which may indicate the presence of multiple non-productive ternary complexes. Crystal structures of ternary hPol η·DNA·dNTP complexes with templates containing O4MedT or O4EtdT at the insertion stage To visualize the O4MedT and O4EtdT lesions at the active site of hPol η trapped at the insertion stage, we determined four crystal structures of ternary complexes with the Pol bound to a 12mer template strand with the incorporated lesion and paired to an ; 8mer primer and incoming purine nucleoside triphosphate. For details regarding the crystallization, data collection and structure determination and refinement procedures please see the ESI.† Selected crystal data, data collection and refinement parameters and examples of the quality of the final electron density for all structures are summarized and depicted in the Table S4.† The two complexes with O4MedT-containing templates and incoming dATP or dGMPNPP reveal similar orientations of the lesion (Fig. 3, PDB ID codes ; 5DLF and ; 5DLG, respectively). Instead of pairing with the incoming nucleotide, O4MedT is lodged near the ceiling of the active site. Thus, its base portion is nestled against Trp-64 (base stacking interaction), Met-63 and Ser-62 (hydrophobic contacts between O4Me and both Cα and C( 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 1111111111111111111111111111111111 1111111111111111111111111111111111 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 1111111111111111111111111111111111 1111111111111111111111111111111111 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 0000000000000000000000000000000000 O) from the two residues) from a loop region in the finger domain and Gly-46 from an adjacent β-strand that together form the roof of the active site (Fig. 3a and c). In addition, O2 of O4MedT and the amino group of Asn-38 are H-bonded (Fig. 3b and d). This position of the O4MedT lesion at the entrance of the active site places it quite far away from the incoming nucleotide triphosphate. The distance between its O4 atom and N6 of dATP is 9 Å (8.2 Å between O4 of O4MedT and O6 of dGMPNPP). Fig. 3 Detached arrangement of incoming purine nucleotide triphosphate and O4MedT in two hPol η insertion-stage complexes. (a) Active site conformation in the complex with dATP opposite O4MedT viewed into the DNA major groove, and (b), rotated by 90° and viewed perpendicular to the adenine plane. (c) Active site conformation in the complex with dGMPNPP opposite O4MedT, viewed into the DNA major groove, and (d), rotated by 90° and viewed perpendicular to the guanine plane. Carbon atoms of O4MedT, the incoming nucleotide triphosphate, and selected hPol η amino acid side chains are colored in maroon, orange and magenta, respectively. Mg2+ and Ca2+ ions are cyan and pink spheres, respectively, and selected H-bonds are shown as dashed lines. Closer inspection of the positions of the incoming nucleotides shows that adenine stacks on the adjacent t(emplate)A:p(rimer)T pair, with the side chain of Arg-61 from the finger domain hovering closely above the adenine plane and engaged in H-bonds to the α- and β-phosphates of dATP via its guanidino moiety (Fig. 3b). By comparison, guanine is shifted into the minor groove and the stacking interaction with the adjacent tA:pT pair is slightly less favorable. The shift is likely a consequence of the altered orientation of the Arg-61 side chain that is extended in the structure of the complex with dGMPNPP, resulting in formation of H-bonds between its guanidino moiety and guanine O6 and N7 (Fig. 3d). The particular orientation of the incoming dG brings it closer to Asn-38 from the finger domain, but the distance of 3.64 Å between N2 of the former and the Oε1 oxygen of asparagine is slightly too long for formation of a H-bond. A surface rendering of the hPol η active site in the O4MedT insertion-stage complex with the incoming dATP indicates that the nucleobase moiety of the lesion fits snugly into the gap between Trp-64 and Ser-62 (Fig. S49†). It is clear that O4EtdT (with the ethyl group in the syn conformation) cannot be accommodated in the same fashion, as the longer substituent would clash with residues from the finger domain. Instead, the O4EtdT lesion has been pulled inside the active site and pairs with incoming dAMPNPP or dGMPNPP via bifurcated H-bonds in the two crystal structures of insertion-stage complexes with this lesion (Fig. 4a–d, PDB ID codes ; 5DQG and ; 5DQH, respectively). As in the complexes with O4MedT, Asn-38 forms a H-bond to O2 of O4EtdT in the minor groove. However, unlike in the O4MedT complex with incoming dGMPNPP, the side chain of Arg-61 in the corresponding complex with O4EtdT does not adopt an extended conformation to contact the major groove edge of guanine. As can be seen in Fig. 4, Arg-61 is directed toward the triphosphate moiety and forms a salt bridge with the α-phosphate group in both insertion-stage complexes. The methylene group (C1) of the O4 substituent in O4EtdT adopts an anti conformation in the complex with dAMPNPP (torsion angle C1–O4–C4–N3 = –142°) and a syn conformation in the complex with dGMPNPP (torsion angle C1–O4–C4–N3 = +62°). This is a clear difference to the structures of complexes with O4MedT, both of which show the lesion adopting a syn conformation (torsion angles C1–O4–C4–N3 of +33° and +25° in the dATP and dGMPNPP complexes, respectively). A further difference between the O4MedT and O4EtdT complexes is constituted by the orientations of template nucleotides 5′-adjacent to the lesions. In the former, A2 and T3 form a stack with Trp-42 outside the active site (Fig. 3). In the O4EtdT complexes, A2 is located outside the active site and forms a stacking interaction with Trp-42. However, T3 sits inside the active site and stacks onto O4EtdT (dAMPNPP complex; Fig. 4a). In the complex with dGMPNPP, T3 juts into the major groove (Fig. 4c). Thus, neither orientation adopted by T3 in these complexes resembles that of O4MedT, lodged near the ceiling of the active site and stacked onto Trp-64. Fig. 4 Pairing between incoming purine nucleotide triphosphate and O4EtdT in two hPol η insertion-stage complexes (a–d). (a) Major groove view of the Watson–Crick pair between dAMPNPP and O4EtdT with formation of bifurcated H-bonds. (b) The active site rotated by 90° relative to A, with the base pair viewed perpendicular to the adenine plane. (c) Major groove view of the sheared configuration between dGMPNPP and O4EtdT, with formation of bifurcated H-bonds. (d) The active site rotated by 90° relative to C, with the base pair viewed perpendicular to the guanine plane. Carbon atoms of O4EtdT, the incoming nucleotide triphosphate, and selected hPol η amino acid side chains are colored in maroon, orange and magenta, respectively. Mg2+ ions are cyan spheres and selected H-bonds are shown as dashed lines. Active site configuration in a ternary hPol η extension-step complex with O4EtdT opposite primer dA, followed by template dG opposite incoming dCTP (e and f). (e) Major groove view of the active site, with carbon atoms of O4EtdT and the paired dA colored in maroon and light blue, respectively. Carbon atoms of the nascent dG:dCTP pair are colored in orange, carbon atoms of selected hPol η amino acid side chains are colored in magenta, and those of the 3′-terminal primer dC are colored in purple. (f) The active site rotated by 180° relative to A and viewed into the minor groove. Crystal structure of a ternary hPol η·DNA·dCTP extension-stage complex with O4EtdT paired opposite primer dA The structure of a complex with O4EtdT paired to dA at the –1 position followed by template dG opposite incoming dCTP was determined at 2.05 Å resolution (Fig. 4e and f, PDB ID codes ; 5DQI). The geometry of the O4EtdT:dA pair replicates that seen in the insertion complex with O4EtdT opposite incoming dAMPNPP (Fig. 4a, b, e and f). As in the case of the latter, the ethyl group has moved outside the thymine plane and adopts an anti orientation (torsion angle C1–O4–C4–N3 = –135°). The base pair itself adopts a Watson–Crick like geometry with formation of a single H-bond; the adjacent dG:dCTP pair displays a standard geometry with three H-bonds. Arg-61 is directed toward the phosphate moieties of the incoming nucleotide and forms two salt bridges with the α- and β-phosphate groups, and Asn-38 is engaged in two H-bonds with N3 and O4′ of template dG. The most unusual feature of the extension-stage structure is the presence of an additional nucleotide at the 3′-end of the primer (Fig. 4f). Because the crystallization solutions contained dCTP and the residual electron density is consistent with a pyrimidine, we extended the primer by dC (Fig. S50†). We suspect that hPol η possesses weak catalytic activity with Ca2+ as the cofactor or that traces of Mg2+ present in the crystallizations led to primer extension in situ (even a very low activity could result in extension by a single nucleotide over the course of two weeks). We showed earlier that the translesion DNA polymerase Dpo4 from Sulfolobus solfataricus is able to catalyze nucleotide insertion with Ca2+, although the activity is far below that seen with Mg2+ as the prosthetic group.21 The additional dC stacks against the backbone of the template strand in the minor groove and its position is further stabilized by two H-bonds between N3 and O2 and the guanidino moiety of Arg-111 (Fig. 4f). Discussion The known toxicity of alkylated adducts at the O4-position of thymidine prompted us to explore the influence of restricting orientation of the alkyl group around the C4–O4 bond to an anti conformation in translesion synthesis catalyzed by hPol η. In these studies, the bicyclic pyrimidine analogs DFP and TTP, which link the C5 and O4 atoms with a di- or trimethylene linker, were evaluated in addition to O4MedT and the bulkier O4EtdT lesion. Conformationally locked analogs of damage that can occur at the nucleobase have been previously synthesized and employed in studies which have provided insights into the requirements for DNA repair processes.22,23 UV thermal denaturation studies of oligomers containing DFP and TTP revealed similar influences on duplex stability to both complementary and mismatched nucleobases compared to O4MedT and O4EtdT. The most stable pairing of either DFP or TTP was with dG, also observed with O4MedT and O4EtdT. NMR studies of a duplex containing an O4MedT·dG pair revealed, in addition to the O4-methyl group adopting a syn-conformation, that the base formed a Watson–Crick “like” pairing with a single hydrogen bond.24 In this structure, the syn-orientation of the O4–Me group influences the hydrogen bond between the imino proton of dG and the N3 atom of O4MedT by increasing the distance between the O6 and O4 atoms of dG and O4MedT, respectively. Limiting the orientation of the methylene group at the O4-atom to the anti-conformation, in the case of the DFP and TTP modifications, appears to have a minimal impact on the interaction with dG and duplex stability. In pairing with dA, a similar drop in duplex stability compared with dT is observed for oligonucleotides containing the O4MedT, O4EtdT, DFP, and TTP modifications. The NMR structure of a duplex containing an O4MedT·dA pair indicated that the O4–Me group is syn and that the bases adopt a wobble alignment with one hydrogen bond formed between the imino nitrogen of O4MedT and the amino group of dA.15 The restricted anti-orientation of the methylene group for DFP or TPP modifications does not significantly impact duplex stability compared to O4MedT or O4EtdT. Steady-state kinetics of individual nucleotide incorporation opposite the DFP and TTP modifications by hPol η demonstrated preferred insertion of purine nucleotides relative to the pyrimidines, similar to O4MedT and O4EtdT. The efficiency of nucleotide insertion (kcat/Km) for the correct nucleotide (dAMP) across the lesions followed the order DFP > TPP > O4MedT > O4EtdT. For dGMP, a similar efficiency of nucleotide insertion occurred for DFP, TPP, and O4MedT whereas a drop was observed for O4EtdT. In agreement with studies involving Saccharomyces cerevisiae DNA polymerase η (yPol η), a reduction in incorporation efficiency due to the presence of an O4MedT insert was observed.17 However, whereas the yeast homolog revealed a significant preference for dGMP, which was incorporated approximately 80 times more efficiently than dAMP,17 hPol η displayed almost equal selectivity at incorporating dAMP (f = 0.94) as dGMP opposite O4MedT. A comparable 80-fold preference for dGMP over dAMP was exhibited by yPol η for the bulkier O4-carboxymethylthymidine lesion.25 The rationale for the preferred incorporation of dGMP opposite O4MedT by yPol η was attributed to a dG·O4MedT wobble base pairing. Differences observed for nucleotide incorporation opposite O4MedT by the yeast and human homologs of Pol η may be influenced in part by different sequence contexts, as previously observed.5,26 In addition, homologs of Pol η have exhibited differences in nucleotide incorporation across some types of DNA damage. For example, yPol η accurately inserts dCMP across 8-oxodG whereas hPol η is less accurate, inserting some dAMP across this lesion as well.5,27 Interestingly, similar misinsertion profiles have been observed in bypass experiments of hPol η and yPol η with O6MedG, a lesion which also protrudes in the major groove of the DNA duplex.28 For the bulkier O4EtdT and conformationally restricted analogs DFP and TTP, a preference for nucleotide incorporation of dAMP over dGMP was observed. For O4EtdT, hPol η was more proficient at incorporating dAMP over dGMP with catalytic efficiencies of 0.10 and 0.06 μM–1 s–1, respectively. These values are approximately two-fold lower compared to those observed for the O4MedT-containing template, but can be rationalized by the increased bulk of the ethyl group, which may influence dNTP incorporation in the hPol η active site. Ethylation of the O4-position of dT has been shown to stall the human Y-family DNA polymerases hPol κ and hPol ι but not hPol η (although steady-state analysis was not reported for oligonucleotides containing O4EtdT).18 Bypass of O4EtdT by hPol η revealed dGMP misincorporation at 55% compared to 19% for dAMP insertion, in agreement with our data despite different sequence contexts. For the conformationally restricted DFP and TPP modifications, incorporation efficiency was observed to be ∼1.5-fold higher for dAMP (0.27 and 0.24 μM–1 s–1, respectively) and comparable for dGMP relative to O4MedT. These results demonstrate that hPol η is more proficient at incorporating both the correct (dAMP) and incorrect (dGMP) nucleotides across from these more conformationally restricted lesions. In addition, the increase in steric bulk from the DFP to TPP slightly decreases incorporation efficiency. Exposure of the hydrogen bonding face of the DFP or TPP modifications may have a greater influence on stabilizing the wobble alignment geometry that has been suggested for the O4MedT·dA pairing. The conformational restriction of the alkyl group to an anti-orientation around the C4–O4 bond, as in the DFP and TPP modifications, would direct the O4-methylene group away from the amino group of dA, which could account for the enhancement of incorporation of the correct nucleotide (dAMP) compared to O4MedT. Incorporation of dGMP may not be as influenced by orientation of the alkyl group around the C4–O4 bond as the proposed hydrogen bonding interaction, based on the NMR structure of the duplex containing the O4MedT·dG, which occurs between the amino group of dG and O2-atom of O4MedT.24 In the case of O4EtdT, the combination of the syn-orientation and the size of the ethyl group may both contribute to the reduced efficiency of nucleotide insertion of dAMP and dGMP in this series. Primer extension reactions in the presence of all four dNTPs for templates containing the O4MedT, O4EtdT, DFP, and TTP modifications demonstrated that hPol η was proficient at incorporating nucleotides across and past the adducted site. However, both O4MedT and O4EtdT exhibited a greater accumulation of non-full length oligonucleotide products at reduced reaction times (30 and 60 min), which was not observed for bicyclic DFP or TPP analogs. The LC-MS/MS analyses of the extension products from the in vitro primer bypass studies revealed that dGMP incorporation across the lesion was preferred over dAMP in all cases except the control (dT). The ratio of dGMP : dAMP incorporation by hPol η, assessed from the extension products, was found to decrease in the series TPP (4.6 : 1) > DFP (3.3 : 1) ≈ O4EtdT (3.2 : 1) > O4MedT (2.1 : 1). The presence of the larger alkyl group for O4EtdT or the analogs with the O4-methylene group in an anti-conformation (TPP and DFP) clearly promotes dGMP misincorporation in the presence of all four nucleotides. In addition, the DFP and TPP modification were not found to induce a significant amount of frameshifts in the products compared to O4MedT. Differences in fidelity observed between the steady-state kinetic and LC-MS/MS full-length experiments have been observed previously.17,19 The variance may be attributed to accommodation of the incoming dGTP relative to dATP for these modifications, highlighting that adduct size and the conformation of the O4-methylene group can influence interactions in the active site of hPol η. It should be noted, however, that other steric and/or stereoelectronic effects may have an impact on hPol η bypass processivity of the conformationally locked analogues relative to O4MedT or O4EtdT, respectively. In the case of the analogs investigated, hPol η continued extension of the primer in an error-free manner after incorporation of dATP or dGTP across from the damaged site on the template. Several observations based on nucleotide incorporation profiles attest to the distinct effects on hPol η bypass synthesis exerted by the O4MedT and O4EtdT lesions. These concern (i) the more error-prone bypass caused by O4MedT, i.e. dGTP is favored relative to dATP (Fig. 2a), (ii) increased accumulation of the +1 product in the full-length extension reaction for O4MedT (Fig. 1), and (iii) significantly more frameshift products caused by the O4MedT lesion (Fig. 2b). Interestingly, the structural data for insertion-stage hPol η complexes with either O4MedT or O4EtdT in the template strand reveal starkly different orientations of the two adducted nucleotides at the active site. O4MedT is trapped in an orientation that keeps it at a considerable distance from the incoming purine nucleotide triphosphates. Conversely, O4EtdT pairs opposite both dATP and dGTP with formation of bifurcated H-bonds (whereby the latter pair features a sheared orientation of the two partners, with G being pushed toward the minor groove). The increased proclivity for insertion of dG opposite O4MedT compared to O4EtdT is not surprising if one considers the strict preference by the O4-methyl substituent for a syn conformation. The syn conformation precludes adoption of an O4MedT:dA pair with standard Watson–Crick geometry, but the sheared pairing mode seen in the case of O4EtdT:dG(MPNPP) (Fig. 4c and d), also presumably adopted by the O4MedT:dG pair, is compatible with a syn conformation of the substituent. This conclusion is borne out by the observations that the ethyl moiety in the O4EtdT:dA(MPNPP) pairs assumes an anti conformation (Fig. 4a, b, e and f), whereas its conformation is syn in the O4EtdT:dGMPNPP pair. Furthermore, the TPP adduct opposite dGMPNPP was modeled from the O4EtdT:dG(MPNPP) ternary crystal structure coordinates. The configuration of the adduct seen in the model is consistent with the enhanced incorporation of dG observed in the primer extension experiments since the constrained anti conformation of the bicyclic system does not hinder the guanine nucleobase from shifting towards the major groove and potentially form two H-bonds with TPP (Fig. S51†). On one hand, one could argue that the higher fraction of frameshifts for O4MedT relative to O4EtdT is consistent with the structural data that show the former is not engaged opposite the incoming nucleotide but trapped adjacent to the ‘entrance’ of the active site. Perhaps the more pronounced accumulation of the +1 product in the case of the full-length extension reactions opposite O4MedT compared to the other O4 adducts tested here are the result of non-templated insertion. Thus, purine nucleoside triphosphates would be favored and their incorporation would not be affected by the particular conformation of the O4-methyl group, syn or anti. This scenario is certainly not inconsistent with the structural data that reveal no interaction between the O4MedT lesion and the incoming dATP or dGMPNPP. Clearly, it is intriguing that both activity and structural data show distinct consequences of the O4MedT and O4EtdT lesions for bypass by hPol η. However, it is important to note that the position of the O4MedT lesion at the active site, unique among all crystal structures of hPol η complexes analyzed to date, represents one state during bypass. Perhaps other orientations and interactions of the adducted nucleotide occur during bypass, which precludes all steps involved in the mechanism of O4-alkyl bypass synthesis by hPol η. All results from the study (kinetic evaluation, full extension assays and crystal structures) may be integrated into one potential extension model. The crystal structure supports the notion that the O4MedT nucleobase is indeed nestled at the ceiling for an undefined period of time. The purine nucleoside triphosphates could then be imported into the active site, and would be subject to a template gap, analogous to being opposite an abasic site. According to full extension assays, there are approximately twice as many dG inserted by hPol η relative to dA. The lack of O4MedT – incoming dNTP clash in the active site may explain higher kp values observed for dATP and dGTP for O4MedT compared to O4EtdT (by 8.8 or 9.3 fold respectively). This may also aid in explaining the preference for purine insertion, in an approximate 2 : 1 dG : dA ratio across O4MedT, whereas a larger ratio is observed for all the other modifications as discussed above. Perhaps the other modifications prevent the modified nucleobases from accessing the conformation at the top of the active site. The ethyl group, albeit bulkier, can populate the other conformations (syn versus anti) as observed in the crystallographic data, which may contribute to the higher correct dA insertion relative to the bicyclic analogues which are locked in an anti conformation. Increase in bulk (from O4MedT to O4EtdT and DFP to TPP) leads to an increase of incorrect dG insertion. Eventually the O4MedT is required to move from the top of the active site back to the 0 or –1 (post-replicative) position(s). It is possible that this mobility causes the frameshift adduct formation observed in the full extension assays. As a result, we have provided intriguing insights on a potentially different bypass mechanism of O4MedT in comparison to the larger O4EtdT adduct. Conclusions Oligonucleotides containing DFP and TPP, designed as analogs of O4-alkylated thymidine, were synthesized to explore the influence of limiting the O4-alkyl lesion to an anti-orientation on nucleotide incorporation by hPol η. These modifications were shown to destabilize the DNA duplex, based on UV thermal denaturation studies, regardless of the base-pairing partner (A, G, T, or C), similar to O4MedT and O4EtdT. Primer extension assays demonstrated that these pyrimidyl modifications hindered nucleotide incorporation by hPol η. Single nucleotide incorporation studies revealed increased selectivity towards dAMP over dGMP that followed the order O4EtdT > DFP ≈ TPP. A slight preference for dGMP over dAMP incorporation was observed for O4MedT. LC-MS/MS analysis of primer extension studies (in the presence of all four dNTPs) revealed that hPol η incorporated dGMP over dAMP across the lesions in the order TPP > DFP ≈ O4EtdT > O4MedT. These trends suggest that limiting the orientation of the O4-alkylene group enhances the proficiency of dNTP incorporation by hPol η across O4-alkylated dT damage. In the presence of all four dNTPs, error-prone nucleotide incorporation by hPol η is enhanced by restricting the O4-lesion to an anti-orientation. This study exemplifies how restricting a lesion's conformational freedom impacts bypass profile by hPol η. Moreover, our results provide mechanistic insights into the mutagenicity of the biologically relevant O4MedT and O4EtdT DNA adducts. Supplementary Material Supplementary informationClick here for additional data file. We thank the Natural Sciences and Engineering Research Council of Canada (Grant No. 299384-2011 to C. J. W.), the Canada Research Chair Program (Grant No. 950-213807 to C. J. W.) and the US NIH for financial support (Grants No. ES010375 to F. P. G. and M. E., Grant No. CA160032 to M. E., and ES010546 to F. P. G.). D. K. O. thanks NSERC for a Canada Graduate Scholarship (CGS) and Michael Smith Foreign Study Supplement Scholarship, and the NSERC CREATE Program in Bionanomachines for support of an exchange with Vanderbilt University. Vanderbilt University is a member institution of the Life Sciences Collaborative Access Team at sector 21 of the Advanced Photon Source, Argonne, IL. Use of the Advanced Photon Source at Argonne National Laboratory was supported by the United States Department of Energy, Office of Basic Energy Sciences (grant DE-AC02-06CH11357). ==== Refs Shrivastav N. Li D. Essigmann J. M. Carcinogenesis 2010 31 59 70 19875697 Stone M. P. Huang H. Brown K. L. Shanmugam G. Chem. Biodiversity 2011 8 1571 1615 Sale J. E. Lehmann A. R. Woodgate R. Nat. Rev. Mol. Cell Biol. 2012 13 141 152 22358330 McCulloch S. D. Kokoska R. J. Masutani C. Iwai S. Hanaoka F. Kunkel T. A. Nature 2004 428 97 100 14999287 Patra A. Nagy L. D. Zhang Q. Su Y. Muller L. Guengerich F. P. Egli M. J. Biol. Chem. 2014 289 16867 16882 24759104 Albertella M. R. Green C. M. Lehmann A. R. O'Connor M. J. Cancer Res. 2005 65 9799 9806 16267001 Singer B. Nature 1976 264 333 339 1004554 Tannenbaum S. R. Lancet 1983 1 629 632 6131307 Bronstein S. M. Skopek T. R. Swenberg J. A. Cancer Res. 1992 52 2008 2011 1551130 Altshuler K. B. Hodes C. S. Essigmann J. M. Chem. Res. Toxicol. 1996 9 980 987 8870985 Singer B. Spengler S. J. Fraenkel-Conrat H. Kusmierek J. T. Proc. Natl. Acad. Sci. U. S. A. 1986 83 28 32 3455756 Preston B. D. Singer B. Loeb L. A. Proc. Natl. Acad. Sci. U. S. A. 1986 83 8501 8505 3464967 Singer B. Cancer Res. 1986 46 4879 4885 3530434 Swenberg J. A. Dyroff M. C. Bedell M. A. Popp J. A. Huh N. Kirstein U. Rajewsky M. F. Proc. Natl. Acad. Sci. U. S. A. 1984 81 1692 1695 6584902 Kalnik M. W. Kouchakdjian M. Li B. F. Swann P. F. Patel D. J. Biochemistry 1988 27 100 108 2831951 Brennan R. G. Pyzalska D. Blonski W. J. Hruska F. E. Sundaralingam M. Biochemistry 1986 25 1181 1185 3964665 Andersen N. Wang J. Wang P. Jiang Y. Wang Y. Chem. Res. Toxicol. 2012 25 2523 2531 23113558 Andersen N. Wang P. Wang Y. Chem. Res. Toxicol. 2013 26 1730 1738 24134187 Choi J. Y. Chowdhury G. Zang H. Angel K. C. Vu C. C. Peterson L. A. Guengerich F. P. J. Biol. Chem. 2006 281 38244 38256 17050527 Patra A. Zhang Q. Lei L. Su Y. Egli M. Guengerich F. P. J. Biol. Chem. 2015 290 8028 8038 25666608 Madril A. C. Johnson R. E. Washington M. T. Prakash L. Prakash S. J. Biol. Chem. 2001 276 42857 42862 11551952 Daniels D. S. Woo T. T. Luu K. X. Noll D. M. Clarke N. D. Pegg A. E. Tainer J. A. Nat. Struct. Mol. Biol. 2004 11 714 720 15221026 Abdu K. Aiertza M. K. Wilkinson O. J. Grasby J. A. Senthong P. Povey A. C. Margison G. P. Williams D. M. Chem. Commun. 2012 48 11214 11216 Kalnik M. W. Kouchakdjian M. Li B. F. Swann P. F. Patel D. J. Biochemistry 1988 27 108 115 3349021 Swanson A. L. Wang J. Wang Y. Biochemistry 2011 50 7666 7673 21809836 Rao S. Chenna A. Slupska M. Singer B. Mutat. Res. 1996 356 179 185 8841483 Haracska L. Yu S. L. Johnson R. E. Prakash L. Prakash S. Nat. Genet. 2000 25 458 461 10932195 Haracska L. Prakash S. Prakash L. Mol. Cell. Biol. 2000 20 8001 8007 11027270

PMC005xxxxxx/PMC5001693.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756401410.1371/journal.pone.0161831PONE-D-15-53089Research ArticleBiology and Life SciencesAnatomyDigestive SystemGastrointestinal TractMedicine and Health SciencesAnatomyDigestive SystemGastrointestinal TractMedicine and Health SciencesOncologyCancers and NeoplasmsHematologic Cancers and Related DisordersLymphomasMedicine and Health SciencesHematologyHematologic Cancers and Related DisordersLymphomasMedicine and Health SciencesOncologyCancer TreatmentMedicine and Health SciencesSurgical and Invasive Medical ProceduresBiology and Life SciencesAnatomyDigestive SystemGastrointestinal TractSmall IntestineMedicine and Health SciencesAnatomyDigestive SystemGastrointestinal TractSmall IntestineBiology and Life SciencesAnatomyHistologyMedicine and Health SciencesAnatomyHistologyBiology and Life SciencesImmunologyImmune System ProteinsImmune ReceptorsT Cell ReceptorsMedicine and Health SciencesImmunologyImmune System ProteinsImmune ReceptorsT Cell ReceptorsBiology and Life SciencesBiochemistryProteinsImmune System ProteinsImmune ReceptorsT Cell ReceptorsBiology and Life SciencesCell BiologySignal TransductionImmune ReceptorsT Cell ReceptorsBiology and Life SciencesBiochemistryProteinsT Cell ReceptorsMedicine and Health SciencesOncologyCancer TreatmentChemotherapyMedicine and Health SciencesClinical MedicineClinical OncologyChemotherapyMedicine and Health SciencesOncologyClinical OncologyChemotherapyMedicine and Health SciencesPharmaceuticsDrug TherapyChemotherapyPrimary Intestinal Extranodal Natural Killer/T-Cell Lymphoma, Nasal Type: A Comprehensive Clinicopathological Analysis of 55 Cases Primary Intestinal ENKTCLYu Bao-Hua 12Shui Ruo-Hong 12Sheng Wei-Qi 12Wang Chao-Fu 12Lu Hong-Fen 12Zhou Xiao-Yan 12Zhu Xiong-Zeng 12Li Xiao-Qiu 12*1 Department of Pathology, Fudan University Shanghai Cancer Center, Shanghai, China2 Department of Oncology, Shanghai Medical College, Fudan University, Shanghai, ChinaBusson Pierre EditorGustave Roussy, FRANCECompeting Interests: The authors have declared that no competing interests exist. Conceived and designed the experiments: BHY XQL. Performed the experiments: BHY RHS XYZ. Analyzed the data: WQS CFW HFL XYZ XZZ XQL. Wrote the paper: BHY XQL. * E-mail: leexiaoqiu@hotmail.com26 8 2016 2016 11 8 e016183110 12 2015 19 7 2016 © 2016 Yu et al2016Yu et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Purpose To investigate the clinicopathological features, survival and prognostic factors of primary intestinal extranodal natural killer/T-cell lymphoma, nasal type (PI-ENKTCL). Methods Clinical and histological characteristics of PI-ENKTCL cases were retrospectively evaluated. Immunohistochemical phenotype and status of Epstein-Barr virus (EBV) and T-cell receptor (TCR) gene rearrangement were examined. The overall survival and prognostic parameters were also analyzed. Results Fifty-five (2.7%) cases with PI-ENKTCL were identified out of 2017 archived ENKTCL cases, with a median age of 39 years and a male to female ratio of 2.1:1. The most common symptom was abdominal pain (90.9%), accompanied frequently with fever and less commonly with intestinal perforation or B symptoms. Small intestine (50.9%) was the most common site to be involved. 47.3% and 36.4% cases presented with stage I and II diseases, respectively. Histologically, most cases displayed characteristic morphologic changes of ENKTCL. Cytoplasmic CD3, TIA-1 and CD56 expression was found in 100%, 94.5% and 89.1% of cases, respectively. In situ hybridization detection for EBV demonstrated positive results in all cases. Monoclonal TCR gene rearrangement was found in 52.9% of tested cases. Chemotherapy with a DICE or L-asparaginase/peg-asparginase-containing regimen was most often employed. Both advanced tumor stage and B symptoms were independent inferior prognostic factors (p = 0.001 and p = 0.010). Noticeably, 6 cases demonstrated a CD4-positive phenotype. These cases featured a relatively older median age (58 years), predominance of small/medium-sized neoplastic cells, a higher rate of TCR rearrangement and slightly favorable outcome. Conclusion We reported by far the largest series of PI-ENKTCL, and demonstrated its heterogeneity, aggressive clinical behavior and unsatisfying response to the current therapeutic strategies. Those CD4-positive cases might represent a unique subtype of PI-ENKTCL or distinct entity. Further investigations are required for the better understanding and management of this unusual disease. The authors received no specific funding for this work. Data AvailabilityAll relevant data are within the paper and its Supporting Information files.Data Availability All relevant data are within the paper and its Supporting Information files. ==== Body Introduction Extranodal natural killer (NK)/T-cell lymphoma, nasal type (ENKTCL), a rare distinct malignancy that comprises 3–8% of all lymphomas, is most prevalent in Asian and Central and South American populations [1]. This tumor predominantly involves extranodal sites, and features pathologically vascular invasion, prominent necrosis, cytotoxic phenotype and association with Epstein-Barr virus (EBV) [2, 3]. While most ENKTCL cases derived from NK cells, some show a cytotoxic T-cell phenotype [3, 4]. Approximately 80% of ENKTCL cases occur in the upper aerodigestive tract, with the nasal cavity being mostly affected [5, 6], other preferential sites of involvement include the skin, soft tissue, gastrointestinal (GI) tract and the gonad [7, 8]. Primary intestinal ENKTCL (PI-ENKTCL) is rare, which accounts for 3.1% of all intestinal non-Hodgkin lymphoma cases according to the literature [9]. Due to its rarity, the clinical and pathological features of intestinal ENKTCL have not been well illustrated, which may lead to dilemmas not only in the diagnosis but also the treatment of this disease. Furthermore, data regarding the therapeutic strategies and prognostic factors of this peculiar lymphoma is still limited. We thus retrospectively analyzed by far the largest series of PI-ENKTCL, for the purpose of better understanding the clinicopathological features of this rare tumor, which is of paramount importance for the accurate diagnosis and appropriate treatment. Materials and Methods Case selection Altogether 55 cases with PI-ENKTCL, diagnosed between January 2007 and August 2015, were retrieved from the files of Department of Pathology, Fudan University Shanghai Cancer Center (Shanghai, China). For all cases, a primary intestinal manifestation of the disease was confirmed by a precise staging work-up through a computed tomography (CT) staging and/or PET-CT scan. And those with a probable secondary involvement of the intestine were not included. Pathological diagnosis was made by two of the authors (BHY and XQL) according to the criteria described in the WHO classification of tumors of the haematopoietic and lymphoid tissues [2]. Clinical data including the follow up information were also collected and analyzed. Histology and immunohistochemistry Archival formalin-fixed paraffin-embedded tumor tissues were recut for a routine hematoxylin and eosin (H&E) stain and the immunohistochemical procedure. Histological characteristics, including the cytological details of the tumor, the presence of necrosis and ulceration, angiocentricity and angiodestruction, admixed inflammatory infiltrates, and the depth of the neoplastic infiltration, were reviewed and assessed by three of the authors (BHY, XYZ and XQL) under a multi-headed microscope. The immunohistochemical study was performed using a Ventana Bench Mark ultra autostainer (Ventana Medical System Inc., Roche Tuson, AZ, USA) and the Ventana ultra view universal DAB detection kit. The primary antibodies against CD20, CD2, cytoplasmic CD3 [CD3 (epsilon)], CD4, CD5, CD7, CD8, CD30, CD56, T-cell-restricted intracellular antigen-1 (TIA-1), granzyme B (GrB), perforin, ALK1 and Ki-67 were employed in the present study. Except for TIA-1 (Dako, Glostrup, Denmark), all of the above-mentioned antibodies were commercial products by Roche Ventana. For each stain, a parallel stain using appropriate positive and negative controls was performed. The Ki-67 labeling index was estimated to the closest decile. In situ hybridization (ISH) detection for EBV-encoded small RNA (EBER) The status of EBV infection was assessed by an ISH detection for EBER on paraffin-embedded tissue sections using fluorescein-labeled oligonucleotide probes (INFROM EBER Probe, Ventana), as previously described [10]. The visualization system used was the Bench Mark XT with enzymatic digestion (ISH Protease 2, Ventana) and the iVIEW Blue v3 detection kit (Ventana). Appropriate positive and negative control sections were included for each run. Polymerase chain reaction (PCR) assays for T-cell receptor (TCR) gene rearrangement Genomic DNA was extracted from formalin-fixed paraffin-embedded tissues, using the QIAamp mini kit (Qiagen, GmbH, Germany), and the concentration of DNA was measured by a spectrophotometer. Rearrangement of TCR-β, γ, δ genes was detected by multiplex PCR assays according to standard techniques, as described previously [11]. Amplifiability of the DNA was confirmed by concurrent PCR amplification of the β-globin sequence. Each PCR study was carried out in duplicate and included positive, negative, and no-template controls. The PCR products were analyzed by capillary electrophoresis as previously documented [11], using the ABI PRISM 310 Genetic Analyzer (Applied Biosystems, CA, USA). Statistical analysis Overall survival (OS) was defined as the interval from the initial diagnosis to the date of death from any cause or the last contact. The OS was estimated using the Kaplan-Meier method and was compared by means of the log-rank test. Multivariate analyses were also carried out using the Cox proportional hazard regression model to identify prognostic factors. The clinicopathologic parameters for assessment included age, tumor location, stage of disease, B symptoms [including fever (temperature >38°C) for 3 consecutive days, night sweats, and/or weight loss exceeding 10% of body weight in 6 months], intestinal perforation, size and immunophenotype of tumor cells, as well as TCR gene rearrangement status. All the statistical analyses were conducted using the SPSS software package (SPSS version 19.0; SPSS Inc., Chicago IL, USA). A p value of <0.05 was considered statistically significant. Ethics statement This study was approved by the Institutional Review Board of Fudan University Shanghai Cancer Center (Shanghai Cancer Center Ethical Committee). The patient records/information was anonymized and de-identified prior to analysis. Results Clinical findings Altogether 2017 ENKTCL cases were documented in our laboratory database during the period between 2007 and 2015, among which only 55 (2.7%) were proved to present with primary intestinal lesions. There were 37 male and 18 female patients, with a male-to-female ratio of 2.1:1. The average and median age at diagnosis was 43 and 39 years, respectively (range, 14–75 years). The most common symptom at diagnosis was abdominal pain (50 patients, 90.9%). Thirty-one patients (56.4%) had fever, and 10 (18.2%) presented with lower GI bleeding or fecal occult blood. Intestinal perforation and B symptoms were observed in 18 (32.7%) and 19 (34.5%) patients, respectively. Other concomitant clinical manifestations included diarrhea, nausea, vomiting, abdominal mass, intestinal obstruction and constipation. With regard to the anatomic sites of involvement, 28 (50.9%) tumors occurred in the small intestine (including the duodenum, jejunum and ileum). Eleven (20.0%) and 13 (23.6%) were located in the ileocecal junction and colon, respectively. And the remaining 3 (5.5%) had multifocal lesions involving at least two different intestinal segments. Mesentery lymph node involvement was observed in 14 (36.8%) out of 38 cases. According to the Lugano staging system, 26 cases (47.3%) in the current series presented with stage I diseases, and 20 (36.4%) and 9 (16.4%) with stage II and advanced stage (stage III/IV) diseases, respectively. Histological findings Overall, the specimens from the 55 cases comprised 42 resected tumors and 13 endoscopical biopsies. Ulceration and necrosis of the overlying mucosa, ranging from focal to extensive, were noticed in all cases. The atypical lymphoid cells infiltrated and effaced the mucosal architecture, which is more easily to be appreciated in the resected tumors (Fig 1A). Geographic necrosis was also frequently observed. Angiocentric /angiodestructive growth pattern of the tumor cells was another common feature noticed in these specimens (Fig 1B). Of the 42 cases with resected specimens, tumor invaded muscularis propria in 2 (4.8%), subserosa in 30 (71.4%), and penetrated the intestinal wall and spread beyond in 10 (23.8%) cases. Cytologically, the neoplastic lymphoid cells exhibited a broad spectrum in their size and appearance. Twenty-four cases (43.6%) displayed a mixed population of small, medium to large cells, 18 cases (32.7%) were composed predominantly of small to medium-sized cells, 12 cases (21.8%) predominated by medium-sized cells, and the remaining one (1.8%) was composed of uniform small cells. The small and intermediate cells often had oval or slightly irregular convoluted nuclei, hyperchromatic or granulated chromatin and inconspicuous nucleoli. The larger cells might have exceedingly irregular nuclei with vesicular or coarsely clumped chromatin and prominent nucleoli. Moderate to abundant, clear or lightly eosinophilic cytoplasm was seen in most cases. In occasional cases, the tumor cells showed a striking atypia with noticeable horseshoe- or kidney-shaped nuclei (Fig 1C). In addition, mononucleated or multinucleated giant cells with striking bizarre appearance were identified in 5 (9.1%) cases. Mitotic figures and varying amounts of apoptotic bodies were easily identified in most cases. Inflammatory infiltrates, consisting of small lymphocytes, plasma cells, histiocytes and eosinophils, were admixed with the tumor cell populations in variable proportions from case to case (Fig 1D). In one case, deposits of calcified schistosome ova were noticed in the intestinal wall and lymph nodes. 10.1371/journal.pone.0161831.g001Fig 1 Histological features of PI-ENKTCL. (A) A low power view of the H&E stained section showed the overlying mucosa was partially effaced and the neoplastic cells extensively infiltrated the muscular wall. (B) A blood vessel with invasion by the neoplastic cells was shown. (C) A high power view demonstrated that the tumor was composed of a mixture of small, medium and large atypical lymphoid cells, with some hallmark cells identified. (D) Numerous eosinophils and plasma cells were intermingled with the tumor cells. Immunohistochemistry, ISH for EBER, and TCR gene rearrangement Immunohistochemically, tumor cells of all cases were positive, either diffusely or partially, for cytoplasmic CD3. Positivity for CD4 and CD8 of tumor cells was found in 14.0% (6/43) and 2.4% (1/41) of the tested cases, respectively, but none expressed both antigens. Other T or NK markers, such as CD2, CD5 and CD7, were expressed in a variable proportion of all cases. Positive CD56 immunostaining was observed in 89.1% (49/55) of cases. Tumor cells were consistently positive for the cytotoxic marker TIA-1 (52/55, 94.5%), less frequently, the expression of GrB (32/45, 71.1%) and perforin (31/42, 73.8%) was noticed. CD30 was expressed, at least partially, in 40.6% (13/32) of the cases. Ki-67 proliferation index varied from 50% to 90%, with an average of 72.7%. In all cases, the tumor cells expressed neither ALK1 nor CD20. (Fig 2) 10.1371/journal.pone.0161831.g002Fig 2 Immunophenotypic features of PI-ENKTCL. The neoplastic cells were positive for CD3 (A), CD56 (B) and TIA-1 (C). Positivity for CD30 (D) was seen in some larger atypical cells. Positive ISH signals of EBER were identified in tumor cells in all cases. Of the 17 tested cases, 9 (52.9%) demonstrated monoclonal TCR gene rearrangement, with a phenotype of TCR γδ+, αβ+ and αβ/γδ+ in 3 patients each. Treatment and outcome Forty-two patients received intestinal segment resection. A total of 35 patients (63.6%) were treated with systemic chemotherapy. Another one (1.8%) received chemoradiotherapy and subsequent autologous hematopoietic stem cell transplantation. Eleven (20.0%) failed to receive any intervention due to poor health condition. And no information was available for the remaining 8 (14.6%) patients. The chemotherapy regimens varied considerably, whereas a DICE regimen (dexamethasone, ifosfamide, cisplatin and etoposide) or L-asparaginase/peg-asparginase-containing regimens, such as SMILE (dexamethasone, methotrexate, ifosfamide, L-asparaginase and etoposide) and P-GEMOX (peg-asparginase, gemcitabine and oxaliplatin), were most frequently employed. Other regimens including the CHOP (cyclophosphamide, doxorubicin, vincristine and prednisone), CVAD (cyclophosphamide, vincristine, doxorubicin and dexamethasone), MINE (mesna, ifosfamide, mitoxantrone and etoposide) and EPOCH (etoposide, prednisone, vincristine, cyclophosphamide and doxorubicin) were variably adopted. The follow-up information was available in 37 patients with a follow-up time averaged 12.7 months (range, 1–66 months). Three cases relapsed 16–20 months after diagnosis, 19 patients died of the disease and 18 were alive at the last contact. Kaplan-Meier analysis revealed that the stage of disease at diagnosis and B symptoms were statistically associated with the OS (p = 0.001, p = 0.010, respectively) (Fig 3). Patients at earlier stage and those without B symptoms had relatively longer OS. Further multivariate analysis confirmed that both the stage and B symptoms were independent prognostic indicators. In addition, the prognosis of patients with intestinal perforation or old age (≧40 years) tended to be less favorable. Interestingly, patients with a T-lineaged tumor tended to have a better survival. Nevertheless, none of the above-mentioned findings could be proven statistical significant (p>0.05). Other parameters, including the tumor location, size of tumor cells, CD30 or CD56 positivity and Ki-67 index, did not show any correlation with the prognosis. 10.1371/journal.pone.0161831.g003Fig 3 Kaplan-Meier survival curves of patients with PI-ENKTCL. (A) Patients with early stage disease had superior survival compared with those with advanced stage disease (p = 0.001). (B) Patients with B symptoms were associated with a much poorer survival than those without B symptoms (p = 0.010). A distinct subgroup of intestinal CD4-positive ENKTCL Worthy to be noticed, 6 tumors in the current series demonstrated a CD4-positive phenotype, characterized by a diffuse and intensive positive staining for this marker in almost all the tumor cells (Table 1, Fig 4). All tumors arose from the small intestine with one exception that involved the ileocecal region. Four patients were male and 2 were female, with a median age of 58 years (range, 37–65 years). Morphologically, these tumors were composed predominantly of a proliferation of small to medium-sized atypical lymphoid cells with mild pleomorphism. Large cells were rarely seen, usually accounting for less than 10% of the tumor cell population. All the cases were positive for CD56, cytotoxic molecules and EBER, and clonal TCR gene rearrangements were demonstrated in all tested cases. Two patients died of disease at 18 and 12 months after diagnosis, respectively, and the remaining 4 were alive at the last contact. The one-year OS of these CD4-positive cases was 80.0%, better than that of CD4-negative ones (47.1%). Kaplan-Meier analysis revealed a superior OS of these CD4-positive ENKTCL cases compared with those CD4-negative ones, although the difference did not reach statistical significance (p = 0.192, Fig 5). 10.1371/journal.pone.0161831.g004Fig 4 A representative case of primary intestinal CD4+ ENKTCL involving the ileum. (A) The neoplastic cells were predominantly small to medium-sized ones with slightly irregular nuclei. Tumor cells were diffusely positive for (B) CD56 and (C) CD4. (D) In situ hybridization for EBER demonstrated positive signals. 10.1371/journal.pone.0161831.g005Fig 5 Kaplan-Meier survival curves of patients with PI-ENKTCL according to CD4 expression. Patients with a CD4-positive tumor had a slightly better OS compared with those with CD4-negative ones, although the difference was not statistically significant (p = 0.192). 10.1371/journal.pone.0161831.t001Table 1 The clinicopathological features of primary intestinal CD4+ ENKTCL. Case No. Sex/Age (years) Site Stage Status Follow-up time (months) Tumor cell size CD4 CD8 CD56 Ki-67 (%) EBER TCR gene rearrangement Treatment 1 Female/37 Ileum IIE Alive 3 Small + - + 70 + Monoclonal Resection +chemotherapy 2 Male/62 Ileum IE Alive 14 Medium + - + 60 + Monoclonal Resection +chemotherapy 3 Male/55 Small intestine IE Alive 18 Small/medium + - + 80 + ND Resection +chemotherapy 4 Male/65 Small intestine IE Dead 12 Small/medium, scattered large cells + - + 60 + Monoclonal Resection* 5 Male/60 Duodenum IE Alive 34 Small/medium, scattered large cells + - + 80 + Monoclonal Chemotherapy 6 Female/39 ileocecal region IIE Dead 18 Small/medium, scattered large cells + - + 70 + ND Resection+chemotherapy ND, not done. *, information on subsequent treatment of this patient is unavailable. Discussion The GI tract is one of the most common extranodal sites that can be involved by lymphomas [12]. However, ENKTCL originating from the intestine is extremely rare. So far only few studies with limited number of cases were documented in the English literature [13, 14]. Due to its rarity, the clinical and pathological features of PI-ENKTCL are less well defined. We thus collected a large cohort of cases to observe the clinicopathological characteristics of this uncommon disease. To the best of our knowledge, the current data set may represent by far the largest one focusing on intestinal ENKTCL. Based on our findings, PI-ENKTCL more frequently affected middle-aged adults in their forties with a male predominance, which is in consistent with several previous studies [12, 15]. Regarding the sites of involvement, small intestine is most commonly affected, followed by colon and the ileocecal junction. This result is in good concordance with the findings by Kim et al [16], but differs with some other studies, which demonstrated that large intestine is much more commonly involved by this type of lymphoma [12, 17]. The most common presenting symptom is abdominal pain, with or without fever, as same as Fang et al indicated before [13]. Endoscopically, most intestinal ENKTCL presents as ulcerative lesions, being prone to perforation, rather than a polypoid mass, suggesting a potential more aggressive behavior of this tumor [12]. The variety of histologic appearance of intestinal ENKTCL might have reflected the striking heterogeneity of this tumor. In terms of immunophenotype, CD56 appears to be a useful marker aiding in the diagnosis, with an expression rate varying between 60% and 100% in different series [3, 8, 18]. It has been proposed that CD56 might be expressed more frequently in NK-derived tumors than T-cell-originated ones [4]. Our results, together with those of some other studies, however, did not reveal such a difference [3, 8, 18]. The expression of one or more cytotoxic molecules, including TIA-1, GrB and perforin, is invariably present in every case, with TIA-1 expression being more frequently detected [4, 18]. As a marker related to cell activation, CD30 was frequently expressed by the tumor cells in the current series, which presumably reflected the fact that some ENKTCL cases feature an activated phenotype. The relatively high positivity for CD30 in intestinal ENKTCL is also in good keeping with that a higher percentage of CD30 expression was noticed in extra-nasal ENKTCL lesions than those nasal ones [4, 19]. ENKTCL is almost constantly associated with EBV, which is suspected to play an important role in the oncogenesis of this disease [20]. Identification of EBV-encoded EBER is therefore essential for the establishment of a diagnosis of ENKTCL. In the current series, nearly half of the intestinal ENKTCL cases that submitted for TCR gene rearrangement detection demonstrated a germline configuration of the TCR genes, confirming a genuine NK-cell origin of the tumor, however, T-cell-derived tumors accounted for 52.9% of all tested cases, an incidence slightly higher than those previously reported ranging from 0%-46% [4, 5, 10, 14, 19, 21–23]. The clinical significance of cell origin of an intestinal ENKTCL remains largely unknown, although in general, ENKTCL cases of different lineage do not show distinct clinopathological features including the survival [4, 10, 19]. We noticed a trend of a better survival in patients with T-cell-derived intestinal ENKTCLs than those with NK-lineaged ones, although the survival advantage was not statistically significant. Pongpruttipan et al seemed to have similar findings although the tumors in their series were mostly located in the upper aerodigestive tract instead of the intestine [4]. Chuang and colleagues also noticed a superior survival of those EBV-related cytotoxic T-cell lymphomas arising in the intestine compared with their NK-cell counterparts, but they referred to the former as “NK-like peripheral T-cell lymphomas”, possibly an entity distinct from the genuine PI-ENKTCLs or enteropathy-associated T-cell lymphomas (EATL) [24]. Intestinal ENKTCL must be distinguished from its morphological mimics. It is not uncommon that some intestinal ENKTCL cases are misdiagnosed as peripheral T-cell lymphoma, not otherwise specified (PTCL, NOS) due to CD3 positivity and a high proliferation rate of the tumor cells. However, the later usually lacks the characteristic massive necrosis and angiodestructive growth pattern. In general, EBER is absent in PTCL, NOS. CD56 expression is also less commonly seen in this tumor compared with ENKTCL. Type II EATL might be confused with intestinal ENKTCL due to their common CD56+ and cytotoxic phenotype [4]. Nevertheless, the former is composed of monotonous medium-sized tumor cells with marked epitheliotropism. Necrosis is uncommon in type II EATL, too. The tumor cells typically co-express CD8 but lack EBER [25], although it still remains controversial that whether some EATL-appearing but EBER-positive intestinal T-cell lymphomas can be labeled as an ENKTCL. Occasionally, PI-ENKTCLs consisting mainly of large or anaplastic cells may show CD30 positivity and thus, might be potentially misinterpreted as anaplastic large cell lymphoma (ALCL) or the classical form of EATL. Detection for EBER is critical for the distinction between these diseases. Rare examples of indolent T- or NK-cell lymphoproliferative disorders of the GI tract, such as the NK-cell lymphomatoid gastroenteropathy and indolent T-cell lymphoproliferative disorder of the GI tract, may morphologicaly or phenotypically mimic ENKTCL, too [26, 27]. These lesions, however, differ from ENKTCL by their characteristic indolent clinical course and an absence of EBV in the lesional cells. Given the rarity of PI-ENKTCL and its clinicohistological heterogeneity, there is currently no consensus approached on the proper treatment strategy [16, 28]. A timely surgical operation might be necessary since it has been found that patients receiving surgery before perforation usually feature an improved outcome compared to those undergoing surgery later or without surgery [12]. Moreover, surgery with subsequent chemotherapy might be more beneficial than initially treatment with chemotherapy for patients eligible for surgery [16]. Nevertheless, Fang et al presumed that operation might accelerate deterioration of the disease [13]. On the other hand, so far there is no standard chemotherapeutic regimen for ENKTCL [29]. It has been reported that patients with GI ENKTCL who received nonanthracycline-based or intensified regimens did not show a significant survival difference compared with those who received CHOP or CHOP-like regimens regardless of surgery [16]. Recently, the efficacy of L-asparaginase or peg-asparginase in the treatment of ENKTCL has been confirmed, and clinical trials using L-asparaginase or peg-asparginase-based chemotherapeutic regimens have achieved promising results [29]. Yet, about half of our patients died during the follow-up period irrespective of the administration of comprehensive treatment. Novel effective approaches are thus warranted to improve the survival of patients with this rare but fatal disease. Both our study and some previous ones had demonstrated a large proportion of patients with intestinal ENKTCL presented with early stage diseases [13], these patients, however, tended to exhibit highly aggressive behavior and thus, conferred an inferior prognosis as compared to their nasal counterparts [16]. For such a contradiction, one reasonable explanation lies that the overt extra-nasal tumors in some patients might represent actually the extra-nasal dissemination of an occult nasal lesion. And the primary nasal lesions are easily to be missed or overlooked only because of their small sizes [3]. Therefore, a PET-CT scan, nasal panendoscopy and careful imaging of the aerodigestive tract, with biopsy when necessary, are mandatory for each new case of ENKTCL, irrespective of the initial presentation sites, as indicated previously [1, 3, 30]. However, all cases enrolled in the current study have been submitted for a precise staging work-up and hence secondary onset can be excluded. Thus there might be some other reasons accounting for the poor outcome of PI-ENKTCL. It has been suggested that old age, intestinal perforations, B symptoms are independent prognostic factors indicating a poor outcome of patients with GI-ENKTCL, and surgery prior to the perforation is another key factor that may influence the survival [12]. In our study, the results have confirmed that both disease stage and B symptoms are significant independent prognostic factors for this rare type of tumor. The prognostic significance concerning CD56 expression in ENKTCL is controversial [31, 32]. We have failed to find any correlations between CD56 expression and OS, probably due to the limited number of CD56-negative cases. Of interest, we noticed and described for the first time 6 cases of intestinal T-cell lymphomas characterized by their predilection for small intestine involvement, predominance of small to medium-sized tumor cells, a CD4+, CD56+, EBER+ T-cell phenotype and importantly, relatively favorable outcome. We tentatively labeled these lesions as ENKTCL mainly based on their cytotoxic T-cell phenotype and EBV positivity. CD4-positive ENKTCL is extremely rare. In a large cohort of 67 cases with ENKTCL, Pongpruttipan et al identified only one CD4-positive case, which might be of αβ T-cell origin [4]. We suppose these peculiar CD4+, CD56+, EBER+ tumors of the small intestine may represent a special variant of ENKTCL or even novel entity due to their unique clinical and pathological features. Recognition of these CD4-positive tumors may aid in predicting the prognosis and tailoring a personalized treatment strategy to the patients. Further work is needed to elucidate the biological nature of the disease and its relationship with the classical form of intestinal ENKTCL. In conclusion, PI-ENKTCL is a rare condition. The disease predominantly affects middle-aged males and features frequently aggressive clinical course and poor outcome. Independent inferior prognostic indicators may include the stage of disease and B symptoms. A subgroup of cases with a CD4-positive phenotype might represent a unique variant or clinicopathological entity. Further molecular and clinical studies are encouraged for the better characterization and management of the disease. Supporting Information S1 Table The clinicopathological characteristics of patients with primary intestinal extranodal natural killer/T-cell lymphoma, nasal type. (XLSX) Click here for additional data file. ==== Refs References 1 Hawkes EA , Wotherspoon A , Cunningham D . Diagnosis and management of rare gastrointestinal lymphomas . Leuk Lymphoma . 2012 ; 53 :2341 –2350 . 10.3109/10428194.2012.695780 22616672 2 Chan JKC , Quintanilla-Martinez L , Ferry JA , Peh SC . Extranodal NK/T-cell lymphoma, nasal type In: Swerdlow SH , Campo E , Harris NL , Jaffe ES , Pileri SA , Stein H , et al, editors. World Health Organization Classification of Tumours Pathology and Genetics of Tumours of Haematopoietic and Lymphoid Tissues . Lyon, France : IARC Press ; 2008 pp. 285 –288 . 3 Jhuang JY , Chang ST , Weng SF , Pan ST , Chu PY , Hsieh PP , et al Extranodal natural killer/T-cell lymphoma, nasal type in Taiwan: a relatively higher frequency of T-cell lineage and poor survival for extranasal tumors . Human Pathology . 2015 ; 46 :313 –321 . 10.1016/j.humpath.2014.11.008 25554090 4 Pongpruttipan T , Sukpanichnant S , Assanasen T , Wannakrairot P , Boonsakan P , Kanoksil W , et al Extranodal NK/T-cell lymphoma, nasal type, includes cases of natural killer cell and αβ, γδ, and αβ/γδ T-cell origin: a comprehensive clinicopathologic and phenotypic study . Am J Surg Pathol . 2012 ;36 :481 –499 . 10.1097/PAS.0b013e31824433d8 22314189 5 Takata K , Hong ME , Sitthinamsuwan P , Loong F , Tan SY , Liau JY , et al Primary cutaneous NK/T-cell lymphoma, nasal type and CD56-positive peripheral T-cell lymphoma: a cellular lineage and clinicopathologic study of 60 patients from Asia . Am J Surg Pathol . 2015 ;39 :1 –12 . 10.1097/PAS.0000000000000312 25188863 6 Bautista-Quach MA , Ake CD , Chen M , Wang J . Gastrointestinal lymphomas: Morphology, immunophenotype and molecular features . J Gastrointest Oncol . 2012 ;3 :209 –225 . 10.3978/j.issn.2078-6891.2012.024 22943012 7 Aozasa K , Zaki MA . Epidemiology and pathogenesis of nasal NK/T-cell lymphoma: a mini-review . Scientific World Journal . 2011 ; 14 :422 –428 . 8 Gualco G , Domeny-Duarte P , Chioato L , Barber G , Natkunam Y , Bacchi CE . Clinicopathologic and molecular features of 122 Brazilian cases of nodal and extranodal NK/T-cell lymphoma, nasal type, with EBV subtyping analysis . Am J Surg Pathol . 2011 ;35 :1195 –1203 . 10.1097/PAS.0b013e31821ec4b5 21716086 9 Kim SJ , Choi CW , Mun YC , Oh SY , Kang HJ , Lee SI , et al Multicenter retrospective analysis of 581 patients with primary intestinal non-hodgkin lymphoma from the Consortium for Improving Survival of Lymphoma (CISL) . BMC Cancer . 2011 ;29 : 321 . 10 Richendollar BG , Hsi ED , Cook JR . Extramedullary plasmacytoma-like posttransplantation lymphoproliferative disorders: clinical and pathologic features . Am J Clin Pathol . 2009 ;132 :581 –588 . 10.1309/AJCPX70TIHETNBRL 19762536 11 Ng SB , Lai KW , Murugaya S , Lee KM , Loong SL , Fook-Chong S , et al Nasal-type extranodal natural killer/T-cell lymphomas: a clinicopathologic and genotypic study of 42 cases in Singapore . Mod Pathol . 2004 ; 17 :1097 –1107 . 15195107 12 Jiang M , Chen X , Yi Z , Zhang X , Zhang B , Luo F , et al Prognostic characteristics of gastrointestinal tract NK/T-cell lymphoma: an analysis of 47 patients in China . J Clin Gastroenterol . 2013 ;47 :e74 –e79 . 10.1097/MCG.0b013e31829e444f 23948755 13 Fang JC , Xia ZX , Wang CN and Li Z . Clinicopathologic and Immunophenotypic features of primary intestinal extranodal NK/T-cell lymphoma, nasal type . Int J Surg Pathol . 2015 ;23 :609 –616 . 10.1177/1066896915595863 26183848 14 Zheng S , Ouyang Q , Li G , Xu H , Jiang M , Cui D , et al Primary intestinal NK/T cell lymphoma: a clinicopathologic study of 25 Chinese cases . Arch Iran Med . 2012 ;15 :36 –42 . doi: 012151/AIM.0011 22208442 15 Gou HF , Zang J , Jiang M , Yang Y , Cao D , Chen XC . Clinical prognostic analysis of 116 patients with primary intestinal non-Hodgkin lymphoma . Med Oncol . 2012 ; 29 : 227 –234 . 10.1007/s12032-010-9783-x 21193968 16 Kim SJ , Jung HA , Chuang SS , Hong H , Guo CC , Cao J , et al Extranodal natural killer/T-cell lymphoma involving the gastrointestinal tract: analysis of clinical features and outcomes from the Asia Lymphoma Study Group . J Hematol Oncol . 2013 ;6 :86 10.1186/1756-8722-6-86 24238138 17 Sun j , Lu Z , Yang D , Chen J . Primary intestinal T-cell and NK-cell lymphomas: a clinicopathological and molecular study from China focused on type II enteropathy-associated T-cell lymphoma and primary intestinal NK-cell lymphoma . Mod Pathol . 2011 ;24 :983 –992 . 10.1038/modpathol.2011.45 21423155 18 Pongpruttipan T , Kummalue T , Bedavanija A , Khuhapinant A , Ohshima K , Arakawa F , et al Aberrant antigenic expression in extranodal NK/T-cell lymphoma: a multi-parameter study from Thailand . Diagn Pathol . 2011 ; 25 :79 . 19 Au WY , Weisenburger DD , Intragumtornchai T , Nakamura S , Kim WS , Sng I , et al Clinical differences between nasal and extranasal natural killer/T-cell lymphoma: a study of 136 cases from the International Peripheral T-Cell Lymphoma Project . Blood . 2009 ; 113 :3931 –3937 . 10.1182/blood-2008-10-185256 19029440 20 Aozasa K , Zaki MA . Epidemiology and pathogenesis of nasal NK/T-cell lymphoma: a mini-review . Scientific World Journal . 2011 ; 14 :422 –428 . 21 Ko YH , Cho EY , Kim JE , Lee SS , Huh JR , Chang HK , et al NK and NK-like T-cell lymphoma in extranasal sites: a comparative clinicopathological study according to site and EBV status . Histopathology . 2004 ;44 :480 –489 . 15139996 22 Li S , Feng X , Li T , Zhang S , Zuo Z , Lin P , et al Extranodal NK/T-cell lymphoma, nasal type: a report of 73 cases at MD Anderson Cancer Center . Am J Surg Pathol . 2013 ; 37 : 14 –23 . 10.1097/PAS.0b013e31826731b5 23232851 23 Suzuki R , Suzumiya J , Yamaguchi M , Nakamura S , Kameoka J , Kojima H , et al Prognostic factors for mature natural killer (NK) cell neoplasms: aggressive NK cell leukemia and extranodal NK cell lymphoma, nasal type . Ann Oncol . 2010 ;21 :1032 –1040 . 10.1093/annonc/mdp418 19850638 24 Chuang SS , Chang ST , Chuang WY , Huang WT , Hsieh PP , Tsou MH , et al NK-cell lineage predicts poor survival in primary intestinal NK-cell and T-cell lymphomas . Am J Surg Pathol . 2009 ; 33 : 1230 –1240 . 10.1097/PAS.0b013e3181a95c63 19561449 25 Medeiros LJ , Miranda RN , Wang SA , Vega F , Muzzafar T , Yin CC , et al Diagnostic Pathology: Lymph Nodes and Spleen with Extranodal Lymphomas . Altona : Amirsys ; 2011 . 26 Takeuchi K , Yokoyama M , Ishizawa S , Terui Y , Nomura K , Marutsuka K ,et al lymphomatoid gastropathy: a distinct clinicopathologic entity of self-limited pseudomalignant NK-cell proliferation . Blood . 2010 ;116 :5631 –5637 . 10.1182/blood-2010-06-290650 20829373 27 Perry AM , Warnke RA , Hu Q , Gaulard P , Copie-Bergman C , Alkan S , et al Indolent T-cell lymphoproliferative disease of thegastrointestinal tract . Blood .2013 ;122 :3599 –3606 . 10.1182/blood-2013-07-512830 24009234 28 Jo JC , Yoon DH , Kim S , Lee BJ , Jang YJ , Park CS , et al Clinical features and prognostic model for extranasal NK/T-cell lymphoma . Eur J Haematol . 2012 ; 89 :103 –110 . 10.1111/j.1600-0609.2012.01796.x 22553935 29 Yong W . Clinical study of l-asparaginase in the treatment of extranodal NK/T-cell lymphoma, nasal type . Hematol Oncol . 2015 4 21 10.1002/hon.2207 [Epub ahead of print] 30 Tse E , Kwong YL . How I treat NK/T-cell lymphomas . Blood . 2013 ;121 : 4997 –5005 . 10.1182/blood-2013-01-453233 23652805 31 Li YX , Wang H , Feng XL , Liu QF , Wang WH , Lv N , et al Immunophenotypic characteristics and clinical relevance of CD56+ and CD56- extranodal nasal-type natural killer/T-cell lymphoma . Leuk Lymphoma . 2011 ;52 :417 –424 . 10.3109/10428194.2010.543718 21323521 32 Wang L , Wang Z , Xia ZJ , Lu Y , Huang HQ , Zhang YJ . CD56-negative extranodal NK/T cell lymphoma should be regarded as a distinct subtype with poor prognosis . Tumor Biol . 2015 ; 36 :7717 –7723 .

PMC005xxxxxx/PMC5001694.txt

==== Front PLoS Negl Trop DisPLoS Negl Trop DisplosplosntdsPLoS Neglected Tropical Diseases1935-27271935-2735Public Library of Science San Francisco, CA USA 2756371810.1371/journal.pntd.0004966PNTD-D-16-00972Research ArticleBiology and Life SciencesAgricultureLivestockSwineBiology and Life SciencesOrganismsAnimalsVertebratesAmniotesMammalsSwineBiology and Life SciencesEvolutionary BiologyPopulation GeneticsBiology and Life SciencesGeneticsPopulation GeneticsBiology and Life SciencesPopulation BiologyPopulation GeneticsMedicine and Health SciencesPathology and Laboratory MedicinePathogenesisHost-Pathogen InteractionsBiology and Life SciencesGeneticsGenetic LociBiology and Life SciencesOrganismsAnimalsVertebratesAmniotesMammalsPrimatesBiology and Life SciencesBiogeographyPhylogeographyEcology and Environmental SciencesBiogeographyPhylogeographyEarth SciencesGeographyBiogeographyPhylogeographyBiology and Life SciencesEvolutionary BiologyPopulation GeneticsPhylogeographyBiology and Life SciencesGeneticsPopulation GeneticsPhylogeographyBiology and Life SciencesPopulation BiologyPopulation GeneticsPhylogeographyMedicine and Health SciencesInfectious DiseasesZoonosesBiology and Life SciencesEvolutionary BiologyEvolutionary GeneticsClonal Evolution of Enterocytozoon bieneusi Populations in Swine and Genetic Differentiation in Subpopulations between Isolates from Swine and Humans Clonal Evolution and Subdivisions of E. bieneusi Population in SwineWan Qiang 1Xiao Lihua 2Zhang Xichen 3Li Yijing 1Lu Yixin 1Song Mingxin 1http://orcid.org/0000-0002-4264-1864Li Wei 1*1 College of Veterinary Medicine, Northeast Agricultural University, Harbin, Heilongjiang, China2 Division of Foodborne, Waterborne and Environmental Diseases, National Center for Emerging and Zoonotic Infectious Diseases, Centers for Disease Control and Prevention, Atlanta, Georgia, United States of America3 College of Veterinary Medicine, Jilin University, Changchun, Jilin, ChinaEscalante Ananias A. EditorTemple University, UNITED STATESThe authors have declared that no competing interests exist. Conceptualization: WL. Formal analysis: QW. Funding acquisition: WL. Investigation: QW WL. Methodology: WL QW. Resources: XZ YLi YLu MS. Writing – original draft: QW WL. Writing – review & editing: LX WL. * E-mail: neaulw@gmail.com; liwei@neau.edu.cn26 8 2016 8 2016 10 8 e000496625 5 2016 9 8 2016 © 2016 Wan et al2016Wan et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Enterocytozoon bieneusi is a widespread parasite with high genetic diversity among hosts. Its natural reservoir remains elusive and data on population structure are available only in isolates from primates. Here we describe a population genetic study of 101 E. bieneusi isolates from pigs using sequence analysis of the ribosomal internal transcribed spacer (ITS) and four mini- and microsatellite markers. The presence of strong linkage disequilibrium (LD) and limited genetic recombination indicated a clonal structure for the population. Bayesian inference of phylogeny, structural analysis, and principal coordinates analysis separated the overall population into three subpopulations (SP3 to SP5) with genetic segregation of the isolates at some geographic level. Comparative analysis showed the differentiation of SP3 to SP5 from the two known E. bieneusi subpopulations (SP1 and SP2) from primates. The placement of a human E. bieneusi isolate in pig subpopulation SP4 supported the zoonotic potential of some E. bieneusi isolates. Network analysis showed directed evolution of SP5 to SP3/SP4 and SP1 to SP2. The high LD and low number of inferred recombination events are consistent with the possibility of host adaptation in SP2, SP3, and SP4. In contrast, the reduced LD and high genetic diversity in SP1 and SP5 might be results of broad host range and adaptation to new host environment. The data provide evidence of the potential occurrence of host adaptation in some of E. bieneusi isolates that belong to the zoonotic ITS Group 1. Author Summary This study explored the genetic characteristics of the ITS and four mini- and microsatellite markers and assessed the population structure and substructures in 101 E. bieneusi isolates from pigs in China. The measures of LD and recombination events supported the occurrence of clonal evolution among the isolates from four study areas of China. Three subpopulations (SP3 to SP5) with potentially varied host ranges were identified within the isolates, which were genetically differentiated from two existing primate subpopulations SP1 and SP2. Population genetic analysis indicated that the isolates in SP1 and SP5 with clonal structure might be responsible for the cross-species transmission and thus have zoonotic potential, while the isolates in SP2, SP3, and SP4 with epidemic structure or host-adapted traits might colonize specific hosts. The data revealed the presence of clonality, potential host adaptation, and population differentiation of E. bieneusi in different hosts. the National Natural Science Foundation of China31302081http://orcid.org/0000-0002-4264-1864Li Wei the University Nursing Program for Young Scholars with Creative Talents in Heilongjiang ProvinceUNPYSCT-2015008http://orcid.org/0000-0002-4264-1864Li Wei This study was supported by the National Natural Science Foundation of China (no. 31302081, http://www.nsfc.gov.cn/) and the University Nursing Program for Young Scholars with Creative Talents in Heilongjiang Province (no. UNPYSCT-2015008, https://www.hljedu.gov.cn/). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityNucleotide sequences of mini- and microsatellite loci MS1, MS3, MS4, and MS7 obtained from 101 E. bieneusi isolates from pigs are available in the GenBank database under the accession numbers KU212402 to KU212502, KU212503 to 212603, KU212604 to KU212704, and KU212705 to KU212805, respectively.Data Availability Nucleotide sequences of mini- and microsatellite loci MS1, MS3, MS4, and MS7 obtained from 101 E. bieneusi isolates from pigs are available in the GenBank database under the accession numbers KU212402 to KU212502, KU212503 to 212603, KU212604 to KU212704, and KU212705 to KU212805, respectively. ==== Body Introduction Microsporidia are obligate intracellular eukaryotic parasites that infect a wide range of animals and are closely related to fungi [1,2]. Genome analyses conducted in several recent studies strongly suggest that some species of microsporidia could have a diploid or polyploid stage and a sexual cycle, and might be true Fungi [3–8]. Nevertheless, the ploidy level of Enterocytozoon bieneusi and whether it undergoes mating and a meiotic cycle are still unclear. E. bieneusi is the most common human microsporidian species and can colonize a variety of other mammals and birds [2,9]. This ubiquitous pathogen causes diarrhea of various severity and duration in relation to host immune status [2,10]. Genotyping of isolates has improved our understanding of the genetic characteristics and the potential transmission modes of E. bieneusi among hosts. E. bieneusi exhibits high genetic diversity among isolates from different hosts [11,12]. Over 200 E. bieneusi genotypes have been identified in humans, companion animals, livestock, horses, birds, and wildlife based on DNA sequence analysis of the ribosomal internal transcribed spacer (ITS) and the established naming convention [13,14]. The genotypes form several genetically isolated clusters (Groups 1 to 8) in phylogenetic analysis, with some found in specific host groups [11,12,15]. Humans and pigs are mainly infected with the zoonotic Group 1 genotypes, ruminants with host-adapted Group 2 genotypes, and dogs with the genotypes in an outlier group [2,16–22]. The public health potential of E. bieneusi in animals has been assessed in numerous studies and pigs were recognized as the most significant reservoir [2,17,23]. However, this was based on results of sequence analysis of a single ITS marker (392 bp in length), which may not adequately represent the evolutionary history of the E. bieneusi genome with a length of about 6 Mb [24]. Several mini- and microsatellites with sufficient resolution have been available to infer subgroup-level phylogenies [25]. Coupled with the ITS locus, they have been used effectively in characterizations of population structures and substructures of E. bieneusi in primates [26–28]. In these studies, a clonal structure was found in human E. bieneusi populations from Peru, India, and Nigeria and no apparent geographic segregation of the isolates was observed. Nevertheless, two genetically isolated subpopulations (SP1 and SP2) were identified within the overall population. Significant linkage disequilibrium (LD) and limited recombination in SP1 support a clonal population structure, while the rapid expansion of some specified multilocus genotypes (MLGs) in SP2 obscures the limited genetic exchange because of its epidemic population structure [26,27]. Although the isolates used in SP1 and SP2 mainly belong to ITS Group 1 genotypes D, IV, and A with zoonotic potential, some of them displayed host-specific features when mini- and microsatellites were considered in the analysis [26,27]. Population genetic traits of E. bieneusi in non-human primates from China were similar to those in humans [28]. These observations on the population genetics of E. bieneusi in primates need to be substantiated in other hosts. The objectives of this study were to characterize 101 pig E. bieneusi isolates that belong to nine genotypes in zoonotic Group 1 at the subtype level at five genetic loci, to assess the population structure and substructures, and to compare them with similar data from E. bieneusi subpopulations SP1 and SP2 in primates to examine the occurrence and extent of host segregation in different E. bieneusi subpopulations. Materials and Methods Ethics statement This study was performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the Ministry of Health, China. Prior to experiments, the protocol of the current study was reviewed and approved by the Institutional Animal Care and Use Committee of Northeast Agricultural University (approved protocol number SRM-08). For specimen collection, we obtained permission from animal owners. No specific permits were required for the described field studies, and the locations where we sampled are not privately owned or protected in any way. The field studies did not involve endangered or protected species. Study area and parasite sampling Isolates of E. bieneusi were obtained from pigs in cities Changchun, Daqing, Harbin, and Qiqihar in northeast China. All of them were previously genotyped by PCR and sequence analysis of the ITS locus [17,29]. A total of 101 isolates of ITS genotypes CHN7, CS-4, EbpA, EbpB, EbpC, Henan-I, Henan-IV, O, and PigEBITS3 in zoonotic Group 1 were selected for population genetic analysis in this study. The number of isolates and their ITS genotype designations by city are shown in Table 1. For comparative purposes, population genetic data from 101 E. bieneusi isolates in humans from India, Peru, and Nigeria and 5 isolates in captive baboons from Kenya were included in this analysis (Table 1) [26,27]. 10.1371/journal.pntd.0004966.t001Table 1 Enterocytozoon bieneusi isolates analyzed in this study and their genotypes based on the ribosomal internal transcribed spacer (ITS) sequences. Location Host No. ITS genotype (no., city) Reference China Pigs 101 CHN7 (3, Changchun), CS-4 (27, Harbin), EbpA (18, Changchun; 4, Qiqihar), EbpB (12, Daqing), EbpC (11, Changchun; 10, Daqing; 2, Harbin; 9, Qiqihar), Henan-I (1, Changchun), Henan-IV (2, Qiqihar), O (1, Qiqihar), PigEBITS3 (1, Changchun) [17,29] India HIV+ children 14 A (3), D (5), PigEBITS7 (6) [26] Nigeria HIV+ adults 15 A (6), D (2), IV (6), Nig2 (1) [26] Peru HIV+ adults 72 A (30), D (8), EbpC (1), IV (16), Peru7 (5), Peru8 (1), Peru10 (3), Peru11 (4), WL11 (4) [27] Kenya Olive baboons 5 D (5) [26] PCR and sequencing Genomic DNAs extracted from 101 pig fecal specimens were used for PCR amplification of a minisatellite marker MS4 and three microsatellite markers (MS1, MS3, and MS7) as described [25]. The secondary PCR products with the expected sizes (bp) of approximately 676 for MS1, 537 for MS3, 885 for MS4, and 471 for MS7 were sequenced in both directions at the Beijing Genomics Institute, China. Raw sequences were assembled and edited with the software Chromas Pro version 1.33 (Technelysium Pty. Ltd., Helensvale, Queensland, Australia). The sequences obtained were compared to the sequence data of each target available in GenBank [26,27] using the software MAFFT version 7.300 (http://mafft.cbrc.jp/alignment/software/) [30]. Genetic analysis The software DnaSP version 5.10.01 (http://www.ub.edu/dnasp/) was used to determine E. bieneusi genotypes at each of the five markers and the MLGs with consideration of both single base nucleotide substitutions and short insertions and deletions (indels) polymorphisms [31]. Intragenic LD and recombination rates for individual locus and concatenated multilocus data set were estimated from the segregating sites without consideration of indels using DnaSP [31]. Recombination rates were further assessed using the methods GENECONV, MaxChi, and SiScan implemented in the software RDP version 3.44 (http://darwin.uvigo.es/rdp/rdp.html) [32]. Tests for genetic diversity and neutrality (Fu’s Fs and Tajima’s D) were run on the concatenated contigs using DnaSP (based on segregating sites) and the software Arlequin 3.5.1.2 (http://cmpg.unibe.ch/software/arlequin35/; based on both segregating sites and indels) [31,33]. The different nucleotide sequences (considering both substitutions and indels) were assigned as distinct alleles and the alleles at each of the five loci defined the allelic profile or sequence type. We measured the pairwise intergenic LD on annotated allelic profiles using the exact test and Markov chain parameters implemented in Arlequin [33]. Values of standardized index of association (ISA) were calculated with LIAN 3.5 (http://guanine.evolbio.mpg.de/cgi-bin/lian/lian.cgi.pl/query) on five-loci haplotypes [34]. We assessed population structure of E. bieneusi by analyzing the intragenic and intergenic LD, ISA, neutrality, and recombination events (Rms). Wright’s fixation index (FST) calculated using Arlequin and gene flow (Nm) calculated using DnaSP were applied to evaluate the degree of genetic differentiation between E. bieneusi populations [31,33]. Phylogenetic and structure sub-clustering analysis A Bayesian analysis implemented in the software MrBayes version 3.2.1 (http://mrbayes.sourceforge.net/) was used in clustering nucleotide sequences using Markov chain Monte Carlo (MCMC) methods [35]. The general time reversible model (GTR+G) was determined to be the best-fit nucleotide substitution model with the program ModelTest 3.7 (http://www.molecularevolution.org/software/phylogenetics/modeltest) [36]. An MCMC-based analysis of phylogeny was conducted using the GTR+G model and the default parameters settings as described [26]. The maximum clade credibility tree generated by these analyses was visualized and edited using the software FigTree version 1.3.1 (http://tree.bio.ed.ac.uk/software/figtree/). Pairwise distance matrices among nucleotide sequences of MLGs were calculated using eachgap calculating method with the dist.seqs command in the software MOTHUR version 1.24.1 (http://www.mothur.org/wiki/Download_mothur) [37]. A principal coordinates analysis (PCoA) via covariance matrix with data standardization was performed on the generated matrices with the software GENALEX version 6.501 (http://biology-assets.anu.edu.au/GenAlEx) [38]. A Bayesian cluster analysis was performed on the allelic profile data using the software STRUCTURE version 2.3.1 (http://pritch.bsd.uchicago.edu/software.html) to assess the presence of distinct subpopulations [39]. We also constructed haplotype networks using the median-joining method implemented in the software Network version 4.6.1.1 (http://www.fluxus-engineering.com/sharenet_rn.htm) to estimate the genetic segregation and evolutionary trend of E. bieneusi isolates [40]. Results Genotyping and LD at individual loci Sequence analysis of 101 pig E. bieneusi isolates identified 9, 15, 5, 15, and 11 genotypes at the loci ITS, MS1, MS3, MS4, and MS7, respectively. Single nucleotide polymorphisms (SNPs) were the only source of genetic diversity at the ITS locus, while genetic variation at the other four loci included the lengths of trinucleotide TAC and TAA repeats at MS1, dinucleotide TA repeats at MS3, tetranucleotide GGTA repeats at MS4, and TAA repeats at MS7 and the SNPs outside the tandem repeat regions. Some MS4 fragments also carried isostructural GG to AA substitutions in the first tetranucleotide repeat. Gene diversity (Hd) at individual loci was calculated using DnaSP and is shown in Table 2. The number of genotypes and Hd value of each locus were also measured for 106 primate E. bieneusi isolates (Table 2). Generally, markers MS1 and MS4 had higher resolution than the other ones (Table 2). The intragenic LD among segregating sites for each locus was calculated based on a linear regression analysis in DnaSP. The markers MS1, MS3, and MS7 had complete LD (LD = 1) when only the pig E. bieneusi isolates were analyzed, whereas only MS1 and MS3 had complete LD in the analysis of all 207 isolates (Table 3). Table 3 displays the number of pairwise comparisons and the number of significant pairwise comparisons after Fisher’s exact test and Bonferroni correction. The occurrence of intragenic recombination was assessed using DnaSP. As shown in Table 3, genetic recombination was only detected in markers with incomplete LD. 10.1371/journal.pntd.0004966.t002Table 2 Number of Enterocytozoon bieneusi genotypes based on combined sequence length and nucleotide polymorphism at four mini- and microsatellite loci. City/subpopulation No. of isolate No. of genotypes (gene diversity) MS1 MS3 MS4 MS7 ITS Changchun 34 6 (0.69) 4 (0.61) 7 (0.76) 6 (0.77) 5 (0.62) Daqing 22 3 (0.62) 2 (0.52) 5 (0.65) 3 (0.56) 2 (0.52) Harbin 29 4 (0.53) 3 (0.31) 3 (0.26) 3 (0.14) 2 (0.13) Qiqihar 16 4 (0.59) 1 (0) 3 (0.70) 4 (0.35) 4 (0.64) Total (4 cities) 101 15 (0.90) 7 (0.66) 15 (0.85) 11 (0.81) 9 (0.77) SP1 66 35 (0.96) 10 (0.78) 9 (0.79) 15 (0.87) 9 (0.79) SP2 39 4 (0.28) 2 (0.15) 5 (0.33) 3 (0.23) 1 (0) SP3 35 9 (0.77) 5 (0.62) 9 (0.76) 7 (0.83) 3 (0.50) SP4 41 9 (0.74) 4 (0.27) 5 (0.56) 6 (0.56) 5 (0.52) SP5 26 5 (0.66) 4 (0.40) 7 (0.80) 5 (0.57) 2 (0.21) Total (5 subpopulations) 207 54 (0.94) 17 (0.84) 29 (0.92) 24 (0.91) 19 (0.89) 10.1371/journal.pntd.0004966.t003Table 3 Intragenic linkage disequilibrium and recombination events at individual genetic loci. Locus S P F B LD (|D'|) Rm MS1a 3 3 1 1 1 0 MS3a 3 3 0 0 1 0 MS4a 21 210 170 137 0.9716–0.1250X 5 MS7a 3 3 1 1 1 0 ITSa 12 66 36 32 0.9924–0.0137X 2 MS1b 7 21 4 4 1 0 MS3b 8 28 9 5 1 0 MS4b 35 595 331 279 0.9951–0.0317X 4 MS7b 20 190 88 79 0.9724 + 0.1318X 2 ITSb 20 190 81 52 0.9641 + 0.0053X 3 S: number of segregating sites; P: number of pairwise comparisons; F: number of significant pairwise comparisons by Fisher’s exact test; B: number of significant comparisons after the Bonferroni correction; LD (|D’|): linkage disequilibrium between sites and X is the nucleotide distance (measured in kilobases; kb); Rm: minimum number of recombination events. aAmong 101 E. bieneusi isolates from pigs. bAmong 207 E. bieneusi isolates from pigs, humans, and baboons. Multilocus sequence typing (MLST) and analysis To investigate the genetic diversity and population characteristics of E. bieneusi, the five loci were concatenated into a single multilocus contig of 2,128 bp in length. The contigs from 101 pig isolates include 159 polymorphic sites (41 segregating sites and 118 indel sites). Due to the difference in substitution rate between SNPs and indels, two methods were used to estimate genetic diversity. The finite population genetic variance estimates that consider both SNPs and indels allowed identification of a total of 44 MLGs with a Hd value of 0.95 and a nucleotide diversity (Pi) value of 0.0197 (Table 4). In contrast, the use of infinite population genetic variance estimates that consider only SNPs led to reduced genetic diversity (MLGs = 37, Hd = 0.93, Pi = 0.0065) (Table 4). The genetic diversity was also estimated for each of the four cities in northeast China (Table 4). 10.1371/journal.pntd.0004966.t004Table 4 Genetic diversity in Enterocytozoon bieneusi based on the analysis of concatenated sequences from five genetic loci. Taxa Test model Variability of multilocus gene sequences N Hd k Pi Theta (k) Fs (Obs) P (Fs ≤ Obs) D (Obs) P (D ≤ Obs) LD (|D'|) Rm Changchun F 19 0.93 25 0.0117 25 2.14 0.830 2.19 0.994 I 16 0.91 13 0.0063 13 0.84 0.655 2.21 0.993 1.0024–0.1373X 4 Daqing F 8 0.82 56 0.0263 56 19.23 1.000 1.16 0.908 I 7 0.75 11 0.0056 11 6.24 0.980 1.16 0.913 0.9635–0.1451X 4 Harbin F 8 0.61 9 0.0041 9 4.86 0.961 −1.12 0.121 I 6 0.37 3 0.0015 3 1.84 0.813 −1.10 0.146 0.9988–0.1797X 3 Qiqihar F 9 0.86 29 0.0138 29 5.66 0.993 1.61 0.977 I 8 0.85 10 0.0050 10 2.37 0.843 1.52 0.961 1.0282–0.2990X 2 Total (4 cities) F 44 0.95 42 0.0197 42 3.26 0.835 2.10 0.978 I 37 0.93 13 0.0065 13 −3.38 0.248 2.07 0.983 0.9352–0.2020X 11 SP1 F 51 0.99 31 0.0142 31 −10.35 0.025 0.04 0.583 I 33 0.97 6 0.0029 6 −15.73 0.000 0.05 0.580 0.9473–0.1084X 5 SP2 F 9 0.52 80 0.0388 80 ∞ N/A −0.97 0.188 I 7 0.51 8 0.0051 8 7.68 0.983 −0.92 0.175 1.0165–0.1078X 2 SP3 F 18 0.91 53 0.0250 53 9.33 0.994 1.55 0.957 I 17 0.91 11 0.0055 11 −0.21 0.510 1.66 0.963 0.9724–0.0385X 5 SP4 F 14 0.79 16 0.0071 16 4.80 0.972 −0.96 0.168 I 11 0.67 4 0.0018 4 −0.68 0.427 −0.87 0.206 0.9844–0.0339X 3 SP5 F 13 0.90 15 0.0069 15 2.18 0.840 0.01 0.556 I 10 0.84 4 0.0020 4 −0.60 0.416 0.01 0.546 0.9399 + 0.0413X 3 Total (5 subpopulations) F 105 0.97 72 0.0328 72 −1.64 0.483 −0.55 0.349 I 76 0.96 13 0.0080 13 −27.05 0.001 −0.54 0.339 0.9773–0.0401X 13 F: finite population genetic variance estimates; I: infinite population genetic variance estimates; N: number of multilocus genotypes; Hd: gene diversity; k: mean number of pairwise differences; Pi: nucleotide diversity (average over loci); Theta (k): gene variance based on mean number of pairwise differences; Fs/D (Obs): observed value of Fu’s/Tajima’s statistic testing selective neutrality based on allele frequency; P (Fs/D ≤ Obs): probability of obtaining Fs/D value equal or lower than the observed; LD (|D'|): linkage disequilibrium between sites and X is the nucleotide distance (measured in kilobases; kb); Rm: minimum number of recombination events. Tests for intragenic LD and recombination among segregating sites were performed on combined multilocus contigs. The overall population and individual populations of E. bieneusi in pigs from four geographic locations all had significant but incomplete LD (Table 4). The negative slope returned by LD score regression is indicative of LD index declines with increasing nucleotide distance, implying the potential occurrence of recombination. A varying number of Rms were detected in the overall and individual pig E. bieneusi populations from four cities using DnaSP (Table 4). The occurrence of Rms was also confirmed using the GENECOV, MaxChi, and SiScan methods in RDP4 (S1 Table). The neutrality tests conducted with DnaSP (using SNPs only) and Arlequin (using both SNPs and indels) were both significant, rejecting the null hypothesis of a neutral population at mutation-drift equilibrium (Table 4). The negative Fs and D values (highlighted in bold in the center of Table 4) obtained in the tests of the pig E. bieneusi populations implied an excess of low frequency polymorphisms, as would be expected from a recent population expansion. We also calculated ISA and compared the values of VD (variance of pairwise differences) and L (the 95% critical value for VD relative to the null hypothesis of panmixia) to assess the population structure of E. bieneusi using allelic profile data [34]. As presented in Table 5, significant positive ISA values (at least 0.2733 in Qiqihar, PMC < 0.001) were obtained for the overall and individual pig E. bieneusi populations. The value of VD was greater than that of L for each data set as well. Thus, the populations tested were all in strong LD. Calculation of ISA and comparison of VD and L were also applied for the analysis of MLGs to avoid the possibility that LD would result from a clonal expansion of one or more MLGs which might mask the underlying equilibrium. The analysis showed the overall pig E. bieneusi population retained LD (ISA = 0.1441, PMC < 0.001; VD > L) when the isolates with the same MLG were treated as one individual (Table 5). Pairwise intergenic analysis of the five loci using the allelic profile data revealed strong LD as well (S2 Table). We also estimated the effective migration rate (Nm) using the FST method. Pairwise analysis between geographic populations yielded FST values ranging from 0.327 to 0.506 and Nm values ranging from 0.24 to 0.48 (Table 6). Thus, geographic segregation of the isolates and limited gene flow occurred among the pig E. bieneusi populations from four cities. 10.1371/journal.pntd.0004966.t005Table 5 Results of linkage disequilibrium analysis based on allelic profile data from five genetic loci. City/subpopulation No. Hd ISA PMC VD L VD > L Changchun 34 0.6913 ± 0.0326 0.3763 < 0.001 2.62 1.1664 Y Daqing 22 0.5740 ± 0.0267 0.6389 < 0.001 4.2961 1.4005 Y Harbin 29 0.2739 ± 0.0737 0.4417 < 0.001 2.4508 1.1817 Y Qiqihar 16 0.4717 ± 0.1335 0.2733 < 0.001 1.8621 1.1226 Y Total (4 cities) 101 0.7986 ± 0.0406 0.3768 < 0.001 1.9338 0.8009 Y Total (4 cities)a 44 0.8408 ± 0.0334 0.1441 < 0.001 1.0197 0.7022 Y SP1 66 0.8366 ± 0.0351 0.1191 < 0.001 0.9725 0.7001 Y SP2 39 0.1976 ± 0.0577 0.3729 < 0.001 1.8093 0.9471 Y SP3 35 0.6971 ± 0.0605 0.4709 < 0.001 2.8327 1.0886 Y SP4 41 0.5276 ± 0.0748 0.4866 < 0.001 3.3424 1.3644 Y SP5 26 0.5292 ± 0.1023 0.2535 < 0.001 3.3424 1.3644 Y Total (5 subpopulations) 207 0.8988 ± 0.0180 0.4666 < 0.001 1.2845 0.4563 Y Total (5 subpopulations)a 105 0.9178 ± 0.0193 0.1577 < 0.001 0.6030 0.3821 Y Hd: mean genetic diversity; ISA: standardized index of association calculated using the program LIAN 3.5; PMC: significance for obtaining this value in 1000 simulations using the Monte Carlo method; VD: variance of pairwise differences; L: 95% critical value for VD; VD > L indicates linkage disequilibrium. a Considering isolates with the same MLG as one individual. 10.1371/journal.pntd.0004966.t006Table 6 Pairwise genetic distance (FST, lower diagonal, P < 0.001) and gene flow (Nm, upper diagonal) between Enterocytozoon bieneusi populations. Changchun Daqing Harbin Qiqihar SP1 SP2 SP3 SP4 SP5 Changchun 0.48 0.32 0.51 SP1 0.39 0.46 0.24 1.10 Daqing 0.344 0.24 0.46 SP2 0.391 0.38 0.30 0.36 Harbin 0.436 0.506 0.41 SP3 0.352 0.398 0.48 0.39 Qiqihar 0.327 0.353 0.379 SP4 0.515 0.454 0.342 0.20 SP5 0.185 0.412 0.393 0.559 Population genetic analysis was also conducted when the MLST data from 106 E. bieneusi isolates from humans and baboons were included. Significant LD and limited recombination were obtained in the analysis of a total of 207 isolates (Tables 4 and 5). The test of selective neutrality revealed a nonneutral structure for the population (Table 4). The negative Fs and D statistics (highlighted in bold at the bottom of Table 4) obtained from this test signified potential epidemic expansion of some MLGs and genetic subdivision. Subpopulations and population genetic analysis A Bayesian method was used to infer phylogenetic relationships among E. bieneusi isolates from pigs, humans, and baboons (Fig 1A). The isolates were divided into two major phylogenetic clusters (one includes 44 MLGs from pigs and 1 MLG from a Peruvian adult with HIV infection and the other one includes 56 MLGs from humans and 4 MLGs from baboons) (Fig 1A). Additional subdivision within the two clusters led to the formation of five genetically isolated subgroups (Fig 1A). The same clustering patterns appeared in the 3D image of PCoA with two main clusters (blue balls represent the isolates from humans and baboons and red balls from pigs) and five genetic subdivisions generated (Fig 1C). Considering the high concordance of the grouping patterns of MLGs formed in Bayesian inference and PCoA, we defined two known primate E. bieneusi subpopulations as SP1 and SP2 and three novel pig E. bieneusi subpopulations as SP3 to SP5. Subpopulations SP1 to SP5 contained 51, 9, 18, 14, and 13 MLGs derived from 66, 39, 35, 41, and 26 E. bieneusi isolates, respectively (Tables 2 and 4, Fig 1D). In general, subpopulations SP2 to SP4 had higher frequency of MLGs than SP1 and SP5 (Fig 1D). 10.1371/journal.pntd.0004966.g001Fig 1 Comprehensive phylogenetic and structural analyses of Enterocytozoon bieneusi isolates from various hosts and locations. Panel A: Bayesian phylogenetic analysis of 105 unique multilocus genotypes (MLGs) from 207 isolates. Among them, 44 MLGs are from pigs, 61 from humans, and 4 from baboons, which are indicated by P, H, and B before the specimen numbers, respectively. The letters I, N, P, K, and C (C, D, H, and Q) followed indicate the isolates are from India, Nigeria, Peru, Kenya, and China (Changchun, Daqing, Harbin, and Qiqihar), respectively. ITS genotypes are labeled at the ends. Panel B: Subpopulation structure of all 207 E. bieneusi isolates. Various subpopulation patterns were obtained when different K values (2 to 6) were used. Panel C: Results of the principal coordinates analysis of 105 unique E. bieneusi MLGs based on pairwise distances. Red and blue balls represented the isolates from pig and primate populations, respectively. Panel D: Frequency (%) of E. bieneusi MLGs in the five subpopulations determined in structural analysis. We also performed substructure analysis based on allelic profile data using STRUCTURE. The initial run with K = 2 identified two major subclusters (Fig 1B). One subcluster in red included all isolates from pigs and one isolate from a human and the other subcluster in green contained the isolates from humans and baboons (Fig 1B), corresponding to the two major substructures generated in Bayesian inference and PCoA. The following runs at K = 3 and 4 yielded various intermediate patterns of population subdivision. The run at K = 5 showed the presence of five clear and robust subpopulations (Fig 1B). The isolates in each of the five subpopulations agreed with those in SP1 to SP5 except that several isolates from SP4 were clustered into SP5 (Fig 1). We also measured the alpha (α) values generated in the substructure analysis. When there are firm subdivisions in a population, the α values are held constant and commonly range from 0 to 0.2 in different runs. The runs at K = 2 to 5 in this study yielded consistent α values around 0.03, which supported the robustness of the substructure formation. However, when the analysis was performed at K = 6 or more, a fairly mixed and confused scene of clustering was observed (Fig 1B). Taken together, these results suggest that the run with K = 5 provided the best fit to our data. Based on the results from Bayesian phylogeny, PCoA, and substructure analysis, it was apparent that five genetically isolated subdivisions were present in the total population. Distribution preference of the pig E. bieneusi isolates from Changchun and Daqing in SP3 and SP5 and those from Harbin and Qiqihar in SP4 suggested the presence of genetic segregation at some geographic level (S3 Table). Genetic diversity, intragenic LD, and Rms were measured based on multilocus sequences for each of the five subpopulations. Pairwise LD comparisons among the subpopulations showed that the clonality of E. bieneusi isolates in SP2 to SP4 was stronger than that in SP1 and SP5 (Table 4). In agreement with this, higher ISA values were generated in the analysis of SP2 to SP4 than in SP1 and SP5 (Table 5). The measurement of population divergence among SP1 to SP5 was performed by the analysis of FST and gene flow (Table 6). SP1 and SP5 were shown to have a close genetic relationship (FST = 0.185 and Nm = 1.10) (Table 6). Nevertheless, in comparative analysis of SP1 or SP5 to any other three subpopulations, the FST values of at least 0.342 and Nm values of at most 0.48 indicated the presence of significant population differentiation and very limited gene flow. Median-joining networks were used to infer the relationships between MLGs (Fig 2). The analysis showed the existence of five clusters marked in blue, green, yellow, purple, and red, which corresponded to SP1 to SP5, respectively (Fig 2). As central haplotypes are generally considered possible ancestors of the peripheral ones [26,41], the MLGs in SP2 to SP4 might have derived from the central ones in SP1 and SP5 (Fig 2). In addition, high dimensional networks in SP1 suggested the presence of significant recombination (Fig 2). 10.1371/journal.pntd.0004966.g002Fig 2 Median-joining network for inferring intraspecific phylogenies of 207 Enterocytozoon bieneusi isolates from pigs in China, humans in India, Nigeria, and Peru, and baboons in Kenya. The size of the circles is proportional to the frequency of each of the 76 multilocus genotypes obtained based on segregating sites. The red, black, blue, yellow, and green colors in circles represent the isolates from China, India, Kenya, Nigeria, and Peru, respectively. ITS genotypes are labeled besides the circles. The black branches connecting multilocus genotypes have a length proportional to the number of single-nucleotide polymorphisms (SNPs), while the red branches having pairwise differences greater than 12 SNPs are shortened for better presentation. Discussion Despite advances in defining E. bieneusi ITS genotypes from different hosts and geographical regions, the relationship between genotypes and phenotypic traits such as host specificity and zoonotic potential remains unclear [2,11]. Analysis of population structure can help us understand the epidemiology and evolution of parasites [42]. MLST analysis has provided substantial new insights into the population genetics of E. bieneusi in humans and non-human primates [26–28]. Herein, we evaluated intragenic and intergenic LD, ISA, neutrality, and Rms to determine the genetic structure and substructures in E. bieneusi populations from pigs using both sequence and allelic profile data from five genetic loci. The presence of strong LD and very limited recombination supported the existence of significant clonal structure in the overall and individual pig E. bieneusi populations from four cities, northeast China. In addition, as illustrated in Fig 2, the results of several neutrality tests suggest that selection acting in pig E. bieneusi populations has led to the expansion of several dominant MLGs, which might play a role in enhancing their adaptation to specific hosts. Two recent studies described that two other microsporidian species known to infect honeybees (Nosema apis and Nosema ceranae) were also under selective pressure and experienced a population expansion [43,44]. The estimates of FST and Nm and the distribution of the pig E. bieneusi isolates revealed the existence of geographic segregation among the cities surveyed. Pigs are considered a potential reservoir for human microsporidiosis based on genotypic features of E. bieneusi isolates at the ITS locus [2,17,23]. The ITS genotypes of pig E. bieneusi isolates used for MLST analysis in this study all belong to phylogenetic Group 1 with zoonotic potential [2,17,23]. The ITS genotypes of primate E. bieneusi isolates used for comparative analysis are also Group 1 members [26,27]. These Group 1 isolates formed several genetically isolated subpopulations (two existing primate SP1 and SP2 and three novel pig SP3 to SP5) in MLST analysis of five genetic loci. SP1 to SP5 probably have different phenotypic traits as reflected in the distribution of E. bieneusi isolates in these subpopulations. As observed in Fig 1A and S4 Table, SP1 and SP5 are comprised mainly of the isolates pertaining to zoonotic ITS genotypes D, IV, and EbpC that are found in a wide range of hosts and regions around the world. In contrast, SP2 to SP4 contain mainly isolates belonging to ITS genotypes A, EbpA, and EbpB that have narrow host and geographic ranges. The stronger LD and higher occurrence of specific MLGs observed in SP2 to SP4 than SP1 and SP5 suggest the presence of higher clonality of E. bieneusi isolates in the former three subpopulations. The high diversity of E. bieneusi isolates in SP1 and SP5 may enable responses to environmental challenges and adaptations to new hosts [45]. Thus, MLGs in SP1 and SP5 might be responsible for cross-species E. bieneusi infections and have zoonotic potential. This is supported by the broad host range of E. bieneusi ITS genotypes D, IV, and EbpC in the two subpopulations. Reduced gene flow between primate SP1 and SP2 and between pig SP5 and SP3 might promote the emergence of advantageous haplotypes in SP2 and SP3 and allow these haplotypes to remain intact despite the possibility of recombination [46]. These processes might enable adaptation to specific host niches and initiate allopatric speciation [46]. In particular, this may have led to the adaptation of ITS genotype A in SP2 to humans and genotypes EbpA and EbpB in SP3 to pigs. Thus, MLGs in SP2 and SP3 are probably involved in host-specific colonization. SP2 was previously reported to have an epidemic population structure [26,27]. Likewise, genetic structure of SP3 with host-adapted features can also be considered to be epidemic. SP4 consists of E. bieneusi isolates belonging to the ITS genotypes CS-4, Henan-I, Henan-IV, PigEBITS3, and EbpC. The former four genotypes were shown to infect a very limited number of host species as shown in S4 Table, while EbpC has a wide host range. Thus, the subpopulation probably represents an evolutionary intermediate between SP3 with host-adapted traits and SP5 with cross-species capability. The placement of one human E. bieneusi isolate with ITS genotype EbpC in SP4 supported the zoonotic nature of some E. bieneusi isolates. Population genetic analysis is an indirect but powerful way to assess reproductive modes that are often difficult to uncover in some microorganisms [42]. Sexual and asexual reproduction alternates in many protozoan species under environmental pressure [47]. It is generally accepted that microsporidia undergo sexual reproduction, whereas some species or genotypes possibly have switched to obligate asexuality [48]. This is supported by the existence of clonal population structure in SP1 and SP5 and epidemic population structure or host-adapted traits in SP3 to SP4. Although a sexual phase might be rare or virtually absent in some microsporidian species, there are footprints of recombination in their genomes [47,48]. This could be responsible for the small number of recombination inferred in subpopulations SP1 through SP5. Although limited recombination would not constitute sufficient evidence for the presence of sexuality in SP1 and SP5, the weakened LD and high levels of MLG diversity might facilitate E. bieneusi to cope with host variations and environmental challenges. The haplotype network suggests that clusters SP3/SP4 may have been derived from SP5 and that SP2 may have arisen from SP1. Thus, SP2 to SP4 with host-specific features might serve as recurrent transitions to asexuality from otherwise SP1 and SP5 with sexual potential and some of the isolates in SP2 to SP4 might have outcompeted or partially displaced their relatives in SP1 and SP5 under certain ecological conditions. This process would limit host range and result in the emergence of highly successful E. bieneusi genotypes. It has been suggested that some asexual microsporidian populations might originate independently several times from their sexual ancestors [48,49]. These results indicate that E. bieneusi might have a sexual phase in its life cycle, sex could be lost or cryptic, and the parasite could switch to obligate asexuality when the population structure becomes epidemic. In conclusion, we have shown an overall clonal population structure of E. bieneusi in pigs. Combined with the MLST data from primate E. bieneusi isolates, five distinct subpopulations have been defined. Among them, the very strong LD and low genetic recombination are indicative of the epidemic or host-adapted characteristics of SP2 to SP4, whereas the weakened LD and higher genetic diversity in SP1 and SP5 may represent higher potential for cross-species transmission of E. bieneusi infections. These data demonstrate the existence of genetic structure within E. bieneusi ITS Group 1 and the evolutionary potential for adaptation to host species in some of the subpopulations. Nevertheless, additional MLST data from other hosts including humans in China are needed for in-depth assessment of the potential for zoonotic transmission, host adaptation, and population differentiation of E. bieneusi isolates in different hosts. Supporting Information S1 Table Recombination detection. Recombination events assessed using the methods GENECONV, MaxChi, and SiScan. (DOC) Click here for additional data file. S2 Table Intergenic linkage disequilibrium. Pairwise P values for intergenic linkage disequilibrium among genetic loci based on allelic profile data from pig Enterocytozoon bieneusi populations. (DOC) Click here for additional data file. S3 Table Geographical distribution of specimens in genetic subdivisions. Proportion of the specimens from each city that belong to each of the three pig-derived Enterocytozoon bieneusi subpopulations. (DOC) Click here for additional data file. S4 Table Distribution of ITS genotypes by host and location. Host range and geographical distribution of ITS genotypes of Enterocytozoon bieneusi referred in this study. (DOC) Click here for additional data file. We would like to thank all the persons who provided kind helps and suggestions to this work and the manuscript. The findings and conclusions in this report are those of the author(s) and do not necessarily represent the official position of the U.S. Centers for Disease Control and Prevention (CDC). ==== Refs References 1 Matos O , Lobo ML , Xiao L . Epidemiology of Enterocytozoon bieneusi infection in humans . J Parasitol Res . 2012 ; 981424 10.1155/2012/981424 23091702 2 Santin M , Fayer R . Microsporidiosis: Enterocytozoon bieneusi in domesticated and wild animals . Res Vet Sci . 2011 ; 90 : 363 –371 . 10.1016/j.rvsc.2010.07.014 20699192 3 Capella-Gutierrez S , Marcet-Houben M , Gabaldon T . Phylogenomics supports microsporidia as the earliest diverging clade of sequenced fungi . BMC Biol . 2012 ; 10 : 47 10.1186/1741-7007-10-47 22651672 4 Lee SC , Corradi N , Byrnes EJ 3rd, Torres-Martinez S , Dietrich FS , et al Microsporidia evolved from ancestral sexual fungi . Curr Biol . 2008 ; 18 : 1675 –1679 . 10.1016/j.cub.2008.09.030 18976912 5 Pelin A , Moteshareie H , Sak B , Selman M , Naor A , et al The genome of an Encephalitozoon cuniculi type III strain reveals insights into the genetic diversity and mode of reproduction of a ubiquitous vertebrate pathogen . Heredity (Edinb) . 2016 ; 116 : 458 –465 .26837273 6 Pelin A , Selman M , Aris-Brosou S , Farinelli L , Corradi N . Genome analyses suggest the presence of polyploidy and recent human-driven expansions in eight global populations of the honeybee pathogen Nosema ceranae . Environ Microbiol . 2015 ; 17 : 4443 –4458 . 10.1111/1462-2920.12883 25914091 7 Desjardins CA , Sanscrainte ND , Goldberg JM , Heiman D , Young S , et al Contrasting host-pathogen interactions and genome evolution in two generalist and specialist microsporidian pathogens of mosquitoes . Nat Commun . 2015 ; 6 : 7121 10.1038/ncomms8121 25968466 8 Cuomo CA , Desjardins CA , Bakowski MA , Goldberg J , Ma AT , et al Microsporidian genome analysis reveals evolutionary strategies for obligate intracellular growth . Genome Res . 2012 ; 22 : 2478 –2488 . 10.1101/gr.142802.112 22813931 9 Mathis A , Weber R , Deplazes P . Zoonotic potential of the microsporidia . Clin Microbiol Rev . 2005 ; 18 : 423 –445 . 16020683 10 Didier ES , Weiss LM . Microsporidiosis: not just in AIDS patients . Curr Opin Infect Dis . 2011 ; 24 : 490 –495 . 10.1097/QCO.0b013e32834aa152 21844802 11 Thellier M , Breton J . Enterocytozoon bieneusi in human and animals, focus on laboratory identification and molecular epidemiology . Parasite . 2008 ; 15 : 349 –358 . 18814706 12 Guo Y , Alderisio KA , Yang W , Cama V , Feng Y , et al Host specificity and source of Enterocytozoon bieneusi genotypes in a drinking source watershed . Appl Environ Microbiol . 2014 ; 80 : 218 –225 . 10.1128/AEM.02997-13 24141128 13 Li W , Li Y , Song M , Lu Y , Yang J , et al Prevalence and genetic characteristics of Cryptosporidium, Enterocytozoon bieneusi and Giardia duodenalis in cats and dogs in Heilongjiang province, China . Vet Parasitol . 2015 ; 208 : 125 –134 . 10.1016/j.vetpar.2015.01.014 25665462 14 Santin M , Fayer R . Enterocytozoon bieneusi genotype nomenclature based on the internal transcribed spacer sequence: a consensus . J Eukaryot Microbiol . 2009 ; 56 : 34 –38 . 10.1111/j.1550-7408.2008.00380.x 19335772 15 Karim MR , Wang R , Dong H , Zhang L , Li J , et al Genetic polymorphism and zoonotic potential of Enterocytozoon bieneusi from nonhuman primates in China . Appl Environ Microbiol . 2014 ; 80 : 1893 –1898 . 10.1128/AEM.03845-13 24413605 16 Ma J , Li P , Zhao X , Xu H , Wu W , et al Occurrence and molecular characterization of Cryptosporidium spp. and Enterocytozoon bieneusi in dairy cattle, beef cattle and water buffaloes in China . Vet Parasitol . 2015 ; 207 : 220 –227 . 10.1016/j.vetpar.2014.10.011 25541482 17 Wan Q , Lin Y , Mao Y , Yang Y , Li Q , et al High prevalence and widespread distribution of zoonotic Enterocytozoon bieneusi genotypes in swine in northeast China: implications for public health . J Eukaryot Microbiol . 2016 ; 63 : 162 –170 . 10.1111/jeu.12264 26333563 18 Jiang Y , Tao W , Wan Q , Li Q , Yang Y , et al Zoonotic and potentially host-adapted Enterocytozoon bieneusi genotypes in sheep and cattle in northeast China and an increasing concern about the zoonotic importance of previously considered ruminant-adapted genotypes . Appl Environ Microbiol . 2015 ; 81 : 3326 –3335 . 10.1128/AEM.00328-15 25746997 19 Karim MR , Dong H , Yu F , Jian F , Zhang L , et al Genetic diversity in Enterocytozoon bieneusi isolates from dogs and cats in China: host specificity and public health implications . J Clin Microbiol . 2014 ; 52 : 3297 –3302 . 10.1128/JCM.01352-14 24989604 20 Yang J , Song M , Wan Q , Li Y , Lu Y , et al Enterocytozoon bieneusi genotypes in children in northeast China and assessment of risk of zoonotic transmission . J Clin Microbiol . 2014 ; 52 : 4363 –4367 . 10.1128/JCM.02295-14 25274994 21 Wang L , Zhang H , Zhao X , Zhang L , Zhang G , et al Zoonotic Cryptosporidium species and Enterocytozoon bieneusi genotypes in HIV-positive patients on antiretroviral therapy . J Clin Microbiol . 2013 ; 51 : 557 –563 . 10.1128/JCM.02758-12 23224097 22 Wang L , Xiao L , Duan L , Ye J , Guo Y , et al Concurrent infections of Giardia duodenalis, Enterocytozoon bieneusi, and Clostridium difficile in children during a cryptosporidiosis outbreak in a pediatric hospital in China . PLoS Negl Trop Dis . 2013 ; 7 : e2437 10.1371/journal.pntd.0002437 24069491 23 Fiuza VR , Oliveira FC , Fayer R , Santin M . First report of Enterocytozoon bieneusi in pigs in Brazil . Parasitol Int . 2015 ; 64 : 18 –23 . 24 Widmer G , Akiyoshi DE . Host-specific segregation of ribosomal nucleotide sequence diversity in the microsporidian Enterocytozoon bieneusi . Infect Genet Evol . 2010 ; 10 : 122 –128 . 10.1016/j.meegid.2009.11.009 19931647 25 Feng Y , Li N , Dearen T , Lobo ML , Matos O , et al Development of a multilocus sequence typing tool for high-resolution genotyping of Enterocytozoon bieneusi . Appl Environ Microbiol . 2011 ; 77 : 4822 –4828 . 10.1128/AEM.02803-10 21622791 26 Li W , Cama V , Akinbo FO , Ganguly S , Kiulia NM , et al Multilocus sequence typing of Enterocytozoon bieneusi: lack of geographic segregation and existence of genetically isolated sub-populations . Infect Genet Evol . 2013 ; 14 : 111 –119 . 10.1016/j.meegid.2012.11.021 23262155 27 Li W , Cama V , Feng Y , Gilman RH , Bern C , et al Population genetic analysis of Enterocytozoon bieneusi in humans . Int J Parasitol . 2012 ; 42 : 287 –293 . 10.1016/j.ijpara.2012.01.003 22534008 28 Karim MR , Wang R , He X , Zhang L , Li J , et al Multilocus sequence typing of Enterocytozoon bieneusi in nonhuman primates in China . Vet Parasitol . 2014 ; 200 : 13 –23 . 10.1016/j.vetpar.2013.12.004 24388499 29 Li W , Li Y , Li W , Yang J , Song M , et al Genotypes of Enterocytozoon bieneusi in livestock in China: high prevalence and zoonotic potential . PLoS One . 2014 ; 9 : e97623 10.1371/journal.pone.0097623 24845247 30 Katoh K , Standley DM . MAFFT multiple sequence alignment software version 7: improvements in performance and usability . Mol Biol Evol . 2013 ; 30 : 772 –780 . 10.1093/molbev/mst010 23329690 31 Rozas J , Sanchez-DelBarrio JC , Messeguer X , Rozas R . DnaSP, DNA polymorphism analyses by the coalescent and other methods . Bioinformatics . 2003 ; 19 : 2496 –2497 . 14668244 32 Martin D , Rybicki E . RDP: detection of recombination amongst aligned sequences . Bioinformatics . 2000 ; 16 : 562 –563 . 10980155 33 Excoffier L , Lischer HE . Arlequin suite ver 3.5: a new series of programs to perform population genetics analyses under Linux and Windows . Mol Ecol Resour . 2010 ; 10 : 564 –567 . 10.1111/j.1755-0998.2010.02847.x 21565059 34 Haubold B , Hudson RR . LIAN 3.0: detecting linkage disequilibrium in multilocus data. Linkage Analysis . Bioinformatics . 2000 ; 16 : 847 –848 . 11108709 35 Ronquist F , Teslenko M , van der Mark P , Ayres DL , Darling A , et al MrBayes 3.2: efficient Bayesian phylogenetic inference and model choice across a large model space . Syst Biol . 2012 ; 61 : 539 –542 . 10.1093/sysbio/sys029 22357727 36 Posada D . ModelTest Server: a web-based tool for the statistical selection of models of nucleotide substitution online . Nucleic Acids Res . 2006 ; 34 : W700 –703 . 16845102 37 Schloss PD , Westcott SL , Ryabin T , Hall JR , Hartmann M , et al Introducing mothur: open-source, platform-independent, community-supported software for describing and comparing microbial communities . Appl Environ Microbiol . 2009 ; 75 : 7537 –7541 . 10.1128/AEM.01541-09 19801464 38 Peakall R , Smouse PE . GenAlEx 6.5: genetic analysis in Excel. Population genetic software for teaching and research—an update . Bioinformatics . 2012 ; 28 : 2537 –2539 . 22820204 39 Pritchard JK , Stephens M , Donnelly P . Inference of population structure using multilocus genotype data . Genetics . 2000 ; 155 : 945 –959 . 10835412 40 Bandelt HJ , Forster P , Rohl A . Median-joining networks for inferring intraspecific phylogenies . Mol Biol Evol . 1999 ; 16 : 37 –48 . 10331250 41 Vidigal PM , Mafra CL , Silva FM , Fietto JL , Silva Junior A , et al Tripping over emerging pathogens around the world: a phylogeographical approach for determining the epidemiology of Porcine circovirus-2 (PCV-2), considering global trading . Virus Res . 2012 ; 163 : 320 –327 . 10.1016/j.virusres.2011.10.019 22056846 42 Tibayrenc M , Ayala FJ . Cryptosporidium,Giardia, Cryptococcus, Pneumocystis genetic variability: cryptic biological species or clonal near-clades? PLoS Pathog . 2014 ; 10 : e1003908 10.1371/journal.ppat.1003908 24722548 43 Roudel M , Aufauvre J , Corbara B , Delbac F , Blot N . New insights on the genetic diversity of the honeybee parasite Nosema ceranae based on multilocus sequence analysis . Parasitology . 2013 ; 140 : 1346 –1356 . 10.1017/S0031182013001133 23880415 44 Maside X , Gomez-Moracho T , Jara L , Martin-Hernandez R , De la Rua P , et al Population genetics of Nosema apis and Nosema ceranae: one host (Apis mellifera) and two different histories . PLoS One . 2015 ; 10 : e0145609 10.1371/journal.pone.0145609 26720131 45 Heitman J . Evolution of eukaryotic microbial pathogens via covert sexual reproduction . Cell Host Microbe . 2010 ; 8 : 86 –99 . 10.1016/j.chom.2010.06.011 20638645 46 Rundle HD , Nosil P . Ecological speciation . Ecol Lett . 2005 ; 8 : 336 –352 . 47 Haag KL , Sheikh-Jabbari E , Ben-Ami F , Ebert D . Microsatellite and single-nucleotide polymorphisms indicate recurrent transitions to asexuality in a microsporidian parasite . J Evol Biol . 2013 ; 26 : 1117 –1128 . 10.1111/jeb.12125 23530861 48 Haag KL , Traunecker E , Ebert D . Single-nucleotide polymorphisms of two closely related microsporidian parasites suggest a clonal population expansion after the last glaciation . Mol Ecol . 2013 ; 22 : 314 –326 . 10.1111/mec.12126 23163569 49 Ironside JE . Multiple losses of sex within a single genus of microsporidia . BMC Evol Biol . 2007 ; 7 : 48 17394631

PMC005xxxxxx/PMC5001695.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756470310.1371/journal.pone.0161816PONE-D-15-49425Research ArticleResearch and Analysis MethodsModel OrganismsAnimal ModelsMouse ModelsBiology and Life SciencesCell BiologyCellular TypesAnimal CellsNeuronsMotor NeuronsBiology and Life SciencesNeuroscienceCellular NeuroscienceNeuronsMotor NeuronsBiology and Life SciencesPhysiologyPhysiological ParametersBody WeightMedicine and Health SciencesPhysiologyPhysiological ParametersBody WeightBiology and Life SciencesBiochemistryMetabolismMetabolic ProcessesCitric Acid CycleBiology and Life SciencesBiochemistryBioenergeticsEnergy-Producing OrganellesMitochondriaBiology and Life SciencesCell BiologyCellular Structures and OrganellesEnergy-Producing OrganellesMitochondriaBiology and Life SciencesAnatomyMusculoskeletal SystemMusclesGastrocnemius MusclesMedicine and Health SciencesAnatomyMusculoskeletal SystemMusclesGastrocnemius MusclesBiology and Life SciencesAnatomyNervous SystemNeuroanatomySpinal CordMedicine and Health SciencesAnatomyNervous SystemNeuroanatomySpinal CordBiology and Life SciencesNeuroscienceNeuroanatomySpinal CordMedicine and Health SciencesNeurologyNeurodegenerative DiseasesMotor Neuron DiseasesAmyotrophic Lateral SclerosisTriheptanoin Protects Motor Neurons and Delays the Onset of Motor Symptoms in a Mouse Model of Amyotrophic Lateral Sclerosis Triheptanoin and Amyotrophic Lateral SclerosisTefera Tesfaye W. 1Wong Yide 1¤aBarkl-Luke Mallory E. 1Ngo Shyuan T. 12¤bThomas Nicola K. 1McDonald Tanya S. 1Borges Karin 1*1 School of Biomedical Sciences, The University of Queensland, Brisbane, Queensland, Australia2 University of Queensland Centre for Clinical Research, The University of Queensland, Brisbane, Queensland, AustraliaKobeissy Firas H EditorUniversity of Florida, UNITED STATESCompeting Interests: We have applied for a US patent regarding the treatment of ALS (Neurodegenerative disorders and methods of treatment and diagnosis thereof. US provisional patent application 2013; US 61/904,365). Sasol and Ultragenyx Pharmaceuticals Inc. donated triheptanoin. There are no additional patents, products in development or marketed products to declare. This does not alter our adherence to PLOS ONE policies on sharing data and materials. Conceptualization: KB STN. Data curation: TWT YW MEB NKT TSM. Formal analysis: TWT YW KB. Funding acquisition: KB. Investigation: TWT YW MEB NKT TSM. Methodology: STN TSM. Project administration: KB. Resources: KB. Supervision: KB. Validation: TWT KB. Visualization: TWT YW MEB KB. Writing – original draft: TWT YW. Writing – review & editing: TWT YW MEB NKT STN TSM KB. ¤a Current address: Queensland Institute of Medical Research, Brisbane, Queensland, Australia ¤b Current address: Queensland Brain Institute, The University of Queensland, Brisbane, Queensland, Australia * E-mail: k.borges@uq.edu.au26 8 2016 2016 11 8 e016181612 11 2015 14 8 2016 © 2016 Tefera et al2016Tefera et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.There is increasing evidence that energy metabolism is disturbed in Amyotrophic Lateral Sclerosis (ALS) patients and animal models. Treatment with triheptanoin, the triglyceride of heptanoate, is a promising approach to provide alternative fuel to improve oxidative phosphorylation and aid ATP generation. Heptanoate can be metabolized to propionyl-CoA, which after carboxylation can produce succinyl-CoA and thereby re-fill the tricarboxylic acid (TCA) cycle (anaplerosis). Here we tested the hypothesis that treatment with triheptanoin prevents motor neuron loss and delays the onset of disease symptoms in female mice overexpressing the mutant human SOD1G93A (hSOD1G93A) gene. When oral triheptanoin (35% of caloric content) was initiated at P35, motor neuron loss at 70 days of age was attenuated by 33%. In untreated hSOD1G93A mice, the loss of hind limb grip strength began at 16.7 weeks. Triheptanoin maintained hind limb grip strength for 2.8 weeks longer (p<0.01). Loss of balance on the rotarod and reduction of body weight were delayed by 13 and 11 days respectively (both p<0.01). Improved motor function occurred in parallel with alterations in the expression of genes associated with muscle metabolism. In gastrocnemius muscles, the mRNA levels of pyruvate, 2-oxoglutarate and succinate dehydrogenases and methyl-malonyl mutase were reduced by 24–33% in 10 week old hSOD1G93A mice when compared to wild-type mice, suggesting that TCA cycling in skeletal muscle may be slowed in this ALS mouse model at a stage when muscle strength is still normal. At 25 weeks of age, mRNA levels of succinate dehydrogenases, glutamic pyruvic transaminase 2 and the propionyl carboxylase β subunit were reduced by 69–84% in control, but not in triheptanoin treated hSOD1G93A animals. Taken together, our results suggest that triheptanoin slows motor neuron loss and the onset of motor symptoms in ALS mice by improving TCA cycling. http://dx.doi.org/10.13039/501100000925Australian National Health and Medical Research Council1044407Borges Karin STN is currently a Scott Sullivan Research Fellow funded by the MND and Me Foundation, The Queensland Brain Institute and The Royal Brisbane and Women’s Hospital Foundation. KB is funded by the Australian National Health and Medical Research Council (Grant 1044407). TWT and TSM receive UQI and APA scholarship support, respectively. Sasol and Ultragenyx Pharmaceuticals Inc. donated triheptanoin. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAll relevant data are within the paper.Data Availability All relevant data are within the paper. ==== Body Introduction Amyotrophic Lateral Sclerosis (ALS) is a progressive neurodegenerative disorder characterized by selective degeneration of motor neurons in the brain and spinal cord that leads to muscle weakness, paralysis and death usually due to respiratory failure [1]. The typical age of onset for most forms of ALS is between 50 to 60 years and most patients die 3 to 5 years after symptom onset [2]. Ninety percent of ALS cases are sporadic (SALS), while 10% are familial (FALS) [3]. Mutations in genes such as superoxide dismutase 1 (SOD1) [4], TAR DNA-Binding Protein (TDP 43) [5], fused in sarcoma (FUS) [6], Ubiquilin2 (UBQLN2) [7], and C9ORF72 [8,9] are implicated in genetic causes of ALS. Twenty percent of familial ALS cases are linked to mutations in the SOD1 gene (most commonly occurring mutation in patients with FALS) and this accounts for 1 to 2% of all forms of ALS while FUS and TDP-43 mutations account for 5% of FALS cases [3,10]. SOD1 mutations result in a toxic gain-of-function and are associated with misfolding and mislocalization of the SOD1 protein. While the normal SOD1 protein is usually found in the cytosol, mutant SOD1 accumulates within mitochondria and appears to contribute to many of the mitochondrial perturbations observed in ALS [11–15]. The exact mechanisms underlying selective motor neuron degeneration in ALS are unclear. The mechanisms causing motor neuron loss are multifactorial and may not be mutually exclusive [16]. Among many pathogenic mechanisms, major key pathological processes have been identified including oxidative stress [17], glutamate excitotoxicity [18], inflammation [19], abnormal protein aggregation [4,20], impaired axonal transport [21] and abnormalities in energy metabolism [22,23]. In addition, pathogenic processes in muscle seem to contribute to the progression of disease [24]. ALS and Energy Metabolism Mitochondria are the main sites of energy production. The tricarboxylic acid (TCA) cycle together with the electron transport chain produce ATP, the primary cellular energy source that is necessary for cell function and survival. Functional and morphological abnormalities in mitochondria have been shown in the brain, spinal cord and muscles of patients with ALS and in mouse models of ALS [25–35]. Defects in mitochondrial respiration, the electron transport chain as well as ATP synthesis in the spinal cords of hSOD1G93A mice [34], and reduced cellular ATP in the skeletal muscles and cerebral cortex of hSOD1G93A mice have also been observed [36,37]. In this regard, impaired mitochondrial function in ALS is likely to underpin defective energy metabolism and a reduction in the capacity to produce ATP. Reduced glucose metabolism has been shown in in various brain regions [38–42] and in the spinal cord [37,43] and muscles [36,44] of patients with ALS and mouse models of ALS. Indeed, the expression of several genes that encode for proteins and enzymes involved in glucose uptake, glycolysis, TCA cycle and the electron transport chain were found to be altered in fibroblasts cultured from ALS patients, in the motor cortex of ALS patients, and in the muscle and spinal cords of SOD1 mice [45–49]. Furthermore, there are indications that suggest reduction in levels of TCA cycle intermediates in SOD1 mouse brain and spinal cord and cellular models of ALS [50–52]. Collectively, these findings indicate that impaired ATP production, β-oxidation and TCA cycling contribute to metabolic aberrations in ALS. Reduced levels of aspartate, glutamine and GABA in the spinal cord of hSOD1G93A mice during early disease stages suggest that C4 intermediate levels of the TCA cycle are decreased [51]. In the CNS, glutamine supplies C4 carbons to glutamate and subsequently GABA and 2-oxoglutarate (α-ketoglutarate), while aspartate can be deaminated to form oxaloacetate (Fig 1). Low levels of oxaloacetate that can occur due to low supply of C4 TCA cycle intermediates and metabolites, including aspartate and glutamine, may impair the entry of acetyl-CoA generated from fuels into the TCA cycle. Therefore, the use of an alternative and additional fuel source that can “re-fill” the C4 carbon deficient TCA cycle intermediates (anaplerosis) may be of benefit in alleviating metabolic defects in ALS. When needed, anaplerosis is likely to improve ATP production in both the CNS and muscle especially during periods of energetic stress [53]. Thus, we suggest anaplerosis as a potential treatment approach for ALS. 10.1371/journal.pone.0161816.g001Fig 1 Simplified TCA cycle and anaplerosis in CNS and muscle. Red numbers indicate anaplerotic pathways, that can refill the levels of C4 intermediates of the cycle: 1 pyruvate carboxylase (mostly in CNS), 2 propionyl-CoA carboxylase, and 3 glutamic pyruvic transaminases (very important in muscle). C5 ketones, branched chain amino acids and heptanoate, are metabolized to propionyl-CoA and can therefore be anaplerotic via the propionyl-CoA carboxylation pathway. OAA–oxaloacetate, 2-OG– 2-oxoglutarate, Ile–isoleucine, Val—valine. Anaplerosis and Triheptanoin as a Treatment Approach Anaplerosis serves to “re-fill” the levels of C4 carbon-deficient TCA cycle intermediates to improve energy supply during periods of increased energy need [53]. Anaplerotic enzymes include pyruvate carboxylase (Pcx), which produces oxaloacetate and glutamic pyruvic transaminases 1 and 2 (Gpt1 and 2), which catalyze the reaction pyruvate + glutamate < = > α-ketoglutarate + alanine (Fig 1). Triheptanoin, the triglyceride of heptanoic (C7) acid, not only provides an alternative fuel source in the form of medium chain fats, that are typically absent in a standard normal diet, but it can also be anaplerotic. It is a novel metabolic therapeutic that is being used in clinical pilot studies to treat patients with various disorders associated with metabolic dysfunction. This includes genetic metabolic disorders of fatty acid oxidation [53,54] and neurological and muscular disorders [55–57]. Triheptanoin provides the body with heptanoate, which diffuses into the mitochondria of cells in both the CNS and peripheral tissues to be metabolized to acetyl-CoA and propionyl-CoA by β-oxidation. Also, the liver converts heptanoate into the “C5 ketones”, β-hydroxypentanoate and β-ketopentanoate, which are then released into blood. After being taken up via monocarboxylate transporters into cells of various tissues, including the CNS, “C5 ketones” are also β-oxidized to acetyl-CoA and propionyl-CoA. Carboxylation of propionyl-CoA produces succinyl-CoA, a C4 TCA cycle intermediate via propionyl-CoA carboxylation, an anaplerotic pathway that has been described in various tissues [58–60] (Fig 1). An adequate supply of C4 TCA cycle intermediates is important for optimal oxidation of fuels by the TCA cycle [53,61]. There is now increasing evidence that triheptanoin improves energy metabolism in the CNS and muscle of patients and in animal models of different diseases, including brain and muscle in Huntington’s Disease patients [55,62] and models of epilepsy [60,63]. Triheptanoin also was found to provide alternative fuel to the brains of patients and mice with glucose transporter 1 deficiency, which show impaired glucose uptake into the CNS [56,59]. In addition, there is evidence that triheptanoin can prevent cell loss in the brain in mouse models of stroke [64] and Canavan disease [65] by improving mitochondrial respiration. Despite the beneficial effects observed following triheptanoin treatment in multiple models of disease and neurodegeneration, the use of triheptanoin as a potential therapeutic in ALS remains unexplored. Thus, we used hSOD1G93A mice to test the hypotheses that 1) triheptanoin attenuates motor neuron death and delays the onset of motor symptoms, and that 2) the expression of enzymes involved in energy metabolism is diminished in muscle and can be rescued by triheptanoin. Materials and Methods Animals All experiments were approved by the University of Queensland Animal Ethics Committee and followed the guidelines of the Queensland Animal Care and Protection Act 2001. Wild-type and hSOD1G93A mice (B6.Cg-Tg(SOD1*G93A)1Gur/J, stock no. 004435, Jackson laboratory, Bar Harbor, ME, USA), were generated by mating hSOD1G93A males with C57BL/6 wild-type females (University of Queensland). Mice were housed in a 12 hour light, 12 hour dark cycle, and had free access to food and water. Experimenters were blinded to animal genotype (until the mice started expressing the ALS phenotype) and treatment. Triheptanoin Treatment Starting at P35, female wild-type and hSOD1G93A mice were given either control or triheptanoin-containing diet treatment until they were sacrificed. Untreated mice received SF11-027, a diet with ingredients typical to other mouse diets (Specialty Feeds, Glen Forrest, WA, AUS). Treated wild-type and hSOD1G93A mice received a matched formulation (SF11-028, Specialty Feeds) in which 35% of the calories were from triheptanoin. This dose was chosen, as it is also used in the clinic to treat patients with metabolic disorders and has shown beneficial metabolic and protective effects in several animal models of various disorders [55,56,60,62–64,66]. The diets were matched in protein, mineral, antioxidant and vitamin content relative to their caloric densities [66]. Triheptanoin replaced sucrose, some of the complex carbohydrates and long chain fats. For the motor neuron count experiments, triheptanoin was provided by Ultragenyx Pharmaceutical Inc. (Novato, CA, USA). All other experiments were performed with triheptanoin obtained from Sasol GmbH (Germany). Motor Neuron Counts Mice were deeply anesthetized with pentobarbital (120 mg/Kg i.p. Provet, Northgate, QLD, AUS) and euthanized by decapitation. Spinal cords were flushed out of the spinal cavity using a cold phosphate buffered saline (PBS) filled syringe that was fitted with a blunted 23-gauge needle. Spinal cords were immediately fixed in 4% paraformaldehyde (PFA, pH 7.4) for 24 to 48 hours and then embedded in liquid paraffin. Serial, transverse sections (16μm) were obtained using a Leica Rotary Microtome. Motor neurons were identified after staining with thionine (0.1%) in acetate buffer (pH = 3.9). The L4 to L5 regions were identified according to standard anatomical guidelines described in [67] using an Olympus BX61 upright light microscope with 10x, 20x and 40x objectives. Briefly, the L4 and L5 sections were identified according to the area of the gracile fasciculus and dorsal corticospinal tract relative to the spinal cord central canal. In L3 sections, the dorsal corticospinal tract is smaller than in L4 and L5. Motor neuron numbers in L4 and L5 were determined using stereological principles based on the counting of every 10th section with a total of 11 to 16 sections counted [68]. Every tenth consecutive L4 and L5 spinal cord section was counted until the sacral dorsal commissural nucleus appeared concomitant with a decrease in size of the gracile fasciculus and dorsal corticospinal tract, which are markers for L6. Motor neurons were identified by a large, darkly stained cell body, a pale nucleus with a continuous boundary and one or more darkly stained nucleoli [67]. Blood ketone and glucose measurements Blood plasma was collected in EDTA-coated tubes from the mice used for motor neuron counting, centrifuged at 2,000g for 10 mins. It was then stored at -80°C until analysis. Plasma levels of the ketone β-hydroxybutyrate and glucose were determined using Cayman colorimetric assay kits (Ann Arbor, MI, USA), according to the manufacturer’s instructions. Behavioral Testing and Observation Animals underwent behavioral testing approximately 3–4 hours into the light cycle. All behavioral testing was conducted in an environment with minimal stimuli so as to minimize any possible effects caused by changes in external stimuli. Animals were weighed before every test session. Mice were observed, and disease progression tracked and graded according to a neurological score sheet [69] to ensure that any non-ALS related deaths were excluded from the study. In accordance with ethical requirements, hSOD1G93A mice that became too weak to reach the food hoppers were provided with wet chow on the floor of the cages. The endpoint of the study was defined as the mouse being unable to right itself within 15 seconds after being placed on its back. Upon reaching end-stage or 25 weeks of age, transgenic mice and their respective wild-type littermates were euthanized with pentobarbital (120 mg/kg, i.p., Provet). Tissues, including the gastrocnemius muscle, were collected for subsequent analysis. To measure the time point when body weight loss started, we defined the day where a loss of more than 10% in an individual mouse occurred relative to its mean body weight from week 12 to 17. Also all subsequent three body weight measurements were ≤ 90% of the original mean weight. Hind Limb Grip Strength Test Hind limb grip strength tests were conducted twice a week using a T-bar force transducer (Ugo Basile, Varese, ITA). The animal was held by the tail, ensuring its hind limbs were gripping the T-bar before being pulled downwards at a 60° angle. The reading on the force transducer was taken only if both hind limbs released the bar at the same time. The average of 10 trials per mouse was recorded for each training session [70]. To compare time points of grip strength loss, we determined the time point where a grip strength loss of more than 30% occurred in an inidividual mouse relative to its mean grip strength of week 9 to 13 and the subsequent three measurements were ≤ 70% of the original average strength. Rotarod Test Rotarod tests were conducted with 10 repeats once a week using a rotarod designed for mice (Ugo Basile). Animals were placed on the rod, which was then rotated for 3 min at 25 revolutions per minute [70]. The time at which the animal fell off was recorded. We defined the age of balance loss on the rotarod when this time was zero. RNA Extraction, cDNA Synthesis, Quantitative Real Time PCR Assay After euthanasia, the gastrocnemius muscle was quickly removed and frozen in liquid nitrogen. To extract RNA, muscle samples were pulverized in liquid nitrogen with a cold mortar and pestle, dissolved with TRI reagent (Life Technologies, Carlsbad, CA, USA) and extracted according to the manufacturer’s instructions. Contaminating DNA was removed by DNAse I treatment and cDNA was synthesized using the Tetro cDNA synthesis kit (Bioline, London, UK) according to the manufacturer’s instructions. The expression of several metabolic genes was assayed (Table 1) by quantitative real time PCR. 10.1371/journal.pone.0161816.t001Table 1 Gene names, symbol, forward and reverse primer sequences and of primers used for the gene expression studies of metabolic genes. Gene Symbol Sequence 5’ to 3’ β2-microglobulin B2m F AGACTGATACATACGCCTGCR ATCACATGTCTCGATCCCAG Hydroxymethylbilane synthase Hmbs F AAGGGCTTTTCTGAGGCACCR AGTTGCCCATCTTTCATCACTG TATA binding protein Tbp F TTCTCGAAAGAATTGCGCTGTR GCCTTGTGAGTCATTTCAGTG Propionyl-CoA Carboxylase (Subunit α) Pcca F AGAATTGCAAGGGAAATTGG R CTAAAGCCATCCCTGGTCTC Propionyl-CoA Carboxylase (Subunit β) Pccb F AGCCTACAACATGCTGGACA R GGTCCTCCCATTCATTCTTG Methylmalonyl-CoA mutase Mut F CCAAACACTGACCGTTCTCA R GGAATGTTTAGCTGCTTCAGG Pyruvate carboxylase Pcx F GAGCTTATCCCGAACATCCC R TCCATACCATTCTCTTTGGCC Pyruvate dehydrogenase E1 α 1 Pdha1 F AACTTCTATGGAGGCAACGG R CTGACCCTGATTAGCAGCAC Glyceraldehyde-3-phosphate dehydrogenase Gapdh F ATACGGCTACAGCAACAGGG R TCTTGCTCAGTGTCCTTGCT Oxoglutarate dehydrogenase Ogdh F TGCAGATGTGCAATGATGAC R GCAGCACATGGAAGAAGTTG Succinate dehydrogenase complex (Subunit A) Sdha F GGAACACTCCAAAAACAGACCT R CCACCACTGGGTATTGAGTAGAA Glutamate-pyruvate transaminase 1 Gpt1 F TGAGGTTATCCGTGCCAATAR GTCCGGACTGCTCAGAAGAT Glutamate-pyruvate transaminase 2 (alanine aminotransferase) Gpt2 F GCGACGGTATTTCTACAATCCR CGCGGAGTACAAGGGATACT All primer pairs were evaluated for efficiency using a 4 fold serial dilution series of muscle cDNA. The derived slope of each primer pair was used to calculate the efficiency by applying the formula, 4[(-1/slope)-1]*100. Reactions consisting of diluted cDNA, 5μl of SYBR Green master mix and 8μM of forward and reverse primers were amplified after an initial hot start. The thermal profile for the assay was an initial hot start of 95°C for 10 minutes, followed by 40 cycles of 95°C for 30 seconds, 60°C for 1 minute and 72°C for 30 seconds (ABI 7900HT Fast Real-Time PCR system, Applied Biosystems). Lastly the melt curve was generated by heating to 95°C for 2 minutes, cooling to 60°C for 15 seconds and a final 2% heating ramp to 95°C for 15 seconds. Samples without reverse transcriptase treatment were included to ensure that samples were free from DNA contamination. To select reference genes for normalization, expression of ten housekeeping genes in the different experimental groups were analyzed using GeNorm function. TATA binding protein (Tbp),β2-microglobulin (B2m) and Hydroxymethylbilane synthase (Hmbs) were chosen because they were the least changed in the experimental groups. The fold expression (ΔCT) of the gene of interest (goi) relative to the geometric mean of housekeeping genes (Tbp, B2m and Hmbs) were calculated with a formula adapted from [71] taking into consideration the individual efficiencies (E) of each primer pair. ΔCTgoi = 2−(CTgoiLog2Egoi−3CTTBPLog22.03CTB2mLog22.03CTHMBSLog21.86] Enzyme Assay Mitochondrial extracts were prepared from the gastrocnemius muscle of male wild-type and hSOD1G93A mice at different disease stages and homogenized with a glass-teflon homogenizer in 500 μL ice cold extraction buffer (0.32 M sucrose, 1 mM EDTA and 10 mM Tris-HCl, pH 7.4). Samples were centrifuged at 1000 g for 10 min at 4°C and the supernatant was then centrifuged at 12,000 g for 15 min at 4°C. The pellet was washed in 500 μL extraction buffer followed by another centrifugation at 12,000g for 15 min and then resuspended in cold extraction buffer with 0.1% Triton X-100. The ratio of buffer to tissue was 2.5 mL/g. The maximal activities of 2-oxoglutarate dehydrogenase in these enriched mitochondrial extracts were determined via the reduction of nicotinamide adenine dinucleotide (β-NAD+) in buffer (75 mM Tris HCl (pH 8), 1 mM ethylenediaminetetraacetic acid, 0.5 mM thiamine pyrophosphate, 1.5 mM Coenzyme A, 4 mM β-NAD+, 1mM DTT, 2 mM calcium chloride) initiated with 15 mM 2-oxoglutarate, with background activity measured when no substrate was added [72]. The rate of NAD+ reduction was measured with a spectrophotometric plate reader (Tecan, Mannedorf, CH), background activity was subtracted and the resultant activity was normalized to protein content measured with a Pierce Bicinchoninic acid assay (ThermoFisher Scientific, Scoresby, VIC, AUS). The amount of protein used was 10 μg per well. Data Analysis Statistical analyses were performed with Graphpad Prism (versions 5.03 and 6.0) using two-way ANOVAs followed by Bonferroni multiple comparisons post-hoc tests for analysis of several groups. For the comparison of the onset of body weight loss and the area under the curve (AUC) for hind limb grip strength, two-tailed unpaired t-tests were employed. Data are represented as mean ± SEM. Using the variation of our motor neuron numbers in our colony determined in a preliminary experiment, power analysis showed that to be able to observe a treatment effect of 30% with α = 0.05 and β = 80%, n = 8–9 mice were necessary, an sample size which is higher than in the guidelines [73]. In addition, power analysis using the average standard deviations for the onset of loss of grip strength and balance on rotarod showed that n = 5 was sufficient to observe changes between means by 2.3 and 1.5 weeks respectively with 80% power at the 0.05 significance level. We are aware that our study was underpowered for survival analyses based on the guidelines for preclinical ALS research [73]. Results Triheptanoin Protected Against Motor Neuron Loss in hSOD1G93A Mice A defining hallmark of ALS is progressive motor neuron loss. Here, we confirmed that relative to wild-type mice, 38% of lumbar L4-L5 motor neurons were lost by 70 days of age in hSOD1G93A mice (Fig 2, two-way ANOVA p<0.0001 for genotype, Bonferroni multiple comparisons post hoc test p<0.0001, n = 7–10 per group). To determine if triheptanoin can protect motor neurons, we initiated triheptanoin treatment at P35 in female hSOD1G93A mice at 35% of caloric intake until P70. This dose was chosen, as it is also used in the clinic to treat patients with metabolic disorders and has shown beneficial metabolic and protective effects in several animal models of various disorders [55,56,60,62–64,66]. Starting triheptanoin treatment at P35 in hSOD1G93A mice resulted in higher motor neuron numbers by 18% relative to untreated hSOD1G93A mice (two-way ANOVA p = 0.0126 for treatment, Bonferroni multiple comparisons post hoc test p<0.05), amounting to a 33% protection against motor neuron loss. 10.1371/journal.pone.0161816.g002Fig 2 Triheptanoin preserves motor neurons. Starting at P35, female wild-type and hSOD1G93A mice were either treated with triheptanoin (TRIH) or control (CON) diet until P70. (A) Stereologically counted motor neuron numbers in the L4-L5 segments of these 70 day old mice (n = 7–10) revealed a 38% loss of motor neurons in control-fed hSOD1G93A mice. Triheptanoin provided a 33% protection against motor neuron loss. Two way ANOVA p<0.0001 for genotype, p = 0.0126 for treatment, the stars indicate results from a Bonferroni multiple comparisons post hoc tests if significant (**** p<0.0001, * p<0.05) (B-D) Representative thionine stained spinal cord sections from an untreated wild-type mouse (B) and untreated (C) and triheptanoin-treated (D) hSOD1G93A mice with arrows pointing to the motor neurons counted. Scale bar 100 μm. Triheptanoin Delayed the Onset of the Loss of Hind Limb Grip Strength and Balance in hSOD1G93A Mice We investigated the extent to which triheptanoin delays the onset of the progressive loss of muscle strength in hSOD1G93A mice. Hind limb grip strength tests were used to assess the course of disease progression in treated vs. untreated mice. There was no observable difference between the mean hind limb grip strength of the treated (n = 15) vs. untreated (n = 12) wild-type mice. Mean hind limb grip strength for both groups of wild-type mice consistently fell between 300 and 600 mN (Fig 3A). The hind limb grip strength of hSOD1G93A mice on both treatments never exceeded 400 mN (Fig 3B). The time at which reduced hind limb grip strength became apparent was significantly later in triheptanoin treated hSOD1G93A mice when compared to untreated hSOD1G93A mice (n = 8 for triheptanoin treated, n = 5 for untreated mice; Fig 3B, p = 0.04, two way ANOVA). Bonferroni’s multiple comparisons tests indicated that at 18 and 19.5 weeks of age, triheptanoin treated hSOD1G93A mice had higher hind limb grip strength when compared to untreated hSOD1G93A mice (Fig 3B, p<0.05). Untreated hSOD1G93A mice began to lose hind limb grip strength at 16.7 weeks of age, but the time of onset of the loss of hind limb grip strength was delayed by 2.8 weeks in hSOD1G93A mice that received oral triheptanoin (p = 0.002, Fig 3C). The area under the curve for the grip strength over time for each mouse treated with triheptanoin was increased by 38% (p = 0.024, t-test, Fig 3D). 10.1371/journal.pone.0161816.g003Fig 3 Triheptanoin treatment delays the loss of hind limb grip strength in hSOD1G93A mice. Starting at P35, female wild-type and hSOD1G93A mice were treated with triheptanoin (TRIH) or control (CON) diet until they were sacrificed. (A) No differences in grip strength was observed between triheptanoin (green open triangles, n = 15) and control treated wild-type mice (black filled squares, n = 12). (B) The grip strength over time differed in triheptanoin treated (red crosses, n = 8) vs. untreated (blue empty circles, n = 5) hSOD1G93A mice (p = 0.04, two way ANOVA), with treated mice having higher grip strength at 18 and 19.5 weeks (p<0.05 Bonferroni post-hoc test). (C) The onset of hind limb grip strength loss was delayed by 2.8 weeks in triheptanoin treated hSOD1G93A mice when compared to untreated hSOD1G93A mice (p = 0.002, t-test). (D) Overall hind limb grip strength shown as the area under the curve over time was increased in triheptanoin treated hSOD1G93A mice compared to control treated hSOD1G93A mice (p = 0.02, t-test). (E) The onset of balance loss in triheptanoin treated hSOD1G93A mice was significantly delayed by 13 days (p = 0.0016, t-test). (F) Body weights over time were significantly different between triheptanoin treated vs. untreated wild-type mice and treated vs. untreated hSOD1G93A mice. (G) The onset of body weight loss in triheptanoin treated vs. untreated hSOD1G93A mice (n = 5) was delayed (p = 0.007, t-test). * p<0.05, ** p<0.01. The onset of body weight loss was defined as the day where a loss of more than 10% in an individual mouse occurred relative to the mean body weight from week 12 to 17 was observed. Also all subsequent three body weight measurements were ≤ 90% of the original mean weight. In the rotarod test, behavior of the hSOD1G93A mice varied widely. Many mice seemed “unmotivated” to perform on the rotating rod and no satisfying rotarod baseline performance was reached. Therefore we could only assess the time point when mice were unable to stay on the rod. Triheptanoin treatment delayed the time of onset of loss of balance on the rod by 13 days (p = 0.0016; Fig 3E). Body Weight Loss and Survival in hSOD1G93A Mice with and without Triheptanoin Body weight is another indicator of disease progression in hSOD1G93A mice. When compared to untreated wild-type mice (n = 12), wild-type mice on oral triheptanoin (n = 15) gained weight at a slower rate (p<0.0001, Fig 3F). At 14 weeks of age, triheptanoin treated wild-type mice were approximately 3g lighter than untreated wild-type mice (p<0.05). This weight difference increased to approximately 4g at 22 weeks of age. Both treated and untreated hSOD1G93A mice (n = 5–8) were lighter when compared to untreated wild-type mice (p<0.0001). There were no differences in weight between hSOD1G93A mice with or without triheptanoin treatment over the full time period. However, compared to control diet, triheptanoin fed hSOD1G93A mice showed a trend of reduced body weight gain from 7–16 weeks. After 20 weeks of age, the body weight of hSOD1G93A mice on the control and triheptanoin diet became similar (Fig 3F). The onset of body weight loss in hSOD1G93A mice was delayed by 1.6 weeks (11 days) in triheptanoin treated mice (Fig 3G, p = 0.007, unpaired two-tailed t-test). Based on power analysis and the guidelines for ALS research [73], this study of motor symptoms was too small to assess survival with adequate power. No differences were seen in the number of days taken to reach end-stage when comparing triheptanoin treated (n = 7) to untreated (n = 5) hSOD1G93A mice (174.9±3.5 vs. 172.4±3.9, p = 0.653), indicating that in this small study, triheptanoin treatment did not alter survival. Expression and Enzyme Activity Studies Given the effect of triheptanoin on motor function in hSOD1G93A mice, we next aimed to evaluate the extent to which TCA cycle metabolism or anaplerosis may be impaired in the gastrocnemius muscle of hSOD1G93A mice in the early and late stages of the disease. We used quantitative real time PCR to compare the expression of genes encoding for enzymes involved in glycolysis (glyceraldehyde-3-phosphate dehydrogenase—Gapdh), the TCA cycle (pyruvate–Pdha1, 2-oxoglutarate—Ogdh and succinate dehydrogenases—Sdha, Fig 4) and anaplerotic pathways of the muscle. The latter include pyruvate carboxylase (Pcx) producing oxaloacetate, glutamic pyruvic transaminase 1 and 2 (Gpt1 and 2, Fig 5), and the enzymes of the propionyl-CoA carboxylation pathway (Fig 6), namely the α and β subunits of propionyl-carboxylase (Pcca and Pccb) and methylmalonyl mutase (Mut), which together metabolize propionyl-CoA to succinyl-CoA. Triheptanoin treated mice were included in this analysis to investigate the effect of triheptanoin on the expression of these metabolic enzymes. 10.1371/journal.pone.0161816.g004Fig 4 Gapdh, Pdha, Ogdh and Sdha mRNA expression in gastrocnemius muscle after triheptanoin treatment. Starting at P35, female wild-type and hSOD1G93A mice were treated with triheptanoin (TRIH) or control (CON) diet until they were sacrificed. Quantitative real time PCR analysis of Gapdh, Pdha1, Ogdh and Sdha of the gastrocnemius muscle of 10 and 25 week old wild-type and hSOD1G93A mice untreated or treated with triheptanoin relative to house keeping genes. N-numbers of each group used in all graphs are indicated in the top bar graphs. The insets above each graph show the p-values for the effects of genotype in two-way ANOVAs, while the effect of diet was p>0.05 for each bar graph. When significant, the results of Bonferroni post tests are shown by stars (* p<0.05, ** p<0.01), showing that decreases of mRNA levels of several enzymes in hSOD1G93A mice were not apparent with triheptanoin treatment. 10.1371/journal.pone.0161816.g005Fig 5 Relative expression of the anaplerotic genes, pyruvate carboxylase Pcx and glutamic pyruvic transferases Gpt1 and 2. Starting at P35, female wild-type and hSOD1G93A mice were either treated with triheptanoin (TRIH) or control (CON) diet until they were sacrificed. Gene expresssion is compared in the gastrocnemius muscle of 10 and 25 week old triheptanoin treated vs. untreated wild-type and hSOD1G93A mice relative to housekeeping genes. N-numbers of each group used in all graphs are indicated in the top bar graphs. The insets above each graph show the p-values for the effects of genotype in two-way ANOVAs, while the effect of diet was p>0.05 for each bar graph. When significant, the results of Bonferroni post tests are shown by a star (* p<0.05), indicating that triheptanoin treatment prevented the decrease of expression in Gpt2 mRNA. 10.1371/journal.pone.0161816.g006Fig 6 Gene expression of enzymes involved in the propionyl-CoA carboxylase pathway. Starting at P35, female wild-type and hSOD1G93A mice were treated with triheptanoin (TRIH) or control (CON) diet until they were sacrificed. Relative expression of the α (Pcca) and β (Pccb) subunits of propionyl-CoA carboxylase and methylmalonyl mutase (Mut). Expresssion is compared in the gastrocnemius muscle of 10 and 25 week old wild-type and hSOD1G93A mice untreated or treated with triheptanoin relative to housekeeping genes. N-numbers used for each group throughout the experiments are indicated in the top bar graphs. The insets above each graph show the p-values for the effects of genotype in two-way ANOVAs, while the effect of diet was p>0.05 for each bar graph. When significant, the results of Bonferroni post tests are indicated by a star (* p<0.05), indicating that triheptanoin treatment in hSOD1G93A mice protected against lowered expression of Pccb and Mut mRNA. Investigating the effects of genotype and treatment on the expression of key metabolic enzymes, two way ANOVAs revealed that the genotype was linked to the variations observed in the expression of several enzymes relative to housekeeping genes at 10 and/or 25 weeks of age (p shown in insets of bar graphs, Figs 4–6, n = 4–6). When compared to wild-type mice, 10 week old hSOD1G93A mice showed a reduction in the expression of several dehydrogenases and methyl-malonyl mutase; namely the mRNA levels of the A1 subunit of pyruvate dehydrogenase (Pdha1) were reduced by 24%, 2-oxoglutarate dehydrogenase (Ogdh) by 30%, the subunit A of succinate dehydrogenase (Sdha) by 23%, and methyl-malonyl mutase (Mut) by 27.5% in hSOD1G93A mice (Figs 4 and 6; all p<0.05 in Bonferroni post-test, n = 4–6 mice per group). Triheptanoin attenuated the reduction in the expression of pyruvate and succinate dehydrogenases in hSOD1G93A mice. However, triheptanoin had no effect on the expression of 2-oxoglutarate dehydrogenase and methylmalonyl mutase. No alterations of mRNA levels were evident in the other investigated genes, including glycolytic Gapdh and the genes involved in anaplerosis, Pcx, Gpt1 and 2, Pcca and Pccb (Figs 4–6). At 25 weeks of age, the end-stage of disease [67,74], and relative to untreated wild-type mice, hSOD1G93A mice showed reduced gene expression for succinate dehydrogenase (Sdha, 70%), Gpt2 (84%) and the β subunit of propionyl carboxylase (Pccb, 64%) (all p<0.05 in post test, Figs 4–6). Triheptanoin prevented the reduction in the expression of these genes, indicating that it can preserve muscle energy metabolism. To investigate the extent to which these mRNA changes result in functional changes, we measured the maximal enzyme activity of 2-oxoglutarate dehydrogenase in extracts of gastrocnemius muscle in hSOD1G93A compared to wild-type mice at different stages of disease (n = 4–6 mice each group). The two-way ANOVA analysis revealed that the maximal activity of this enzyme was significantly altered dependent on genotype (p<0.028) and disease stage (p<0.0007). At mid-stage (P110-130), there was a trend of a reduction by 25%, while the 45% loss of maximal 2-oxoglutarate dehydrogenase activity at end-stage (P150-175) was statistically significant by a Bonferroni multiple comparison post test (p<0.05, Fig 7). 10.1371/journal.pone.0161816.g007Fig 7 Lower maximal activities of 2-oxoglutarate dehydrogenase (OGDH) in hSOD1G93A gastrocnemius muscle. The maximal activities of OGDH in extracts from gastrocnemius muscle of male wild-type and hSOD1G93A mice at different disease stages are compared. Stages are defined as presymptomatic (days 35–36), onset (days 63–75), mid-stage (days 110–130) and end-stage (days 150–175). The inset above the graph shows the two-way ANOVA p-values for the effects of genotype and disease stage, indicating that 2-oxoglutarate activity in hSOD1G93A mice declines with progression of disease. The star denotes significance in the Bonferroni post test (* p<0.05). Triheptanoin slightly increased plasma ketone (β-hydroxybutyrate) levels Triheptanoin treated hSOD1G93A mice showed a 56% increase in levels of plasma β-hydroxybutyrate compared to those treated with control diet (One-way ANOVA p = 0.042; p<0.05 Fisher’s LSD post test, Fig 8). Similarly, triheptanoin also increased plasma β-hydroxybutyrate levels in wild-type mice by 96% (p = 0.0008, p<0.001) (Fig 8). However, in both wild-type and hSOD1G93A mice following triheptanoin treatment, no differences were seen in plasma glucose levels, with average levels between 213–219 mg/dl (One-way ANOVA, p = 0.99). 10.1371/journal.pone.0161816.g008Fig 8 Triheptanoin treatment increased levels of plasma β-hydroxybutyrate. Plasma β-hydroxybutyrate levels (mM) in 70 days old wild-type and hSOD1G93A mice fed with either triheptanoin (TRIH) or control (CON) diet from day 35 to 70 (One-way ANOVA p = 0.042; Fisher’s LSD post test: *p<0.05, ***p<0.001, n = 10–12). Discussion The most important findings of this study are that 1) triheptanoin attenuated motor neuron loss in hSOD1G93A mice, 2) triheptanoin delayed the onset of motor symptoms in hSOD1G93A mice, and 3) decreased expression of muscle enzyme mRNA involved in TCA cycling in hSOD1G93A mice was attenuated by triheptanoin. The hypothesized mechanisms by which triheptanoin protects motor neurons and delays muscle wasting is summarized in Fig 9. 10.1371/journal.pone.0161816.g009Fig 9 Hypothesized mechanisms of triheptanoin. Triheptanoin is the triglyceride of heptanoate, which is metabolized to acetyl-CoA as well as propionyl-CoA providing alternative and anaplerotic fuel. Following carboxylation propionyl-CoA produces succinyl-CoA (anaplerosis), which via metabolism to oxaloacetate can increase ATP production and aid in further acetyl-CoA oxidation. Thus triheptanoin can improve mitochondrial energy production and thereby protect neurons and muscle against degeneration. Neuroprotective Effects of Triheptanoin In our study, triheptanoin treated hSOD1G93A mice showed 33% less motor neuron loss in the L4-L5 spinal cord. Similarly, in two other neurological disorders models the same dose of triheptanoin was found to protect cells. In nur7 mutant mice, a model for Canavan disease, triheptanoin treatment prevented the loss of oligodendrocytes [65]. In a mouse middle cerebral artery occlusion (MCAO) model of stroke, triheptanoin pre-treatment reduced the infarct area and mitochondrial function was preserved in mitochondria isolated from brains 1 h after onset of MCAO [64]. Based on the previous findings that triheptanoin diminished abnormalities in brain energy metabolism in different models and patients with CNS disorders [55,56,59,60] it is likely that improvements in ATP production contribute to triheptanoin’s neuroprotective effects. In addition, in an Alzheimer’s Disease model, triheptanoin in the context of a ketogenic diet increased the expression of the mRNA levels of Sirt1, Pparg, Sod1 and Sod2 [75]. Sirtuin 1 and PPAR-γ are regulators of lipid and glucose metabolism as well as mitochondrial respiration and oxidative stress. Thus, these studies indicate that the neuroprotective effects of triheptanoin may be mediated, in part, by limiting oxidative stress and by preserving mitochondrial function, which together with triheptanoin’s direct provision of alternative and anaplerotic fuel will contribute to improved energy supply. Clinical Importance of Triheptanoin’s Protective Effects on ALS Symptoms In addition to its neuroprotective effects, triheptanoin delayed the onset of the loss of grip strength by 2.8 weeks in hSOD1G93A mice. This corresponds to 11% of the life span and 30% of the “symptomatic” time in these mice. Thus, our data show an improvement in disease symptoms when treatment was initiated prior to the onset of disease symptoms in hSOD1G93A mice. This raises hope that in ALS patients, triheptanoin may be able to preserve motor neurons and the function of muscles when treatment is initiated at an early stage of disease. On the other hand, presymptomatic hSOD1G93A mice already show loss of crural flexor motor neurons [67]. This indicates that in ALS mouse models, pathological neuron loss does not necessarily translate to discernable motor deficits in mice, while motor deficits in patients may be present at a similar pathological stage. Thus, while triheptanoin was given at a presymptomatic stage of disease in mice, it is difficult to correlate our findings to the clinical course of disease in patients. Given the difficulty in diagnosing ALS, future studies are required to evaluate the extent to which triheptanoin can preserve motor neurons and motor function when treatment is initiated in ALS models with obvious muscle and/or neuronal impairments. As a medium chain triglyceride, triheptanoin quickly provides heptanoic acid and C5 ketones to the blood. These molecules cross the blood brain barrier and are expected to also be metabolized by muscle. Therefore, it is anticipated that the protective effects of triheptanoin will begin quickly after the initiation of treatment. Because triheptanoin was safe in pilot studies with patients with disorders involving severe metabolic deficiencies [54,76,77], CNS disorders [55,56,76], Pompe’s Disease and Inclusion Body Myositis [57], there are relatively few hurdles to overcome to start a phase I clinical trial to assess safety and tolerability for patients with ALS. It will be important to rule out potential and known metabolic contraindications in study subjects and to provide professional dietary advice in addition to triheptanoin treatment. To date, no central or cardiovascular side effects have been reported for triheptanoin. For medium chain fats, the most common side effects consist of abdominal pain, diarrhea and nausea, which can largely be controlled by mixing the treatment with food and slowly increasing the dose. Given previous reports of gastric dysfunction in ALS patients [78,79], it is also pertinent that the impact of triheptanoin on gut function in ALS be investigated. As a means to circumvent potential side effects of oral triheptanoin on gut function, triheptanoin could be delivered intravenously in an emulsion. This route of delivery would circumvent the known side effects on the GI tract and supply heptanoate quickly to various tissues before it reaches the liver. The liver can convert heptanoate to glucose via gluconeogenesis and thus reduces the amounts of heptanoate and C5 ketones that can reach other tissues. Such a triheptanoin emulsion is currently being developed by industry. Triheptanoin is an ideal precursor to supply heptanoic acid and C5 ketones to the body, although these metabolites should also be directly effective for patients. However, large amounts of heptanoic acid and C5 ketones or their sodium salts are unsuitable for direct application, as they would add unphysiological levels of acid or salt to the body. It is now thought that the majority of clinical trials addressing the efficacy of new treatments in ALS have failed due to the lack of solid preclinical data. Consequently, new guidelines have been developed for laboratory work before phase II trials [73], including power analyses and high animal numbers when claiming effects on survival. While our power analyses regarding loss of grip strength and balance show that our study used adequate numbers of mice regarding these analyses of motor symptoms, a larger and more complete preclinical study of triheptanoin following the mentioned guidelines would be ideal to test its effect on survival in multiple animal models of ALS. Also, treatment initiation during the symptomatic stage and assessments of muscle function and the integrity of neuromuscular junctions will be important to inform and motivate clinical trials. Metabolic Changes in hSOD1G93A Mouse Muscle When compared to healthy wild-type mice, we found that the expression of several enzyme mRNA involved in glycolysis, the TCA cycle and anaplerosis were significantly reduced in the gastrocnemius muscle of hSOD1G93A mice at 10 weeks of age (Figs 4–6), a time when hind limb grip strength is still normal (Fig 3B). We also demonstrate that the maximal 2-oxoglutarate dehydrogenase activity was significantly reduced based on genotype and disease stage in gastrocnemius muscle in hSOD1G93A compared to wild-type mice (Fig 7), indicating that the observed mRNA changes result in functional alterations. Taken together this suggests that ATP production is insufficient in the muscle for the function and maintenance of tissue at an early stage of the disease. Similar changes in mRNA levels were seen at the end-stage of disease at 25 weeks of age (Figs 4–6), highlighting that an impairment of TCA cycling may persist throughout the progression of disease. Numerous studies have alluded to altered energy metabolism during the early and later stages of disease progression in muscle in ALS patients and in ALS mouse models [25,29–33,35]. In the majority of sporadic ALS patients, respiratory chain defects and multiple deletions in mitochondrial DNA have been observed in muscle biopsies [24], suggesting that metabolic abnormalities occur in muscle and that muscle pathology may contribute to ALS. Moreover, citrate synthase activity was decreased by 43% in muscle from ALS patients when compared to healthy controls [32]. Consistent with this, ALS patients displayed increased blood lactate levels, especially after exercise, indicating that acetyl-CoA is not metabolized well by the TCA cycle [80]. Furthermore, SOD1 mouse models of ALS showed reduced levels of ATP in skeletal muscle [35,36]. Moreover, oxidative stress, which is commonly observed in ALS (reviewed in [12,81]), is well known to diminish the activities of CNS and muscle aconitase and 2-oxoglutarate dehydrogenase, both enzymes of the TCA cycle [82,83]. At 25 weeks, we found reduced levels of the main anaplerotic enzyme mRNA of muscle, glutamic pyruvic transaminase 2, suggesting diminished levels of TCA cycle intermediates. Triheptanoin prevented the downregulation of several enzyme mRNA observed in hSOD1G93A when compared to wild-type mice. This finding implies that normalization of energy metabolism by triheptanoin prevents downregulation of certain metabolism genes, which in turn would help to maintain a healthy metabolism to optimize energy supply, function and survival of tissue. When considering these collective metabolic alterations, it is not surprising that triheptanoin, which is currently the most effective treatment to produce anaplerotic propionyl-CoA slowed the onset of motor symptoms in hSOD1G93A mice. Triheptanoin and Body Weight Loss of body weight is a concern in many ALS patients and is an indicator for ALS progression during the later stages of disease in the hSOD1G93A mouse. In this study, triheptanoin slowed body weight increases in the wild-type C57BL/6 mice. This effect has been previously mentioned for medium chain triglycerides, which in contrast to long chain fats, do not accumulate in adipose tissue (reviewed by [84]). However, in patients with muscular and neurological disorders, treatment with triheptanoin has not been reported to induce a loss in body weight. Rather increases in weight were seen in some patients [57]. Also, mice from two outbred strains gained weight equally with and without triheptanoin treatment [63,66]. In spite of a trend to lower body weight in this study, mice on triheptanoin showed similar hind limb grip strength when compared to untreated mice, indicating that lower body weight does not impair muscle function. In addition, despite this potential confounding factor, triheptanoin treated mice showed a delay in the loss of body weight and similar survival when compared to untreated mice. If triheptanoin treatment is tested in ALS models on background strains that do not exhibit a loss in body weight, it is likely that the disease modifying effects of triheptanoin will be further enhanced, potentially improving survival. Comparison to Other Metabolic Treatment Approaches for ALS Other metabolic treatment approaches have been tested in hSOD1G93A mice. Ketogenic therapy, which increases blood C4 ketone levels is commonly thought to improve the production of ATP by mitochondria [85]. When tested in hSOD1G93A mice, a ketogenic diet preserved motor performance and delayed weight loss and loss of motor neurons in the lumbar spinal cord [86]. Similarly, trioctanoin (caprylic triglyceride, the triglyceride of octanoate), which also increases blood C4 ketone levels, increased oxygen consumption rates in mitochondria of the spinal cord, delayed onset of motor symptoms and protected motor neurons [87]. However, both treatments did not extend survival in hSOD1G93A mice and ketogenic therapy is contraindicated in ALS as it typically leads to weight loss in adults. Our recent study indicated that chronic trioctanoin, but not triheptanoin, treatment can decrease glycolysis and the levels of TCA cycle intermediates in mouse brain tissue [88]. Thus, trioctanoin treatment alone might interfere with the generation of energy from glucose, which may worsen metabolic problems in ALS (reviewed in [89]). In the same study triheptanoin feeding to CD1 mice increased brain β-hydroxybutyrate levels 1.7-fold, which may contribute to its beneficial effects. Similarly, in the current study plasma levels of the C4 ketone β-hydroxybutyrate were increased with triheptanoin treatment in hSOD1G93A mice by 1.6-fold (Fig 8). In addition, triheptanoin is expected to be superior to any fuel that produces C4 ketones, because triheptanoin is anaplerotic. Dichloroacetate, a metabolic treatment that increases pyruvate dehydrogenase activity, has been found to improve survival time and/or delay the loss of motor function in SOD1 mice [49,90]. An increase in pyruvate dehydrogenase kinase activity is expected to allow more efficient production of energy from glucose, if there is sufficient TCA cycling. However, studies are needed to determine the impact of dichloroacetate on TCA cycle metabolism in order to allow a better understanding of this potential treatment approach in ALS. There is evidence that some ALS patients benefitted from the “Deanna Protocol”, a complex combination of many metabolic supplements, in addition to medium chain triglyceride-rich coconut oil [91]. More recently, [92] showed the beneficial effects of a simplified “Deanna Protocol” in hSOD1G93A mice. Specifically, a combination of arginine α-ketoglutarate, β-phenyl-γ-amino butyrate, ubiquinone and medium even chain triglyceride oil, which is largely a mixture of triglycerides of octanoate and decanoate, extended survival. Providing that α-ketoglutarate or a C5 or C4 metabolite of this compound enters mitochondria, the TCA cycle metabolite pool could be (re-)filled and this treatment can be considered as being “anaplerotic” similar to heptanoate. This would aid the metabolism of the co-supplied alternative medium chain fatty acids, octanoate and decanoate, which taken together closely resembles the proposed mechanism of action of triheptanoin. Despite current reports of beneficial outcomes in response to dietary intervention strategies in mouse models of ALS, the extent to which dichloroacetate, the Deanna protocol and triheptanoin are efficacious in controlled clinical trials remains to be investigated. Conclusion This study revealed that triheptanoin is a promising new treatment approach for ALS to delay motor neuron loss and the onset of motor symptoms. Further studies with increased animal numbers and treatment initiation in symptomatic stages, and those that use 13C-labeled glucose or heptanoate are necessary to elucidate triheptanoin’s metabolic fate and mechanism of action in ALS. Such studies will be important to inform and optimize future clinical trials. The authors wish to thank Paula and Kenneth Anderson for help with experiments and the Queensland Brain Institute for providing animals. ==== Refs References 1 Baumer D , Talbot K , Turner MR (2014 ) Advances in motor neurone disease . Journal of the Royal Society of Medicine 107 : 14 –21 . 10.1177/0141076813511451 24399773 2 Cleveland DW , Rothstein JD (2001 ) From Charcot to Lou Gehrig: deciphering selective motor neuron death in ALS . Nature reviews Neuroscience 2 : 806 –819 . 11715057 3 Chen S , Sayana P , Zhang X , Le W (2013 ) Genetics of amyotrophic lateral sclerosis: an update . Molecular neurodegeneration 8 : 28 10.1186/1750-1326-8-28 23941283 4 Rosen DR , Siddique T , Patterson D , Figlewicz DA , Sapp P , Hentati A , et al (1993 ) Mutations in Cu/Zn superoxide dismutase gene are associated with familial amyotrophic lateral sclerosis . Nature 362 : 59 –62 . 8446170 5 Kabashi E , Valdmanis PN , Dion P , Spiegelman D , McConkey BJ , Vande Velde C , et al (2008 ) TARDBP mutations in individuals with sporadic and familial amyotrophic lateral sclerosis . Nature genetics 40 : 572 –574 . 10.1038/ng.132 18372902 6 Kwiatkowski TJ Jr., Bosco DA , Leclerc AL , Tamrazian , Vanderburg CR , Russ C , et al (2009 ) Mutations in the FUS/TLS gene on chromosome 16 cause familial amyotrophic lateral sclerosis . Science 323 : 1205 –1208 . 10.1126/science.1166066 19251627 7 Deng HX , Chen W , Hong ST , Boycott KM , Gorrie GH , Siddique N , et al (2011 ) Mutations in UBQLN2 cause dominant X-linked juvenile and adult-onset ALS and ALS/dementia . Nature 477 : 211 –215 . 10.1038/nature10353 21857683 8 DeJesus-Hernandez M , Mackenzie IR , Boeve BF , Boxer AL , Baker M , Rutherford NJ , et al (2011 ) Expanded GGGGCC hexanucleotide repeat in noncoding region of C9ORF72 causes chromosome 9p-linked FTD and ALS . Neuron 72 : 245 –256 . 10.1016/j.neuron.2011.09.011 21944778 9 Renton AE , Majounie E , Waite A , Simon-Sanchez J , Rollinson S , Gibbs JR , et al (2011 ) A hexanucleotide repeat expansion in C9ORF72 is the cause of chromosome 9p21-linked ALS-FTD . Neuron 72 : 257 –268 . 10.1016/j.neuron.2011.09.010 21944779 10 Kiernan MC , Vucic S , Cheah BC , Turner MR , Eisen A , Hardiman O , et al (2011 ) Amyotrophic lateral sclerosis . Lancet 377 : 942 –955 . 10.1016/S0140-6736(10)61156-7 21296405 11 Milani P , Gagliardi S , Cova E , Cereda C (2011 ) SOD1 Transcriptional and Posttranscriptional Regulation and Its Potential Implications in ALS . Neurol Res Int 2011 : 458427 10.1155/2011/458427 21603028 12 Cozzolino M , Carri MT (2012 ) Mitochondrial dysfunction in ALS . Prog Neurobiol 97 : 54 –66 . 10.1016/j.pneurobio.2011.06.003 21827820 13 Turner BJ , Talbot K (2008 ) Transgenics, toxicity and therapeutics in rodent models of mutant SOD1-mediated familial ALS . Prog Neurobiol 85 : 94 –134 . 10.1016/j.pneurobio.2008.01.001 18282652 14 Vucic S , Kiernan MC (2009 ) Pathophysiology of neurodegeneration in familial amyotrophic lateral sclerosis . Curr Mol Med 9 : 255 –272 . 19355908 15 Shi P , Wei Y , Zhang J , Gal J , Zhu H (2010 ) Mitochondrial dysfunction is a converging point of multiple pathological pathways in amyotrophic lateral sclerosis . J Alzheimers Dis 20 Suppl 2 : S311 –324 . 10.3233/JAD-2010-100366 20463400 16 Ferraiuolo L , Kirby J , Grierson AJ , Sendtner M , Shaw PJ (2011 ) Molecular pathways of motor neuron injury in amyotrophic lateral sclerosis . Nature reviews Neurology 7 : 616 –630 . 10.1038/nrneurol.2011.152 22051914 17 Barber SC , Mead RJ , Shaw PJ (2006 ) Oxidative stress in ALS: a mechanism of neurodegeneration and a therapeutic target . Biochimica et biophysica acta 1762 : 1051 –1067 . 16713195 18 Rothstein JD , Tsai G , Kuncl RW , Clawson L , Cornblath DR , Drachman DB , et al (1990 ) Abnormal excitatory amino acid metabolism in amyotrophic lateral sclerosis . Annals of neurology 28 : 18 –25 . 2375630 19 McGeer PL , McGeer EG (2002 ) Inflammatory processes in amyotrophic lateral sclerosis . Muscle & nerve 26 : 459 –470 .12362410 20 Bruijn LI , Houseweart MK , Kato S , Anderson KL , Anderson SD , Ohama E , et al (1998 ) Aggregation and motor neuron toxicity of an ALS-linked SOD1 mutant independent from wild-type SOD1 . Science 281 : 1851 –1854 . 9743498 21 Julien JP , Beaulieu JM (2000 ) Cytoskeletal abnormalities in amyotrophic lateral sclerosis: beneficial or detrimental effects? Journal of the neurological sciences 180 : 7 –14 . 11090858 22 Dupuis L , Oudart H , Rene F , Gonzalez de Aguilar JL , Loeffler JP (2004 ) Evidence for defective energy homeostasis in amyotrophic lateral sclerosis: benefit of a high-energy diet in a transgenic mouse model . Proceedings of the National Academy of Sciences of the United States of America 101 : 11159 –11164 . 15263088 23 Dupuis L , Pradat PF , Ludolph AC , Loeffler JP (2011 ) Energy metabolism in amyotrophic lateral sclerosis . The Lancet Neurology 10 : 75 –82 . 10.1016/S1474-4422(10)70224-6 21035400 24 Vielhaber S , Kunz D , Winkler K , Wiedemann FR , Kirches E , Feistner H , et al (2000 ) Mitochondrial DNA abnormalities in skeletal muscle of patients with sporadic amyotrophic lateral sclerosis . Brain 123 (Pt 7 ): 1339 –1348 . 10869047 25 Crugnola V , Lamperti C , Lucchini V , Ronchi D , Peverelli L , Prelle A , et al (2010 ) Mitochondrial respiratory chain dysfunction in muscle from patients with amyotrophic lateral sclerosis . Archives of neurology 67 : 849 –854 . 10.1001/archneurol.2010.128 20625092 26 Kong J , Xu Z (1998 ) Massive mitochondrial degeneration in motor neurons triggers the onset of amyotrophic lateral sclerosis in mice expressing a mutant SOD1 . Journal of Neuroscience 18 : 3241 –3250 . 9547233 27 Sasaki S , Iwata M (1996 ) Ultrastructural study of synapses in the anterior horn neurons of patients with amyotrophic lateral sclerosis . Neuroscience letters 204 : 53 –56 . 8929976 28 Wong PC , Pardo CA , Borchelt DR , Lee MK , Copeland NG , Jenkins NA , et al (1995 ) An adverse property of a familial ALS-linked SOD1 mutation causes motor neuron disease characterized by vacuolar degeneration of mitochondria . Neuron 14 : 1105 –1116 . 7605627 29 Vielhaber S , Winkler K , Kirches E , Kunz D , Buchner M , Feistner H , et al (1999 ) Visualization of defective mitochondrial function in skeletal muscle fibers of patients with sporadic amyotrophic lateral sclerosis . Journal of the neurological sciences 169 : 133 –139 . 10540022 30 Wiedemann FR , Winkler K , Kuznetsov AV , Bartels C , Vielhaber S , Feistner H , et al (1998 ) Impairment of mitochondrial function in skeletal muscle of patients with amyotrophic lateral sclerosis . Journal of the neurological sciences 156 : 65 –72 . 9559989 31 Luo G , Yi J , Ma C , Xiao Y , Yi F , Yu T , et al (2013 ) Defective mitochondrial dynamics is an early event in skeletal muscle of an amyotrophic lateral sclerosis mouse model . PloS one 8 : e82112 10.1371/journal.pone.0082112 24324755 32 Russell AP , Wada S , Vergani L , Hock MB , Lamon S , Leger B , et al (2013 ) Disruption of skeletal muscle mitochondrial network genes and miRNAs in amyotrophic lateral sclerosis . Neurobiology of disease 49 : 107 –117 . 10.1016/j.nbd.2012.08.015 22975021 33 Echaniz-Laguna A , Zoll J , Ponsot E , N'Guessan B , Tranchant C , Loeffler JP , et al (2006 ) Muscular mitochondrial function in amyotrophic lateral sclerosis is progressively altered as the disease develops: a temporal study in man . Experimental neurology 198 : 25 –30 . 16126198 34 Mattiazzi M , D'Aurelio M , Gajewski CD , Martushova K , Kiaei M , Beal MF , et al (2002 ) Mutated human SOD1 causes dysfunction of oxidative phosphorylation in mitochondria of transgenic mice . The Journal of biological chemistry 277 : 29626 –29633 . 12050154 35 Dupuis L , di Scala F , Rene F , de Tapia M , Oudart H , Pradat PF , et al (2003 ) Up-regulation of mitochondrial uncoupling protein 3 reveals an early muscular metabolic defect in amyotrophic lateral sclerosis . FASEB J 17 : 2091 –2093 . 14500553 36 Derave W , Van Den Bosch L , Lemmens G , Eijnde BO , Robberecht W , Hespel P (2003 ) Skeletal muscle properties in a transgenic mouse model for amyotrophic lateral sclerosis: effects of creatine treatment . Neurobiology of disease 13 : 264 –272 . 12901841 37 Browne SE , Bowling AC , Baik MJ , Gurney M , Brown RH Jr, Beal MF (1998 ) Metabolic dysfunction in familial, but not sporadic, amyotrophic lateral sclerosis . Journal of neurochemistry 71 : 281 –287 . 9648876 38 Dalakas MC , Hatazawa J , Brooks RA , Di Chiro G (1987 ) Lowered cerebral glucose utilization in amyotrophic lateral sclerosis . Annals of neurology 22 : 580 –586 . 3501273 39 Hatazawa J , Brooks RA , Dalakas MC , Mansi L , Di Chiro G (1988 ) Cortical motor-sensory hypometabolism in amyotrophic lateral sclerosis: a PET study . Journal of computer assisted tomography 12 : 630 –636 . 3260610 40 Ludolph AC , Langen KJ , Regard M , Herzog H , Kemper B , Kuwert T , et al (1992 ) Frontal lobe function in amyotrophic lateral sclerosis: a neuropsychologic and positron emission tomography study . Acta neurologica Scandinavica 85 : 81 –89 . 1574993 41 Cistaro A , Valentini MC , Chio A , Nobili F , Calvo A , Moglia C , et al (2012 ) Brain hypermetabolism in amyotrophic lateral sclerosis: a FDG PET study in ALS of spinal and bulbar onset . European journal of nuclear medicine and molecular imaging 39 : 251 –259 . 10.1007/s00259-011-1979-6 22089661 42 Pagani M , Chio A , Valentini MC , Oberg J , Nobili F , Calvo A , et al (2014 ) Functional pattern of brain FDG-PET in amyotrophic lateral sclerosis . Neurology 83 : 1067 –1074 . 10.1212/WNL.0000000000000792 25122207 43 Miyazaki K , Masamoto K , Morimoto N , Kurata T , Mimoto T , Obata T , et al (2012 ) Early and progressive impairment of spinal blood flow-glucose metabolism coupling in motor neuron degeneration of ALS model mice. Journal of cerebral blood flow and metabolism: official journal of the International Society of Cerebral Blood Flow and Metabolism 32 : 456 –467 . 44 Smittkamp SE , Morris JK , Bomhoff GL , Chertoff ME , Geiger PC , Stanford JA (2014 ) SOD1-G93A mice exhibit muscle-fiber-type-specific decreases in glucose uptake in the absence of whole-body changes in metabolism . Neuro-degenerative diseases 13 : 29 –37 . 10.1159/000351606 24021858 45 Raman R , Allen SP , Goodall EF , Kramer S , Ponger LL , Heath PR , et al (2015 ) Gene expression signatures in motor neurone disease fibroblasts reveal dysregulation of metabolism, hypoxia-response and RNA processing functions . Neuropathology and applied neurobiology 41 : 201 –226 . 10.1111/nan.12147 24750211 46 Lederer CW , Torrisi A , Pantelidou M , Santama N , Cavallaro S (2007 ) Pathways and genes differentially expressed in the motor cortex of patients with sporadic amyotrophic lateral sclerosis . BMC genomics 8 : 26 17244347 47 Wang XS , Simmons Z , Liu W , Boyer PJ , Connor JR (2006 ) Differential expression of genes in amyotrophic lateral sclerosis revealed by profiling the post mortem cortex . Amyotrophic lateral sclerosis: official publication of the World Federation of Neurology Research Group on Motor Neuron Diseases 7 : 201 –210 . 48 Saris CG , Groen EJ , van Vught PW , van Es MA , Blauw HM , Veldink JH , et al (2013 ) Gene expression profile of SOD1-G93A mouse spinal cord, blood and muscle . Amyotrophic lateral sclerosis & frontotemporal degeneration 14 : 190 –198 .23298163 49 Palamiuc L , Schlagowski A , Ngo ST , Vernay A , Dirrig-Grosch S , Henriques A , et al (2015 ) A metabolic switch toward lipid use in glycolytic muscle is an early pathologic event in a mouse model of amyotrophic lateral sclerosis . EMBO Mol Med 7 : 526 –546 . 10.15252/emmm.201404433 25820275 50 Choi JK , Kustermann E , Dedeoglu A , Jenkins BG (2009 ) Magnetic resonance spectroscopy of regional brain metabolite markers in FALS mice and the effects of dietary creatine supplementation . The European journal of neuroscience 30 : 2143 –2150 . 10.1111/j.1460-9568.2009.07015.x 19930399 51 Niessen HG , Debska-Vielhaber G , Sander K , Angenstein F , Ludolph AC , Hilfert L , et al (2007 ) Metabolic progression markers of neurodegeneration in the transgenic G93A-SOD1 mouse model of amyotrophic lateral sclerosis . Eur J Neurosci 25 : 1669 –1677 . 17432958 52 D'Alessandro G , Calcagno E , Tartari S , Rizzardini M , Invernizzi RW , Cantoni L (2011 ) Glutamate and glutathione interplay in a motor neuronal model of amyotrophic lateral sclerosis reveals altered energy metabolism . Neurobiol Dis 43 : 346 –355 . 10.1016/j.nbd.2011.04.003 21530659 53 Brunengraber H , Roe CR (2006 ) Anaplerotic molecules: Current and future . J Inherit Metab Dis 29 : 327 –331 . 16763895 54 Roe CR , Mochel F (2006 ) Anaplerotic diet therapy in inherited metabolic disease: Therapeutic potential . J Inherit Metab Dis 29 : 332 –340 . 16763896 55 Adanyeguh IM , Rinaldi D , Henry PG , Caillet S , Valabregue R , Durr A , et al (2015 ) Triheptanoin improves brain energy metabolism in patients with Huntington disease . Neurology 84 : 490 –495 . 10.1212/WNL.0000000000001214 25568297 56 Pascual JM , Liu P , Mao D , Kelly DI , Hernandez A , Sheng M , et al (2014 ) Triheptanoin for glucose transporter type I deficiency (G1D): modulation of human ictogenesis, cerebral metabolic rate, and cognitive indices by a food supplement . JAMA Neurol 71 : 1255 –1265 . 10.1001/jamaneurol.2014.1584 25110966 57 Corbett A , Garg N , Crosbie A , Borges K (2015 ) A Pilot Study of Triheptanoin Treatment in Sporadic Inclusion Body Myositis (sIBM) . Annals of Neurology 78 : S102 . 58 Gu L , Zhang GF , Kombu RS , Allen F , Kutz G , Brewer WU , et al (2010 ) Parenteral and enteral metabolism of anaplerotic triheptanoin in normal rats. II. Effects on lipolysis, glucose production, and liver acyl-CoA profile . Am J Physiol Endocrinol Metab 298 : E362 –371 . 10.1152/ajpendo.00384.2009 19903863 59 Marin-Valencia I , Good LB , Ma Q , Malloy CR , Pascual JM (2013 ) Heptanoate as a neural fuel: energetic and neurotransmitter precursors in normal and glucose transporter I-deficient (G1D) brain . J Cereb Blood Flow Metab 33 : 175 –182 . 10.1038/jcbfm.2012.151 23072752 60 Hadera MG , Smeland OB , McDonald TS , Tan KN , Sonnewald U , Borges K (2014 ) Triheptanoin partially restores levels of tricarboxylic acid cycle intermediates in the mouse pilocarpine model of epilepsy . J Neurochem 129 : 107 –119 . 10.1111/jnc.12610 24236946 61 Borges K , Sonnewald U (2012 ) Triheptanoin—a medium chain triglyceride with odd chain fatty acids: a new anaplerotic anticonvulsant treatment? Epilepsy Res 100 : 239 –244 . 10.1016/j.eplepsyres.2011.05.023 21855298 62 Mochel F , Duteil S , Marelli C , Jauffret C , Barles A , Holm J , et al (2010 ) Dietary anaplerotic therapy improves peripheral tissue energy metabolism in patients with Huntington's disease . Eur J Hum Genet 18 : 1057 –1060 . 10.1038/ejhg.2010.72 20512158 63 Willis S , Stoll J , Sweetman L , Borges K (2010 ) Anticonvulsant effects of a triheptanoin diet in two mouse chronic seizure models . Neurobiol Dis 40 : 565 –572 . 10.1016/j.nbd.2010.07.017 20691264 64 Schwarzkopf TM , Koch K , Klein J (2015 ) Reduced severity of ischemic stroke and improvement of mitochondrial function after dietary treatment with the anaplerotic substance triheptanoin . Neuroscience 300 : 201 –209 . 10.1016/j.neuroscience.2015.05.014 25982559 65 Francis JS , Markov V , Leone P (2014 ) Dietary triheptanoin rescues oligodendrocyte loss, dysmyelination and motor function in the nur7 mouse model of Canavan disease . J Inherit Metab Dis 37 : 369 –381 . 10.1007/s10545-013-9663-6 24288037 66 Thomas NK , Willis S , Sweetman L , Borges K (2012 ) Triheptanoin in acute mouse seizure models . Epilepsy Res 99 : 312 –317 . 10.1016/j.eplepsyres.2011.12.013 22260920 67 Ngo ST , Baumann F , Ridall PG , Pettitt AN , Henderson RD , Bellingham MC , et al (2012 ) The relationship between Bayesian motor unit number estimation and histological measurements of motor neurons in wild-type and SOD1(G93A) mice . Clin Neurophysiol 123 : 2080 –2091 . 10.1016/j.clinph.2012.01.028 22521362 68 Coggeshall RE , Lekan HA (1996 ) Methods for determining numbers of cells and synapses: a case for more uniform standards of review . J Comp Neurol 364 : 6 –15 . 8789272 69 Gill A , Kidd J , Vieira F , Thompson K , Perrin S (2009 ) No benefit from chronic lithium dosing in a sibling-matched, gender balanced, investigator-blinded trial using a standard mouse model of familial ALS . PLoS One 4 : e6489 10.1371/journal.pone.0006489 19649300 70 Lee JD , Kamaruzaman NA , Fung JN , Taylor SM , Turner BJ , Atkin JD , et al (2013 ) Dysregulation of the complement cascade in the hSOD1G93A transgenic mouse model of amyotrophic lateral sclerosis . J Neuroinflammation 10 : 119 10.1186/1742-2094-10-119 24067070 71 Vandesompele J , De Preter K , Pattyn F , Poppe B , Van Roy N , De Paepe A , et al (2002 ) Accurate normalization of real-time quantitative RT-PCR data by geometric averaging of multiple internal control genes . Genome Biol 3 : RESEARCH0034 12184808 72 Lai JCK , Cooper AJL (1986 ) Brain α-Ketoglutarate Dehydrogenase Complex: Kinetic Properties, Regional Distribution, and Effects of Inhibitors . J Neurochem 47 : 1376 –1386 . 3760866 73 Ludolph AC , Bendotti C , Blaugrund E , Chio A , Greensmith L , Loeffler JP , et al (2010 ) Guidelines for preclinical animal research in ALS/MND: A consensus meeting . Amyotroph Lateral Scler 11 : 38 –45 . 10.3109/17482960903545334 20184514 74 Steyn FJ , Ngo ST , Lee JD , Leong JW , Buckley AJ , Veldhuis JD , et al (2012 ) Impairments to the GH-IGF-I axis in hSOD1G93A mice give insight into possible mechanisms of GH dysregulation in patients with amyotrophic lateral sclerosis . Endocrinology 153 : 3735 –3746 . 10.1210/en.2011-2171 22621959 75 Aso E , Semakova J , Joda L , Semak V , Halbaut L , Calpena A , et al (2013 ) Triheptanoin supplementation to ketogenic diet curbs cognitive impairment in APP/PS1 mice used as a model of familial Alzheimer's disease . Curr Alzheimer Res 10 : 290 –297 . 23131121 76 Roe CR , Bottiglieri T , Wallace M , Arning E , Martin A (2010 ) Adult Polyglucosan Body Disease (APBD): Anaplerotic diet therapy (Triheptanoin) and demonstration of defective methylation pathways . Mol Genet Metab 101 : 246 –252 . 10.1016/j.ymgme.2010.06.017 20655781 77 Roe CR , Yang BZ , Brunengraber H , Roe DS , Wallace M , Garritson BK (2008 ) Carnitine palmitoyltransferase II deficiency: successful anaplerotic diet therapy . Neurology 71 : 260 –264 . 10.1212/01.wnl.0000318283.42961.e9 18645163 78 Camilleri M , Balm RK , Low PA (1993 ) Autonomic dysfunction in patients with chronic intestinal pseudo-obstruction . Clin Auton Res 3 : 95 –100 . 8324379 79 Toepfer M , Folwaczny C , Klauser A , Riepl RL , Muller-Felber W , Pongratz D (1999 ) Gastrointestinal dysfunction in amyotrophic lateral sclerosis . Amyotroph Lateral Scler Other Motor Neuron Disord 1 : 15 –19 . 12365061 80 Vadakkadath Meethal S , Atwood CS (2012 ) Lactate dyscrasia: a novel explanation for amyotrophic lateral sclerosis . Neurobiol Aging 33 : 569 –581 . 10.1016/j.neurobiolaging.2010.04.012 20541840 81 Barber SC , Shaw PJ (2010 ) Oxidative stress in ALS: key role in motor neuron injury and therapeutic target . Free Radic Biol Med 48 : 629 –641 . 10.1016/j.freeradbiomed.2009.11.018 19969067 82 Andersson U , Leighton B , Young ME , Blomstrand E , Newsholme EA (1998 ) Inactivation of aconitase and oxoglutarate dehydrogenase in skeletal muscle in vitro by superoxide anions and/or nitric oxide . Biochem Biophys Res Commun 249 : 512 –516 . 9712727 83 Tretter L , Adam-Vizi V (2000 ) Inhibition of Krebs cycle enzymes by hydrogen peroxide: A key role of [alpha]-ketoglutarate dehydrogenase in limiting NADH production under oxidative stress . J Neurosci 20 : 8972 –8979 . 11124972 84 Bach AC , Babayan VK (1982 ) Medium-chain triglycerides: an update . Am J Clin Nutr 36 : 950 –962 . 6814231 85 Bough KJ , Rho JM (2007 ) Anticonvulsant mechanisms of the ketogenic diet . Epilepsia 48 : 43 –58 . 17241207 86 Zhao Z , Lange DJ , Voustianiouk A , MacGrogan D , Ho L , Suh J , et al (2006 ) A ketogenic diet as a potential novel therapeutic intervention in amyotrophic lateral sclerosis . BMC neuroscience 7 : 29 16584562 87 Zhao W , Varghese M , Vempati P , Dzhun A , Cheng A , Wang J , et al (2012 ) Caprylic triglyceride as a novel therapeutic approach to effectively improve the performance and attenuate the symptoms due to the motor neuron loss in ALS disease . PloS one 7 : e49191 10.1371/journal.pone.0049191 23145119 88 McDonald TS , Tan KN , Hodson MP , Borges K (2014 ) Alterations of hippocampal glucose metabolism by even versus uneven medium chain triglycerides . J Cereb Blood Flow Metab 34 : 153 –160 . 10.1038/jcbfm.2013.184 24169853 89 Ngo ST , Steyn FJ (2015 ) The interplay between metabolic homeostasis and neurodegeneration: insights into the neurometabolic nature of amyotrophic lateral sclerosis . Cell Regen (Lond) 4 : 5 .26322226 90 Miquel E , Cassina A , Martinez-Palma L , Bolatto C , Trias E , Gandelman M , et al (2012 ) Modulation of astrocytic mitochondrial function by dichloroacetate improves survival and motor performance in inherited amyotrophic lateral sclerosis . PloS one 7 : e34776 10.1371/journal.pone.0034776 22509356 91 ALS Untangled Group (2013 ) ALS Untangled No. 20: the Deanna protocol . Amyotroph Lateral Scler Frontotemporal Degener 14 : 319 –323 . 10.3109/21678421.2013.788405 23638638 92 Ari C , Poff AM , Held HE , Landon CS , Goldhagen CR , Mavromates N , et al (2014 ) Metabolic therapy with Deanna Protocol supplementation delays disease progression and extends survival in amyotrophic lateral sclerosis (ALS) mouse model . PLoS One 9 : e103526 10.1371/journal.pone.0103526 25061944

PMC005xxxxxx/PMC5001696.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756468310.1371/journal.pone.0161906PONE-D-16-21560Research ArticlePhysical SciencesPhysicsClassical MechanicsDeformationPhysical SciencesPhysicsClassical MechanicsDamage MechanicsDeformationMedicine and Health SciencesSurgical and Invasive Medical ProceduresMedicine and Health SciencesRheumatologyKyphosisBiology and Life SciencesAnatomyMusculoskeletal SystemSpineMedicine and Health SciencesAnatomyMusculoskeletal SystemSpineBiology and Life SciencesAnatomyMusculoskeletal SystemSpineVertebraeMedicine and Health SciencesAnatomyMusculoskeletal SystemSpineVertebraePhysical SciencesMathematicsGeometryCurvatureResearch and Analysis MethodsImaging TechniquesNeuroimagingComputed Axial TomographyBiology and Life SciencesNeuroscienceNeuroimagingComputed Axial TomographyMedicine and Health SciencesDiagnostic MedicineDiagnostic RadiologyTomographyComputed Axial TomographyResearch and Analysis MethodsImaging TechniquesDiagnostic RadiologyTomographyComputed Axial TomographyMedicine and Health SciencesRadiology and ImagingDiagnostic RadiologyTomographyComputed Axial TomographyResearch and Analysis MethodsMathematical and Statistical TechniquesStatistical MethodsRegression AnalysisLinear Regression AnalysisPhysical SciencesMathematicsStatistics (Mathematics)Statistical MethodsRegression AnalysisLinear Regression AnalysisEffects of Multilevel Facetectomy and Screw Density on Postoperative Changes in Spinal Rod Contour in Thoracic Adolescent Idiopathic Scoliosis Surgery Postoperative Changes in Spinal Rod Contour in AIS SurgeryKokabu Terufumi 1Sudo Hideki 1*Abe Yuichiro 2Ito Manabu 3Ito Yoichi M. 4Iwasaki Norimasa 11 Department of Orthopaedic Surgery, Hokkaido University Hospital, Sapporo, Hokkaido, Japan2 Eniwa Hospital, Eniwa, Hokkaido, Japan3 Department of Spine and Spinal Cord Disorders, Hokkaido Medical Center, Sapporo, Hokkaido, Japan4 Department of Biostatistics, Hokkaido University Graduate School of Medicine, Sapporo, Hokkaido, JapanNazarian Ara EditorHarvard Medical School/BIDMC, UNITED STATESCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: HS. Data curation: HS YA. Formal analysis: TK HS YMI. Investigation: TK HS. Methodology: TK HS YA. Project administration: HS. Resources: HS MI. Software: TK YMI. Supervision: HS MI NI. Validation: TK HS YMI. Visualization: TK. Writing – original draft: TK HS. Writing – review & editing: HS. * E-mail: hidekisudo@yahoo.co.jp26 8 2016 2016 11 8 e016190628 5 2016 12 8 2016 © 2016 Kokabu et al2016Kokabu et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Flattening of the preimplantation rod contour in the sagittal plane influences thoracic kyphosis (TK) restoration in adolescent idiopathic scoliosis (AIS) surgery. The effects of multilevel facetectomy and screw density on postoperative changes in spinal rod contour have not been documented. This study aimed to evaluate the effects of multilevel facetectomy and screw density on changes in spinal rod contour from before implantation to after surgical correction of thoracic curves in patients with AIS prospectively. The concave and convex rod shapes from patients with thoracic AIS (n = 49) were traced prior to insertion. Postoperative sagittal rod shape was determined by computed tomography. The angle of intersection of the tangents to the rod end points was measured. Multiple stepwise linear regression analysis was used to identify variables independently predictive of change in rod contour (Δθ). Average Δθ at the concave and convex side were 13.6° ± 7.5° and 4.3° ± 4.8°, respectively. The Δθ at the concave side was significantly greater than that of the convex side (P < 0.0001) and significantly correlated with Risser sign (P = 0.032), the preoperative main thoracic Cobb angle (P = 0.031), the preoperative TK angle (P = 0.012), and the number of facetectomy levels (P = 0.007). Furthermore, a Δθ at the concave side ≥14° significantly correlated with the postoperative TK angle (P = 0.003), the number of facetectomy levels (P = 0.021), and screw density at the concave side (P = 0.008). Rod deformation at the concave side suggests that corrective forces acting on that side are greater than on the convex side. Multilevel facetectomy and/or screw density at the concave side have positive effects on reducing the rod deformation that can lead to a loss of TK angle postoperatively. The authors received no specific funding for this work. Data AvailabilityAll relevant data are within the paper.Data Availability All relevant data are within the paper. ==== Body Introduction Restoration and maintenance of the normal sagittal contour as well as coronal correction of the thoracic curve is an important surgical strategy in patients with thoracic adolescent idiopathic scoliosis (AIS), because these patients typically have a hypokyphotic thoracic spine compared with nonscoliosis patients [1]. Currently, posterior segmental pedicle screw (PS) instrumentation and fusion has become one of the most common surgical treatments. However, recent studies have reported that PS constructs to maximize scoliosis correction can cause further lordosis of the thoracic spine [2–4]. These patients exhibit a flat back, leading to progressive decompensation and sagittal imbalance [1,5]. Preservation of thoracic kyphosis (TK) is also critical to maintain lumbar lordosis after surgical treatment of AIS [1]. To overcome these issues, Ito et al. [6] and Sudo et al. [7–9] recently developed a very simple surgical technique called the simultaneous double-rod rotation technique (SDRRT) for correcting AIS. In this technique, two rods are connected to the screw heads and are simply rotated simultaneously to correct the scoliosis, while TK is maintained or improved. Moreover, hypokyphotic rod deformation is prevented with dual-rod derotation instead of single-rod derotation [6–9]. Some studies have investigated the correlation between AIS curve correction and destabilization procedures such as multilevel facetectomy [10] or the number of fixation anchors, such as PS density [11–14]. Implant rod curvature will also influence the postoperative TK. The initial shape of the rod could lead to a certain sagittal outcome. However, it has been recognized that rods bent by surgeons prior to implantation tend to flatten after surgery [15,16]. The postoperative implant rod deformation as a “spring-back” effect can alter the sagittal alignment of the spine and consequently the clinical outcome [17]. Until now, there has been no consensus on what possible factors can alter the shape of the rod. Based on the biomechanical point of view, the comprehensive effects of the surgical strategies on postoperative TK remain unknown. This study aimed to evaluate the effects of multilevel facetectomy and/or screw density on the change in the rod contour and TK in patients with thoracic AIS. Materials and Methods Patients This study was an investigator-initiated observational cohort study conducted at a single medical center and approved by institutional review board of Hokkaido University Hospital (approval number: 014–0370). A written informed consent was obtained from all participants. Data from 49 patients (1 male, 48 female) with Lenke type 1 or type 2 AIS curves who underwent posterior thoracic curve correction between June 2009 and April 2016 were evaluated at our institution. Exclusion criteria included syndromic, neuromuscular, and congenital scoliosis and the presence of other double or triple major AIS curves, as well as thoracolumbar and lumbar AIS curves. The average age and Risser sign at surgery were 15.5 ± 2.2 years (range, 12–20) and 3.9 ± 1.1 (range, 1–5; Table 1), respectively. 10.1371/journal.pone.0161906.t001Table 1 Disease characteristics and clinical features of the subjects. Mean ± standard deviation Range Body mass index (kg/m2) 18.8 ± 2.4 12.4 to 24.2 Risser sign (grade) 3.9 ± 1.1 1 to 5 Preoperative main thoracic Cobb angle (°) 59.5 ± 10.2 46 to 88 Postoperative main thoracic Cobb angle (°) 13.3 ± 7.3 1 to 36 Preoperative thoracic kyphosis angle (°) 11.7 ± 7.8 -4 to 34 Postoperative thoracic kyphosis angle (°) 21.1 ± 6.3 7 to 33 Number of vertebrae in fusion (no.) 10.8 ± 1.6 7 to 14 Number of facetectomy levels (no.) 6.5 ± 3.4 0 to 12 Screw density at concave side (no. of screws / level instrumented) 0.89 ± 0.14 0.5 to 1 Screw density at convex side (no. of screws / level instrumented) 0.80 ± 0.16 0.4 to 1 Standing long-cassette posteroanterior and lateral radiographs were evaluated for multiple parameters before and at the 2-week follow-up. Coronal and sagittal Cobb angle measurements of the main thoracic (MT) curves were obtained. The end vertebrae levels were determined on preoperative radiographs and measured on subsequent radiographs to maintain consistency for statistical comparisons [7,8]. Sagittal measurements included the TK (T5–T12) angle [7,8]. The number of facetectomy levels was counted, and screw density was expressed as the number of screws per level instrumented for each patient. In this study, the number of hooks in the instrumented level was not counted. Surgical Technique Six-millimeter diameter titanium-alloy implant rods and polyaxial PSs (USS II Polyaxial, DePuy Synthes, Raynham, MA, USA) were used to correct the scoliosis deformity. All rods were prebent only at a single plane. Rods and screws were surgically implanted via the double rod rotation technique [6–9]. In this technique, two implant rods were inserted into the polyaxial screw heads. The polyaxial screw heads remained unfastened until the completion of rod rotation, allowing the rods to rotate and translate freely inside the screw head. A torque was applied to the rod-rotating device to rotate the rods simultaneously, transferring the previous curvature of the rod at the coronal plane to the sagittal plane. Additional in situ bending or other reduction maneuvers were not performed in all cases. All polyaxial screws were carried upward and medially to the concave side of the curve by the rotation of the rods, which did not exert any downward force on the vertebral body [6–9]. Both polyaxial screw heads and simultaneous double-rod rotation were key to the current technique. Frictional force at the screw–rod interface was decreased, and there was little chance of screw cut-out laterally[9]. This technique provided derotation of the apical vertebra as well as restoration of TK, leading to rib hump correction without additional costoplasty [9]. Rod Analysis The implant rod angle of curvature was used to evaluate implant rod deformation. Prior to implantation, following the intraoperative contouring of the rods, the surgeon traced the rod shapes on paper [15]. The angle between the proximal and distal tangential line was measured as the rod angle before implantation (θ1) as previously described [15]. Postoperative implant rod geometry was obtained a maximum of 2 weeks after the surgical operation using computed tomography (Aquilion 64 CT scan; Toshiba Medical Systems Corporation, Tokyo, Japan). Digital Imaging and Communications in Medicine (DICOM) data were obtained to reconstruct new images by DICOM viewer software (OsiriX Imaging Software; Pixmeo Labs., Geneva, Switzerland). The reconstructed sagittal images of the implanted rods were obtained, and the angle between the proximal tangential line and the distal tangential line was measured (θ2) (Fig 1). In cases in which the rod shape had both thoracic and lumbar curvature, the distal tangential line was determined based on the inflection point. The angle of rod deformation (Δθ) was defined as the difference between θ1 and θ2 (θ1–θ2). The angles θ1, θ2, and Δθ were obtained from the rods at both the concave and convex sides. 10.1371/journal.pone.0161906.g001Fig 1 Rod angle before and after implantation. (A) Prior to implantation, the surgeon traced the rod shapes on paper. The angle between the proximal and distal tangential line was measured (θ1). (B) Postoperative implant rod geometry (θ2) was obtained after the surgical operation using computed tomography. Statistical Analysis Bivariate statistical analysis was performed between the change in TK (postoperative TK–preoperative TK) and the Δθ at the concave or convex side using the Wilcoxon rank sum test. Pearson’s correlation coefficient analysis was used to assess relationships between independent variables. Stepwise linear regression analysis was applied to control for possible confounding variables and to identify variables independently predictive of Δθ both at the concave and convex side. Patients’ age and disease characteristics were included in the variables: age, body mass index [weight (kg)/height(m)2], Risser sign, preoperative MT Cobb angle, postoperative MT Cobb angle, preoperative TK angle, postoperative TK angle, number of facetectomy levels, and screw density at both the concave and convex side. Significant multivariate predictors are reported with their respective predictive equations, including the intercept and regression coefficients (β). Model fit was assessed by using the goodness-of-fit F test and R2 statistic. Data analyses were performed using JMP statistical software for Windows (version 12; SAS, Inc., Cary, NC, USA). P < 0.05 was considered statistically significant. All data are expressed as mean ± standard deviation. Results Disease characteristics are summarized in Table 1. On average, 10.8 ± 1.6 vertebrae were instrumented in the 49 patients. The average preoperative MT curve was 59.5° ± 10.2°. Postoperative radiographs showed an average MT curve of 13.3° ± 7.3°. Sagittal plane analysis revealed that the average preoperative TK was 11.7° ± 7.8°, which improved significantly to 21.1° ± 6.3° (P < 0.0001). The preoperative θ1 and postoperative θ2 implant rod angle of curvatures at the concave and convex sides of the deformity are listed in Table 2. 10.1371/journal.pone.0161906.t002Table 2 Implant rod angle of curvature at the concave and convex side of deformity. Mean ± standard deviation Range Preoperative rod angle (θ1) at concave side (°) 41.8 ± 7.1 22.3 to 66.5 Preoperative rod angle (θ1) at convex side (°) 38.4 ± 9.5 19.5 to 69.9 Postoperative rod angle (θ2) at concave side (°) 28.2 ± 9.1 9.2 to 48.5 Postoperative rod angle (θ2) at convex side (°) 34.1 ± 8.2 15.0 to 55.8 Rod deformation (Δθ) at concave side (°) 13.6 ± 7.5 -0.3 to 36.5 Rod deformation (Δθ) at convex side (°) 4.3 ±4.8 -6.8 to 17.8 The θ2 was significantly lower than the θ1 at the concave side (P < 0.001 at the concave side, P = 0.019 at the convex side, respectively). The Δθ at the concave side was significantly greater than that of the convex side (P < 0.0001) (Fig 2). 10.1371/journal.pone.0161906.g002Fig 2 Implant rod angle of curvature at the concave and convex sides of the deformity. (A) θ1 and θ2 at the concave side of each patients. (B) θ1 and θ2 at the convex side of each patients. (C) Comparison between θ1 and θ2 at the concave side. (D) Comparison between θ1 and θ2 at the convex side. (E) Comparison between Δθ at the concave side and Δθ at the convex side. Postoperative TK was significantly correlated with the postoperative θ2 implant rod angle at both sides, particularly at the concave side (concave: r = –0.415, P = 0.003; convex: r = –0.321, P = 0.025, respectively) (Fig 3). 10.1371/journal.pone.0161906.g003Fig 3 Correlation analysis between the postoperative rod angle and the thoracic kyphosis angle. (A) concave side. (B) convex side. In multiple stepwise linear regression analysis, 4 variables were independent predictive factors for Δθ at the concave side: Risser sign (P = 0.032), the preoperative MT Cobb angle (P = 0.031), the preoperative TK angle (P = 0.012, and the number of facetectomy levels (P = 0.007). The model fit the data well (goodness-of-fit F test = 7.05, R2 = 0.50, P = 0.0001) (Table 3). 10.1371/journal.pone.0161906.t003Table 3 Associations between various factors and rod deformation at the concave side (°) using multiple stepwise linear regression analysis. Regression Coefficient Standard Error 95% Confidence Interval t Standardized β P Constant -10.787 8.57 (-28.081, 6.509) -1.26 - 0.215 Risser sign (grade) 1.668 0.753 (0.148, 3.188) 2.21 0.249 0.032 Preoperative main thoracic Cobb angle (°) 0.265 0.085 (0.095, 0.436) 3.14 0.362 0.031 Preoperative thoracic kypohosis angle(°) -0.279 0.106 (-0.494, -0.064) -2.62 -0.292 0.012 Number of facetectomy levels (no.) -0.716 0.253 (-1.225, -0.206) -2.83 -0.325 0.007 Screw density at convex side (no. of screws / level instrumented) 10.372 5.205 (-0.133, 20.876) 1.99 0.223 0.053 P < 0.05 was considered statistically significant Conversely, for Δθ at the convex side, 3 variables emerged as predictors: the number of vertebrae in fusion (standardized β = –0.596, P = 0.0003), the number of facetectomy levels (standardized β = 0.578, P = 0.0006), and the Risser sign (standardized β = –0.292, P = 0.026). However, R2 was low (goodness-of-fit F test = 5.67, R2 = 0.34, P = 0.0009), indicating that only 34% of the variation in Δθ was explained by these 3 predictors. Subgroup Analysis To determine whether Δθ affects postoperative TK, the total cohort was then divided into 2 groups on the basis of the mean Δθ at the concave side. The Δθ ≥ 14° group was defined by Δθ above the mean degree (13.6° ± 7.5°) at the concave side and further analyzed. The average age (n = 23) were 15.4 ± 2.1 years (range, 12–20). Disease characteristics and rod data in the group of ≥ 14° rod deformation are summarized in Table 4. 10.1371/journal.pone.0161906.t004Table 4 Disease characteristics and rod data in the group of ≥ 14° rod deformation at the concave side. Mean ± standard deviation Range Body mass index (kg/m2) 18.6 ± 2.5 13.1 to 23.4 Risser sign (grade) 4.0 ± 0.9 1 to 5 Preoperative main thoracic Cobb angle (°) 61.7 ± 11.7 46 to 88 Postoperative main thoracic Cobb angle (°) 14.0 ± 6.3 1 to 27 Preoperative thoracic kyphosis angle (°) 7.6 ±5.5 -4 to 23 Postoperative thoracic kyphosis angle (°) 19.6 ± 6.0 7 to 33 Number of vertebrae in fusion (no.) 10.5 ± 1.6 7 to 13 Number of facetectomy levels (no.) 5.6 ± 3.4 0 to 11 Screw density at concave side (no. of screws / level instrumented) 0.89 ± 0.15 0.56 to 1 Screw density at convex side (no. of screws / level instrumented) 0.84 ± 0.15 0.5 to 1 Preoperative rod angle (θ1) at concave side (°) 43.4 ± 8.0 29.6 to 66.5 Preoperative rod angle (θ1) at convex side (°) 37.9 ± 11.3 19.5 to 69.9 Postoperative rod angle (θ2) at concave side (°) 23.7 ± 9.6 9.2 to 48.5 Postoperative rod angle (θ2) at convex side (°) 32.5 ± 9.3 15.0 to 55.8 Rod deformation (Δθ) at concave side (°) 19.7 ± 5.3 14.1 to 36.5 Rod deformation (Δθ) at convex side (°) 5.5 ± 6.0 -6.8 to 17.8 Pearson’s correlation coefficient analysis showed that in the group of Δθ ≥ 14°, Δθ at the concave side had significant correlation with the postoperative TK angle (r = −0.590, P = 0.003), the number of facetectomy levels (r = −0.479, P = 0.021), and screw density at the concave side (r = −0.537, P = 0.008)(Table 5). 10.1371/journal.pone.0161906.t005Table 5 Correlation analysis between rod deformation and variable in patients with rod deformation ≥14 ° at the concave side. Pearson’s correlation coefficients Variable Correlation coefficient 95% CI Statistical significance Age at surgery (yrs) r = -0.017 (-0.398, 0.427) P = 0.937 Body mass index (kg/m2) r = -0.207 (-0.570, 0.225) P = 0.344 Risser sign (grade) r = -0.084 (-0.479, 0.340) P = 0.705 Preoperative main thoracic Cobb angle (°) r = 0.142 (-0.287, 0.524) P = 0.518 Postoperative main thoracic Cobb angle (°) r = 0.396 (-0.019, 0.695) P = 0.061 Preoperative thoracic kyphosis angle (°) r = -0.286 (-0.625, 0.143) P = 0.186 Postoperative thoracic kyphosis angle (°) r = -0.590 (-0.806, -0.235) P = 0.003 Number of vertebrae in fusion (no.) r = -0.324 (-0.649, 0.102) P = 0.132 Number of facetectomy levels (no.) r = -0.479 (-0.744, -0.083) P = 0.021 Screw density at concave side (no. of screws / level instrumented) r = -0.537 (-0.777, -0.160) P = 0.008 Screw density at convex side (no. of screws / level instrumented) r = 0.350 (-0.073, 0.666) P = 0.102 Rod deformation (Δθ) at convex side (°) r = 0.014 (-0.424, 0.400) P = 0.948 P < 0.05 was considered statistically significant Discussion Careful investigation of the changes in implant rod geometry is important to fully understand the biomechanics of scoliosis correction [16]. However, there have been few studies examining the relationship between rod deformation and sagittal alignment of the thoracic spine [15,16,18]. Cidambi et al. [15] documented that a significant difference was observed between pre- and postoperative rod contour, particularly for concave rods, and that the resulting deformations were likely associated with substantial in vivo deforming forces. Similarly, Salmingo et al. [16] reported that implant rods at the concave side were significantly deformed after surgery, whereas rods at the convex side had no significant deformation. Abe et al. [18] suggested that the corrective force during scoliosis surgery was 4 times greater at the concave side than at the convex side. The present study also showed that there was a significant positive relationship between postoperative TK and the postoperative implant rod angle of curvature, indicating that implant rod curvature influences sagittal curve correction. In addition, rod deformation at the concave side was significantly greater than that of the convex side. Removing the facets and soft tissues between the posterior elements has been shown to allow greater distraction abilities along the length of the posterior column [1]. Destabilization of the posterior spinal segment by releasing soft tissue or facet joints could be important to prevent implant breakage or pedicle fracture during maneuver in more severe curve corrections [18]. However, it is still unclear whether these posterior releases positively affect the TK, especially with a hyphokyphotic thoracic spine [1,9]. Recently, Sudo et al. [9] documented that in patients with a hypokyphotic thoracic spine < 15°, a significant correlation was found between the change in TK and the number of facetectomy levels, indicating that multilevel facetectomy is an important factor to restore TK in patients with hypokyphotic thoracic spines. In the present study, there was a significant negative correlation between preoperative TK and rod deformation, indicating that the rod deformation was greater in patients with preoperative hypokyphotic thoracic spines. In addition, the deformation could be decreased by increasing the number of facetectomy levels. Screw density may be also a possible factor in optimizing restoration of TK. However, the effect of implant density on sagittal plane correction and TK restoration has been reported in only a few studies, and the results have been controversial [12,14,19]. Larson et al. [12] demonstrated that decreased TK was correlated with increased screw density for Lenke type 1 and 2 curves. Conversely, Liu et al. [14] documented that higher screw density provided better TK restoration than low screw density. Recently, Sudo et al. [9] also documented that in patients with preoperative TK < 15°, a significant positive correlation was found between the change in TK and screw density, whereas no correlation was found in patients with TK ≥15°, suggesting that screw density had a positive effect on TK restoration in patients with hypokyphotic thoracic spines. Their results indicate that screw density at the concave side has an impact not only on scoliosis correction but also on TK restoration. In the present study, in patients with rod deformation at the concave side ≥ 14°, there were significant negative correlations between rod deformation at the concave side and postoperative TK or screw density at the concave side. These results suggest that rod deformation ≥ 14° at the concave side significantly decreases postoperative TK. However, this rod deformation could be decreased by increasing screw density at the concave side. Hence, the current results biomechanically supported the results presented by Sudo et al.[9], documenting that in patients with preoperative hypokyphotic thoracic spines, increasing screw density at the concave side is important for optimizing postoperative TK. There were limitations to this study. First, we evaluated deformity surgery with the use of titanium rods. The module of elasticity of the titanium alloy is much less than either stainless steel or cobalt chrome implants [16]. Second, we did not analyze the effects of multilevel osteotomy on the in vivo flexibility of the thoracic spine. We are now measuring in vivo force acting at the vertebrae before and after multilevel osteotomies in order to investigate the biomechanical effects of spinal flexibility provided by multilevel facet osteotomies on rod deformation. Third, resisting forces from the deformed spine might be different between males and females and this would need to be addressed in our predominantly female cohort. However, there were no effects of gender on thoracic hypokyphosis postoperatively (data not shown). Last, the relationships between rod deformation and clinical symptoms remain unclear. Conclusion The present study showed that there was a significant relationship between postoperative TK and the postoperative implant rod angle of curvature. In addition, the rod at the concave side was significantly deformed after the surgical treatment. The rod deformation at the concave side suggests that corrective forces acting on that side are greater than on the convex side. Multilevel facetectomy and/or screw density at the concave side have positive effects on reducing the rod deformation that can lead to a loss of TK angle postoperatively. ==== Refs References 1 Newton PO , Yaszay B , Upasani VV , Pawelek JB , Bastrom TP , Lenke LG , et al Preservation of thoracic kyphosis is critical to maintain lumbar lordosis in the surgical treatment of adolescent idiopathic scoliosis . Spine (Phila Pa 1976) 2010 ;35 :1365 –70 .20505560 2 Lowenstein JE , Matsumoto H , Vitale MG , Weidenbaum M , Gomez JA , Lee FY , et al Coronal and sagittal plane correction in adolescent idiopathic scoliosis: a comparison between all pedicle screw versus hybrid thoracic hook lumbar screw constructs . Spine (Phila Pa 1976) 2007 ;32 :448 –52 .17304136 3 Winter RB , Lovell WW , Moe JH . Excessive thoracic lordosis and loss of pulmonary function in patients with idiopathic scoliosis . J Bone Joint Surg Am 1975 ;57 :972 –7 . 1184646 4 Kim YJ , Lenke LG , Kim J , Bridwell KH , Cho SK , Cheh G , et al Comparative analysis of pedicle screw versus hybrid instrumentation in posterior spinal fusion of adolescent idiopathic scoliosis . Spine (Phila Pa 1976) 2006 ;31 :291 –8 .16449901 5 Roussouly P , Nnadi C . Sagittal plane deformity: an overview of interpretation and management . Eur Spine J 2010 ;19 :1824 –36 . 10.1007/s00586-010-1476-9 20567858 6 Ito M , Abumi K , Kotani Y , Takahata M , Sudo H , Hojo Y , et al Simultaneous double-rod rotation technique in posterior instrumentation surgery for correction of adolescent idiopathic scoliosis . J Neurosurg Spine 2010 ;12 :293 –300 . 10.3171/2009.9.SPINE09377 20192630 7 Sudo H , Ito M , Abe Y , Abumi K , Takahata M , Nagahama K , et al Surgical treatment of Lenke 1 thoracic adolescent idiopathic scoliosis with maintenance of kyphosis using the simultaneous double-rod rotation technique . Spine (Phila Pa 1976) 2014 ;39 :1163 –9 .24732855 8 Sudo H , Abe Y , Abumi K , Iwasaki N , Ito M , et al Surgical treatment of double thoracic adolescent idiopathic scoliosis with a rigid proximal thoracic curve . Eur Spine J 2016 ; 25 :569 –77 . 10.1007/s00586-015-4139-z 26195082 9 Sudo H , Abe Y , Kokabu T , Ito M , Abumi K , Ito YM , et al Correlation analysis between change in thoracic kyphosis and multilevel facetectomy/ screw density in main thoracic adolescent idiopathic scoliosis surgery . Spine J , Epub ahead of print. 10 Halanski MA , Cassidy JA . Do multilevel Ponte osteotomies in thoracic idiopathic scoliosis surgery improve curve correction and restore thoracic kyphosis? J Spinal Disord Tech 2013 ;26 :252 –5 . 10.1097/BSD.0b013e318241e3cf 22198324 11 Bharucha NJ , Lonner BS , Auerbach JD , Kean KE , Trobisch PD , et al Low-density versus high-density thoracic pedicle screw constructs in adolescent idiopathic scoliosis: do more screws lead to a better outcome? Spine J 2013 ;13 :375 –81 . 10.1016/j.spinee.2012.05.029 22901787 12 Larson AN , Polly DW Jr, Diamond B , Ledonio C , Richards BS 3rd, Emans JB , et al Does higher anchor density result in increased curve correction and improved clinical outcomes in adolescent idiopathic scoliosis? Spine (Phila Pa 1976) 2014 ;39 :571 –8 .24430717 13 Le Navéaux F , Aubin CÉ , Larson AN , Polly DW Jr, Baghdadi YM , Labelle H . Implant distribution in surgically instrumented Lenke 1 adolescent idiopathic scoliosis: does it affect curve correction? Spine (Phila Pa 1976) 2015 ;40 :462 –8 .25608243 14 Liu H , Li Z , Li S , Zhang K , Yang H , Wang J , et al Main thoracic curve adolescent idiopathic scoliosis: association of higher rod stiffness and concave-side pedicle screw density with improvement in sagittal thoracic kyphosis restoration . J Neurosurg Spine 2015 ;22 :259 –66 . 10.3171/2014.10.SPINE1496 25525960 15 Cidambi KR , Glaser DA , Bastrom TP , Nunn TN , Ono T , Newton PO . Postoperative changes in spinal rod contour in adolescent idiopathic scoliosis: an in vivo deformation study . Spine (Phila Pa 1976) 2012 ;37 :1566 –72 .22426445 16 Salmingo RA , Tadano S , Abe Y , Ito M . Influence of implant rod curvature on sagittal correction of scoliosis deformity . Spine J 2014 ;14 :1432 –9 . 10.1016/j.spinee.2013.08.042 24275616 17 Delorme S , Labelle H , Poitras B , Rivard CH , Coillard C , Dansereau J . Pre-, intra-, and postoperative three-dimensional evaluation of adolescent idiopathic scoliosis . J Spinal Disord 2000 ;13 :93 –101 . 10780682 18 Abe Y , Ito M , Abumi K , Sudo H , Salmingo R , Tadano S . Scoliosis corrective force estimation from the implanted rod deformation using 3D-FEM analysis . Scoliosis 2015 ;10 (Suppl 2 ):S2 10.1186/1748-7161-10-S2-S2 25810754 19 Lonner BS , Lazar-Antman MA , Sponseller PD , Shah SA , Newton PO , Betz R , et al Multivariate analysis of factors associated with kyphosis maintenance in adolescent idiopathic scoliosis . Spine (Phila Pa 1976) 2012 ;37 :1297 –302 .22228329

PMC005xxxxxx/PMC5001698.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756444810.1371/journal.pone.0161791PONE-D-16-18493Research ArticleBiology and Life SciencesPlant SciencePlant AnatomyFruit and Seed AnatomyPlant Embryo AnatomyHypocotylBiology and Life SciencesPlant SciencePlant AnatomyPlant Embryo AnatomyHypocotylBiology and Life SciencesDevelopmental BiologyEmbryogenesisPlant EmbryogenesisPlant Embryo AnatomyHypocotylBiology and Life SciencesDevelopmental BiologyPlant Growth and DevelopmentPlant DevelopmentPlant EmbryogenesisPlant Embryo AnatomyHypocotylBiology and Life SciencesPlant SciencePlant Growth and DevelopmentPlant DevelopmentPlant EmbryogenesisPlant Embryo AnatomyHypocotylBiology and Life SciencesPlant SciencePlant AnatomyLeavesBiology and Life SciencesOrganismsPlantsSeedlingsBiology and Life SciencesOrganismsPlantsFlowering PlantsResearch and Analysis MethodsMathematical and Statistical TechniquesStatistical MethodsRegression AnalysisPhysical SciencesMathematicsStatistics (Mathematics)Statistical MethodsRegression AnalysisBiology and Life SciencesGeneticsPhenotypesBiology and Life SciencesOrganismsPlantsBrassicaArabidopsis ThalianaResearch and Analysis MethodsModel OrganismsPlant and Algal ModelsArabidopsis ThalianaBiology and Life SciencesGeneticsPlant GeneticsBiology and Life SciencesPlant SciencePlant GeneticsPIF4 and ELF3 Act Independently in Arabidopsis thaliana Thermoresponsive Flowering ELF3/PIF4 Independence in Plant Adult Thermal Responseshttp://orcid.org/0000-0001-9233-2157Press Maximilian O. 1Lanctot Amy 2Queitsch Christine 1*1 University of Washington Department of Genome Sciences, Seattle, United States of America2 University of Washington Molecular and Cellular Biology Program, University of Washington Department of Biology, Seattle, United States of AmericaSomers David E. EditorOhio State University, UNITED STATESCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: MOP AL CQ. Formal analysis: MOP. Funding acquisition: CQ. Investigation: MOP AL. Methodology: MOP CQ. Supervision: CQ. Validation: MOP AL. Visualization: MOP. Writing – original draft: MOP AL CQ. Writing – review & editing: MOP CQ. * E-mail: queitsch@uw.edu26 8 2016 2016 11 8 e016179127 5 2016 11 8 2016 This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.Plants have evolved elaborate mechanisms controlling developmental responses to environmental stimuli. A particularly important stimulus is temperature. Previous work has identified the interplay of PIF4 and ELF3 as a central circuit underlying thermal responses in Arabidopsis thaliana. However, thermal responses vary widely among strains, possibly offering mechanistic insights into the wiring of this circuit. ELF3 contains a polyglutamine (polyQ) tract that is crucial for ELF3 function and varies in length across strains. Here, we use transgenic analysis to test the hypothesis that natural polyQ variation in ELF3 is associated with the observed natural variation in thermomorphogenesis. We found little evidence that the polyQ tract plays a specific role in thermal responses beyond modulating general ELF3 function. Instead, we made the serendipitous discovery that ELF3 plays a crucial, PIF4-independent role in thermoresponsive flowering under conditions more likely to reflect field conditions. We present evidence that ELF3 acts through the photoperiodic pathway, pointing to a previously unknown symmetry between low and high ambient temperature responses. Moreover, in analyzing two strain backgrounds with different thermal responses, we demonstrate that responses may be shifted rather than fundamentally rewired across strains. Our findings tie together disparate observations into a coherent framework in which multiple pathways converge in accelerating flowering in response to temperature, with some such pathways modulated by photoperiod. http://dx.doi.org/10.13039/100000052NIH Office of the DirectorDP2OD008371Queitsch Christine http://dx.doi.org/10.13039/100000001National Science FoundationMCB-1516701Queitsch Christine This work was supported by National Institutes of Health grant NIH New Innovator Award DP2OD008371 and NSF MCB-1516701 to CQ (https://commonfund.nih.gov/newinnovator/index). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityAnalysis scripts and data are provided at https://figshare.com/articles/elf3_pif4_data_code_v2/3398353, https://dx.doi.org/10.6084/m9.figshare.3398353.v1.Data Availability Analysis scripts and data are provided at https://figshare.com/articles/elf3_pif4_data_code_v2/3398353, https://dx.doi.org/10.6084/m9.figshare.3398353.v1. ==== Body Introduction The responses of plants to temperature variation are of central importance to food security in a changing world [1]. Therefore, the elucidation of the genetic pathways underlying these responses has been a key mission of plant science [2]. Many previous studies examined the phenomena of circadian temperature compensation [3–5], thermoresponsive flowering [6–10], and temperature effects on plant morphology [11–16]. Several have converged on PIF4 as a master regulator of temperature responses, and ELF3 as an input to PIF4 integration, among many other genes and pathways [9,11,14–16]. Given known regulatory interactions between ELF3 and PIF4 [17–20], it is reasonable to predict that both operate in the same pathway for thermal response phenotypes [21]. Recent reports focusing on one such phenotype, hypocotyl elongation, support this expectation [14–16]. ELF3 serves to repress hypocotyl elongation by reducing PIF4 levels. This repression of PIF4 occurs at both the transcriptional level, through the role of ELF3 in the Evening Complex (EC) [17,19], and at the post-translational level, through PIF4 destabilization by phytochrome phyB [22]. Light sensing enforces circadian oscillations of the EC and other components, leading to calibration of the circadian clock [23,24], resulting in diurnal repression of hypocotyl elongation through repression of PIF4 and PIF5 [17,19]. ELF3 also plays a crucial role as a flowering repressor [25]. Consequently, elf3 null mutants show elongated hypocotyls even in the light, and flower early. PIF4 is one of a family of basic helix-loop-helix (bHLH) “phytochrome-interacting factors” (PIFs), transcription factors with overlapping functions promoting skotomorphogenesis. Under dark conditions, the PIFs act to target phyB for ubiquitin-mediated degradation by the E3 ubiquitin ligase COP1, thereby repressing photomorphogenesis [26]. Under light conditions, degradation of PIFs is mediated by direct interactions with photoactivated phyB [22]. PIF4 is distinct from the other PIFs in having specific roles in temperature sensing and flowering [27]. pif4 null mutants show short hypocotyls with photomorphogenic attributes even in the dark [28]. At elevated ambient temperatures (27°-29°) the wiring of these signaling pathways changes. Several independent studies have recently found that elevated temperatures, specifically during dark periods [29], inhibit the activity of the EC by an unknown mechanism [14–16], leading to increased expression of PIF4 and its targets [11,27]. This increased PIF4 activity leads to several morphological temperature responses through various signaling pathways [13,27]. PIF4 is also required for the acceleration of flowering at 27°C under short photoperiods [9,29], though these observations have been disputed [30,31]. While PIF4 action alone (among PIFs) is essentially sufficient for most described thermomorphogenic responses [11,27], there is evidence for a limited role of PIF5 (though not other PIFs) in thermoresponsive flowering under short days (SDs) [30,31]. In contrast, under continuous light, pif4 null mutants have an intact temperature-dependent acceleration of flowering [11]. Lastly, pif4 null mutants lose the normal elongation of petioles under high temperatures [11]. It is unclear why PIF4 does not affect thermoresponsive flowering under continuous light; yet, this phenomenon may reflect low PIF4 levels under these conditions due to inhibition by phyB. Under longer photoperiods and higher temperature a flowering acceleration still exists [7,11], which suggests a PIF4-independent thermoresponsive flowering pathway. Nonetheless, recent reviews of the literature tend to emphasize the primacy of PIF4 in this response [10,32,33], although the condition of elevated temperature with short photoperiods is probably rare in the field. Recent studies have identified ELF3 as a plausible upstream regulator of PIF4 in thermal responses [14–18]. However, others have implicated different candidates, such as FCA [13], and mathematical modeling has suggested that ELF3/EC complex regulation alone is insufficient to explain PIF4 thermal regulation [14,34]. The exact mechanisms of this response have yet to be unraveled. Specifically, the mechanism by which EC/ELF3 activity is reduced under elevated temperatures (“temperature sensing”) is not known. We recently used transgenic experiments to demonstrate that ELF3 function is dependent on the unit copy number of its C-terminal polyglutamine (polyQ) tract [35]. This domain is likely disordered, and disordered domains evince structural changes in response to physical parameters such as temperature [36]. Thermal remodeling of this polyQ tract is a plausible mechanism by which ELF3 activity could be modulated through temperature. This polyQ tract also shows substantial natural variation [35], potentially serving as a factor underlying natural variation in thermoresponsive phenotypes. For example, in flies, variable repeats are associated with local temperature compensation adaptations [37]. In short, the ELF3-polyQ is an attractive candidate for adaptive variation in the ecologically relevant trait of temperature response [38]. In this study, we used transgenic polyQ variants of ELF3 in two A. thaliana genetic backgrounds to dissect the contribution of the polyQ tract to temperature response. We show that polyQ repeat copy number modulates temperature sensing by affecting overall ELF3 function. Surprisingly, we found that ELF3’s role in thermoresponsive flowering appears to be entirely independent of PIF4. We postulate that ELF3’s primary role in thermoresponsive flowering is PIF4-independent and occurs through the photoperiodic pathway, and that this role is in turn dependent on the genetic background. Results The hypocotyl elongation temperature response is modulated by the ELF3 polyQ tract affecting overall gene function Many recent studies noted the involvement of ELF3 in temperature-dependent hypocotyl elongation [14–16,39], concluding that ELF3 protein activity is reduced under elevated temperatures, thereby relieving ELF3 repression of PIF4. PIF4 up-regulation then leads to the observed hypocotyl elongation. We examined whether polyQ tract variation in ELF3 in two backgrounds affects hypocotyl elongation at 27° in short days (Fig 1), a condition previously shown to require ELF3 for thermal responses [14]. We previously showed that ELF3 polyQ variation has pleiotropic background-dependent effects, with nonlinear associations between polyQ tract length and quantitative phenotypes (including hypocotyl elongation at 22°C; ref. [35]). Certain variants (16Q for Ws, >20Q for Col) generally complemented elf3 null mutant phenotypes in Col and Ws A. thaliana strains, whereas other variants complemented only specific phenotypes or behaved as hypomorphs across all tested phenotypes. Here, we observed similar trends for thermoresponsive hypocotyl elongation (Fig 1). For example, in the Ws background (Fig 1A), the endogenous ELF3 variant (16Q) partially complements the elf3 null mutant; another variant (9Q) fully complements the hypocotyl temperature response. Other polyQ variants behaved as hypomorphs in Ws. In the Col background (Fig 1B), the endogenous 7Q variant, among other variants, failed to rescue the response, agreeing with our previous observation that these transgenic lines are hypomorphic in this background [35]. Deleting the entire polyQ tract eliminated thermoresponsive hypocotyl elongation in both Col and Ws backgrounds. We next addressed whether the observed phenotypic variation among polyQ variants was due to variation in thermosensing or variation in general ELF3 function. We found that robust thermal responses were strongly correlated with the overall functionality of each ELF3 variant in hypocotyl elongation (Fig 1C), such that variants with intact thermal responses exhibited short hypocotyls at 22°C, whereas ELF3 variants with defective thermal responses exhibited elongated hypocotyls regardless of temperature. Furthermore, this ELF3 functionality effect is dependent on genetic background (comparing for instance the 16Q and 20Q responses). Together, these results suggest that the ELF3 polyQ tract controls repression of hypocotyl elongation regardless of temperature, rather than sensing temperature specifically. Nonetheless, our transgenic ELF3 polyQ lines remain informative as an allelic series of ELF3 function to understand the role of ELF3 in the de-repression of PIF4, which is thought to underlie thermomorphogenesis [14–16,40,41]. 10.1371/journal.pone.0161791.g001Fig 1 Response to elevated temperature (27°, relative to 22°) among transgenic lines expressing ELF3-polyQ variants. Mean response and error were estimated by regression, based on two independently-generated transgenic lines for each genotype, with n > = 30 seedlings of each genotype in each condition (S1 Table). WT = Ws, elf3 = elf3 mutant+vector control, 0Q = elf3 mutant+ELF3 transgene lacking polyQ, etc. Error bars indicate standard error of the mean. (A): Ws (Wassilewskija) strain background. Lines are generated in an elf3-4 background. (B): Response in the Col (Columbia) strain background, lines were generated in an elf3-200 background. In both (A) and (B), response is defined as the change in hypocotyl length in mm; **: Bonferroni-corrected p < 0.01, *: Bonferroni-corrected p < 0.05,.: Bonferroni-corrected p < 0.1 in testing the interaction term (different response from WT, Ws or Col). (C): Temperature response is a function of ELF3 functionality (repression of hypocotyl elongation at 22°). Simple means of 22° hypocotyl length, regression estimates of temperature response. PCC = Pearson correlation coefficient; p-value is from a Pearson correlation test. Expression of PIF4 and PIF4 targets as a function of temperature and ELF3 To evaluate the hypothesis that the thermal response defects in the transgenic lines were due to up-regulation of PIF4 and PIF4 targets, we measured transcript levels of PIF4 and its target AtHB2 in seedlings of selected lines from both backgrounds at 22°C and 27°C (S1 Fig). Like others [14–16], we observed an inverse relationship between ELF3 functionality and transcript levels of PIF4 and AtHB2, with larger effects on PIF4 expression. The ELF3 lines with the strongest thermal response (e.g. 16Q in the Ws background) showed the most robust de-repression of PIF4 at elevated temperature. However, elf3 null mutants retained some PIF4 up-regulation under these conditions, especially in the Ws background. We conclude that ELF3-mediated de-repression of PIF4 is involved in thermal responses as suggested by prior studies [14–16]; however, de-repression of PIF4 and its targets may not be sufficient to explain the entirety of thermal response defects in elf3 null mutants. ELF3 polyQ variation affects thermoresponsive adult morphology and flowering time Following the expectation that ELF3’s thermal response acts through PIF4, we reasoned that ELF3 should also play a role in other PIF4-dependent thermal responses. One well-known response to elevated temperature is adult petiole elongation, which has been demonstrated at day lengths longer than SD (9hr to continuous light) [11,42,13]. Consequently, we considered petiole elongation under LD. pif4 mutants fail to show this response when grown at elevated temperatures [11]. We measured petiole length in the ELF3 polyQ transgenic lines, expecting that, due to general PIF4 de-repression, poorly-functioning ELF3 polyQ lines would show no response (perhaps due to constitutively elongated petioles, similar to hypocotyls; Fig 2). In stark contrast to this expectation, we found that all lines had a robust petiole response to temperature (Fig 2A and 2B). This effect was apparent in both Ws (Fig 2A) and Col backgrounds (Fig 2B). Moreover, this response was actually accentuated in elf3 null mutants and in poorly-functioning ELF3 polyQ variants (Fig 2A and 2B). 10.1371/journal.pone.0161791.g002Fig 2 Adult plant responses to elevated temperature (27°, relative to 22°) in long days among transgenic lines expressing different ELF3-polyQ variants. (A) and (C): Response in the Ws (Wassilewskija) strain background. Lines are in an elf3-4 background. (B) and (D): Response in the Col (Columbia) strain background, lines are in an elf3-200 background. (A) and (B) display PL:LL temperature response as differences in the PL:LL ratio between temperatures, (C) and (D) display RLN temperature response as differences in the number of leaves between temperatures. Average responses and errors were estimated in a regression model accounting for variation between experiments (S2 Table), based on two to three independently-generated transgenic lines for each genotype. n > = 24 plants overall for each genotype in each condition. PL:LL = petiole to leaf length ratio at 25 days post germination, RLN = rosette leaf number at flowering, WT = wild type, elf3 = elf3 mutant+vector control, 0Q = elf3 mutant+ELF3 transgene with entire polyglutamine removed, etc. Error bars indicate standard error. In each case, **: Bonferroni-corrected p < 0.01, *: Bonferroni-corrected p < 0.05,.: Bonferroni-corrected p < 0.1 in testing the interaction term (different response from WT, Ws or Col). Further, we measured flowering time in transgenic lines as the number of rosette leaves at flowering (Fig 2C and 2D). PIF4 is not required for the accelerated flowering temperature response under longer photoperiods [11]. Hence, we expected that loss of ELF3 function should not affect thermoresponsive flowering if ELF3’s thermal signaling role acts through PIF4. In contrast to this expectation, in the Col background, elf3 mutants had an abrogated flowering response to elevated temperature (Fig 2D). Moreover, most variants in the Col background entirely failed to rescue this phenotype, and even the endogenous 7Q showed only a partial rescue. While elf3 mutants show abrogated flowering response to lower temperatures [8], it was not expected that a similar role would extend to the elevated temperature flowering response, which is usually considered to be dominated by PIF4 [9], gibberellin signaling [30], and other transcriptional regulators of FT such as SVP [31]. Unlike Col, Ws lacks a robust flowering response to elevated temperature under these conditions [43], and indeed, variants in the Ws background generally showed no thermoresponsive flowering (Fig 2C). Thus, ELF3 polyQ variation does not suffice to enhance the negligible thermoresponsive flowering in the Ws background under these conditions. In light of this data, the roles of ELF3 and PIF4 in the elevated temperature response appear to be independent of one another under these experimental conditions and for these traits. These results are intriguing, given that the PIF4 pathway is the best-recognized mechanism for thermoresponsive flowering at high temperatures [9,10,32,33]. Therefore, we suggest that ELF3 acts in a PIF4-independent pathway for thermoresponsive flowering at high temperatures. ELF3 regulates thermoresponsive flowering under long days, and is not required for PIF4-dependent adult thermomorphogenesis Our results with ELF3-polyQ variants suggested that ELF3 dysfunction does not meaningfully affect PIF4-dependent seedling traits in short days, but does affect PIF4-independent traits in adult plants in long days. However, these results may be due to subtle differences in conditions between our approach and those used by previous investigators. We therefore directly addressed the relationship of ELF3 and PIF4 in adult thermoresponsive phenotypes by growing pif4 and elf3 mutants with various thermal treatments. Previous experiments with pif4 mutants used different conditions from ours, specifically a later transfer to elevated temperature [11]. Hence, it was possible that the observed inconsistencies between elf3 and pif4 effects on adult thermoresponsive phenotypes were a trivial consequence of experimental conditions. Specifically, the effects of elevated temperature during the early seedling stages (the conditions we use) may induce pathways irrelevant to treatments at later, vegetative stages. Thus, we tested both transfer conditions under long days (Fig 3). We found that the effect of different experimental conditions is negligible, though the earlier 27°C treatment showed a slightly stronger morphological response (Fig 3A and 3B). Thus, the timing of the 27°C treatment (early seedling vs. vegetative stage) does not substantially affect adult thermoresponsive traits. Further, our results under long days were similar to previous observations under continuous light [11], showing that PIF4 is essential for petiole elongation (Fig 3B), but dispensable for thermoresponsive flowering (Fig 3C). Our PIF4 results were in direct contrast to ELF3, which was dispensable for petiole elongation (Fig 3B), but essential for thermoresponsive flowering (Fig 3C). These results indicate the apparent independence of ELF3 and PIF4 in these specific responses, and suggest that seedling thermomorphogenesis, adult thermomorphogenesis, and thermoresponsive flowering constitute three independent developmental responses. 10.1371/journal.pone.0161791.g003Fig 3 elf3 and pif4 null mutant phenotypes are independent under LD treatments and robust to conditions. (A), (B), and (C): 22°: constant 22° LD growth; 27° 14d: transfer from 22° to 27° at 14 days post-germination; 27° 1d: transfer from 22° to 27° at 1 day post-germination. (A): Col (WT), elf3-200, and pif4-2 plants grown under long days with three different temperature regimes were photographed at 20 days post germination. Experiment was repeated with similar results. (B and C): Petiole elongation responses of the indicated genotypes, measured by ratio of petiole to whole leaf length at 25 days post germination. Regression analysis of data in S3 Table. In each case, **: Bonferroni-corrected p < 0.01, *: Bonferroni-corrected p < 0.05, in testing whether the genotype x environment interaction term (difference of 22°-27 response from the Col 22°-27° response) differs from zero. Outliers (defined as >1.5 interquartile ranges away from the median) of each distribution are indicated as points. One open question was whether the dispensability of ELF3 for petiole elongation reflected increased importance of other inputs to PIF4, such as FCA, which is involved in PIF4-dependent thermoresponsive petiole elongation in 7-day-old seedlings [13]. We therefore measured adult thermoresponsive petiole elongation in fca mutants (S2A Fig), and unexpectedly found no substantial difference between fca mutants and WT Col. Regulatory rewiring across development may remove FCA and ELF3 as inputs to PIF4-dependent thermomorphogenesis in 25-day-old adult plants. A second question was whether loss of ELF3 function can affect thermoresponsive flowering in the Ws strain under other temperature conditions. We therefore assayed flowering in Ws and the Ws-derived null mutant elf3-4 at 16°C and 22°C (S2B Fig). Under these conditions, Ws robustly accelerated flowering at 22°C relative to 16°C, whereas elf3-4 showed no perceptible difference in flowering between the two temperatures. Thus, ELF3’s role in thermoresponsive flowering is not restricted to the Col strain or a certain temperature, but rather is necessary for whatever thermoresponsive reaction norm a strain may have for flowering. ELF3 and PIF4 regulate adult thermoresponsive phenotypes independently If ELF3 and PIF4 were independent in controlling thermal responses of adult phenotypes under long days, then elf3 pif4 double mutants would show approximately additive phenotypes. We generated elf3 pif4 double mutants and subjected them to the same experiments as above. Our results indicated that flowering and petiole elongation constitute independent temperature responses, with ELF3 controlling the former and PIF4 controlling the latter in additive fashions (Fig 4). That is, elf3 pif4 double mutants showed negligible thermoresponsive flowering like elf3, and a negligible petiole response like pif4. Additionally, elf3 pif4 flowered slightly later than elf3 at 22°, while maintaining a negligible thermal response in flowering, indicating that elf3 mutants are not simply restricted by a physiological limit of early flowering. The additivity of these phenotypes establishes that, under these conditions, ELF3 and PIF4 likely operate in separate thermal response pathways. 10.1371/journal.pone.0161791.g004Fig 4 Double mutant analysis confirms PIF4 and ELF3 independence in adult temperature responses and non-redundancy of PIF4 with PIF5. (A): Col, elf3-200, pif4-2, and elf3-200 pif4-2 plants grown under long days with two different temperature regimes were photographed at 25 days post germination. (B): Petiole elongation responses of the indicated genotypes, measured by ratio of petiole to whole leaf length at 25 days post germination. (C) and (D): Flowering temperature response of indicated genotypes, measured by rosette leaf number (RLN) at flowering. (B) and (C): n > 8 plants for each genotype in each treatment. All “27°” plants were seeded and incubated one day at 22° before transfer to 27°. Experiments were repeated with similar results. Regression analysis of data reported in S6 and S7 Tables. In each case, **: Bonferroni-corrected p < 0.01, *: Bonferroni-corrected p < 0.05, in testing whether the genotype x environment interaction term (difference of 22°-27° response from the Col 22°-27° response) differs from zero. Outliers (defined as >1.5 interquartile ranges away from the median) of each distribution are indicated as points. Previous studies have indicated that other members of the PIF family have negligible or minor (in the case of PIF5) roles in these same thermal response phenotypes [11,27,44]. For instance, under SD, pif1 pif3 pif4 pif5 mutants behave essentially identical to pif4 pif5 mutants in flowering response and FT expression, which in turn show only a very slight abrogation of these responses relative to pif4 mutants [31]. pif4 pif5 double mutants do show slightly abrogated thermoresponsive flowering under 12 hour light: 12 hour dark photoperiods relative to single mutants [29], similar to other thermally responsive phenotypes [11,29–31]. These previous findings, combined with the completely intact flowering response of pif4 mutants, suggest that redundancy between PIFs plays little meaningful role in this response. However, to directly address this possibility, we evaluated thermoresponsive flowering in pif5 and pif4 pif5 mutants (Fig 4D), because PIF5 is most often considered to act redundantly with PIF4 [20,29,31,45], and the only other PIF to show any small contribution to thermoresponsive flowering [29–31]. As expected, both pif5 single mutants and pif4 pif5 double mutants demonstrate intact thermoresponsive flowering. These observations indicate that redundancy with other PIFs is not responsible for the apparent independence of PIF4 and ELF3. Notably, petiole elongation at elevated temperatures is equally disrupted in pif4 and pif4 pif5 mutants, but intact in pif5 single mutants (S3 Fig), reproducing the known dependence of this trait upon PIF4 alone [11]. Consequently, our results support the previously-suggested dominance of thermomorphogenesis by PIF4 rather than other PIFs, and the irrelevance of PIF4 (and most likely other PIFs as well) to thermoresponsive flowering under LD. Overall, the strong photoperiod-dependence of PIF4-related thermoresponsive flowering necessitates the existence of some pathway or pathways independent of PIF4 under long days, given the persistence of the phenomenon under these conditions. Based on our data, ELF3 acts in one such pathway. Thermoresponsive flowering under long days can operate through the photoperiodic pathway ELF3 operates in thermoresponsive flowering at low ambient temperatures via the photoperiodic pathway, through repressing GI expression, after which GI in turn directly activates FT [46–48]. It has also been argued that one important consequence of increased temperature in the circadian clock is the expansion of GI’s nighttime expression peak into the early morning [15], and GI and CO are de-repressed in the early morning in elf3 null mutants [48]. To evaluate whether this pathway might explain our results, we measured transcript levels of GI and CO in wild-type and elf3 mutants under 22°C and 27°C at ZT0 (Fig 5A). We found that GI is strongly up-regulated in elf3 null mutants of Col and Ws backgrounds, confirming previous reports in Col [39,47,48]. Further, wild-type Ws showed higher basal GI levels compared to Col, which did not increase at higher temperatures. In contrast, Col showed very low basal GI levels that increased at higher temperatures to approximately the same levels as Ws. CO levels, however, were not substantially increased by either elf3 mutation or increased temperature, consistent with previous reports [8,47]. Thus, robust thermoresponsive flowering was correlated with low basal levels of GI, and with temperature-dependent GI up-regulation, as observed in Col. The ELF3-dependent thermal responsiveness of GI expression confirms previous reports [15,39], though the among-strain variation in responsiveness appears to be novel and correlated specifically with flowering induction (but not hypocotyl or petiole elongation, Figs 1 and 2). High basal GI levels in Ws may be associated with other thermoresponsive deficiencies at high temperatures in this strain [43,49,50]. These observations support the model under which ELF3 acts in the photoperiodic pathway to engender thermoresponsive flowering, just as it does in response to lower ambient temperatures [8,47]. 10.1371/journal.pone.0161791.g005Fig 5 ELF3 and GI regulate thermoresponsive flowering. (A): Temperature-responsive expression of photoperiodic pathway components at ZT0. Expression of each gene is quantified relative to levels in Col at 22° (Col 22 = 1.0). Error bars represent SEM across three biological replicates. elf3-4: elf3 null in Ws background; elf3-200: elf3 null in Col background. (B): Thermoresponsive flowering in various flowering mutants. LD RLN = rosette leaf number at flowering under long days. *: Bonferroni-corrected p < 0.05 in testing whether the genotype x environment interaction term (difference of 22°-27° response from the Col 22°-27° response) differs from zero; details of regression model in S9 Table. (C) Thermoresponsive petiole elongation in various flowering mutants. For (B) and (C), n > = 8 plants of each genotype in each condition; white boxes indicate measurements at 22°, red boxes indicate measurements at 27°. gi: gi-2, co: co-101, spy: spy-3, soc1: soc1 T-DNA insertion, elf3: elf3-200. Outliers (defined as >1.5 interquartile ranges away from the median) of each distribution are indicated as points. This experiment was repeated with similar results. (D): Models of thermoresponsive flowering under long and short photoperiods. Dashed edges indicate speculated temperature sensing mechanisms. Edges with increased weight indicate relative increases of influence between conditions. Pathways are indicated, along with other important actors reported elsewhere. We attempted to measure FT transcript levels in these samples, expecting that they would be elevated in the early-flowering elf3 and 27°C conditions (S4 Fig). However, while FT levels may increase slightly in the elf3 mutants, FT appears dramatically down-regulated in all 27°C samples. This finding is difficult to interpret in light of the phenotypic data, as most models of thermoresponsive flowering agree that signaling operates through FT [7–9,29–31], suggesting rather that these 7-day-old seedlings may be too young, or that the ZT0 time point employed may not be informative [48] for measuring physiologically relevant FT expression differences under these conditions. If the photoperiodic pathway contributes to thermoresponsive flowering at elevated ambient temperatures in long days (LD), we would expect mutants in this pathway to show abrogated thermal responses, as they do under short days (SD), along with members of the autonomous pathway [7]. These two pathways also contribute independently to thermoresponsive flowering at low temperatures (16°C vs. 23°C) [6,8]. Altogether, we would expect that a photoperiodic thermoresponsive flowering pathway would operate independently of both PIF4 and the autonomous pathways in long days. It is not clear whether the autonomous pathway would be independent of PIF4, given known regulatory interactions between FCA and PIF4 [13]. To evaluate whether these past results under other conditions also apply to long days and elevated temperatures, we measured flowering time at 22°C and 27°C in mutants in the photoperiodic pathway (gi, co, Fig 5B). We also tested mutants of the gibberellin pathway (spy), and a terminal floral integrator (soc1), which we do not expect to be necessary for thermoresponsive flowering. We found robust thermal responses in all mutants except elf3 and gi, similar to previous results under different conditions [7,8,46,47]. All of these mutants retained intact thermoresponsive petiole elongation (Fig 5C). These results emphasize once again that differences in thermoresponsive flowering are not generalizable between photoperiods, as it has recently been shown that co mutants have a partial flowering acceleration defect under SD [31]. These results implicate GI (but not CO) as an actor in thermoresponsive flowering at elevated temperatures. Collectively, these experiments suggest that the photoperiod pathway is necessary in promoting thermoresponsive flowering in long days, and expression data in this and other studies suggests that ELF3 is likely to act within this pathway. Discussion ELF3 and PIF4 are both crucial integrators of temperature and light signaling in controlling A. thaliana development. Recent literature has emphasized the centrality of PIF4-dependent thermoresponsive regulation in a variety of phenotypes, including in flowering [9,10,32]. Here, we show that PIF4 is dispensable for thermoresponsive flowering under long photoperiod conditions [11], and that ELF3 is essential for thermoresponsive flowering under these conditions. Our results integrate previous knowledge about thermoresponsive flowering, and identify at least one pathway for this response that does not involve PIF4. Moreover, we show that while polyQ variation in ELF3 affects ELF3 function, the polyQ tract is unlikely a temperature-responsive component in itself. Our results allow us to integrate the many disparate findings of current studies into classic models of thermal responses in A. thaliana, allowing a comprehensive view of the genetic underpinnings of this agronomically crucial plant trait. ELF3 polyglutamine variation appears to affect thermoresponsive traits by modulating overall ELF3 activity In previous work, we demonstrated that polyQ variation in ELF3 is (i) common, (ii) affects many known ELF3-dependent phenotypes, and (iii) is dependent on the genetic background [35]. Following the recent discoveries that ELF3 is involved with thermal response [14–16], we confirmed that ELF3 polyQ variation also affects thermal response phenotypes in a background-dependent fashion. However, we found little support for the hypothesis that the polyQ tract has a special role in temperature sensing. Instead, as was the case for other ELF3-dependent phenotypes, ELF3 polyQ variation appeared to affect overall ELF3 functionality, with less functional ELF3 variants lacking robust temperature responses. However, a more exhaustive series of polyQ variants may be required for revealing polyQ-specific effects, in particular because the molecular mechanism(s) by which polyQ variation affects ELF3 functionality remain unknown. ELF3-PIF4 relationship in thermomorphogenesis One question that remains unanswered is to what extent ELF3 participates in PIF4-dependent thermoresponsive morphologies. While our study and previous work [14–16] support a PIF4-ELF3 link in thermoresponsive hypocotyl elongation, this relationship disappears in the analogous case of thermoresponsive petiole elongation. These results can be explained by many hypotheses. For instance, it is possible that ELF3 regulation of PIF4 is only relevant at the early seedling stage. Another possible hypothesis is that ELF3 regulation of PIF4 in some instances is sufficient but not necessary for thermal responses. More studies are needed to understand the mechanistic details of the ELF3 and PIF4 relationship in thermomorphogenesis. Natural variation in temperature response Several studies have found that different A. thaliana strains respond to temperature differently, either shifting or inverting the reaction norm of the phenotype in question [43,49,50]. Ws has a shifted reaction norm with respect to temperature compared to Col for photoperiod-related phenotypes, including flowering. For instance, Ws displays accelerated flowering at 23°C vs. 16°C [43], but accelerates flowering no further at 27°C. Here, we show that this acceleration requires ELF3, like the elevated temperature acceleration in Col. Another example of differential mutational effects among strains is that gi mutants in the Ler background display robust thermoresponsive flowering [6,7]. It is unclear whether this finding is due to altered wiring of pathways between these backgrounds. Thermoresponsive flowering requires either PIF4 or ELF3, depending on photoperiod Under various conditions, both ELF3 and PIF4 have been found to be crucial for thermoresponsive flowering. Other members of the autonomous and the photoperiodic pathways have also been implicated in thermoresponsive flowering [6–8] (besides other pathways, [51]). Consequently, some combination of these pathways, modulated by experimental conditions, must require ELF3 and/or PIF4. We and others [11,29] have observed that PIF4 and its paralogs are not required for proper thermoresponsive flowering under longer photoperiods. Furthermore, we and others [8,47] have shown that ELF3 and the photoperiod pathway (excluding CO) are essential for proper thermoresponsive flowering under long days. It has been previously shown that PIF4 and the photoperiodic pathway contribute to thermoresponsive flowering via independent pathways [9], suggesting that under longer photoperiods PIF4 activity is inhibited, allowing other mechanisms to dominate thermoresponsive flowering. We propose a model of thermoresponsive flowering, in which PIF4, ELF3, the photoperiodic pathway, and other pathways interact depending upon condition and genetic background (Fig 5D). Under short days or other short photoperiods, phyB activity is down-regulated, leading to up-regulation of PIF4 [22,52–54], which at high levels occupies the promoter of the flowering integrator FT and induces flowering [9]. However, under longer photoperiods, phyB up-regulation leads to an attenuation of PIF4 activity, and consequently the role of PIF4 and other PIFs becomes negligible [11]. This allows canonical ambient temperature responses (such as the photoperiodic pathway, including ELF3, [8,47]) to take a dominant role in thermoresponsive flowering. Constitutive overexpression of either PIF4, PIF5, or PIF3 under long day conditions induces early flowering [30], supporting the hypothesis that differences in PIF levels underlie the photoperiod-dependence of PIF4’s role. We have not formally excluded the possibility that members of the large PIF family other than PIF4 and PIF5 might contribute to the phenotype; however, there is no evidence at present to suggest that they might [11,27,30,31]. Several reports have indicated that GI and COP1, but not CO, are involved in thermoresponsive flowering [7,8,47], with GI directly binding the FT promoter [47]. Under each of these conditions, FT-induced flowering is activated by a different signaling cascade. This interpretation leads to a coherent view of how light and temperature responses are integrated in this important plant trait. To summarize, at least three independent mechanisms have been described that promote thermoresponsive flowering in any context. These include the photoperiodic pathway (PHYB/ELF3/GI/COP1), the autonomous pathway (PHYA/FCA/FVE/TFL1/FLC), and the PIF4-dependent pathway (PIF4/H2A.Z/gibberellin), all of which converge by regulating FT (although the last pathway may also act through other integrators [29,30]). The collective results of our experiments and previous work suggest that the first two pathways are necessary but not sufficient for thermoresponsive flowering, and that the third (PIF4) is sufficient but not necessary for thermoresponsive flowering. Further study will be necessary in understanding the interdependencies of the three pathways. For instance, it has been suggested that PIF4 binding to the FT promoter is dependent on cooperativity with a second photoperiod-controlled actor [34]. In conclusion, we observe that ELF3 is involved in the hypocotyl response to elevated temperature as reported previously, and that this response can be abrogated by poorly-functioning ELF3 polyQ variants. We further demonstrate that ELF3 has little effect on the petiole temperature response, and is necessary for the flowering temperature response, suggesting that it functions independently of PIF4, potentially in the photoperiodic pathway. These results reiterate the complexity of these crucial environmental responses in plants, and can serve as a basis for further development of our understanding of how plants respond to elevated temperatures. In the context of climatic changes, this understanding will serve those attempting to secure the global food supply. Materials and Methods Plant materials and growth conditions All mutant lines (except pif4-2 elf3-200) were either described previously or obtained as T-DNA insertions from the Arabidopsis Biological Resources Center at Ohio State University [55,56], and are described in S11 Table. pif4-2 elf3-200 was obtained via crossing and genotyping. T-DNA insertions were confirmed with primers described in S10 Table. For hypocotyl assays, seedlings were grown for 15d in incubators set to SD (8h light: 16h dark days, with light supplied at 100 μmol· m-2·s-1 by cool white fluorescent bulbs) on vertical plates as described previously [35]. All plates were incubated at 22° for one day, after which one replicate arm was transferred to an incubator set to 27°, with another replicate arm maintained at 22°. For flowering time assays, plants were stratified 3-5d at 4° in 0.1% agarose and seeded into Sunshine #4 soil in 36-pot or 72-pot flats to germinate at 22° under LD (16h light: 8 hr dark days, with light supplied at 100 μmol·m-2·s-1 by cool white fluorescent bulbs). Replicate arms were subsequently transferred to 27° LD conditions as indicated, with others remaining at 22°. Different temperature treatments of the same experiment were identical with respect to randomization, setup, and format. At 25d, petiole length and whole leaf length (including petiole) of the third leaf were measured, and the ratio of these values was further analyzed. Flowering was defined as an inflorescence ≥1cm tall; at this point, date and rosette leaf number were recorded. Trait data analysis All data analysis was performed using R v3.2.1 [57]. Where indicated, temperature responses were modeled using multiple regression in the form Phenotype ~ μ + βGGenotype + βTTemperature + βGxT(Genotype x Temperature) + βEExperiment + Error. All experiments were included in models for transgenic experiments, and thus the βE term describes systematic variation between experiments, whereas line-specific effects among transgenics should be modeled in the error term. Where temperature responses are directly reported, they consist of the βT + βGxT terms and associated errors (σT2+σGxT2 where σT is the standard error for βT and σGxT2 is the standard error for βGxT), and thus are corrected for systematic experimental variation and temperature-independent genotype effects. Where p-values are reported for genotype x temperature interaction effects, they test the null hypothesis that the βGxT term above is equal to zero, and have been subjected to a Bonferroni correction to adjust for multiple comparisons. Analysis scripts and data are provided at https://figshare.com/articles/elf3_pif4_data_code_v2/3398353. Gene expression analyses Seedlings were grown for 1d under LD at 22°, after which one replicate arm was transferred to LD at 27°, with another replicate arm remaining at 22°, and all seedlings were harvested 6d later at indicated times. At harvest, ~30mg aerial tissue of pooled seedlings was flash-frozen immediately in liquid nitrogen and stored at -80°. RNA extraction, cDNA synthesis, and real-time quantitative PCR were performed as described previously [35], using primers in S10 Table. Transcript levels were quantified using the means of technical triplicates across at least two biological replicates using the ΔΔCt method, assuming 100% primer efficiency [58]. Supporting Information S1 Fig Expression analysis of PIF4 and AtHB2 depends on temperature, genetic background, and ELF3 functionality. Error bars represent the standard deviation across two biological replicates. White bars represent 22° expression, red bars 27° expression for each line. Tissue was collected from 7d seedlings at ZT0. (TIF) Click here for additional data file. S2 Fig Regulation of adult thermoresponsive traits by ELF3 and FCA is independent of PIF4 and modulated by genetic background. Flowering temperature response of indicated genotypes under indicated conditions, measured by petiole length to leaf length ratio at 25 days or rosette leaf number (RLN) at flowering. For each experiment, n > 10 plants for each genotype in each treatment. Outliers (defined as >1.5 interquartile ranges away from the median) of each distribution are indicated as points. Regression analysis of data in S4 and S5 Tables. (TIF) Click here for additional data file. S3 Fig Regulation of adult thermoresponsive petiole elongation traits occurs principally through PIF4. Petiole elongation temperature response of indicated genotypes under indicated conditions, measured by ratio of petiole length to leaf length at 25d. For each experiment, n > 10 plants for each genotype in each treatment. This experiment was repeated with similar results. Outliers (defined as >1.5 interquartile ranges away from the median) of each distribution are indicated as points. Regression analysis of data in S8 Table. (TIF) Click here for additional data file. S4 Fig Expression of FT in 7d seedlings responds to temperature and elf3 status. White bars represent 22° expression, red bars 27° expression for each line. Tissue was collected from 7d seedlings at ZT0. Error bars indicate SEM across three biological replicates. (TIF) Click here for additional data file. S1 Table Regression analysis of hypocotyl elongation temperature response among Col and Ws transgenic lines. (XLSX) Click here for additional data file. S2 Table Regression analysis of petiole: leaf length ratio and rosette leaf number at flowering temperature response among Col and Ws transgenic lines. (XLSX) Click here for additional data file. S3 Table Regression analysis of rosette leaf number at flowering and petiole: leaf length ratio temperature responses in elf3 and pif4. (XLSX) Click here for additional data file. S4 Table Regression analysis of rosette leaf number at flowering temperature response in Ws and elf3-4. (XLSX) Click here for additional data file. S5 Table Regression analysis of petiole: leaf length ratio temperature response in Col and fca mutants. (XLSX) Click here for additional data file. S6 Table Regression analysis of rosette leaf number at flowering temperature response in elf3 pif4 double mutants. (XLSX) Click here for additional data file. S7 Table Regression analysis of rosette leaf number at flowering temperature response in pif4 pif5 double mutants. (XLSX) Click here for additional data file. S8 Table Regression analysis of the petiole elongation temperature response in pif4 pif5 double mutants. (XLSX) Click here for additional data file. S9 Table Regression analysis of rosette leaf number at flowering and petiole: leaf length ratio temperature responses in flowering pathway mutants. (XLSX) Click here for additional data file. S10 Table Primers used in this study. (XLSX) Click here for additional data file. S11 Table Mutant lines used in this study. (XLSX) Click here for additional data file. We thank Philip Wigge and Jaehoon Jung for ideas, helpful conversations, sharing unpublished data, and comments on this manuscript. We thank Evan Eichler for use of the LightCycler instrument. We thank members of the Queitsch lab for helpful discussions. This work was supported by NIH New Innovator Award DP2OD008371 and NSF MCB-1516701 to CQ. ==== Refs References 1 Battisti DS , Naylor RL . Historical warnings of future food insecurity with unprecedented seasonal heat . Science . 2009 ;323 : 240 –4 . 10.1126/science.1164363 19131626 2 Quint M , Delker C , Franklin KA , Wigge PA , Halliday KJ , van Zanten M . Molecular and genetic control of plant thermomorphogenesis . Nat Plants . Nature Publishing Group; 2016 ;2 : 15190 10.1038/nplants.2015.190 27250752 3 Gould PD , Locke JCW , Larue C , Southern MM , Davis SJ , Hanano S , et al The molecular basis of temperature compensation in the Arabidopsis circadian clock . Plant Cell . 2006 ;18 : 1177 –87 . 10.1105/tpc.105.039990 16617099 4 Edwards KD , Anderson PE , Hall A , Salathia NS , Locke JCW , Lynn JR , et al FLOWERING LOCUS C mediates natural variation in the high-temperature response of the Arabidopsis circadian clock . Plant Cell . 2006 ;18 : 639 –50 . 10.1105/tpc.105.038315 16473970 5 Thines B , Harmon FG . Ambient temperature response establishes ELF3 as a required component of the core Arabidopsis circadian clock . Proc Natl Acad Sci U S A . 2010 ;107 : 3257 –62 . 10.1073/pnas.0911006107 20133619 6 Blázquez MA , Ahn JH , Weigel D . A thermosensory pathway controlling flowering time in Arabidopsis thaliana . Nat Genet . 2003 ;33 : 168 –71 . 10.1038/ng1085 12548286 7 Balasubramanian S , Sureshkumar S , Lempe J , Weigel D . Potent induction of Arabidopsis thaliana flowering by elevated growth temperature . PLoS Genet . 2006 ;2 : e106 10.1371/journal.pgen.0020106 16839183 8 Strasser B , Alvarez MJ , Califano A , Cerdán PD . A complementary role for ELF3 and TFL1 in the regulation of flowering time by ambient temperature . Plant J Cell Mol Biol . 2009 ;58 : 629 –40 . 10.1111/j.1365-313X.2009.03811.x 9 Kumar SV , Lucyshyn D , Jaeger KE , Alós E , Alvey E , Harberd NP , et al Transcription factor PIF4 controls the thermosensory activation of flowering . Nature . 2012 ;484 : 242 –245 . 10.1038/nature10928 22437497 10 Song YH , Ito S , Imaizumi T . Flowering time regulation: photoperiod- and temperature-sensing in leaves . Trends Plant Sci . 2013 ;18 : 575 –583 . 10.1016/j.tplants.2013.05.003 23790253 11 Koini MA , Alvey L , Allen T , Tilley CA , Harberd NP , Whitelam GC , et al High temperature-mediated adaptations in plant architecture require the bHLH transcription factor PIF4 . Curr Biol CB . 2009 ;19 : 408 –13 . 10.1016/j.cub.2009.01.046 19249207 12 Johansson H , Jones HJ , Foreman J , Hemsted JR , Stewart K , Grima R , et al Arabidopsis cell expansion is controlled by a photothermal switch . Nat Commun . Nature Publishing Group; 2014 ;5 : 4848 10.1038/ncomms5848 25258215 13 Lee H-J , Jung J-H , Cortés Llorca L , Kim S-G , Lee S , Baldwin IT , et al FCA mediates thermal adaptation of stem growth by attenuating auxin action in Arabidopsis . Nat Commun . 2014 ;5 : 5473 10.1038/ncomms6473 25400039 14 Box MS , Huang BE , Domijan M , Jaeger KE , Khattak AK , Yoo SJ , et al ELF3 Controls Thermoresponsive Growth in Arabidopsis . Curr Biol . 2015 ;25 : 194 –199 . 10.1016/j.cub.2014.10.076 25557663 15 Mizuno T , Nomoto Y , Oka H , Kitayama M , Takeuchi A , Tsubouchi M , et al Ambient temperature signal feeds into the circadian clock transcriptional circuitry through the EC night-time repressor in Arabidopsis thaliana . Plant Cell Physiol . 2014 ;55 : 958 –76 . 10.1093/pcp/pcu030 24500967 16 Raschke A , Ibañez C , Ullrich KK , Anwer MU , Becker S , Glöckner A , et al Natural variants of ELF3 affect thermomorphogenesis by transcriptionally modulating PIF4-dependent auxin response genes . BMC Plant Biol . 2015 ;15 : 197 10.1186/s12870-015-0566-6 26269119 17 Nusinow DA , Helfer A , Hamilton EE , King JJ , Imaizumi T , Schultz TF , et al The ELF4-ELF3-LUX complex links the circadian clock to diurnal control of hypocotyl growth . Nature . 2011 ;475 : 398 –402 . 10.1038/nature10182 21753751 18 Nieto C , López-Salmerón V , Davière J-M , Prat S . ELF3-PIF4 Interaction Regulates Plant Growth Independently of the Evening Complex . Curr Biol . 2014 ;25 : 187 –193 . 10.1016/j.cub.2014.10.070 25557667 19 Nozue K , Covington MF , Duek PD , Lorrain S , Fankhauser C , Harmer SL , et al Rhythmic growth explained by coincidence between internal and external cues . Nature . 2007 ;448 : 358 –61 . 10.1038/nature05946 17589502 20 Filo J , Wu A , Eliason E , Richardson T , Thines BC , Harmon FG . Gibberellin driven growth in elf3 mutants requires PIF4 and PIF5 . Plant Signal Behav . Taylor & Francis; 2015 ; 10.4161/15592324.2014.992707 21 Lorenzo CD , Sanchez-Lamas M , Antonietti MS , Cerdán PD . Emerging Hubs in Plant Light and Temperature Signaling . Photochem Photobiol . 2016 ;92 : 3 –13 . 10.1111/php.12535 26439979 22 Lorrain S , Allen T , Duek PD , Whitelam GC , Fankhauser C . Phytochrome-mediated inhibition of shade avoidance involves degradation of growth-promoting bHLH transcription factors . Plant J Cell Mol Biol . 2008 ;53 : 312 –23 . 10.1111/j.1365-313X.2007.03341.x 23 McWatters HG , Bastow RM , Hall A , Millar AJ . The ELF3 zeitnehmer regulates light signalling to the circadian clock . Nature . Macmillian Magazines Ltd.; 2000 ;408 : 716 –20 . 10.1038/35047079 11130072 24 Covington MF , Panda S , Liu XL , Strayer CA , Wagner DR , Kay SA . ELF3 modulates resetting of the circadian clock in Arabidopsis . Plant Cell . 2001 ;13 : 1305 –15 . 11402162 25 Zagotta MT , Hicks KA , Jacobs CI , Young JC , Hangarter RP , Meeks-Wagner DR . The Arabidopsis ELF3 gene regulates vegetative photomorphogenesis and the photoperiodic induction of flowering . Plant J . 1996 ;10 : 691 –702 . 10.1046/j.1365-313X.1996.10040691.x 8893545 26 Leivar P , Quail PH . PIFs: pivotal components in a cellular signaling hub . Trends Plant Sci . 2011 ;16 : 19 –28 . 10.1016/j.tplants.2010.08.003 20833098 27 Proveniers MCG , van Zanten M . High temperature acclimation through PIF4 signaling . Trends Plant Sci . Elsevier; 2013 ;18 : 59 –64 . 10.1016/j.tplants.2012.09.002 23040086 28 Bai M-Y , Shang J-X , Oh E , Fan M , Bai Y , Zentella R , et al Brassinosteroid, gibberellin and phytochrome impinge on a common transcription module in Arabidopsis . Nat Cell Biol . Nature Publishing Group, a division of Macmillan Publishers Limited. All Rights Reserved.; 2012 ;14 : 810 –7 . 10.1038/ncb2546 22820377 29 Thines BC , Youn Y , Duarte MI , Harmon FG . The time of day effects of warm temperature on flowering time involve PIF4 and PIF5 . J Exp Bot . 2014 ;65 : 1141 –1151 . 10.1093/jxb/ert487 24574484 30 Galvão VC , Collani S , Horrer D , Schmid M . Gibberellic acid signaling is required for ambient temperature-mediated induction of flowering in Arabidopsis thaliana . Plant J Cell Mol Biol . 2015 ;84 : 949 –62 . 10.1111/tpj.13051 31 Fernández V , Takahashi Y , Le Gourrierec J , Coupland G . Photoperiodic and thermosensory pathways interact through CONSTANS to promote flowering at high temperature under short days . Plant J . 2016 ;86 : 426 –440 . 10.1111/tpj.13183 27117775 32 Wigge PA . Ambient temperature signalling in plants . Curr Opin Plant Biol . 2013 ;16 : 661 –6 . 10.1016/j.pbi.2013.08.004 24021869 33 Liu J , Feng L , Li J , He Z . Genetic and epigenetic control of plant heat responses . Front Plant Sci . Frontiers; 2015 ;06 10.3389/fpls.2015.00267 34 Seaton DD , Smith RW , Song YH , MacGregor DR , Stewart K , Steel G , et al Linked circadian outputs control elongation growth and flowering in response to photoperiod and temperature . Mol Syst Biol . EMBO Press; 2015 ;11 : 776 10.15252/msb.20145766 25600997 35 Undurraga SF , Press MO , Legendre M , Bujdoso N , Bale J , Wang H , et al Background-dependent effects of polyglutamine variation in the Arabidopsis thaliana gene ELF3 . Proc Natl Acad Sci U S A . 2012 ;109 : 19363 –7 . 10.1073/pnas.1211021109 23129635 36 Uversky VN . Intrinsically disordered proteins and their environment: effects of strong denaturants, temperature, pH, counter ions, membranes, binding partners, osmolytes, and macromolecular crowding . Protein J . 2009 ;28 : 305 –25 . 10.1007/s10930-009-9201-4 19768526 37 Sawyer LA . Natural Variation in a Drosophila Clock Gene and Temperature Compensation . Science . 1997 ;278 : 2117 –2120 . 10.1126/science.278.5346.2117 9405346 38 Gemayel R , Vinces MD , Legendre M , Verstrepen KJ . Variable tandem repeats accelerate evolution of coding and regulatory sequences . Annu Rev Genet . 2010 ;44 : 445 –77 . 10.1146/annurev-genet-072610-155046 20809801 39 Mizuno T , Kitayama M , Oka H , Tsubouchi M , Takayama C , Nomoto Y , et al The EC night-time repressor plays a crucial role in modulating circadian clock transcriptional circuitry by conservatively double-checking both warm-night and night-time-light signals in a synergistic manner in Arabidopsis thaliana . Plant Cell Physiol . 2014 ;55 : 2139 –51 . 10.1093/pcp/pcu144 25332490 40 Nomoto Y , Nomoto Y , Kubozono S , Miyachi M , Yamashino T , Nakamichi N , et al A circadian clock- and PIF4-mediated double coincidence mechanism is implicated in the thermosensitive photoperiodic control of plant architectures in Arabidopsis thaliana . Plant Cell Physiol . 2012 ;53 : 1965 –73 . 10.1093/pcp/pcs141 23037004 41 Mizuno T , Kitayama M , Takayama C , Yamashino T . Insight into a Physiological Role for the EC Night-Time Repressor in the Arabidopsis Circadian Clock . Plant Cell Physiol . 2015 ;56 : 1738 –47 . 10.1093/pcp/pcv094 26108788 42 van Zanten M , Voesenek LACJ , Peeters AJM , Millenaar FF . Hormone- and light-mediated regulation of heat-induced differential petiole growth in Arabidopsis . Plant Physiol . 2009 ;151 : 1446 –58 . 10.1104/pp.109.144386 19741046 43 Ibañez C , Poeschl Y , Peterson T , Bellstädt J , Denk K , Gogol-Döring A , et al Developmental plasticity of Arabidopsis thaliana accessions across an ambient temperature range . bioRxiv . Cold Spring Harbor Labs Journals; 2015 ; 017285 10.1101/017285 44 Stavang JA , Gallego-Bartolomé J , Gómez MD , Yoshida S , Asami T , Olsen JE , et al Hormonal regulation of temperature-induced growth in Arabidopsis . Plant J Cell Mol Biol . 2009 ;60 : 589 –601 . 10.1111/j.1365-313X.2009.03983.x 45 Sun J , Qi L , Li Y , Zhai Q , Li C . PIF4 and PIF5 transcription factors link blue light and auxin to regulate the phototropic response in Arabidopsis . Plant Cell . 2013 ;25 : 2102 –14 . 10.1105/tpc.113.112417 23757399 46 Sawa M , Kay SA . GIGANTEA directly activates Flowering Locus T in Arabidopsis thaliana . Proc Natl Acad Sci U S A . 2011 ;108 : 11698 –703 . 10.1073/pnas.1106771108 21709243 47 Jang K , Gil Lee H , Jung S-J , Paek N-C , Joon Seo P . The E3 Ubiquitin Ligase COP1 Regulates Thermosensory Flowering by Triggering GI Degradation in Arabidopsis . Sci Rep . Nature Publishing Group; 2015 ;5 : 12071 10.1038/srep12071 26159740 48 Kim W-Y , Hicks KA , Somers DE . Independent roles for EARLY FLOWERING 3 and ZEITLUPE in the control of circadian timing, hypocotyl length, and flowering time . Plant Physiol . 2005 ;139 : 1557 –69 . 10.1104/pp.105.067173 16258016 49 Edwards KD , Lynn JR , Gyula P , Nagy F , Millar AJ . Natural allelic variation in the temperature-compensation mechanisms of the Arabidopsis thaliana circadian clock . Genetics . 2005 ;170 : 387 –400 . 10.1534/genetics.104.035238 15781708 50 Edwards KD , Guerineau F , Devlin PF , Millar AJ . Low-temperature-specific effects of PHYTOCHROME C on the circadian clock in Arabidopsis suggest that PHYC underlies natural variation in biological timing . bioRxiv . Cold Spring Harbor Labs Journals; 2015 ; 030577 10.1101/030577 51 Lee JH , Yoo SJ , Park SH , Hwang I , Lee JS , Ahn JH . Role of SVP in the control of flowering time by ambient temperature in Arabidopsis . Genes Dev . 2007 ;21 : 397 –402 . 10.1101/gad.1518407 17322399 52 Duek PD , Fankhauser C . bHLH class transcription factors take centre stage in phytochrome signalling . Trends Plant Sci . 2005 ;10 : 51 –4 . 10.1016/j.tplants.2004.12.005 15708340 53 Huq E , Quail PH . PIF4, a phytochrome-interacting bHLH factor, functions as a negative regulator of phytochrome B signaling in Arabidopsis . EMBO J . EMBO Press; 2002 ;21 : 2441 –50 . 10.1093/emboj/21.10.2441 12006496 54 Luo Q , Lian H-L , He S-B , Li L , Jia K-P , Yang H-Q . COP1 and phyB Physically Interact with PIL1 to Regulate Its Stability and Photomorphogenic Development in Arabidopsis . Plant Cell . 2014 ;26 : 2441 –2456 . 10.1105/tpc.113.121657 24951480 55 Alonso JM , Stepanova AN , Leisse TJ , Kim CJ , Chen H , Shinn P , et al Genome-wide insertional mutagenesis of Arabidopsis thaliana . Science . 2003 ;301 : 653 –7 . 10.1126/science.1086391 12893945 56 Kleinboelting N , Huep G , Kloetgen A , Viehoever P , Weisshaar B . GABI-Kat SimpleSearch: new features of the Arabidopsis thaliana T-DNA mutant database . Nucleic Acids Res . 2012 ;40 : D1211 –5 . 10.1093/nar/gkr1047 22080561 57 R Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria; 2016. 58 Pfaffl MW . A new mathematical model for relative quantification in real-time RT-PCR . Nucleic Acids Res . 2001 ;29 : e45 11328886

PMC005xxxxxx/PMC5001699.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756432010.1371/journal.pone.0161774PONE-D-16-11044Research ArticleSocial SciencesEconomicsHealth EconomicsHealth InsuranceMedicine and Health SciencesHealth CareHealth EconomicsHealth InsuranceMedicine and Health SciencesPublic and Occupational HealthPeople and PlacesGeographical LocationsAsiaChinaMedicine and Health SciencesHealth CareSocioeconomic Aspects of HealthMedicine and Health SciencesPublic and Occupational HealthSocioeconomic Aspects of HealthSocial SciencesSociologyEducationSchoolsMedicine and Health SciencesHealth CareHealth Care PolicyPeople and PlacesPopulation GroupingsAge GroupsElderlyEarth SciencesGeographyGeographic AreasUrban AreasDeterminants of Health Insurance Coverage among People Aged 45 and over in China: Who Buys Public, Private and Multiple Insurance Determinants of Health Insurance Coverage in ChinaJin Yinzi 1Hou Zhiyuan 2*Zhang Donglan 31 China Center for Health Development Studies, Peking University. 38 Xue Yuan Road, Haidian District, Beijing 100191, China2 Department of Social Medicine, School of Public Health, National Key Laboratory of Health Technology Assessment (Ministry of Health), Collaborative Innovation Center of Social Risks Governance in Health, Fudan University, 138 Yi Xue Yuan Road, Shanghai 200032, China3 Department of Health Policy and Management, College of Public Health, University of Georgia, 100 Foster Road, Wright Hall, Athens, GA, United States of AmericaZhang Harry EditorOld Dominion University, UNITED STATESCompeting Interests: The authors have declared that no competing interests exist. Conceptualization: ZH DZ. Formal analysis: ZH. Funding acquisition: ZH. Methodology: ZH DZ. Writing – original draft: YJ ZH. Writing – review & editing: DZ ZH. * E-mail: zyhou@fudan.edu.cn26 8 2016 2016 11 8 e016177416 3 2016 11 8 2016 © 2016 Jin et al2016Jin et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.Background China is reforming and restructuring its health insurance system to achieve the goal of universal coverage. This study aims to understand the determinants of public, private and multiple insurance coverage among people of retirement-age in China. Methods We used data from the China Health and Retirement Longitudinal Survey 2011 and 2013, a nationally representative survey of Chinese people aged 45 and over. Multinomial logit regression was performed to identify the determinants of public, private and multiple health insurance coverage. We also conducted logit regression to examine the association between public insurance coverage and demand for private insurance. Results In 2013, 94.5% of this population had at least one type of public insurance, and 12.2% purchased private insurance. In general, we found that rural residents were less likely to be uninsured (Relative Risk Ratio (RRR) = 0.40, 95% Confidence Interval (CI): 0.34–0.47) and were less likely to buy private insurance (RRR = 0.22, 95% CI: 0.16–0.31). But rural-to-urban migrants were more likely to be uninsured (RRR = 1.39, 95% CI: 1.24–1.57). Public health insurance coverage may crowd out private insurance market (Odds Ratio = 0.55, 95% CI: 0.48–0.63), particularly among enrollees of Urban Resident Basic Medical Insurance. There exists a huge socioeconomic disparity in both public and private insurance coverage. Conclusion The migrants, the poor and the vulnerable remained in the edge of the system. The growing private insurance market did not provide sufficient financial protection and did not cover the people with the greatest need. To achieve universal coverage and reduce socioeconomic disparity, China should integrate the urban and rural public insurance schemes across regions and remove the barriers for the middle-income and low-income to access private insurance. National Nature Science Foundation of China71403007Hou Zhiyuan Postdoctoral Science Foundation of China2013T60046Hou Zhiyuan Postdoctoral Science Foundation of China2012M520132Hou Zhiyuan This work was supported by National Nature Science Foundation (71403007) and Postdoctoral Science Foundation of China (2012M520132, 2013T60046) for ZH. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data AvailabilityThe data used in this paper is from the China Health and Retirement Longitudinal Study (CHARLS), which is publicly available. The CHARLS data can be downloaded from the following website: http://charls.ccer.edu.cn/en.Data Availability The data used in this paper is from the China Health and Retirement Longitudinal Study (CHARLS), which is publicly available. The CHARLS data can be downloaded from the following website: http://charls.ccer.edu.cn/en. ==== Body Introduction China’s public health insurance system covers the largest number of population in the world [1], however to be precise, the system itself is fragmented rather than integrated. At present, there are three major types of public insurance—the New Rural Cooperative Medical Scheme (NCMS), the Urban Employee Basic Medical Insurance (UEBMI), and the Urban Resident Basic Medical Insurance (URBMI), covering 95% of the entire population in China [2]. The three insurance is designed according to the permanent residence registration system (aka the “Hukou” system) and/or the person’s employment status. The Hukou System classifies people as rural or urban residents based on their places of birth, which is not easily transferable from rural to urban residence [3]. For example, due to the Hukou System, rural people migrating and living in urban areas may still remain in their rural residence status. The three insurance targets different populations, and Hukou becomes the main barrier to shifting across these insurance plans. NCMS, targeting the registered rural population, is a “semi-mandatory” health insurance scheme. It was established in 2003 and administered by the Ministry of Health (currently the State Health and Family Planning Commission of China). NCMS pools its fund at the county level in the rural areas, whereby each rural county can design the benefit package and implementation plan, following the guidelines issued by the national and provincial governments.[4] UEBMI, targeting the urban employees, is legally mandatory that all employers must provide medical insurance to their employees. It was transformed from the previous labor medical insurance during China’s economic reform in 1990s.[5] URBMI, targeting the urban non-employee residents including children and adolescents under age 18, is a voluntary health insurance scheme that has been established since 2007 [6]. The UEBMI and URBMI are both administered by the Ministry of Human Resources and Social Security. UEBMI and URBMI pool their funds at the municipal level in the urban areas.[7] The three public insurance varies significantly in the amount of premiums, benefit packages and reimbursement rates for different health care services. Among the three insurance, UEBMI offers the most generous benefit packages, while NCMS is often considered the most rudimentary type of insurance [8]. In 2013, the annual premium of NCMS and URBMI was about 350 Chinese Renminbi (RMB) per capita, but the annual premium of UEBMI reached 2500 RMB per capita [9]. NCMS only covers 800–1200 types of medications, but UEBMI and URBMI cover more than 2100 types of medications, almost doubled the coverage of NCMS. UEBMI and NCMS reimburse both outpatient and inpatient care, whereas URBMI mainly provides reimbursement for inpatient care [10]. The reimbursement rates for inpatient care reached 50.1%, 53.6% and 68.8% respectively for NCMS, URBMI and UEBMI in 2013 [11]. However, preventive care services and luxury medical services such as high technologies, cosmetic and plastic surgery, as well as services covered by other insurance schemes (e.g., injury protection insurance and maternity insurance) are not reimbursed through any of the three public insurance schemes [7]. There are large variations in the economic development and population size across the rural counties and urban cities. Lack of integration among these three schemes results in non-portability across insurance schemes and geographic areas, which have left certain population without insurance and some with multiple insurance coverage [2, 12]. In particular, a growing number of migrant workers are rushing into urban cities from rural regions [13]. Since public health insurance is pooled and administered at local county- or municipal- level, insurance agencies often designate local health care facilities as their coverage network; in other words, health services received out of the local county or city are possibly not reimbursed by local public insurance. Thus, migrants have incentives to shift their insurance or participate in multiple insurance to obtain reimbursement. Although those rural-to-urban migrants are eligible for both rural and urban public health insurance, in 2014, about 14.9% of migrants did not have any health insurance [14]. But at the same time, some migrants were covered by multiple insurance. A previous study showed that, from 2007 to 2010, 2.9% of migrant workers had been covered by two types of insurance, among whom 1.5% enrolled in NCMS and URBMI, and 1.2% in NCMS and UEBMI [15]. Given that there were 253 million migrants in 2014, it is estimated that 7.3 million have repeatedly enrolled in public health insurance and get duplicated financial subsidies from governments [15, 16]. Therefore, it is important to understand the determinants of insurance coverage particularly among the migrant population. Nevertheless, there has been no research focusing on this issue in China. Another research gap is about the demand for private health insurance. Although public health insurance has covered the majority of population in mainland China, it only offers limited financial protection [17]. While with the rising challenges of chronic diseases [18], private health insurance plays an increasingly critical role to fill the coverage gap and meet the various health care needs of the population [19]. Since formally recognized and regulated by the Chinese government in 1998, private health insurance has remained slow in its development [20]. In 2009, China initiated a national healthcare system reform, and one of the reformed policies was to promote the development of private health insurance. The Chinese government encouraged the insurance companies to provide various insurance products, especially for the elderly, disabled and those with catastrophic diseases, and it also encouraged individuals to purchase these insurance products. During the healthcare reform, various strategies have been released in order to promote its development, such as the exemption of business tax for private insurers, preferential tax rate of health insurance for individuals and allowing UEBMI enrollees to purchase private health insurance with individual medical savings accounts [21]. Private insurers were also allowed to manage the public health insurance plans. It is reported that there were about 100 private insurers offering over 1500 health insurance products in 2012 [17]. The products mainly covered catastrophic diseases with little reimbursement for services like outpatient or nursing home care [22]. The premium from private health insurance had increased from 3.6 billion to 69.2 billion RMB from 1999 to 2011, and its share of the total health care expenditures also increased from 0.9% to 3.4% [23]. Private health insurance is available in both rural and urban China, targeting mainly the high-income population. According to the National Health Services Survey [11], 6.9% of the Chinese population purchased private health insurance in 2013. During the period of 2008 to 2013, the proportion of the population who purchased private insurance had increased from 6.9% to 7.7% in urban areas, but decreased from 6.9% to 6.1% in rural areas. While most studies to date focused on public health insurance in China, few literature has ever examined the private health insurance coverage as well as the dual coverage of public and private health insurance [22]. This paper is among the first to investigate demands of private health insurance among the population of retirement age in China. The present study examined the determinants of public, private and multiple insurance coverage among the Chinese people aged 45 and older. We used the China Health and Retirement Longitudinal Study (CHARLS) to answer two research questions: 1) what are the characteristics of those who were covered by public, private, multiple urban and rural public health insurance, as well as private and public health insurance? And 2) what is the association between public insurance coverage and the demand for private health insurance? Methods Study design and data We used the 2011 and 2013 panel of the CHARLS data, a nationally representative sample of Chinese people aged 45 and over, and their spouses [24]. It was modeled after the Health and Retirement Study in the United States, covering questions on demographics, socioeconomic status, health insurance, and health status and health behaviors. Age 45 was selected as the cutoff age by the CHARLS was because it is the minimum retirement age (the minimum age for receiving pension) in China [25]. We were particularly interested in this middle-aged and elderly population because with the rising prevalence of chronic diseases among this population, they had a higher demand for health insurance [26]. The details of the data have been described elsewhere previously [27]. Ethical consent was approved by the Institutional Review Board of Peking University. The CHARLS national sample was drawn using the stratified four-stage cluster sampling method. In the first sampling stage, 150 county-level units (rural counties or urban districts) were randomly selected by Probability Proportional to Size (PPS) from a sampling frame containing all county-level units of mainland China excluding those in Tibet. Within each county-level unit, 3 primary sampling units (PSUs)–administrative villages in rural areas or neighborhoods in urban areas–were randomly selected by PPS. Within each PSU, 24 households with members aged 45 or older were randomly selected. The member aged 45 or older and his or her spouse (if present) were interviewed face-to-face in each household.[24] The national baseline survey was conducted between June 2011 and March 2012, having a total sample of 17,705 respondents. These respondents were followed up in 2013. The second national survey was finished in 2013, including a total of 18,605 respondents. In this study, we used the CHARLS longitudinal data from 2011 and 2013 waves. Measurements Dependent variables The dependent variable in our analysis was a categorically distributed variable with five discrete outcomes: public insurance only, private insurance only, double coverage of public and private health insurance, double coverage of both rural (NCMS) and urban public health insurance (URBMI or UEBMI) and no insurance. Independent variables The independent variables were carefully chosen based on the Andersen’s Behavioral Model of Health Services Use, a widely-used conceptual framework to investigate demand for health care and insurance [28]. We included variables of five dimensions: policy, health, socioeconomic, risk aversion, and other confounding factors. Policy-related factors: The policy factors reflected the influence of the residence registration system (Hukou) and health insurance policy on the demand for different health insurance. Variables included registration as rural or urban residents and migration status that are closely related to the regional health insurance policy. Migrants are defined as those who live in regions where they are not registered as local residence (Hukou), which may limit their access to local public health insurance and social welfare [13]. Health-related factors: Health status was measured by self-reported health, chronic diseases, and respondents’ age. Self-reported health variable was divided into two groups: good health, and poor/fair health. Chronic diseases were measured by the question “have you been diagnosed with the following chronic conditions by a doctor: hypertension, dyslipidemia, diabetes or high blood sugar, cancer or malignant tumor, cardiovascular disease, stroke, chronic lung diseases, liver disease, kidney disease, stomach or other digestive disease, arthritis or rheumatism, asthma, emotional or psychiatric problems, memory-related disease”. Socioeconomic factors: Variables included household annual income per capita, educational attainment, and employment status. Income was divided into quintiles: the lowest, lower, middle, higher, and highest income groups. Respondents’ education levels were categorized into four groups: no formal education, not complete primary school but capable of reading/writing, completed primary school and junior high school and above. Employment status also contained four categories: farmers, unemployed or retired, informal, and formal employed. Risk aversion factors: Respondents’ risk aversion was measured by whether had health checkup during the past year and smoking status (never or past smokers vs. current smokers). Other confounding factors: These factors included gender, marital status (unmarried, widowed, separated or divorced vs. currently married), number of family members in the household, survey year and province-level fixed effects. Number of family members in the household was categorized into 3 groups: 1–2, 3–4, and 5 or more. Statistical analysis Descriptive analyses were first performed to show the discrete outcomes of public, private health insurance and multiple insurance coverages, as well as the sociodemographic characteristics of the survey respondents. Using pooled data from the CHARLS 2011 and 2013, multinomial logit regression was performed to identify the determinants of public, private and multiple health insurance coverages. To determine whether the model meets the irrelevant alternative assumption (IIA) that the relative probability of choosing among two existing choices is unaffected by the addition or deletion of another alternatives, we performed the Hausman test [29]. The result of the test was insignificant and the coefficients did not change when we eliminated one of the categories, thus the IIA assumption was not violated. All regressions included province dummies to adjust for unobserved province-level fixed effects, and were adjusted for correlation at the individual level between the two panels. We performed regression analysis with the total sample and separately with the rural and urban subsamples. We further estimated the association between public health insurance coverage and the demand for private health insurance. Relative Risk Ratios (RRRs) and Odds ratios (ORs) with 95% confidence intervals (CIs) were reported. Survey weights were applied to account for the multi-stage stratified sampling design in both descriptive and regression analyses. The sampling weights took into consideration of the representativeness of the estimates, and the household and individual non-response biases. All statistical analyses were performed using STATA 12.0 (College Station, TX: Stata Corp LP). Results Public, private and multiple health insurance coverage Table 1 showed the public, private and multiple health insurance coverage among those aged 45 and over between 2011 and 2013 in China. In 2013, about 4.7% of them were uninsured, reducing from 7.4% in 2011. This decrease of the uninsured occurred partially due to an expansion of public health insurance from 91.3% in 2011 to 94.5% in 2013, and partly attributed to a boost of private health insurance from 7.2% in 2011 to 12.2% in 2013. Among those with double/multiple insurance coverage, around 11.4% had both public and private health insurance in 2013, almost doubled from that in 2011 (5.8%). Coverage by both rural (referred to NCMS) and urban public health insurance (referred to UEBMI or URBMI) slightly increased from 0.52% in 2011 to 0.70% in 2013. 10.1371/journal.pone.0161774.t001Table 1 Public, private and multiple health insurance coverage among the population aged 45 and older in China, 2011–2013 CHARLS (%). 2011 * 2013 * Total Rural a Urban a Total Rural Urban Public health insurance only 84.91 87.65 78.03 82.40 86.19 73.38 Rural & urban public health insurance 0.52 0.55 0.45 0.70 0.68 0.72 Private health insurance only 1.36 0.47 3.64 0.86 0.37 1.97 Public & private health insurance 5.83 4.94 8.08 11.36 8.35 18.82 No health insurance 7.38 6.40 9.79 4.67 4.41 5.11 Total N 17,711 13,638 3,804 18,618 13,810 3,928 Note: All statistics adjusted for sampling weights. a Rural and urban was defined by the respondent’s registration status. * The distribution was significantly different between rural and urban areas, P<0.001 calculated from Pearson Chi2 test. We further compared the differences in health insurance coverage between rural and urban registered residents. We noticed that the percentage of people with public insurance only was significantly higher in registered rural residents than in those urban residents (P<0.001), whereas a higher percentage of urban residents purchased private health insurance and enjoyed coverage by both public and private health insurance in 2011 and 2013 panels. Characteristics of the survey respondents Table 2 presented the descriptive statistics for the study population. Overall, the majority of the people (71.51% in 2011 and 71.23% in 2013) were registered rural residents. About 33.78% in 2011 and 27.98% in 2013 of the entire population were rural-to-urban migrants. Around 75% of the sample reported poor or fair health status, with around 65% of the population having at least one chronic condition. The average age was around 60 in both panels, and about half of the respondents were females. Around 85% were currently married and the average number of family members within a household was about 3.50. Average household income increased from 8930 RMB in 2011 to 10110 RMB in 2013. More than half of the sample had finished primary school education. And about 70% of the population in both panels was either self-employed farmers or unemployed/retired. Only less than 20% of the study sample was formally employed. A small proportion of people (18.06% in 2011 and 13.47% in 2013) had health checkup during the previous year and a slightly less than 30% of the sample were current smokers. 10.1371/journal.pone.0161774.t002Table 2 Characteristics of the survey respondents, 2011–2013 CHARLS (%). Characteristics 2011 2013 Rural residents 71.51 71.23 Migrant 33.78 27.98 Poor or fair health 74.06 75.58 Having any chronic disease 66.37 62.35 Age (years)* 59.20(10.15) 60.05(10.24) Female 52.27 52.15 Currently Married 85.58 85.24 Number of family members in a household     1–2 33.37 37.38     3–4 35.35 35.11     5 or more 31.28 27.51 Household income per capita (1000 Yuan)* 8.93(11.63) 10.11(17.29) Education         No formal education 25.61 24.42         Semi-literate but can read/write 16.65 16.49         Primary school 21.31 21.6         Junior high school and above 36.44 37.48 Employment status         Farmers 37.65 37.07     Unemployed or retired 34.35 33.99         Informal employed 8.53 10.2         Formal employed 19.47 18.74 Health checkup during the past year 18.06 13.47 Currently smoking 29.97 29.80 N 17,711 18,618 *Mean, SD Note: All statistics adjusted for sampling weights. Determinants of public, private and multiple health insurance coverage Table 3 showed the determinants of health insurance coverage from the weighted multinomial logit regression. Each insurance outcome was compared to the base outcome of public insurance coverage only. As for the policy-related factors, compared to urban residents, rural registered population were significantly less likely to have both rural and urban public insurance (RRR = 0.52, 95% CI: 0.31–0.87), less likely to buy private insurance (RRR = 0.22, 95% CI: 0.16–0.31), less likely to have coverage by both public and private health insurance (RRR = 0.55, 95% CI: 0.48–0.62), and interestingly, also less likely to be uninsured (RRR = 0.40, 95% CI: 0.34–0.47). Generally, rural residents are expected to have more public insurance coverage as compared to other types of insurance. The likelihood of migrants being covered by multiple insurance, private insurance and being uninsured as compared to public insurance coverage were all significantly higher than local residents (P<0.05). As for the health-related factors, those with worse health status such as self-reported as poor or fair health and older people were less likely to purchase private insurance. They were also less likely to be uninsured as compared to public insurance coverage. Compared to males, females were less likely to be covered by both public and private health insurance (RRR = 0.89, 95% CI: 0.80–1.00). 10.1371/journal.pone.0161774.t003Table 3 Determinants of public, private and multiple health insurance coverage: Results from the multinomial logit model (Base outcome: public insurance only). Multiple coverage: rural & urban public insurance Private insurance only Multiple coverage: public & private insurance No insurance Rural residents 0.51(0.30–0.87)** 0.22(0.16–0.31)*** 0.55(0.48–0.62)*** 0.40(0.34–0.46)*** Migrant 1.85(1.26–2.71)*** 1.74(1.36–2.22)*** 1.09(0.99–1.21)* 1.39(1.24–1.57)*** Poor or fair health 0.95(0.66–1.38) 0.64(0.49–0.82)** 0.85(0.77–0.94)*** 0.89(0.79–1.01)* Any chronic disease 1.55(1.08–2.23)** 0.88(0.69–1.12) 1.05(0.96–1.16) 0.68(0.61–0.77)*** Age (10 years) 0.85(0.68–1.06) 1.07(0.92–1.25) 0.82(0.77–0.87)*** 0.82(0.76–0.88)*** Female 0.54(0.34–0.85)*** 1.11(0.83–1.49) 0.89(0.80–1.00)* 1.06(0.92–1.22) Currently Married 0.97(0.55–1.72) 1.13(0.77–1.64) 0.94(0.80–1.09) 0.52(0.45–0.60)*** Household size (referred to 1–2 family members)     3–4 0.76(0.50–1.15) 1.02(0.78–1.35) 1.15(1.04–1.29)*** 0.92(0.81–1.04)     5 or more 1.06(0.70–1.63) 0.82(0.59–1.14) 1.15(1.02–1.29)** 0.92(0.81–1.06) Income (referred to lowest income)         Lower 0.91(0.49–1.72) 0.77(0.44–1.35) 1.04(0.90–1.20) 0.78(0.68–0.90)***         Middle 1.49(0.84–2.62) 1.22(0.76–1.95) 1.02(0.88–1.17) 0.68(0.58–0.78)***         Higher 1.71(0.98–2.97)* 1.64(1.07–2.53)** 1.10(0.95–1.27) 0.68(0.58–0.78)***         Highest 1.61(0.87–2.97) 1.74(1.12–2.69)** 1.51(1.32–1.74)*** 0.45(0.37–0.54)*** Education (referred to no formal education)         No education but can read/write 0.58(0.29–1.14) 0.90(0.56–1.44) 0.83(0.71–0.97)** 0.80(0.69–0.94)***         Primary school 1.22(0.70–2.13) 0.84(0.54–1.30) 0.97(0.84–1.13) 0.65(0.55–0.76)***         Junior high school and above 1.38(0.76–2.53) 1.13(0.75–1.71) 1.23(1.07–1.42)*** 0.47(0.40–0.56)*** Employment status (referred to farmers)         Unemployed or retired 1.39(0.85–2.28) 1.75(1.19–2.58)*** 0.99(0.88–1.12) 1.40(1.22–1.60)***         Informal employed 1.67(1.00–2.79)* 1.77(1.12–2.79)** 1.00(0.85–1.17) 1.90(1.61–2.25)***         Formal employed 1.76(1.10–2.81)** 1.05(0.69–1.60) 1.05(0.93–1.20) 1.24(1.05–1.47)** Health checkup 1.68(1.16–2.42)*** 0.97(0.71–1.32) 1.16(1.04–1.30)** 0.76(0.66–0.87)*** Currently smoking 0.78(0.53–1.16) 0.95(0.71–1.28) 0.88(0.79–0.99)** 1.08(0.94–1.23) 2013 1.81(1.33–2.44)*** 0.55(0.43–0.69)*** 2.11(1.95–2.29)*** 0.59(0.54–0.65)*** Constant 0.01(0.00–0.12)*** 0.04(0.01–0.20)*** 0.54(0.27–1.08)* 0.98(0.27–3.61) Pseudo R2 0.10 Observations 35,068 Notes: Relative Risk Ratios and 95% confidence intervals were shown. Multinomial logit model was used with public health insurance as the reference group. All models included sampling weights, province dummy and adjusted for clustering at the individual level. Significance level *** p<0.01 ** p<0.05 * p<0.10. As expected, income and education levels were significantly associated with types of insurance coverage. If income level had increased from lowest to highest, the relative risk for private insurance purchase and coverage by both public and private insurance relative to public insurance only would be expected to increase by a factor of 1.74 (95% CI: 1.12–2.69) and 1.51 (95% CI: 1.32–1.74) respectively, whereas the relative risk for no insurance relative to public insurance only would be expected to decrease by a factor of 0.45 (95% CI: 0.37–0.54). A similar pattern was observed when education level had increased from no formal education to junior high school and above. Employment status was an important predictor of insurance coverage. Compared to farmers, the relative risk for coverage by both urban and rural insurance relative to public insurance only was significantly higher among those informally or formally employed, while they were also more likely to be uninsured (RRR = 1.90, 95% CI: 1.61–2.25 for informally employed; RRR = 1.24, 95% CI: 1.05–1.47 for formally employed). We also found evidence that having health checkups, an indicator of risk aversion, was positively associated with multiple insurance coverage, and negatively associated with being uninsured (P<0.001). We further examined the determinants of health insurance coverage among rural and urban residents only (Table 4). The sample for one outcome “multiple coverage of rural & urban public insurance” was not large enough for analysis in the rural or urban subsamples, and was then excluded from these regressions. The results in the rural and urban subsamples were similar to those in the total sample. Among both subsamples, migrants were less likely to participate in public insurance, but were more likely to purchase private insurance. Compared to public insurance, the likelihood of urban residents purchasing only private insurance decreased significantly in 2013 compared to 2011 (RRR = 0.43, 95% CI: 0.32–0.58), but there was no significant change among the rural residents. 10.1371/journal.pone.0161774.t004Table 4 Determinants of public, private and multiple health insurance coverage among rural/ urban residents: Results from the multinomial logit model (Base outcome: public insurance only). Rural residents Urban residents Private insurance only Multiple coverage: public & private insurance No insurance Private insurance only Multiple coverage: public & private insurance No insurance Migrant 1.39 1.13* 1.37*** 2.32*** 0.91 1.31** (0.91–2.12) (1.00–1.27) (1.19–1.58) (1.70–3.16) (0.75–1.11) (1.04–1.64) Poor or fair health 0.53*** 0.80*** 0.85** 0.70** 0.94 1.08 (0.35–0.80) (0.71–0.89) (0.74–0.98) (0.51–0.96) (0.80–1.11) (0.87–1.34) Any chronic disease 0.99 1.00 0.75*** 0.70** 1.08 0.55*** (0.68–1.44) (0.90–1.12) (0.66–0.85) (0.51–0.95) (0.92–1.28) (0.44–0.68) Age (10 years) 0.83 0.74*** 0.86*** 1.30*** 0.82*** 0.75*** (0.64–1.07) (0.69–0.80) (0.79–0.93) (1.08–1.57) (0.74–0.91) (0.65–0.86) Female 0.83 0.85** 1.12 1.44** 0.91 1.18 (0.47–1.46) (0.73–0.99) (0.94–1.35) (1.01–2.04) (0.76–1.10) (0.92–1.51) Currently Married 1.33 0.95 0.50*** 1.08 0.94 0.65*** (0.65–2.73) (0.79–1.13) (0.43–0.59) (0.69–1.71) (0.72–1.22) (0.48–0.86) Household size (referred to 1–2 family members)   3–4 0.90 1.09 0.88* 1.09 1.10 0.89 (0.53–1.51) (0.95–1.24) (0.76–1.02) (0.78–1.51) (0.92–1.32) (0.71–1.12)   5 or more 0.98 1.25*** 0.88* 0.58** 1.07 0.82 (0.61–1.57) (1.09–1.43) (0.76–1.02) (0.36–0.95) (0.86–1.33) (0.62–1.08) Income (referred to lowest income)     Lower 0.85 1.05 0.83** 0.46 1.03 0.45*** (0.45–1.63) (0.90–1.22) (0.71–0.96) (0.13–1.63) (0.64–1.67) (0.30–0.68)     Middle 1.02 1.01 0.73*** 1.19 0.85 0.46*** (0.55–1.87) (0.87–1.17) (0.61–0.86) (0.58–2.44) (0.59–1.21) (0.34–0.63)     Higher 1.59 1.04 0.71*** 1.45 0.93 0.48*** (0.87–2.93) (0.88–1.22) (0.59–0.85) (0.78–2.70) (0.68–1.28) (0.37–0.63)     Highest 2.03** 1.33*** 0.74*** 1.47 1.36** 0.22*** (1.08–3.83) (1.13–1.58) (0.60–0.91) (0.80–2.70) (1.01–1.84) (0.16–0.29) Education (referred to no formal education)     No education but can read/write 0.70 0.76*** 0.88 1.72 1.18 0.84 (0.36–1.37) (0.65–0.90) (0.74–1.05) (0.85–3.48) (0.77–1.80) (0.58–1.22)     Primary school 0.63 0.86* 0.77*** 1.57 1.34 0.72* (0.34–1.17) (0.74–1.01) (0.64–0.92) (0.81–3.02) (0.91–1.98) (0.51–1.02)     Junior high school and above 1.08 0.92 0.65*** 1.76* 1.67*** 0.45*** (0.62–1.89) (0.79–1.08) (0.53–0.79) (0.96–3.23) (1.18–2.37) (0.33–0.63) Employment status (referred to farmers)     Unemployed or retired 2.45*** 1.06 1.59*** 1.24 1.37* 1.00 (1.51–3.96) (0.92–1.21) (1.37–1.84) (0.67–2.32) (0.98–1.91) (0.71–1.39) Informal employed 1.86* 0.93 1.69*** 1.41 1.19 1.61** (1.00–3.45) (0.78–1.12) (1.38–2.06) (0.69–2.90) (0.80–1.77) (1.09–2.39)     Formal employed 1.14 0.81*** 1.27** 0.93 1.75*** 0.82 (0.65–1.99) (0.70–0.94) (1.05–1.54) (0.48–1.80) (1.24–2.47) (0.55–1.23) Health checkup 0.83 1.22*** 0.76*** 0.87 1.07 0.77** (0.48–1.43) (1.06–1.39) (0.64–0.91) (0.60–1.27) (0.89–1.28) (0.61–0.96) Currently smoking 0.74 0.97 1.10 1.19 0.90 1.16 (0.44–1.23) (0.84–1.12) (0.93–1.30) (0.83–1.71) (0.74–1.09) (0.91–1.49) 2013 0.77 1.79*** 0.64*** 0.43*** 2.82*** 0.49*** (0.54–1.12) (1.63–1.96) (0.57–0.71) (0.32–0.58) (2.44–3.26) (0.42–0.58) Constant 0.01*** 0.37*** 0.41*** 0.00*** 0.12*** 5.92*** (0.00–0.14) (0.22–0.64) (0.21–0.80) (0.00–0.02) (0.04–0.31) (1.86–18.89) Pseudo R2 0.03 0.08 Observations 27,352 7,716 Notes: Relative Risk Ratios and 95% confidence intervals were shown. Multinomial logit model was used with public health insurance as the reference group. All models included sampling weights, province dummy and adjusted for clustering at the individual level. Significance level *** p<0.01 ** p<0.05 * p<0.10. Association between public insurance coverage and demand for private health insurance Table 5 further presented the association between coverage by public health insurance, certain types of public insurance and demand for private insurance. We found that having any public health insurance (i.e. NCMS, URBMI or UEBMI) was expected to reduce the likelihood of purchasing private insurance, while keeping other variables constant in the model (OR = 0.55, 95% CI: 0.48–0.63), indicating a possible substitute effect between public and private health insurance. Further separating the sample by rural and urban residents, we found that this substitute effect was only significant for urban residents (OR = 0.33, 95% CI: 0.27–0.40), but not for rural residents (OR = 0.89, 95% CI: 0.71–1.11). While taking a closer look at the types of public health insurance, as compared to coverage by NCMS, having URBMI is negatively associated with demand for private insurance (OR = 0.77, 95% CI: 0.63–0.94). However, it should be noted that having UEBMI significantly increased the odds of purchasing private health insurance (OR = 1.33, 95% CI: 1.12–1.58) than covering by NCMS. And as expected, those without public insurance coverage were significantly more likely to purchase private insurance (OR = 1.91, 95% CI: 1.65–2.21). 10.1371/journal.pone.0161774.t005Table 5 Association between public health insurance coverage, types of public insurance and purchase of private health insurance, results from the logit regressions. Private health insurance Total Rural Urban Total Having public health insurance 0.55(0.48–0.63)*** 0.89(0.71–1.11) 0.33(0.27–0.40)*** Types of public health insurance (referred to NCMS)     Urban Resident Basic Medical Insurance (URBMI) 0.77(0.63–0.94)**     Urban Employee Basic Medical Insurance (UEBMI) 1.33(1.12–1.58)***     No public health insurance 1.91(1.65–2.21)*** Pseudo R2 0.10 0.11 0.08 0.10 Observations 35,314 27,522 7,784 35,314 Notes: Odds ratio and 95% confidence intervals were reported. Logit models were used, including all variables in Table 3. All models included sampling weights, province dummies, and adjusted for clustering at individual level. Significance level *** p<0.01 ** p<0.05 * p<0.10. Discussion This study examined the determinants of public, private and multiple health insurance coverage among the population of retirement-age in China. Given the diversity of the population and substantial socioeconomic (SES) disparity in accessibility and affordability of health care among the Chinese elderly, this is the first study ever undertaken to comprehensively understand the status quo of China’s fragmented insurance system. Zhang and coauthor’s study used CHARLS 2011 and examined the predictors of being covered by UEBMI, URBMI, NCMS, or any insurance among urban and rural residents [26], while their study did not distinguish those covered by multiple insurance or those who purchased private insurance that were growing into an increasingly important component of China’s insurance system. Our analysis showed that, till 2013, 94.5% of this population had at least one type of public insurance, and 12.2% purchased private insurance. In general, we found that (a) compared to urban residents, rural residents were more likely to participate in public health insurance. But rural-to-urban migrants were more likely to be uninsured. (b) Public health insurance coverage may crowd out private insurance market particularly among those urban residents who enrolled in URBMI. There exists a large SES disparity in both public and private insurance coverage. The reason why rural residents had a higher percentage of public insurance coverage than urban residents might be related to the difference in enrollment unit between NCMS and the two urban insurance. The unit of enrollment for the NCMS is at the household level, but is at the individual level for UEBMI and URBMI [6, 30]. Due to the fact that NCMS achieved risk sharing among family members, large-size families with more elderly members may show greater willingness of insurance participation. However, for urban families, people with employment can be covered by UEBMI and other dwellers without formal employment can only enroll in URBMI. Therefore, household members cannot share risks in urban residents. Allowing household coverage in the urban insurance schemes might help to achieve universal coverage among the urban residents. [31] It is worth noting that migration may reduce the likelihood of being covered by public health insurance, but increase the likelihood of multiple coverage. Although the average age of the migrants was around 35 years, the first generation of the migrants has been turning to their 50s since 1980s. In our national sample, migrants accounted for about 30% of the entire population aged 45 and older. With a higher risk for developing chronic diseases, they are in greater need of health insurance coverage than the younger ones. However, a national survey showed that in 2014, the older migrants aged 45 and 60 had similar public insurance coverage with those aged 15–45 years (75.6% vs 75.0%), but had slightly lower private insurance coverage compared with the younger ones (3.9% vs 4.6%) [26]. One explanation to this low coverage was that many of the migrants were employed in informal sectors or small businesses which made them not qualified for UEBMI enrollment [14]. Another explanation was that public insurance, administered at the county or municipal level, usually did not cover health services outside of their local regions. Therefore, migrants had less incentive to participate in public insurance due to little opportunity for reimbursement. In addition, to gain insurance reimbursement from their destination place, migrants who were covered by rural insurance were more likely to also participate in urban public health insurance. But this dual coverage would increase overall premiums for the enrollees and administrative costs for the government agencies. Since 2011, some regions launched pilot policies to integrate NCMS and URBMI into one scheme, which may expand urban insurance coverage for the rural-to-urban migrant workers [15]. However, the lack of integration across regions in the insurance system and the rigid Hukou system would continuously be barriers for rural migrant workers to access urban health care [14]. Thus, the integration of rural and urban public health insurance across regions and the management of both insurance by one government agency would be a better policy solution in the foreseeable future. Understanding the role of private insurance is a little difficult since private insurance can supplement public insurance for China to achieve universal health coverage [19], but it can also be substituted by the expansion of public insurance [32]. Both evidence has been observed in our data and in other countries [22, 32–36]. On the one hand, our results showed that UEBMI enrollees and high SES population were more likely to buy private insurance. A China-based study also found that high-income adults were more likely to purchase private insurance when they were covered by NCMS, but for low-income people, the likelihood of buying private insurance decreased with public insurance coverage [22]. On the other hand, our findings suggested that public insurance coverage was associated with a reduced demand for private insurance, especially for urban residents who were covered by URBMI. This was consistent with a US-based study that found public insurance, subsidy or compensation was associated with a lower likelihood of private insurance purchase among the Medicaid enrollees [36]. However, we did not observe an adverse selection in the demand for private insurance, a phenomenon also documented in Liu and coauthor’s study [22]. There seemed counterintuitive that the people aged 45 and older with poor or fair health had a lower likelihood of purchasing private insurance. Other studies have observed that risk-takers are less likely to maintain good health and buy health insurance [37]. But these results may be understood in the China-specific context. First, health insurance is relatively new to Chinese residents particularly for the rural poor. The high coverage of public insurance was achieved due to the compulsory or semi- compulsory nature of the insurance [38]. But private insurance was much less familiar to the elderly population who received no or limited education. The complex design in the private insurance plans prohibited those who needed insurance but were lacking the cognitive ability to understand the details [38–40]. These double roles private insurance played in China’s health insurance system may exacerbate the SES inequality in access to health insurance among China’s middle-aged and elderly population. There are several limitations in our analysis. First, albeit at the national level, the CHARLS sample only represented the population aged 45 and older in mainland China, except for residents of Tibet. The results can hardly be generalizable to the entire population in China. In particular, the majority of the migrant workers are younger adults, and their health status and demand for insurance could be different from the older adults [41]. Second, we cannot rule out the possibility of misspecification due to self-report bias. For instance, some people may have double rural or urban public insurance coverage–one was obtained from their hometown and the other was offered by the places where they live and work, which may lead to underestimation of the multiple insurance coverage. Third, our sample does not contain information on usual source of care, thus we cannot estimate whether access to health care influenced the take-up of health insurance. However, we expect this bias to be minimal. After the health care reform in 2009, access to health care has been largely improved, and most people can go to a health care clinic/ provider within 15 minutes [9]. In addition, access to health care (such as distance) mainly differs by region, and in our regression analysis, we have controlled for regional variation such as rural/urban areas and provinces. The health checkup variable was also a good proxy for access to care, which was included in the regression. In conclusion, although China’s public insurance schemes have gradually covered the majority of its population in rural and urban areas, the domestic migrants, the poor and the vulnerable remained in the edge of the system. The growing private insurance market did not provide sufficient financial protection and was not accessible for people with the greatest need. With the rapid urbanization and population ageing, China needs to achieve universal coverage as well as reduce SES disparity in access to health insurance. Chinese government should reform the current fragmented insurance system by integrating the urban and rural public insurance schemes across regions and making them managed by one government agency, and removing the barriers for the middle-income and low-income to access private insurance. Such efforts would require a strong partnership across regional and national public sectors, and trust between public and private sectors [19]. Future research will focus on the geographic variation and the SES disparity in insurance integration and coverage, as well as its changes to cover vulnerable groups including those with low SES and chronic diseases. This work was supported by the National Nature Science Foundation (71403007) and Postdoctoral Science Foundation of China (2012M520132, 2013T60046). We thank the China Health and Retirement Longitudinal Study (CHARLS) team for providing the data. We thank Beibei Yuan for her comments. ==== Refs References 1 Meng Q , Fang H , Liu X , Yuan B , Xu J . Consolidating the social health insurance schemes in China: towards an equitable and efficient health system . The Lancet . 2015 ;386 (10002 ):1484 –1492 . 2 Marten R , McIntyre D , Travassos C , Shishkin S , Longde W , Reddy S , et al An assessment of progress towards universal health coverage in Brazil, Russia, India, China, and South Africa (BRICS) . The Lancet . 2014 ;384 (9960 ):2164 –2171 . 3 Zhang Z , Treiman DJ . Social origins, hukou conversion, and the wellbeing of urban residents in contemporary China . Social Science Research . 2013 ;42 (1 ):71 –89 . 10.1016/j.ssresearch.2012.08.004 23146599 4 You X , Kobayashi Y . The new cooperative medical scheme in China . Health Policy . 2009 ;91 (1 ):1 –9 . 10.1016/j.healthpol.2008.11.012 19121873 5 Du J . Economic reforms and health insurance in China . Social Science & Medicine . 2009 ;69 (3 ):387 –395 .19520475 6 Lin W , Liu GG , Chen G . The urban resident basic medical insurance: a landmark reform towards universal coverage in China . Health Economics . 2009 ;18 (S2 ):S83 –S96 .19551750 7 Ramesh M , Wu X , He AJ . Health governance and healthcare reforms in China . Health Policy and Planning . 2014 ;29 (6 ):663 –672 . 10.1093/heapol/czs109 23293100 8 Shi L , Zhang D . China's new rural cooperative medical scheme and underutilization of medical care among adults over 45: Evidence from CHARLS pilot data . The Journal of Rural Health . 2013 ;29 (s1 ):s51 –s61 .23944280 9 National Health and Family Planning Commission . China Statistical Yearbook of Health and Family Planning . Beijing : Peking Union Medical College Press ; 2015 . 10 Yip WC-M , Hsiao WC , Chen W , Hu S , Ma J , Maynard A . Early appraisal of China's huge and complex health-care reforms . The Lancet . 2012 ;379 (9818 ):833 –842 . 11 Center for Health Statistics and Information , National Health and Family Planning Commission . An Analysis Report of National Health Services Survey in China . Beijing : Peking Union Medical College Press ; 2013 . 12 Yip W , Hsiao W . China's health care reform: A tentative assessment . China Economic Review . 2009 ;20 (4 ):613 –619 . 13 Qiu P , Yang Y , Zhang J , Ma X . Rural-to-urban migration and its implication for new cooperative medical scheme coverage and utilization in China . BMC Public Health . 2011 ;11 (1 ):1 .21199570 14 Xu Q , Guan X , Yao F . Welfare program participation among rural‐to‐urban migrant workers in China . International Journal of Social Welfare . 2011 ;20 (1 ):10 –21 . 15 Qin X , Pan J , Liu GG . Does participating in health insurance benefit the migrant workers in China? An empirical investigation . China Economic Review . 2014 ;30 :263 –278 . 16 Hesketh T , Jun YX , Lu L , Mei WH . Health status and access to health care of migrant workers in China . Public Health Reports . 2008 ;123 (2 ):189 –197 . 18457071 17 Li H , Liu GG , Glaetzer C . Financing innovative medicines in mainland China: the role of commercial health insurance . Chinese Studies . 2013 ;2 (03 ):128 . 18 Tang S , Tao J , Bekedam H . Controlling cost escalation of healthcare: making universal health coverage sustainable in China . BMC Public Health . 2012 ;12 (1 ):1 .22214479 19 Bhattacharjya AS , Sapra PK . Health insurance in China and India: segmented roles for public and private financing . Health Affairs . 2008 ;27 (4 ):1005 –1015 . 10.1377/hlthaff.27.4.1005 18607034 20 Ying XH , Hu TW , Ren J , Chen W , Xu K , Huang JH . Demand for private health insurance in Chinese urban areas . Health Economics . 2007 ;16 (10 ):1041 –1050 . 17199233 21 Naughton B , Tsai KS . State capitalism, institutional adaptation, and the Chinese miracle Cambridge University Press ; 2015 . 22 Liu H , Gao S , Rizzo JA . The expansion of public health insurance and the demand for private health insurance in rural China . China Economic Review . 2011 ;22 (1 ):28 –41 . 23 Yan X WH , Chen H , Yang L , Rao K . Commercial health insurance under the new healthcare reform: Status, problems and strategies . Chinese Journal of Health Policy . 2013 ;6 (5 ):50 –54 . 24 Zhao Y , Strauss J , Yang G , Giles J , Hu P , Hu Y , et al China Health and Retirement Longitudinal Study–2011–2012 national baseline users’ guide Beijing : National School of Development, Peking University ; 2013 . 25 Zhang D , Shi L , Tian F , Zhang L . Care utilization with China’s New Rural Cooperative Medical Scheme: Updated evidence from the China Health and Retirement Longitudinal Study 2011–2012 . International Journal of Behavioral Medicine . 2016 10.1007/s12529-016-9560-0 26 Zhang C , Lei X , Strauss J , Zhao Y . Health insurance and health care among the mid‐aged and older Chinese: Evidence from the national baseline survey of CHARLS . Health Economics . 2016 10.1002/hec.3322 27 Zhao Y , Hu Y , Smith JP , Strauss J , Yang G . Cohort profile: the China Health and Retirement Longitudinal Study (CHARLS) . International Journal of Epidemiology . 2014 ;43 (1 ):61 –68 . 10.1093/ije/dys203 23243115 28 Andersen RM . Revisiting the behavioral model and access to medical care: does it matter? Journal of Health and Social Behavior . 1995 ;36 (1 ):1 –10 . 7738325 29 Cheng S , Long JS . Testing for IIA in the multinomial logit model . Sociological Methods & Research . 2007 ;35 (4 ):583 –600 . 30 Liu Y , Rao K . Providing health insurance in rural China: from research to policy . Journal of Health Politics, Policy and Law . 2006 ;31 (1 ):71 –92 . 16484669 31 Xu L , Wang Y , Collins CD , Tang S . Urban health insurance reform and coverage in China using data from National Health Services Surveys in 1998 and 2003 . BMC Health Services Research . 2007 ;7 (1 ):1 .17199886 32 Sekhri N , Savedoff W . Private health insurance: implications for developing countries . Bulletin of the World Health Organization . 2005 ;83 (2 ):127 –134 . 15744405 33 Doiron D , Jones G , Savage E . Healthy, wealthy and insured? The role of self‐assessed health in the demand for private health insurance . Health economics . 2008 ;17 (3 ):317 –334 . 17623485 34 Drechsler D , Jütting J . Different countries, different needs: the role of private health insurance in developing countries . Journal of Health Politics, Policy and Law . 2007 ;32 (3 ):497 –534 . 17519475 35 Liu T-C , Chen C-S . An analysis of private health insurance purchasing decisions with national health insurance in Taiwan . Social Science & Medicine . 2002 ;55 (5 ):755 –774 .12190269 36 Rask KN , Rask KJ . Public insurance substituting for private insurance: new evidence regarding public hospitals, uncompensated care funds, and Medicaid . Journal of Health Economics . 2000 ;19 (1 ):1 –31 . 10947569 37 Hopkins S , Kidd MP . The determinants of the demand for private health insurance under Medicare . Applied Economics . 1996 ;28 (12 ):1623 –1632 . 38 Eggleston K . Health care for 1.3 billion: An overview of China’s health system . SSRN Electronic Journal . 2012 10.2139/ssrn.2029952 39 Rice T , Lavarreda SA , Ponce NA , Brown ER . The impact of private and public health insurance on medication use for adults with chronic diseases . Medical Care Research and Review . 2005 ;62 (2 ):231 –249 . 15750178 40 Liu J , Shi L , Meng Q , Khan MM . Income-related inequality in health insurance coverage: analysis of China Health and Nutrition Survey of 2006 and 2009 . Int J Equity Health . 2012 ;11 (42 ):42 .22891984 41 Cantiello J , Fottler MD , Oetjen D , Zhang NJ . The impact of demographic and perceptual variables on a young adult’s decision to be covered by private health insurance . BMC Health Services Research . 2015 ;15 (1 ):1 .25603697

PMC005xxxxxx/PMC5001700.txt

==== Front PLoS OnePLoS ONEplosplosonePLoS ONE1932-6203Public Library of Science San Francisco, CA USA 2756438310.1371/journal.pone.0161897PONE-D-16-19836Research ArticleMedicine and Health SciencesEndocrinologyEndocrine DisordersDiabetes MellitusMedicine and Health SciencesMetabolic DisordersDiabetes MellitusMedicine and Health SciencesVascular MedicineBlood PressureHypertensionMedicine and Health SciencesPharmacologyDrugsStatinsMedicine and Health SciencesMetabolic DisordersDyslipidemiaMedicine and Health SciencesVascular MedicineBlood PressurePeople and PlacesGeographical LocationsAsiaTaiwanSocial SciencesEconomicsHealth EconomicsHealth InsuranceMedicine and Health SciencesHealth CareHealth EconomicsHealth InsuranceMedicine and Health SciencesInfectious DiseasesInfectious Disease ControlDiabetic Retinopathy in Patients with Diabetic Nephropathy: Development and Progression Diabetic Nephropathy and Risk of Diabetic RetinopathyJeng Chi-Juei 123Hsieh Yi-Ting 3Yang Chung-May 3Yang Chang-Hao 3Lin Cheng-Li 4*http://orcid.org/0000-0002-3045-0614Wang I-Jong 35*1 Department of Ophthalmology, National Taiwan University Hospital Hsin-Chu Branch, Hsinchu City, Taiwan2 Graduate Institute of Clinical Medicine, College of Medicine, National Taiwan University, Taipei, Taiwan3 Department of Ophthalmology, National Taiwan University Hospital, School of Medicine, Taipei, Taiwan4 Management Office for Health Data, China Medical University, Taichung, Taiwan5 Graduate Institute of Clinical Medical Science, China Medical University, Taichung, TaiwanTzekov Radouil EditorRoskamp Institute, UNITED STATESCompeting Interests: The authors have declared that no competing interests exist. Data curation: CJJ CLL IJW. Funding acquisition: CLL IJW. Methodology: CJJ CLL IJW. Project administration: IJW. Resources: IJW CLL. Software: CLL. Supervision: CMY IJW. Validation: YTH CHY CMY. Writing – original draft: CJJ CLL IJW. Writing – review & editing: CJJ CLL IJW YTH CHY CMY. * E-mail: a21467@mail.cmuh.org.tw (CLL); ijong@ms8.hinet.net (IJW)26 8 2016 2016 11 8 e016189717 5 2016 12 8 2016 © 2016 Jeng et al2016Jeng et alThis is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.The purpose of current study aims to investigate the development and progression of diabetic retinopathy (DR) in patients with diabetic nephropathy (DN) in a nationwide population-based cohort in Taiwan. Newly diagnosed DN patients and age- and sex-matched controls were identified from the Taiwanese Longitudinal Health Insurance Database from 2000 to 2010. We studied the effects of age, sex, hypertension, dyslipidemia, diabetic polyneuropathy (DPN), and medications on the development of nonproliferative DR (NPDR), proliferative DR (PDR), and diabetic macular edema (DME) in patients with DN. Cox proportional hazard regression analyses were used to estimate the adjusted hazard ratios (HRs) of the development of DR. Our results show that the adjusted HRs of NPDR and PDR were 5.01 (95% confidence interval (CI) = 4.68–5.37) and 9.7 (95% CI = 8.15–11.5), respectively, in patients with DN as compared with patients in the non-DN cohort. At 5-year follow-up, patients with DN showed an increased HR of NPDR progression to PDR (HR = 2.26, 95% CI = 1.68–3.03), and the major comorbidities were hypertension (HR = 1.23, 95% CI = 1.10–1.38 with NPDR; HR = 1.33, 95% CI = 1.02–1.72 with PDR) and DPN (HR = 2.03, 95% CI = 1.72–2.41 in NPDR; HR = 2.95, 95% CI = 2.16–4.03 in PDR). Dyslipidemia increased the HR of developing NPDR but not PDR or DME. Moreover, DN did not significantly affect DME development (HR = 1.47, 95% CI = 0.87–2.48) or progression (HR = 0.37, 95% CI = 0.11–1.20). We concluded that DN was an independent risk factor for DR development and progression; however, DN did not markedly affect DME development in this study, and the potential association between these disorders requires further investigation. the Taiwan Ministry of Health and Welfare Clinical Trial and Research Center of ExcellenceMOHW105-TDU-B-212-133019Lin Cheng-Li China Medical University Hospital, Academia Sinica Taiwan Biobank Stroke Biosignature ProjectBM10501010037Jeng Chi-Juei NRPB Stroke Clinical Trial ConsortiumMOST 104-2325-B-039-005Jeng Chi-Juei Tseng-Lien Lin FoundationJeng Chi-Juei Taiwan Brain Disease Foundation (Taipei, Taiwan), and Katsuzo and Kiyo Aoshima Memorial Funds (Japan)Jeng Chi-Juei The author(s) received no specific funding for this work. Data AvailabilityAll relevant data are within the paper.Data Availability All relevant data are within the paper. ==== Body Introduction Diabetic retinopathy (DR) is the leading cause of blindness in working-age people [1]. As in the case of the global epidemic, diabetic retinopathy in Taiwan has been reported in 35% of all diabetic patients [2, 3]. In relation to the risk factors identified for DR, epidemiological studies conducted on both type 1 and type 2 diabetes mellitus (DM) patients from the Diabetes Control and Complications Trial (DCCT) and the Action to Control Cardiovascular Risk in Diabetes (ACCORD) Eye Study have revealed the importance of glycemic control in delaying or preventing DR development [4–6]. Moreover, disease duration, elevated blood pressure, lipid profiles, serum levels of advanced glycation end products (AGEs), evidence of early stage atherosclerosis, increased caliber of retinal blood vessels, and several genetic factors (such as those related to enzymes involved in glucose and lipid metabolism) also contribute to the development of DR [4]. Diabetic nephropathy (DN), the primary cause of chronic kidney disease, accounts for 40% of all new cases of end-stage renal disease development recorded annually [7]; DN is characterized by persistent albuminuria, progressive decline of glomerular filtration rate, and elevation of blood pressure [8, 9]. In patients with DN, the presence of albumin in urine not only signifies glomerular injury, but also reflects systemic endothelial abnormalities and vasculopathy, which can represent an independent risk factor for cardiovascular disease [10, 11]. As in the case of DR, the major risk factors identified for DN include prolonged duration of diabetes, poor glycemic control, and hypertension [12]. Furthermore, diabetic patients with proteinuria or on dialysis frequently present with vision-threatening DR and proliferative DR (PDR) [13] and are at risk for developing diabetic macular edema (DME) [14]. However, Man et al. reported, based on a cross-sectional study of 263 patients, that a reduction in glomerular filtration rate (eGFR) is associated with increased severity of DR, but not with DME [15]. Nevertheless, optimizing blood-sugar control together with tightly controlling blood pressure can reduce the risk of developing both DR and DN because the diseases share the same microvascular changes [16, 17]. In DR, chronic hyperglycemia causes endothelial damage, loss of pericytes, basement-membrane thickening, breakdown of the blood-retinal barrier (BRB), platelet aggregation, and leukocyte adhesion in retinal capillaries [18, 19]. The microstructure disarrangement and microcirculation dysfunction lead to vascular hyperpermeability and microaneurysm formation, as observed in nonproliferative DR (NPDR) [20, 21]. Excessive vascular leakage of fluids, proteins, or lipids in the macular area leads to the development of DME [22]. As the disease progresses, capillaries close and arterioles become atrophied, and this matches the nonperfusion areas detected in patients’ fluorescein angiography [23]. Eventually, chronic hypoxia induces the expression of several angiogenic growth factors, which results in retinal neovascularization, as observed in PDR [24, 25]. In DN, chronic hyperglycemia also alters the expression of growth factors and cytokines in renal glomeruli [26–29], and these changes, in turn, result in an imbalance of the hemodynamics in glomerular cells. In the early stages, glomerular hypertrophy and hyperfiltration occur as glomeruli respond to the expression of hyperglycemia. However, increased intraglomerular pressure and increased shear stress following loss of heparin sulfates in the glomeruli eventually lead to the thickening of the glomerular and tubular basement membrane, accumulation of the mesangial matrix, and albuminuria [30–32]. Given the findings of the aforementioned pathophysiological and epidemiological studies, we were intrigued by the association of vision-threatening DR, PDR, and DME with DN development according to the pathogenesis of the diseases, and to investigate the association, we conducted this population-based cohort study. Methods Data source This study was conducted using the claim data obtained from the Longitudinal Health Insurance Database (LHID), which is a database of 1 million insurance claimants from the Taiwan National Health Insurance (Taiwan NHI) program. The Taiwan NHI was established in 1995 and it has served as a nationwide and compulsory health insurance program for Taiwan citizens. The National Health Research Institute (NHRI) established the LHID by randomly selecting 1 million insurance claimants from 1996 to 2000 and collecting their claim data annually. The LHID contains all of the data on claims from the Taiwan NHI, including the registry for beneficiaries, data on ambulatory care and hospital care claims, prescription files, and other medical expenditure files. The disease records in the Taiwan NHI were registered based on International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM). To safeguard the privacy of claimants, the NHRI concealed the original identification numbers and provided scrambled and anonymous numbers before releasing the database. Moreover, this study was approved by the Ethics Review Board of China Medical University (CMUH104-REC2-115). Study population In this study, we compared the risk of newly-diagnosed NPDR, PDR, and DME between DM patients with and without DN between January 1, 2000 and December 31, 2010. The DN cohort was composed of ≥18-year-old patients with DM (ICD-9-CM 250) plus DN (ICD-9-CM 249.4, 250.4), and the non-DN cohort was defined as the diagnosis of DN has not been made in this period. The non-DN cohort selected from the LHID comprised DM patients without a DN diagnosis and was 4-fold frequency matched by age and sex. The outcomes of interest in this study were (1) NPDR (ICD-9-CM 250.5, 362.01, 362.03–06 362.1, 362.81, 362.82), (2) PDR (ICD-9-CM 362.02, 379.23) and administration of panretinal photocoagulation (PRP) treatment, and (3) DME (ICD-9-CM 362.53, 362.83, 362.07) and administration of IVI (intravitreal injection) treatment. The diagnosis of DME or the administration of IVI treatment rely on the results of OCT (ocular computer tomography) or FAG (fluorescein angiography) requested by the Taiwan National Health Insurance Program in insurance claimants on a reimbursement. Each patient included in the study was followed-up for each outcome. DR at the baseline of both DN and non-DN cohorts is excluded to determine the incidence of DR. The follow-up time was defined as the duration from the occurrence of NPDR, PDR, or DME to December 31, 2010. The diagnosis of NPDR, PDR, and DME was made at subsequent two visits with the same diagnosis. We also examined the 5-year risk of PDR and DME in the DN cohort and the occurrence of NPDR events in the non-DN cohort. These NPDR patients were followed-up until the patients withdrew from the health insurance program, the occurrence of PDR or DME events, or end of the 5-year follow-up. We also investigated the influence of comorbidity and medication on the risk of NPDR, PDR, and DME. A patient with an identified comorbidity was defined as a patient with a history of the comorbidity before the index date; the comorbidities included were cerebrovascular accidents (CVA, ICD-9-CM 390–438), diabetic polyneuropathy (DPN, ICD-9-CM 357.2, 249.60, 249.61), hypertension (ICD-9-CM 401–405), and dyslipidemia (ICD-9-CM 272). The mediations considered were statin use, fibrate use, and antihypertension medication (including angiotensin-converting enzyme inhibitors, angiotensin II receptor blockers, α-blockers, β-blockers, calcium channel blockers, thiazides, and diuretics) before the index date. Statistical analysis To compare DM patients with and without DN, we calculated the mean age and the corresponding standard deviation (SD) of the patients in the 2 cohorts and determined the number and percentages of males and females, the comorbidities, and the medications. The distribution difference between the study cohorts was assessed by performing t tests for age and the chi-square test for sex, comorbidities, and medications. The follow-up duration was calculated from the index date to the end of follow-up (person-years), and the incidence density was measured as the total number of events divided by the sum of the follow-up durations. The incidence curve for each cohort was also evaluated using the Kaplan-Meier method, and the differences in the curves were examined using the log-rank test. Moreover, the risk of NPDR, PDR, and DME in DM patients with DN was compared with the corresponding risk in the case of DM patients without DN; for this comparison, the hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated using the crude and adjusted Cox proportional hazard models. We also compared the risk of NPDR, PDR, and DME between DM patients with and without DN after stratifying the patients according to age, sex, comorbidities, and medication. SAS 9.4 software (SAS Institute, Cary, NC, USA) was used to perform statistical analyses and R software was used to plot the incidence curve. The significant level was set at <0.05 for two-sided testing. Results For this study, we included 10692 DM patients with DN and 42761 DM patients without DN (Table 1). The mean age of the patients in the 2 cohorts was approximately 64 years and nearly 50.5% of the study participants were aged ≥ 65 years old, and both cohorts included more males than females (53.3% vs 46.7%). The proportions of patients with hypertension, CVA, DPN, and dyslipidemia were significantly (all P < 0.001) higher in the DN cohort than in the non-DN cohort. Furthermore, statin use, fibrate use, and antihypertension medication were also more frequent in the DN cohort than in the non-DN cohort. 10.1371/journal.pone.0161897.t001Table 1 Demographics and comorbidities of diabetes mellitus patients with and without diabetic nephropathy. Diabetic nephroropathy No (N = 42761) Yes (N = 10692) n(%) n(%) p-value Age, years 0.99  ≤64 21156(49.5) 5289(49.5)  ≥65 21605(50.5) 5403(50.5)  Mean (SD) † 63.5(13.5) 64.0(13.4) 0.003 Gender 0.99  Female 19966(46.7) 4992(46.7)  Male 22795(53.3) 5700(53.3) Comorbidity  Hypertension 24771(57.9) 8279(77.4) <0.001  CVA 27496(64.3) 8859(82.9) <0.001  DPN 283(0.66) 378(3.54) <0.001  Dyslipidemia 16783(39.3) 6463(60.5) <0.001 Medication  Statin 8180(19.1) 3848(36.0) <0.001  Fibrate 6675(15.6) 3235(30.3) <0.001  Antihypertensive medications 22331(52.2) 6945(65.0) <0.001 Chi-square test was used to examine categorical data; †t tests were used to examine continuous data; CVA: cerebrovascular accidents The mean follow-up durations in the cases of the occurrence of NPDR, PDR, and DME were the following (respectively, in years): DN cohort, 4.26 (SD = 3.23), 4.91 (SD = 4.84), and 5.09 (SD = 3.27); non-DN cohort, 5.91 (SD = 3.27), 6.05 (SD = 6.02), and 6.06 (SD = 3.27). The cumulative incidence curves of NPDR (31.9% vs. 7.60%; P< 0.001), PDR (7.90% vs. 0.96%; P< 0.001), and DME (0.55% vs. 0.36%; P = 0.03) were plotted for the DN and non-DN cohorts (Fig 1), and the results revealed significantly higher numbers of NPDR, PDR, and DME events by the end of follow-up in the case of DM patients with DN than in the case of patients without DN. Moreover, we compared the risk of NPDR, PDR, and DME between the DN cohort and the non-DN cohort (Table 2). After adjustment for age, sex, comorbidity, and medication, the DM patients with DN showed, relative to patients without DN, nearly 5-fold higher risk of NPDR (HR = 5.01, 95% CI = 4.68–5.37) and 9.7-fold higher risk of PDR (HR = 9.70, 95% CI = 8.15–11.5); however, the risk of DME did not differ in a statistically significant manner between DM patients with and without DN (HR = 1.49, 95% CI = 0.88–2.51). 10.1371/journal.pone.0161897.g001Fig 1 Cummulative incidences of (A) NPDR, (B) PDR, and (C) DME: comparison between diabetes mellitus patients with and without diabetic nephropathy. 10.1371/journal.pone.0161897.t002Table 2 Hazard ratios of outcomes according to sex, age, and comorbidity, obtained using univariate and multivariate Cox regression models. NPDR PDR DME Variable Crude HR (95% CI) Adjusted HR† (95% CI) Crude HR (95% CI) Adjusted HR† (95% CI) Crude HR (95% CI) Adjusted HR† (95% CI) Diabetic nephroropathy 5.90(5.53, 6.30)*** 5.01(4.68, 5.37)*** 11.1(9.42, 13.0)*** 9.70(8.15, 11.5)*** 1.74(1.06, 2.87)* 1.49(0.88, 2.51) Gender (Women vs Men) 0.96(0.90, 1.02) 0.97(0.91, 1.04) 1.07(0.92, 1.24) 0.99(0.85, 1.15) 1.78(1.12, 2.81)* 1.95(1.23, 3.10)** Age, years 1.00(0.99, 1.00) 0.99(0.99, 1.00)*** 0.97(0.96, 0.97)*** 0.96(0.96, 0.97)*** 0.97(0.97, 0.98)*** 1.02(1.00, 1.04) Baseline comorbidities (yes vs no)  Hypertension 1.77(1.65, 1.90)*** 1.22(1.10, 1.37)*** 1.43(1.22, 1.67)*** 1.33(1.02, 1.73)* 2.30(1.39, 3.82)** 1.47(0.71, 3.04)  CVA 1.77(1.65, 1.91)*** 1.11(0.99, 1.25) 1.39(1.18, 1.64)*** 1.08(0.82,1.41) 2.565(1.48, 4.45)*** 1.86(0.84, 4.10)  DPN 4.19(3.54, 4.96)*** 2.03(1.71, 2.41)*** 6.48(4.76, 8.82)*** 2.95(2.16, 4.04)*** 3.89(1.23, 12.3)* 3.00(0.93, 9.73)  Dyslipidemia 1.95(1.83, 2.09)*** 1.21(1.12, 1.31)*** 1.62(1.39, 1.88)*** 0.86(0.71,1.03) 1.27(0.82, 1.98) 0.96(0.56,1.64) Medication  Statin 1.91(1.78, 2.06)*** 1.17(1.08, 1.27)*** 1.69(1.43, 2.00)*** 1.09(0.90, 1.32) 2.08(1.27, 3.39)** 2.06(1.15, 3.69)*  Fibrate 1.76(1.63, 1.89)*** 1.03(0.94, 1.12) 1.75(1.47, 2.07)*** 1.12(0.92, 1.37) 0.86(0.45, 1.62) 0.56(0.28, 1.12)  Antihypertensive medications 1.44(1.35, 1.54)*** 1.03(0.96, 1.12) 1.09(0.94, 1.26) 0.99(0.83, 1.18) 1.17(0.75, 1.82) 0.61(0.37, 1.01) Crude HR: relative hazard ratio; Adjusted HR†: adjusted hazard ratio after controlling for age, sex, comorbidities (hypertension, CVA, DPN, and dyslipidemia), and medication (use of statin, fibrate, and antihypertension medication); *P < 0.05, **P < 0.01, ***P < 0.001. Next, the risk of NPDR, PDR, and DME was compared between DM patients with and without DN after stratification by age, sex, comorbidity, and medication (Table 3). First, NPDR risk was significantly higher for the DN cohort than for the non-DN cohort following stratification by age, sex, comorbidity, and medication (all P < 0.001). The HR calculated for NPDR reached 9.21 (95% CI = 7.60–11.2) for the DN cohort as compared with non-DN cohort in the case of study patients without any comorbidity; however, the HR was only 4.82 (95% CI = 4.49–5.17) when we compared all patients presenting at least one comorbidity. Furthermore, the HR of NPDR was approximately 5-fold higher for the DN cohort than for the non-DN cohort in the case of participants who did not use any medication (HR = 5.75 for statin nonusers, 5.36 for fibrate nonusers, and 5.35 for antihypertension medication nonusers), and the HR was approximately 4-fold higher in the case of participants who used medications (HR = 3.83 for statin users, 4.38 for fibrate users, and 4.91 for antihypertension medication users). Second, the PDR risk calculated for the DN cohort was also higher than that determined for the non-DN cohort after stratification by age, sex, comorbidity, and medication (all P < 0.001). For the DM patients with DN, the HRs calculated for PDR (relative to patients without DN) were 12.3 (95% CI = 10.1–14.9) and 7.40 (95% CI = 5.27–10.4) in the case of statin nonusers and statin users, respectively. Third, although the overall DME risk did not differ in a statistically significant manner between DM patients with and without DN, in the younger age group (age ≤ 64), the risk of DME was significantly higher in the case of DM patients with DN than without DN (HR = 2.47, 95% CI = 1.21–5.05). 10.1371/journal.pone.0161897.t003Table 3 Incidence and adjusted hazard ratio of NPDR, PDR, and DME according to sex, age, and comorbidity, compared between diabetes patients with and without diabetic nephropathy. Diabetic nephroropathy Compared to Control No Yes Variables Events n PY Rate# Events n PY Rate# Crude HR (95% CI) Adjusted HR† (95% CI) NPDR  All 1777 252522 7.04 1941 45547 42.6 5.90(5.53, 6.30)*** 5.01(4.68, 5.37)***  Gender   Female 877 122096 7.18 949 21672 43.8 5.95(5.43, 6.53)*** 4.87(4.42, 5.37)***   Male 900 130427 6.90 992 23875 41.6 5.86(5.35, 6.42)*** 5.11(4.65, 5.63)***   P for interaction 0.88  Age, years   ≤64 921 135413 6.80 1215 25929 46.9 6.74(6.19, 7.35)*** 5.46(4.98, 5.99)***   ≥65 856 117110 7.31 726 19619 37.0 4.89(4.43, 5.40)*** 4.44(4.01, 4.93)***   P for interaction <0.001  Comorbidity§   No 336 74515 4.51 158 3984 39.7 8.83(7.31, 10.7)*** 9.21(7.60, 11.2)***   Yes 1441 178007 8.10 1783 41563 42.9 5.17(4.82, 5.54)*** 4.82(4.49, 5.17)***   P for interaction <0.001 Medication  Statin   No 1343 215081 6.24 1276 31405 40.6 6.41(5.94, 6.92)*** 5.74(5.30, 6.23)***   Yes 434 37441 11.6 665 14143 47.0 3.96(3.51, 4.47)*** 3.81(3.37, 4.31)***   P for interaction <0.001  Fibrate   No 1437 218233 6.58 1337 32559 41.1 6.11(5.67, 6.58)*** 5.36(4.95, 5.79)***   Yes 340 34290 9.92 604 12989 46.5 4.58(4.01, 5.23)*** 4.38(3.82, 5.01)***   P for interaction <0.001  Antihypertensive medications   No 806 132605 6.08 704 17435 40.4 6.56(5.93, 7.26)*** 5.35(4.79, 5.97)***   Yes 971 119917 8.10 1237 28113 44.0 5.29(4.86, 5.76)*** 4.91(4.50, 5.36)***   P for interaction 0.003 PDR  All 209 258959 0.81 478 52476 9.11 11.1(9.42, 13.0)*** 9.70(8.15, 11.5)***  Gender   Female 100 125145 0.80 219 25355 8.64 10.6(8.37, 13.4)*** 8.69(6.74, 11.2)***   Male 109 133451 0.82 259 27121 9.55 11.5(9.21, 14.4)*** 10.6(8.36, 13.5)***   P for interaction 0.64  Age, years   ≤64 155 138406 1.12 385 30084 12.8 11.2(9.33, 13.6)*** 10.2(8.35, 12.4)***   ≥65 54 120189 0.45 93 22391 4.15 8.96(6.40, 12.5)*** 9.19(6.43, 13.1)***   P for interaction 0.24  Comorbidity§   No 53 75606 0.70 47 4599 10.2 14.6(9.86, 21.6)*** 13.5(9.03, 20.1)***   Yes 156 182989 0.85 431 47877 9.00 10.4(8.65, 12.5)*** 8.99(7.45, 10.8)***   P for interaction 0.14 Medication  Statin   No 165 219790 0.75 332 36024 9.22 12.2(10.1, 14.7)*** 12.3(10.1, 14.9)***   Yes 44 38805 1.13 146 16452 8.87 7.73(5.52, 10.8)*** 7.40(5.27, 10.4)***   P for interaction 0.02  Fibrate   No 177 223163 0.79 333 37422 8.90 11.0(9.19, 13.2)*** 11.0(9.10, 13.4)***   Yes 32 35432 0.90 145 15054 9.63 10.6(7.22, 15.5)*** 10.4(7.08, 15.3)***   P for interaction 0.83  Antihypertensive medications   No 117 135290 0.86 206 19867 10.4 11.8(9.38, 14.8)*** 10.9(8.50, 13.8)***   Yes 92 123305 0.75 272 32608 8.34 11.1(8.72, 14.0)*** 10.8(8.48, 13.8)***   P for interaction 0.71 DME  All 59 259142 0.23 21 54420 0.39 1.74(1.06, 2.87)* 1.49(0.88, 2.51)  Gender   Female 22 125471 0.18 6 26302 0.23 1.38(0.56, 3.39) 1.21(0.46, 3.16)   Male 37 133672 0.28 15 28119 0.53 1.96(1.07, 3.57)* 1.61(0.86, 3.01)   P for interaction 0.48  Age, years   ≤64 20 138870 0.14 16 31692 0.50 3.58(1.85, 6.90)*** 2.49(1.22, 5.10)*   ≥65 39 120273 0.32 5 22729 0.22 0.71(0.28, 1.81) 0.68(0.27, 1.76)   P for interaction 0.006  Comorbidity§   No 8 75760 0.11 0 4811 0.00 - -   Yes 51 183382 0.28 21 49609 0.42 1.55(0.93, 2.57) 1.58(0.94, 2.65)   P for interaction 0.98 Medication  Statin   No 44 220220 0.20 13 37439 0.35 1.76(0.95, 3.27) 1.61(0.85, 3.06)   Yes 15 38922 0.39 8 16982 0.47 1.25(0.53, 2.94) 1.17(0.48, 2.85)   P for interaction 0.51  Fibrate   No 54 223640 0.24 15 38754 0.39 1.64(0.92, 2.90) 1.15(0.64, 2.09)   Yes 5 35502 0.14 6 15665 0.38 2.75(0.84, 9.00) 2.95(0.88, 9.84)   P for interaction 0.44  Antihypertensive medications   No 29 135583 0.21 9 20678 0.44 2.08(0.98, 4.39) 1.41(0.63, 3.14)   Yes 30 123560 0.24 12 33742 0.36 1.49(0.76, 2.91) 1.41(0.71, 2.81)   P for interaction 0.49 PY: person-years; Rate#: incidence rate, per 1000 PY; Crude HR: relative hazard ratio; Adjusted HR†: adjusted hazard ratio after controlling for age, sex, comorbidities (hypertension, CVA, DPN, and dyslipidemia), and medication (use of statin, fibrate, and antihypertension medication); Comorbidity§: the comorbidity group included patients with any one of these comorbidities: hypertension, CVA, DPN, and dyslipidemia; *P < 0.05, **P < 0.01, ***P < 0.001 Lastly, we compared the risk of PDR and DME between DM patients with and without DN after the occurrence of NPDR during the 5-year follow-up period (Table 4). After NPDR occurrence, PDR risk was significantly higher in the case of DM patients with DN than in the case of DM patients without DN (HR = 2.25, 95% CI = 1.68–3.02), but DME risk did not differ significantly between the 2 groups (HR = 0.37, 95% CI = 0.12–1.22). 10.1371/journal.pone.0161897.t004Table 4 Overall PDR and DME events and hazard ratios of PDR and DME measured for NPDR among study participants within a 5-year follow-up period. NPDR Compared to Control No Yes Variables Events n PY Rate# Events n PY Rate# Crude HR (95% CI) Adjusted HR† (95% CI) PDR 66 5316 12.4 177 5892 30.0 2.48(1.87, 3.30)*** 2.25(1.68, 3.02)*** DME 11 5446 2.02 4 6394 0.63 0.33(0.11, 1.04) 0.37(0.12, 1.22) PY: person-years; Rate#: incidence rate, per 1000 PY; Crude HR: relative hazard ratio; Adjusted HR†: adjusted hazard ratio after controlling for age, sex, comorbidities (hypertension, CVA, DPN, and dyslipidemia), and medication (use of statin, fibrate, and antihypertension medication); ***P < 0.001. Discussion DR, DN, and DPN are the most common complications related to small-vessel injuries due to long-term hyperglycemia [33, 34]. Previously, we reported that DPN and DR were correlated: patients with DPN presented an elevated risk of developing DR and PDR [35]. Similarly, Barr et al. reported a correlation between these 3 microvascular complications and indicated that patients with DPN exhibit (relative to control) a 4-fold increase in DR rate and a 2-fold increase in the rate of microalbuminuria [36]. Here, we further demonstrated that the incidences of NPDR, PDR, and DME increased with time in patients with DN, who also presented a higher rate to DR development than did patients without DN (Fig 1). We also demonstrated that relative to patients without DN, patients with DN carried a higher risk of developing NPDR and PDR and progression from NPDR to PDR during the 5-year follow-up (Tables 3 and 4), which agrees with the results of Wisconsin Epidemiologic Study of Diabetic Retinopathy (WESDR) [8, 9]. Intriguingly, the effect of DN on NPDR development was measured to be markedly elevated after we stratified and adjusted the effects of other risk factors and comorbidities in this study. In the case of patients without any comorbidity, DN increased the risk of NPDR development 9.21 fold, whereas the increase was 4.82 fold in patients presenting other comorbidities. The patients with DN also carried an elevated risk of developing PDR, although the independent effect of DN did not differ in a statistically significant manner between patients with and without comorbidities. In the longitudinal 5-year follow-up, DN was also identified to increase the risk of NPDR progression to PDR. However, the effect on DME development was not statistically significant. These findings partly agree with the risk factors for DR occurrence and progression that were reported based on WESDR [8, 9] and the UK Prospective Diabetes Study (UKPDS) [21], which demonstrated that the duration of diabetes, degree of metabolic control, elevated glycosylated hemoglobin A1c levels, severity of DR, hypertension, low socioeconomic status, and older age are DME risk factors [37, 38]. Regarding the duration of DM, our study cannot provide the duration of DM in this health claim database. However, in the Table 1, the difference of the age between DN cohort and non-DN cohort was statistically insignificant. In the Table 2, there was no effect of age on the development of NPDR, PDR and DME. The mean follow-up durations in the cases of the occurrence of NPDR, PDR, and DME were the following: DN cohort, 4.26 years, 4.91 years, and 5.09 years; non-DN cohort, 5.91 years, 6.05 years, and 6.06 years. In spite of the absence of duration of DM in this study, our results revealed that the presence of DN has a significant role in the development of NPDR or PDR, but the effect of age was insignificant. Nevertheless, the duration of diabetes may still have a major impact on DR rates and be a confounder. This limitation should be investigated in future studies. We determined that hypertension was associated in a statistically significant manner with the development of NPDR and PDR but not DME (Table 2); this result is partly in accord with the results of previous studies indicating that increased systolic blood pressure is a major risk factor for DR [39]. Although certain cross-sectional data have suggested that hypertension is associated with DR, longitudinal data have been inconsistent [40–45]. The UKPDS results showed that DR incidence was associated with systolic blood pressure [46], and in WESDR, diastolic blood pressure was identified as a statistically significant independent predictor of DR progression to PDR over a 14-year follow-up period in patients with younger-onset (type 1) DM, regardless of glucose control and proteinuria [47]; however, no association was identified in the case of type 2 DM, which might be due to the original selection of older-onset diabetes and mortality rates [47]. By contrast, in another study, diastolic blood pressure in the fourth-quartile range was identified to be associated with a 3.3-fold increase in the 4-year risk of developing DME as compared with the blood pressure in the first-quartile range in patients with younger-onset DM, and further with a 2.1-fold increase in the risk in the case of patients with older-onset diabetes [48]. Moreover, the results of a randomized clinical trial demonstrated that a lowering blood pressure to below 140/90 mmHg was associated with a substantial reduction in the risk of developing macrovascular and microvascular complications in hypertensive patients with DM [49]. In this study, DN patients with hypertension presented a higher risk of developing NPDR and PDR than did DN patients without hypertension. Medical control of blood pressure exerts a protective effect in the early but not late stage of DR, which agrees with the findings of certain studies showing that blood pressure control might not be able to halt disease progression to the proliferative stage or macular edema development [50]. However, another study showed that the use of an angiotensin-converting enzyme inhibitor for blood pressure control might protect against DR progression [51]. The possible pathogenic mechanisms by which hypertension affects DR are (1) hemodynamic mechanisms (impaired autoregulation and hyperperfusion) and (2) vascular endothelial growth factor (VEGF)-dependent mechanisms, because hypertension independent of hyperglycemia upregulates VEGF expression in retinal endothelial cells and ocular fluids [52]. Therefore, we conclude that DR duration, hypertension, hypertension treatment, and potential ethnic factors led to a nonsignificant effect of hypertension on DME development in this study. CVA was identified here as a comorbidity in the cohort of patients with DN (Table 1), although CVA alone did not increase the HRs of NPDR, PDR, and DME (Table 2). Furthermore, the interactions among hypertension, CVA, dyslipidemia, and DPN increased the adjusted HRs of developing NPDR and PDR, but not DME, in patients with DN (Table 3). Our findings agree with the results of previous studies showing that diabetic patients with DN carry an increased risk of fatal and nonfatal cardiovascular and other complications [53–57], and this risk is also affected by genetic and ethnic predisposition [58, 59]. However, the most notable difference in the case of our results is that these comorbidities did not contribute to the development of DME in the DN cohort. Although the association between DR and cardiovascular outcomes has been extensively studied [60–63] and reviewed [64], the cardiovascular outcomes in DME patients remain inadequately examined; in previously studies that included DME patients, the statistical power was insufficient to characterize the relationship between DR and cardiovascular outcomes [62, 63]. For example, in one study, a large insured population was used to quantify and compare the incidence rates of myocardial infarction or CVA/stroke in hospitalized patients with DME against matched diabetic control patients [65]. The adjusted rate ratio calculated for CVA was 1.98 (95% CI: 1.39–2.83, P < 0.001) for DME patients versus the diabetic controls. By contrast, our cohorts included hospitalized and nonhospitalized patients, and this is likely to be comparatively more representative of patients with DME. With regard to dyslipidemia, our results showed that patients under statin and fibrate medication developed DN more frequently than did patients who did not receive these treatments (Table 1). Moreover, the hazard ratio calculated for NPDR, PDR, and DME were higher in patients who received statins than in patients who received fibrate (Table 2). Similarly, statin and fibrate use in DN patients increased the adjusted HRs of developing NPDR and PDR but not DME (Table 3). These results indicate that dyslipidemia plays a role in the development of DR and DN, and that statin and fibrate use can lower the risk of developing DME in patients with dyslipidemia. Serum lipids have been reported to be a risk factor for DR and DME [66, 67], and permeability changes in the retinal microvasculature have been suggested to result in extravascular accumulation of lipoprotein deposits coupled with a consequent loss of function in the surrounding retinal cells [68, 69]; however, the role of serum lipids in the pathogenesis of DR and DME remains controversial. The Fenofibrate Intervention and Event Lowering in Diabetes study reported that fenofibrate treatment resulted in reduced DR progression and a diminished requirement for laser treatment in type 2 DM in the study participants [70]. The Action to Control Cardiovascular Risk in Diabetes Eye study showed that concomitant use of fenofibrate and statin reduced the requirement for laser treatment by 40% [71], which is compatible with the treatment’s protective effect against NPDR development in DN patients that was observed in this study. However, these findings suggest a complex mechanistic association between serum dyslipidemia and DR and DME, the underlying pathogenetic process of which remains unclear. Although evidence gathered from cohort studies and meta-analyses of case-control studies have suggested a strong relationship between lipid levels and DME, this relationship was not confirmed by a meta-analysis that included only prospective random clinical trials [72]. Thus, the relationship between lipid levels and DME warrants further investigation. The most intriguing result obtained in this study was that the risk of DME development did not differ between patients with and without DN, although patients with DN still showed an increased the risk of developing PDR from NPDR than did the patients without DN. In patients aged less than 64 years old, DN can influence DME development to a certain extent. As mentioned, DME is a complex disease of multifactorial origin that is caused by a disruption of the BRB [73]. Chronic elevation of blood glucose, high cholesterol, accumulation of oxygen free radicals and AGEs/AGE receptors, protein kinase C, and other factors have been implicated in the pathogenesis of DME [22]. These factors ultimately contribute to an increase in VEGF expression, which results in a breakdown of the BRB. Moreover, although reversible, hyperglycemia impairs the function of the retinal pigment epithelium at an early stage of the disease [74]. In addition to the increased permeability of retinal capillaries, the primary retinal change in DR, the closure of retinal capillaries leads to retinal ischemia. Retinal ischemia, in turn, can result in the formation of neovascularization, which might lead to vitreous hemorrhage or traction damage in the retina through the production of various growth factors, including VEGF [75]. In the pathogenesis of DN, as in DME pathogenesis, podocytes secrete increased amounts of VEGF-A [76], tight-junction loss occurs and leads to hyperpermeability, and albuminuria is prevalent [77, 78]. As the nephropathy progresses, DN is eventually associated with capillary nonperfusion, which leads to podocyte death in DN and to increased extracellular matrix deposition and thus a thickening of the glomerular basement membrane [79, 80], as in PDR [81]. These findings could explain the results of our study, which demonstrated that young patients with DN or patients presenting the early events of DN carried an elevated risk of DME, which coexisted with capillary hyperpermeability and the presence of albuminuria. However, these parallel and intercorrelated diseases progressed together, and the DM patients with DN presented an increased risk of PDR in the long-term follow-up. In summary, our findings indicate that patients with DN experience higher incidences of DR and progression to PDR as compared with patients without DN. Moreover, the results confirmed that follow-up duration and hypertension are associated with DR development, and that lipid-modulating agents exert a protective effect during the early stages of DR. Patients with type 2 DM and albuminuria must be carefully monitored for progressive eye disease, and patients with DME must be evaluated for concomitant kidney disease. DME formation (which might be multipathogenic) and its correlation with DN require further investigation. ==== Refs References 1 Aiello LP , Gardner TW , King GL , Blankenship G , Cavallerano JD , Ferris FL 3rd, et al Diabetic retinopathy . Diabetes Care . 1998 ;21 (1 ):143 –56 . 9538986 2 Chang C , Lu F , Yang YC , Wu JS , Wu TJ , Chen MS , et al Epidemiologic study of type 2 diabetes in Taiwan . Diabetes Res Clin Pract . 2000 ;50 Suppl 2 :S49 –59 . 11024584 3 Zheng Y , He M , Congdon N . The worldwide epidemic of diabetic retinopathy . Indian J Ophthalmol . 2012 ;60 (5 ):428 –31 . 10.4103/0301-4738.100542 22944754 4 The Diabetes Control and Complications Trial (DCCT) . Design and methodologic considerations for the feasibility phase. The DCCT Research Group . Diabetes . 1986 ;35 (5 ):530 –45 . 2869996 5 Reichard P , Nilsson BY , Rosenqvist U . The effect of long-term intensified insulin treatment on the development of microvascular complications of diabetes mellitus . N Engl J Med . 1993 ;329 (5 ):304 –9 . 8147960 6 Chew EY , Davis MD , Danis RP , Lovato JF , Perdue LH , Greven C , et al The effects of medical management on the progression of diabetic retinopathy in persons with type 2 diabetes: the Action to Control Cardiovascular Risk in Diabetes (ACCORD) Eye Study . Ophthalmology . 2014 ;121 (12 ):2443 –51 . 10.1016/j.ophtha.2014.07.019 25172198 7 Fong DS , Aiello LP , Ferris FL 3rd, Klein R . Diabetic retinopathy . Diabetes Care . 2004 ;27 (10 ):2540 –53 . 15451934 8 Klein R , Klein BE , Moss SE , Davis MD , DeMets DL . The Wisconsin epidemiologic study of diabetic retinopathy. IV. Diabetic macular edema . Ophthalmology . 1984 ;91 (12 ):1464 –74 . 6521986 9 Klein R , Klein BE , Moss SE , Davis MD , DeMets DL . The Wisconsin epidemiologic study of diabetic retinopathy. III. Prevalence and risk of diabetic retinopathy when age at diagnosis is 30 or more years . Arch Ophthalmol . 1984 ;102 (4 ):527 –32 . 6367725 10 Deckert T , Feldt-Rasmussen B , Borch-Johnsen K , Jensen T , Kofoed-Enevoldsen A . Albuminuria reflects widespread vascular damage. The Steno hypothesis . Diabetologia . 1989 ;32 (4 ):219 –26 . 2668076 11 Ochodnicky P , Henning RH , van Dokkum RP , de Zeeuw D . Microalbuminuria and endothelial dysfunction: emerging targets for primary prevention of end-organ damage . J Cardiovasc Pharmacol . 2006 ;47 Suppl 2 :S151 –62 ; discussion S72-6. 16794452 12 Lewis G , Maxwell AP . Risk factor control is key in diabetic nephropathy . Practitioner . 2014 ;258 (1768 ):13 –7 , 2 24689163 13 Boelter MC , Gross JL , Canani LH , Costa LA , Lisboa HR , Tres GS , et al Proliferative diabetic retinopathy is associated with microalbuminuria in patients with type 2 diabetes . Braz J Med Biol Res . 2006 ;39 (8 ):1033 –9 . 16906278 14 Hammes HP , Welp R , Kempe HP , Wagner C , Siegel E , Holl RW , et al Risk Factors for Retinopathy and DME in Type 2 Diabetes-Results from the German/Austrian DPV Database . PLoS One . 2015 ;10 (7 ):e0132492 10.1371/journal.pone.0132492 26177037 15 Man RE , Sasongko MB , Wang JJ , MacIsaac R , Wong TY , Sabanayagam C , et al The Association of Estimated Glomerular Filtration Rate With Diabetic Retinopathy and Macular Edema . Invest Ophthalmol Vis Sci . 2015 ;56 (8 ):4810 –6 . 10.1167/iovs.15-16987 26218909 16 Ritz E . Hypertension in diabetic nephropathy: prevention and treatment . Am Heart J . 1993 ;125 (5 Pt 2 ):1514 –9 . 8480622 17 Sivaprasad S , Gupta B , Crosby-Nwaobi R , Evans J . Prevalence of diabetic retinopathy in various ethnic groups: a worldwide perspective . Surv Ophthalmol . 2012 ;57 (4 ):347 –70 . 10.1016/j.survophthal.2012.01.004 22542913 18 Cai J , Boulton M . The pathogenesis of diabetic retinopathy: old concepts and new questions . Eye (Lond) . 2002 ;16 (3 ):242 –60 .12032713 19 Heng LZ , Comyn O , Peto T , Tadros C , Ng E , Sivaprasad S , et al Diabetic retinopathy: pathogenesis, clinical grading, management and future developments . Diabet Med . 2013 ;30 (6 ):640 –50 . 10.1111/dme.12089 23205608 20 Kohner EM , Patel V , Rassam SM . Role of blood flow and impaired autoregulation in the pathogenesis of diabetic retinopathy . Diabetes . 1995 ;44 (6 ):603 –7 . 7789621 21 Kohner EM , Stratton IM , Aldington SJ , Turner RC , Matthews DR . Microaneurysms in the development of diabetic retinopathy (UKPDS 42). UK Prospective Diabetes Study Group . Diabetologia . 1999 ;42 (9 ):1107 –12 . 10447523 22 Bhagat N , Grigorian RA , Tutela A , Zarbin MA . Diabetic macular edema: pathogenesis and treatment . Surv Ophthalmol . 2009 ;54 (1 ):1 –32 . 10.1016/j.survophthal.2008.10.001 19171208 23 Boeri D , Maiello M , Lorenzi M . Increased prevalence of microthromboses in retinal capillaries of diabetic individuals . Diabetes . 2001 ;50 (6 ):1432 –9 . 11375345 24 Smith LE , Shen W , Perruzzi C , Soker S , Kinose F , Xu X , et al Regulation of vascular endothelial growth factor-dependent retinal neovascularization by insulin-like growth factor-1 receptor . Nat Med . 1999 ;5 (12 ):1390 –5 . 10581081 25 Boulton M , Foreman D , Williams G , McLeod D . VEGF localisation in diabetic retinopathy . Br J Ophthalmol . 1998 ;82 (5 ):561 –8 . 9713066 26 Han HJ , Kang CW , Park SH . Tissue-specific regulation of insulin-like growth factors and insulin-like growth factor binding proteins in male diabetic rats in vivo and in vitro . Clin Exp Pharmacol Physiol . 2006 ;33 (12 ):1172 –9 . 17184497 27 Vasylyeva TL , Ferry RJ Jr. Novel roles of the IGF-IGFBP axis in etiopathophysiology of diabetic nephropathy . Diabetes Res Clin Pract . 2007 ;76 (2 ):177 –86 . 17011663 28 Lupia E , Elliot SJ , Lenz O , Zheng F , Hattori M , Striker GE , et al IGF-1 decreases collagen degradation in diabetic NOD mesangial cells: implications for diabetic nephropathy . Diabetes . 1999 ;48 (8 ):1638 –44 . 10426384 29 Ways DK , Sheetz MJ . The role of protein kinase C in the development of the complications of diabetes . Vitam Horm . 2000 ;60 :149 –93 . 11037624 30 Osterby R , Parving HH , Hommel E , Jorgensen HE , Lokkegaard H . Glomerular structure and function in diabetic nephropathy. Early to advanced stages . Diabetes . 1990 ;39 (9 ):1057 –63 . 2384188 31 Gruden G , Perin PC , Camussi G . Insight on the pathogenesis of diabetic nephropathy from the study of podocyte and mesangial cell biology . Curr Diabetes Rev . 2005 ;1 (1 ):27 –40 . 18220580 32 Pavenstadt H , Kriz W , Kretzler M . Cell biology of the glomerular podocyte . Physiol Rev . 2003 ;83 (1 ):253 –307 . 12506131 33 Agardh E , Agardh CD , Koul S , Torffvit O . A four-year follow-up study on the incidence of diabetic retinopathy in older onset diabetes mellitus . Diabet Med . 1994 ;11 (3 ):273 –8 . 8033526 34 Cohen O , Norymberg K , Neumann E , Dekel H . Complication-free duration and the risk of development of retinopathy in elderly diabetic patients . Arch Intern Med . 1998 ;158 (6 ):641 –4 . 9521229 35 Lin IC , Wang YH , Lin CL , Chang YJ , Lee SH , Wang IJ . Diabetic polyneuropathy and the risk of developing diabetic retinopathy: a nationwide, population-based study . Acta Ophthalmol . 2015 ;93 (8 ):713 –8 . 10.1111/aos.12746 25989925 36 Barr EL , Wong TY , Tapp RJ , Harper CA , Zimmet PZ , Atkins R , et al Is peripheral neuropathy associated with retinopathy and albuminuria in individuals with impaired glucose metabolism? The 1999–2000 AusDiab. Diabetes Care . 2006 ;29 (5 ):1114 –6 . 16644648 37 Williams R , Airey M , Baxter H , Forrester J , Kennedy-Martin T , Girach A . Epidemiology of diabetic retinopathy and macular oedema: a systematic review . Eye (Lond) . 2004 ;18 (10 ):963 –83 .15232600 38 Girach A , Lund-Andersen H . Diabetic macular oedema: a clinical overview . Int J Clin Pract . 2007 ;61 (1 ):88 –97 . 17229183 39 Leske MC , Wu SY , Hennis A , Hyman L , Nemesure B , Yang L , et al Hyperglycemia, blood pressure, and the 9-year incidence of diabetic retinopathy: the Barbados Eye Studies . Ophthalmology . 2005 ;112 (5 ):799 –805 . 15878059 40 Norgaard K , Feldt-Rasmussen B , Deckert T . Is hypertension a major independent risk factor for retinopathy in type 1 diabetes? Diabet Med . 1991 ;8 (4 ):334 –7 . 1830256 41 Teuscher A , Schnell H , Wilson PW . Incidence of diabetic retinopathy and relationship to baseline plasma glucose and blood pressure . Diabetes Care . 1988 ;11 (3 ):246 –51 . 3416678 42 Haffner SM , Fong D , Stern MP , Pugh JA , Hazuda HP , Patterson JK , et al Diabetic retinopathy in Mexican Americans and non-Hispanic whites . Diabetes . 1988 ;37 (7 ):878 –84 . 3384186 43 Hamman RF , Mayer EJ , Moo-Young GA , Hildebrandt W , Marshall JA , Baxter J . Prevalence and risk factors of diabetic retinopathy in non-Hispanic whites and Hispanics with NIDDM. San Luis Valley Diabetes Study . Diabetes . 1989 ;38 (10 ):1231 –7 . 2792575 44 Schmittdiel JA , Steiner JF , Adams AS , Dyer W , Beals J , Henderson WG , et al Diabetes care and outcomes for American Indians and Alaska natives in commercial integrated delivery systems: a SUrveillance, PREvention, and ManagEment of Diabetes Mellitus (SUPREME-DM) Study . BMJ Open Diabetes Res Care . 2014 ;2 (1 ):e000043 10.1136/bmjdrc-2014-000043 25452877 45 Ballard DJ , Melton LJ 3rd, Dwyer MS , Trautmann JC , Chu CP , O'Fallon WM , et al Risk factors for diabetic retinopathy: a population-based study in Rochester, Minnesota . Diabetes Care . 1986 ;9 (4 ):334 –42 . 3743309 46 Stratton IM , Kohner EM , Aldington SJ , Turner RC , Holman RR , Manley SE , et al UKPDS 50: risk factors for incidence and progression of retinopathy in Type II diabetes over 6 years from diagnosis . Diabetologia . 2001 ;44 (2 ):156 –63 . 11270671 47 Klein R , Klein BE , Moss SE . The Wisconsin epidemiological study of diabetic retinopathy: a review . Diabetes Metab Rev . 1989 ;5 (7 ):559 –70 . 2689119 48 Klein R , Moss SE , Klein BE , Davis MD , DeMets DL . The Wisconsin epidemiologic study of diabetic retinopathy. XI. The incidence of macular edema . Ophthalmology . 1989 ;96 (10 ):1501 –10 . 2587045 49 Volpe M , Battistoni A , Savoia C , Tocci G . Understanding and treating hypertension in diabetic populations . Cardiovasc Diagn Ther . 2015 ;5 (5 ):353 –63 . 10.3978/j.issn.2223-3652.2015.06.02 26543822 50 Do DV , Wang X , Vedula SS , Marrone M , Sleilati G , Hawkins BS , et al Blood pressure control for diabetic retinopathy . Sao Paulo Med J . 2015 ;133 (3 ):278 –9 . 10.1590/1516-3180.20151333T1 26176931 51 Chaturvedi N , Sjolie AK , Stephenson JM , Abrahamian H , Keipes M , Castellarin A , et al Effect of lisinopril on progression of retinopathy in normotensive people with type 1 diabetes. The EUCLID Study Group. EURODIAB Controlled Trial of Lisinopril in Insulin-Dependent Diabetes Mellitus . Lancet . 1998 ;351 (9095 ):28 –31 . 9433426 52 Srivastava BK , Rema M . Does hypertension play a role in diabetic retinopathy? J Assoc Physicians India . 2005 ;53 :803 –8 . 16334627 53 Gaede P , Vedel P , Larsen N , Jensen GV , Parving HH , Pedersen O . Multifactorial intervention and cardiovascular disease in patients with type 2 diabetes . N Engl J Med . 2003 ;348 (5 ):383 –93 . 12556541 54 Hirata-Dulas CA , Rith-Najarian SJ , McIntyre MC , Ross C , Dahl DC , Keane WF , et al Risk factors for nephropathy and cardiovascular disease in diabetic Northern Minnesota American Indians . Clin Nephrol . 1996 ;46 (2 ):92 –8 . 8869785 55 Ritz E , Rychlik I , Locatelli F , Halimi S . End-stage renal failure in type 2 diabetes: A medical catastrophe of worldwide dimensions . Am J Kidney Dis . 1999 ;34 (5 ):795 –808 . 10561134 56 Girach A , Manner D , Porta M . Diabetic microvascular complications: can patients at risk be identified? A review . Int J Clin Pract . 2006 ;60 (11 ):1471 –83 . 17073842 57 Girach A , Vignati L . Diabetic microvascular complications—can the presence of one predict the development of another? J Diabetes Complications . 2006 ;20 (4 ):228 –37 . 16798474 58 Young BA , Maynard C , Boyko EJ . Racial differences in diabetic nephropathy, cardiovascular disease, and mortality in a national population of veterans . Diabetes Care . 2003 ;26 (8 ):2392 –9 . 12882868 59 Cowie CC . Diabetic renal disease: racial and ethnic differences from an epidemiologic perspective . Transplant Proc . 1993 ;25 (4 ):2426 –30 . 8356621 60 Miettinen H , Haffner SM , Lehto S , Ronnemaa T , Pyorala K , Laakso M . Retinopathy predicts coronary heart disease events in NIDDM patients . Diabetes Care . 1996 ;19 (12 ):1445 –8 . 8941482 61 van Hecke MV , Dekker JM , Stehouwer CD , Polak BC , Fuller JH , Sjolie AK , et al Diabetic retinopathy is associated with mortality and cardiovascular disease incidence: the EURODIAB prospective complications study . Diabetes Care . 2005 ;28 (6 ):1383 –9 . 15920056 62 Cheung N , Rogers S , Couper DJ , Klein R , Sharrett AR , Wong TY . Is diabetic retinopathy an independent risk factor for ischemic stroke? Stroke . 2007 ;38 (2 ):398 –401 . 17194880 63 Cheung N , Wang JJ , Klein R , Couper DJ , Sharrett AR , Wong TY . Diabetic retinopathy and the risk of coronary heart disease: the Atherosclerosis Risk in Communities Study . Diabetes Care . 2007 ;30 (7 ):1742 –6 . 17389333 64 Cheung N , Wong TY . Diabetic retinopathy and systemic vascular complications . Prog Retin Eye Res . 2008 ;27 (2 ):161 –76 . 10.1016/j.preteyeres.2007.12.001 18249026 65 Nguyen-Khoa BA , Goehring EL , Werther W , Fung AE , Do DV , Apte RS , et al Hospitalized cardiovascular events in patients with diabetic macular edema . BMC Ophthalmol . 2012 ;12 :11 10.1186/1471-2415-12-11 22646811 66 Golubovic-Arsovska M . Association of dyslipidaemia with macular oedema and hard exudates in diabetic maculopathy . Prilozi . 2007 ;28 (2 ):149 –60 . 18356786 67 Asensio-Sanchez VM , Gomez-Ramirez V , Morales-Gomez I , Rodriguez-Vaca I . [Clinically significant diabetic macular edema: systemic risk factors ]. Arch Soc Esp Oftalmol . 2008 ;83 (3 ):173 –6 . 18311676 68 Panagiotoglou TD , Ganotakis ES , Kymionis GD , Moschandreas JA , Fanti GN , Charisis SK , et al Atorvastatin for diabetic macular edema in patients with diabetes mellitus and elevated serum cholesterol . Ophthalmic Surg Lasers Imaging . 2010 ;41 (3 ):316 –22 . 10.3928/15428877-20100430-04 20507015 69 Chew EY . Diabetic retinopathy and lipid abnormalities . Curr Opin Ophthalmol . 1997 ;8 (3 ):59 –62 . 10168895 70 Keech AC , Mitchell P , Summanen PA , O'Day J , Davis TM , Moffitt MS , et al Effect of fenofibrate on the need for laser treatment for diabetic retinopathy (FIELD study): a randomised controlled trial . Lancet . 2007 ;370 (9600 ):1687 –97 . 17988728 71 Wright AD , Dodson PM . Medical management of diabetic retinopathy: fenofibrate and ACCORD Eye studies . Eye (Lond) . 2011 ;25 (7 ):843 –9 .21436845 72 Das R , Kerr R , Chakravarthy U , Hogg RE . Dyslipidemia and Diabetic Macular Edema: A Systematic Review and Meta-Analysis . Ophthalmology . 2015 ;122 (9 ):1820 –7 . 10.1016/j.ophtha.2015.05.011 26150053 73 Pelzek C , Lim JI . Diabetic macular edema: review and update . Ophthalmol Clin North Am . 2002 ;15 (4 ):555 –63 . 12515087 74 Samuels IS , Bell BA , Pereira A , Saxon J , Peachey NS . Early retinal pigment epithelium dysfunction is concomitant with hyperglycemia in mouse models of type 1 and type 2 diabetes . J Neurophysiol . 2015 ;113 (4 ):1085 –99 . 10.1152/jn.00761.2014 25429122 75 Bandello F , Lattanzio R , Zucchiatti I , Del Turco C . Pathophysiology and treatment of diabetic retinopathy . Acta Diabetol . 2013 ;50 (1 ):1 –20 . 10.1007/s00592-012-0449-3 23277338 76 Tufro A , Veron D . VEGF and podocytes in diabetic nephropathy . Semin Nephrol . 2012 ;32 (4 ):385 –93 . 10.1016/j.semnephrol.2012.06.010 22958493 77 Lang GE . Diabetic macular edema . Ophthalmologica . 2012 ;227 Suppl 1 :21 –9 . 10.1159/000337156 22517122 78 Chen S , Ziyadeh FN . Vascular endothelial growth factor and diabetic nephropathy . Curr Diab Rep . 2008 ;8 (6 ):470 –6 . 18990304 79 Advani A , Gilbert RE . The endothelium in diabetic nephropathy . Semin Nephrol . 2012 ;32 (2 ):199 –207 . 10.1016/j.semnephrol.2012.02.006 22617769 80 Najafian B , Alpers CE , Fogo AB . Pathology of human diabetic nephropathy . Contrib Nephrol . 2011 ;170 :36 –47 . 10.1159/000324942 21659756 81 Arboleda-Velasquez JF , Valdez CN , Marko CK , D'Amore PA . From pathobiology to the targeting of pericytes for the treatment of diabetic retinopathy . Curr Diab Rep . 2015 ;15 (2 ):573 10.1007/s11892-014-0573-2 25620405